Re: [PATCH V3 XRT Alveo 17/18] fpga: xrt: partition isolation platform driver

[Lizhi Hou <lizhi.hou@xilinx.com>]

Hi Moritz,


On 02/21/2021 12:36 PM, Moritz Fischer wrote:
>
> On Wed, Feb 17, 2021 at 10:40:18PM -0800, Lizhi Hou wrote:
>> Add partition isolation platform driver. partition isolation is
>> a hardware function discovered by walking firmware metadata.
>> A platform device node will be created for it. Partition isolation
>> function isolate the different fpga regions
>>
>> Signed-off-by: Sonal Santan <sonal.santan@xilinx.com>
>> Signed-off-by: Max Zhen <max.zhen@xilinx.com>
>> Signed-off-by: Lizhi Hou <lizhih@xilinx.com>
>> ---
>>   drivers/fpga/xrt/include/xleaf/axigate.h |  25 ++
>>   drivers/fpga/xrt/lib/xleaf/axigate.c     | 298 +++++++++++++++++++++++
>>   2 files changed, 323 insertions(+)
>>   create mode 100644 drivers/fpga/xrt/include/xleaf/axigate.h
>>   create mode 100644 drivers/fpga/xrt/lib/xleaf/axigate.c
>>
>> diff --git a/drivers/fpga/xrt/include/xleaf/axigate.h b/drivers/fpga/xrt/include/xleaf/axigate.h
>> new file mode 100644
>> index 000000000000..2cef71e13b30
>> --- /dev/null
>> +++ b/drivers/fpga/xrt/include/xleaf/axigate.h
>> @@ -0,0 +1,25 @@
>> +/* SPDX-License-Identifier: GPL-2.0 */
>> +/*
>> + * Header file for XRT Axigate Leaf Driver
>> + *
>> + * Copyright (C) 2020-2021 Xilinx, Inc.
>> + *
>> + * Authors:
>> + *   Lizhi Hou <Lizhi.Hou@xilinx.com>
>> + */
>> +
>> +#ifndef _XRT_AXIGATE_H_
>> +#define _XRT_AXIGATE_H_
>> +
>> +#include "xleaf.h"
>> +#include "metadata.h"
>> +
>> +/*
>> + * AXIGATE driver IOCTL calls.
>> + */
>> +enum xrt_axigate_ioctl_cmd {
>> +     XRT_AXIGATE_FREEZE = XRT_XLEAF_CUSTOM_BASE, /* See comments in xleaf.h */
>> +     XRT_AXIGATE_FREE,
>> +};
>> +
>> +#endif       /* _XRT_AXIGATE_H_ */
>> diff --git a/drivers/fpga/xrt/lib/xleaf/axigate.c b/drivers/fpga/xrt/lib/xleaf/axigate.c
>> new file mode 100644
>> index 000000000000..382969f9925f
>> --- /dev/null
>> +++ b/drivers/fpga/xrt/lib/xleaf/axigate.c
>> @@ -0,0 +1,298 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +/*
>> + * Xilinx Alveo FPGA AXI Gate Driver
>> + *
>> + * Copyright (C) 2020-2021 Xilinx, Inc.
>> + *
>> + * Authors:
>> + *      Lizhi Hou<Lizhi.Hou@xilinx.com>
>> + */
>> +
>> +#include <linux/mod_devicetable.h>
>> +#include <linux/platform_device.h>
>> +#include <linux/delay.h>
>> +#include <linux/device.h>
>> +#include <linux/io.h>
>> +#include "metadata.h"
>> +#include "xleaf.h"
>> +#include "xleaf/axigate.h"
>> +
>> +#define XRT_AXIGATE "xrt_axigate"
>> +
>> +struct axigate_regs {
>> +     u32             iag_wr;
>> +     u32             iag_rvsd;
>> +     u32             iag_rd;
>> +} __packed;
> Just make them #defines, even more so if there are only 3 of them.
We will use #define and regmap.
>> +
>> +struct xrt_axigate {
>> +     struct platform_device  *pdev;
>> +     void                    *base;
>> +     struct mutex            gate_lock; /* gate dev lock */
>> +
>> +     void                    *evt_hdl;
>> +     const char              *ep_name;
>> +
>> +     bool                    gate_freezed;
>> +};
>> +
>> +/* the ep names are in the order of hardware layers */
>> +static const char * const xrt_axigate_epnames[] = {
>> +     XRT_MD_NODE_GATE_PLP,
>> +     XRT_MD_NODE_GATE_ULP,
>> +     NULL
>> +};
>> +
>> +#define reg_rd(g, r)                                         \
>> +     ioread32((void *)(g)->base + offsetof(struct axigate_regs, r))
>> +#define reg_wr(g, v, r)                                              \
>> +     iowrite32(v, (void *)(g)->base + offsetof(struct axigate_regs, r))
>> +
>> +static inline void freeze_gate(struct xrt_axigate *gate)
>> +{
>> +     reg_wr(gate, 0, iag_wr);
>> +     ndelay(500);
>> +     reg_rd(gate, iag_rd);
>> +}
>> +
>> +static inline void free_gate(struct xrt_axigate *gate)
>> +{
>> +     reg_wr(gate, 0x2, iag_wr);
>> +     ndelay(500);
> Magic constants?
Will use #define for 500
>> +     (void)reg_rd(gate, iag_rd);
> At the very least add a comment on why? Is this for PCI synchronization
> reasons?
Will add comment. Some old board requires this extra read and it will 
not hurt performance.
>
>> +     reg_wr(gate, 0x3, iag_wr);
>> +     ndelay(500);
> Magic constants?
>> +     reg_rd(gate, iag_rd);
> Does it nead a (void) or not? Be consistent, again, why do we read here
> at all?
>> +}
>> +
>> +static int xrt_axigate_epname_idx(struct platform_device *pdev)
>> +{
>> +     int                     i;
>> +     int                     ret;
>> +     struct resource         *res;
> Nope. Indents:
>
> struct resource *res;
> int, i, ret;
Will change this.
>
>> +
>> +     res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>> +     if (!res) {
>> +             xrt_err(pdev, "Empty Resource!");
>> +             return -EINVAL;
>> +     }
>> +
>> +     for (i = 0; xrt_axigate_epnames[i]; i++) {
>> +             ret = strncmp(xrt_axigate_epnames[i], res->name,
>> +                           strlen(xrt_axigate_epnames[i]) + 1);
>> +             if (!ret)
>> +                     break;
>> +     }
>> +
>> +     ret = (xrt_axigate_epnames[i]) ? i : -EINVAL;
> Why not just:
>
>          if (xrt_axigate_epnames[i])
>                  return i;
>
>          return -EINVAL;
Will change this.
>> +     return ret;
>> +}
>> +
>> +static void xrt_axigate_freeze(struct platform_device *pdev)
>> +{
>> +     struct xrt_axigate      *gate;
>> +     u32                     freeze = 0;
> Indents. Fix everywhere.
Will fix this.
>> +
>> +     gate = platform_get_drvdata(pdev);
>> +
>> +     mutex_lock(&gate->gate_lock);
>> +     freeze = reg_rd(gate, iag_rd);
>> +     if (freeze) {           /* gate is opened */
>> +             xleaf_broadcast_event(pdev, XRT_EVENT_PRE_GATE_CLOSE, false);
>> +             freeze_gate(gate);
>> +     }
>> +
>> +     gate->gate_freezed = true;
> s/freezed/frozen
Will change terms to open / close.
>> +     mutex_unlock(&gate->gate_lock);
>> +
>> +     xrt_info(pdev, "freeze gate %s", gate->ep_name);
> debug?
axigate is a very critical part for programming FPGA. We hope to have an 
explicit printk to indicate the axigate open/close.
>> +}
>> +
>> +static void xrt_axigate_free(struct platform_device *pdev)
>> +{
>> +     struct xrt_axigate      *gate;
>> +     u32                     freeze;
>> +
>> +     gate = platform_get_drvdata(pdev);
>> +
>> +     mutex_lock(&gate->gate_lock);
>> +     freeze = reg_rd(gate, iag_rd);
>> +     if (!freeze) {          /* gate is closed */
>> +             free_gate(gate);
>> +             xleaf_broadcast_event(pdev, XRT_EVENT_POST_GATE_OPEN, true);
>> +             /* xrt_axigate_free() could be called in event cb, thus
>> +              * we can not wait for the completes
>> +              */
>> +     }
>> +
>> +     gate->gate_freezed = false;
>> +     mutex_unlock(&gate->gate_lock);
>> +
>> +     xrt_info(pdev, "free gate %s", gate->ep_name);
>> +}
>> +
>> +static void xrt_axigate_event_cb(struct platform_device *pdev, void *arg)
>> +{
>> +     struct platform_device *leaf;
>> +     struct xrt_event *evt = (struct xrt_event *)arg;
>> +     enum xrt_events e = evt->xe_evt;
>> +     enum xrt_subdev_id id = evt->xe_subdev.xevt_subdev_id;
>> +     int instance = evt->xe_subdev.xevt_subdev_instance;
>> +     struct xrt_axigate *gate = platform_get_drvdata(pdev);
>> +     struct resource *res;
> Reverse x-mas tree;
> xxxxxxxxxx
> xxxxxxxxx
> xxxxxxxx
> xxxxxx
Will fix this.
>> +
>> +     switch (e) {
>> +     case XRT_EVENT_POST_CREATION:
>> +             break;
>> +     default:
>> +             return;
>> +     }
>> +
>> +     if (id != XRT_SUBDEV_AXIGATE)
>> +             return;
>> +
>> +     leaf = xleaf_get_leaf_by_id(pdev, id, instance);
>> +     if (!leaf)
>> +             return;
>> +
>> +     res = platform_get_resource(leaf, IORESOURCE_MEM, 0);
>> +     if (!res || !strncmp(res->name, gate->ep_name, strlen(res->name) + 1)) {
>> +             (void)xleaf_put_leaf(pdev, leaf);
>> +             return;
>> +     }
>> +
>> +     /*
>> +      * higher level axigate instance created,
>> +      * make sure the gate is openned. This covers 1RP flow which
>> +      * has plp gate as well.
>> +      */
>> +     if (xrt_axigate_epname_idx(leaf) > xrt_axigate_epname_idx(pdev))
>> +             xrt_axigate_free(pdev);
>> +     else
>> +             xleaf_ioctl(leaf, XRT_AXIGATE_FREE, NULL);
>> +
>> +     (void)xleaf_put_leaf(pdev, leaf);
>> +}
>> +
>> +static int
>> +xrt_axigate_leaf_ioctl(struct platform_device *pdev, u32 cmd, void *arg)
>> +{
>> +     switch (cmd) {
>> +     case XRT_XLEAF_EVENT:
>> +             xrt_axigate_event_cb(pdev, arg);
>> +             break;
>> +     case XRT_AXIGATE_FREEZE:
>> +             xrt_axigate_freeze(pdev);
>> +             break;
>> +     case XRT_AXIGATE_FREE:
>> +             xrt_axigate_free(pdev);
>> +             break;
>> +     default:
>> +             xrt_err(pdev, "unsupported cmd %d", cmd);
>> +             return -EINVAL;
>> +     }
>> +
>> +     return 0;
>> +}
>> +
>> +static int xrt_axigate_remove(struct platform_device *pdev)
>> +{
>> +     struct xrt_axigate      *gate;
>> +
>> +     gate = platform_get_drvdata(pdev);
>> +
>> +     if (gate->base)
>> +             iounmap(gate->base);
>> +
>> +     platform_set_drvdata(pdev, NULL);
>> +     devm_kfree(&pdev->dev, gate);
> No! The point of using devres is so cleanup happens on removal.
> While you're at it, if you move the ioremap to a devres version, this
> function can basically go away entirely.
Will fix this.
>> +
>> +     return 0;
>> +}
>> +
>> +static int xrt_axigate_probe(struct platform_device *pdev)
>> +{
>> +     struct xrt_axigate      *gate;
>> +     struct resource         *res;
>> +     int                     ret;
>> +
>> +     gate = devm_kzalloc(&pdev->dev, sizeof(*gate), GFP_KERNEL);
>> +     if (!gate)
>> +             return -ENOMEM;
>> +
>> +     gate->pdev = pdev;
>> +     platform_set_drvdata(pdev, gate);
>> +
>> +     xrt_info(pdev, "probing...");
>> +     res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>> +     if (!res) {
>> +             xrt_err(pdev, "Empty resource 0");
>> +             ret = -EINVAL;
>> +             goto failed;
>> +     }
>> +
>> +     gate->base = ioremap(res->start, res->end - res->start + 1);
>> +     if (!gate->base) {
>> +             xrt_err(pdev, "map base iomem failed");
>> +             ret = -EFAULT;
>> +             goto failed;
>> +     }
>> +
>> +     gate->ep_name = res->name;
>> +
>> +     mutex_init(&gate->gate_lock);
>> +
>> +     return 0;
>> +
>> +failed:
>> +     xrt_axigate_remove(pdev);
>> +     return ret;
>> +}
>> +
>> +static struct xrt_subdev_endpoints xrt_axigate_endpoints[] = {
>> +     {
>> +             .xse_names = (struct xrt_subdev_ep_names[]) {
>> +                     { .ep_name = "ep_pr_isolate_ulp_00" },
>> +                     { NULL },
>> +             },
>> +             .xse_min_ep = 1,
>> +     },
>> +     {
>> +             .xse_names = (struct xrt_subdev_ep_names[]) {
>> +                     { .ep_name = "ep_pr_isolate_plp_00" },
>> +                     { NULL },
>> +             },
>> +             .xse_min_ep = 1,
>> +     },
>> +     { 0 },
>> +};
>> +
>> +static struct xrt_subdev_drvdata xrt_axigate_data = {
>> +     .xsd_dev_ops = {
>> +             .xsd_ioctl = xrt_axigate_leaf_ioctl,
>> +     },
>> +};
>> +
>> +static const struct platform_device_id xrt_axigate_table[] = {
>> +     { XRT_AXIGATE, (kernel_ulong_t)&xrt_axigate_data },
>> +     { },
>> +};
>> +
>> +static struct platform_driver xrt_axigate_driver = {
>> +     .driver = {
>> +             .name = XRT_AXIGATE,
>> +     },
>> +     .probe = xrt_axigate_probe,
>> +     .remove = xrt_axigate_remove,
>> +     .id_table = xrt_axigate_table,
>> +};
>> +
>> +void axigate_leaf_init_fini(bool init)
>> +{
>> +     if (init) {
>> +             xleaf_register_driver(XRT_SUBDEV_AXIGATE,
>> +                                   &xrt_axigate_driver, xrt_axigate_endpoints);
>> +     } else {
>> +             xleaf_unregister_driver(XRT_SUBDEV_AXIGATE);
>> +     }
>> +}
> This thing is duplicated in every file, maybe a macro would be an idea.
Will define a macro.

Thanks,
Lizhi
>> --
>> 2.18.4
>>
> - Moritz