On 11. 05. 22, 8:21, Hitomi Hasegawa wrote:
> Enable diagnostic interrupts for the Fujitsu A64FX.
>
> Register the NMI/IRQ corresponding to the A64FX's device definition
> dedicated to diagnostic interrupts, so that when this interrupt is
> sent using the BMC, it causes a panic. This can be used to obtain
> a kernel dump.
>
> Signed-off-by: Hitomi Hasegawa <[email protected]>
Hi,
I'm not sure why you cc linux-serial, but anyway, comments below :).
> --- /dev/null
> +++ b/drivers/soc/fujitsu/a64fx-diag.c
> @@ -0,0 +1,155 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * A64FX diag driver.
> + * Copyright (c) 2022 Fujitsu Ltd.
> + */
> +
> +#include <linux/acpi.h>
> +#include <linux/interrupt.h>
> +#include <linux/irq.h>
> +#include <linux/module.h>
> +#include <linux/platform_device.h>
> +
> +#define A64FX_DIAG_IRQ 1
> +#define BMC_DIAG_INTERRUPT_STATUS_OFFSET (0x0044)
> +#define BMC_DIAG_INTERRUPT_ENABLE_OFFSET (0x0040)
> +#define BMC_DIAG_INTERRUPT_MASK BIT(31)
> +
> +struct a64fx_diag_priv {
> + int irq;
> + void __iomem *mmsc_reg_base;
> + bool has_nmi;
There are unnecessary holes in the struct. If you reorder it, you drop
some alignment. Like: pointer, int, bool.
> +};
> +
> +static irqreturn_t a64fx_diag_handler_nmi(int irq, void *dev_id)
> +{
> + nmi_panic(NULL, "a64fx_diag: interrupt received\n");
> +
> + return IRQ_HANDLED;
> +}
> +
> +static irqreturn_t a64fx_diag_handler_irq(int irq, void *dev_id)
> +{
> + panic("a64fx_diag: interrupt received\n");
> +
> + return IRQ_HANDLED;
> +}
> +
> +static void a64fx_diag_interrupt_clear(struct a64fx_diag_priv *priv)
> +{
> + u32 mmsc;
> + void __iomem *diag_status_reg_addr;
I'm not sure what soc/ maintainers prefer, but inverted xmas tree would
look/read better.
> +
> + diag_status_reg_addr = priv->mmsc_reg_base + BMC_DIAG_INTERRUPT_STATUS_OFFSET;
> + mmsc = readl(diag_status_reg_addr);
> + if (mmsc & BMC_DIAG_INTERRUPT_MASK)
> + writel(BMC_DIAG_INTERRUPT_MASK, diag_status_reg_addr);
> +}
> +
> +static void a64fx_diag_interrupt_enable(struct a64fx_diag_priv *priv)
> +{
> + u32 mmsc;
> + void __iomem *diag_enable_reg_addr;
> +
> + diag_enable_reg_addr = priv->mmsc_reg_base + BMC_DIAG_INTERRUPT_ENABLE_OFFSET;
> + mmsc = readl(diag_enable_reg_addr);
> + if (!(mmsc & BMC_DIAG_INTERRUPT_MASK)) {
> + mmsc |= BMC_DIAG_INTERRUPT_MASK;
> + writel(mmsc, diag_enable_reg_addr);
> + }
> +}
> +
> +static void a64fx_diag_interrupt_disable(struct a64fx_diag_priv *priv)
> +{
> + u32 mmsc;
> + void __iomem *diag_enable_reg_addr;
> +
> + diag_enable_reg_addr = priv->mmsc_reg_base + BMC_DIAG_INTERRUPT_ENABLE_OFFSET;
> + mmsc = readl(diag_enable_reg_addr);
> + if (mmsc & BMC_DIAG_INTERRUPT_MASK) {
> + mmsc &= ~BMC_DIAG_INTERRUPT_MASK;
> + writel(mmsc, diag_enable_reg_addr);
> + }
> +}
> +
> +static int a64fx_diag_probe(struct platform_device *pdev)
> +{
> + int ret;
> + unsigned long irq_flags;
> + struct device *dev = &pdev->dev;
> + struct a64fx_diag_priv *priv;
> +
> + priv = devm_kzalloc(dev, sizeof(struct a64fx_diag_priv), GFP_KERNEL);
Don't we prefer sizeof(*priv)?
> + if (priv == NULL)
> + return -ENOMEM;
> +
> + priv->mmsc_reg_base = devm_platform_ioremap_resource(pdev, 0);
> + if (IS_ERR(priv->mmsc_reg_base))
> + return PTR_ERR(priv->mmsc_reg_base);
> +
> + priv->irq = platform_get_irq(pdev, A64FX_DIAG_IRQ);
> + if (priv->irq < 0)
> + return priv->irq;
> +
> + platform_set_drvdata(pdev, priv);
> +
> + a64fx_diag_interrupt_clear(priv);
> + a64fx_diag_interrupt_enable(priv);
> +
> + irq_flags = IRQF_PERCPU | IRQF_NOBALANCING | IRQF_NO_AUTOEN |
> + IRQF_NO_THREAD;
> + ret = request_nmi(priv->irq, &a64fx_diag_handler_nmi, irq_flags,
> + "a64fx_diag_nmi", NULL);
> + if (ret) {
> + ret = request_irq(priv->irq, &a64fx_diag_handler_irq,
> + irq_flags, "a64fx_diag_irq", NULL);
> + if (ret) {
> + dev_err(dev, "cannot register IRQ %d\n", ret);
No a64fx_diag_interrupt_disable()?
> + return ret;
> + }
> + enable_irq(priv->irq);
Hmm...
> + priv->has_nmi = false;
No need to set zeroed priv member to zero.
> + } else {
> + enable_nmi(priv->irq);
Provided the above, I don't immediatelly see, what's the purpose of
IRQF_NO_AUTOEN then?
> + priv->has_nmi = true;
> + }
> +
> + return 0;
> +}
> +
> +static int __exit a64fx_diag_remove(struct platform_device *pdev)
Is __exit appropriate here at all -- I doubt that.
> +{
> + struct a64fx_diag_priv *priv = platform_get_drvdata(pdev);
> +
> + a64fx_diag_interrupt_disable(priv);
> + a64fx_diag_interrupt_clear(priv);
> +
> + if (priv->has_nmi)
> + free_nmi(priv->irq, NULL);
> + else
> + free_irq(priv->irq, NULL);
> +
> + return 0;
> +}
> +
> +static const struct acpi_device_id a64fx_diag_acpi_match[] = {
> + { "FUJI2007", 0 },
> + { },
> +};
> +MODULE_DEVICE_TABLE(acpi, a64fx_diag_acpi_match);
> +
> +
> +static struct platform_driver a64fx_diag_driver = {
> + .driver = {
> + .name = "a64fx_diag_driver",
> + .acpi_match_table = ACPI_PTR(a64fx_diag_acpi_match),
> + },
> + .probe = a64fx_diag_probe,
> + .remove = a64fx_diag_remove,
> +};
> +
> +module_platform_driver(a64fx_diag_driver);
> +
> +MODULE_LICENSE("GPL v2");
> +MODULE_AUTHOR("Hitomi Hasegawa <[email protected]>");
> +MODULE_DESCRIPTION("A64FX diag driver");
--
js
suse labs
Hi Jiri,
> I'm not sure why you cc linux-serial, but anyway, comments below :).
I used sysrq until the last version, so I still included kernel-serial in
the destination. I am not planning to use sysrq now, so I will remove it
from the destination from the next version.
Thank you for your comment.
> > +struct a64fx_diag_priv {
> > + int irq;
> > + void __iomem *mmsc_reg_base;
> > + bool has_nmi;
>
> There are unnecessary holes in the struct. If you reorder it, you drop some
> alignment. Like: pointer, int, bool.
> > + u32 mmsc;
> > + void __iomem *diag_status_reg_addr;
>
> I'm not sure what soc/ maintainers prefer, but inverted xmas tree would look/read
> better.
> > + priv = devm_kzalloc(dev, sizeof(struct a64fx_diag_priv),
> > +GFP_KERNEL);
>
> Don't we prefer sizeof(*priv)?
> > + ret = request_irq(priv->irq, &a64fx_diag_handler_irq,
> > + irq_flags, "a64fx_diag_irq", NULL);
> > + if (ret) {
> > + dev_err(dev, "cannot register IRQ %d\n", ret);
>
> No a64fx_diag_interrupt_disable()?
> > + priv->has_nmi = false;
>
> No need to set zeroed priv member to zero.
I understand. I will fix it as per your comment. Thank you.
> > + enable_nmi(priv->irq);
>
> Provided the above, I don't immediatelly see, what's the purpose of
> IRQF_NO_AUTOEN then?
It seems that request_nmi() requires IRQF_NO_AUTOEN.
> > +static int __exit a64fx_diag_remove(struct platform_device *pdev)
>
> Is __exit appropriate here at all -- I doubt that.
I will remove __exit as it seems unnecessary as you suggested.
Also, I will correct BMC_DIAG_INTERRUPT_STATUS_OFFSET
and BMC_DIAG_INTERRUPT_ENABLE_OFFSET.
Thank you.
Hitomi Hasegawa