Re: [PATCH V3 09/22] LoongArch: Add boot and setup routines

Re: [PATCH V3 09/22] LoongArch: Add boot and setup routines

[Arnd Bergmann]

On Fri, Sep 17, 2021 at 5:57 AM Huacai Chen <chenhuacai@loongson.cn> wrote:
> This patch adds basic boot, setup and reset routines for LoongArch.
> LoongArch uses UEFI-based firmware and uses ACPI as the boot protocol.

This needs to be reviewed by the maintainers for the EFI and ACPI subsystems,
I added them to Cc here. If you add lines like

Cc: Ard Biesheuvel <ardb@kernel.org>
Cc: linux-efi@vger.kernel.org

in the patch description before your Signed-off-by, then git-send-email will
Cc them automatically without you having to spam them with the entire series.

In particular, I know that Ard previously complained that you did not use the
EFI boot protocol correctly, and I want to make sure that he's happy with the
final version.

> +static ssize_t boardinfo_show(struct kobject *kobj,
> +                             struct kobj_attribute *attr, char *buf)
> +{
> +       return sprintf(buf,
> +               "BIOS Information\n"
> +               "Vendor\t\t\t: %s\n"
> +               "Version\t\t\t: %s\n"
> +               "ROM Size\t\t: %d KB\n"
> +               "Release Date\t\t: %s\n\n"
> +               "Board Information\n"
> +               "Manufacturer\t\t: %s\n"
> +               "Board Name\t\t: %s\n"
> +               "Family\t\t\t: LOONGSON64\n\n",
> +               b_info.bios_vendor, b_info.bios_version,
> +               b_info.bios_size, b_info.bios_release_date,
> +               b_info.board_vendor, b_info.board_name);
> +}
> +
> +static struct kobj_attribute boardinfo_attr = __ATTR(boardinfo, 0444,
> +                                                    boardinfo_show, NULL);
> +
> +static int __init boardinfo_init(void)
> +{
> +       if (!efi_kobj)
> +               return -EINVAL;
> +
> +       return sysfs_create_file(efi_kobj, &boardinfo_attr.attr);
> +}
> +late_initcall(boardinfo_init);

I see you have documented this interface for your mips machines,
but nothing else uses it.

I think some of this information should be part of the soc_device,
either in addition to, or in place of this sysfs file.

Isn't there an existing method to do this on x86/arm/ia64 machines?

> +static int constant_set_state_periodic(struct clock_event_device *evt)
> +{
> +       unsigned long period;
> +       unsigned long timer_config;
> +
> +       raw_spin_lock(&state_lock);
> +
> +       period = const_clock_freq / HZ;
> +       timer_config = period & CSR_TCFG_VAL;
> +       timer_config |= (CSR_TCFG_PERIOD | CSR_TCFG_EN);
> +       csr_writeq(timer_config, LOONGARCH_CSR_TCFG);
> +
> +       raw_spin_unlock(&state_lock);

I see this pattern in a couple of places, using a spinlock or raw_spinlock
to guard MMIO access, but on many architectures a register write is
not serialized by the following spin_unlock, unless you insert another
read from the same address in there. E.g. on PCIe, writes are always
posted and it would not work.

Can you confirm that it works correctly on CSR registers in loongarch?

         Arnd

Re: [PATCH V3 09/22] LoongArch: Add boot and setup routines

[Huacai Chen]

Hi, Arnd,

On Fri, Sep 17, 2021 at 4:11 PM Arnd Bergmann <arnd@arndb.de> wrote:
>
> On Fri, Sep 17, 2021 at 5:57 AM Huacai Chen <chenhuacai@loongson.cn> wrote:
> > This patch adds basic boot, setup and reset routines for LoongArch.
> > LoongArch uses UEFI-based firmware and uses ACPI as the boot protocol.
>
> This needs to be reviewed by the maintainers for the EFI and ACPI subsystems,
> I added them to Cc here. If you add lines like
>
> Cc: Ard Biesheuvel <ardb@kernel.org>
> Cc: linux-efi@vger.kernel.org
>
> in the patch description before your Signed-off-by, then git-send-email will
> Cc them automatically without you having to spam them with the entire series.
OK, I will add them.

>
> In particular, I know that Ard previously complained that you did not use the
> EFI boot protocol correctly, and I want to make sure that he's happy with the
> final version.
The Correct way means efistub?  We have investigated for some time and
found that it is very difficult. E.g., our BIOS team said that they
cannot get GOP drivers for graphics cards.

>
> > +static ssize_t boardinfo_show(struct kobject *kobj,
> > +                             struct kobj_attribute *attr, char *buf)
> > +{
> > +       return sprintf(buf,
> > +               "BIOS Information\n"
> > +               "Vendor\t\t\t: %s\n"
> > +               "Version\t\t\t: %s\n"
> > +               "ROM Size\t\t: %d KB\n"
> > +               "Release Date\t\t: %s\n\n"
> > +               "Board Information\n"
> > +               "Manufacturer\t\t: %s\n"
> > +               "Board Name\t\t: %s\n"
> > +               "Family\t\t\t: LOONGSON64\n\n",
> > +               b_info.bios_vendor, b_info.bios_version,
> > +               b_info.bios_size, b_info.bios_release_date,
> > +               b_info.board_vendor, b_info.board_name);
> > +}
> > +
> > +static struct kobj_attribute boardinfo_attr = __ATTR(boardinfo, 0444,
> > +                                                    boardinfo_show, NULL);
> > +
> > +static int __init boardinfo_init(void)
> > +{
> > +       if (!efi_kobj)
> > +               return -EINVAL;
> > +
> > +       return sysfs_create_file(efi_kobj, &boardinfo_attr.attr);
> > +}
> > +late_initcall(boardinfo_init);
>
> I see you have documented this interface for your mips machines,
> but nothing else uses it.
>
> I think some of this information should be part of the soc_device,
> either in addition to, or in place of this sysfs file.
This file list something describe the motherboard, which is different
from SOC. These information are used by some user programs.

>
> Isn't there an existing method to do this on x86/arm/ia64 machines?
>
> > +static int constant_set_state_periodic(struct clock_event_device *evt)
> > +{
> > +       unsigned long period;
> > +       unsigned long timer_config;
> > +
> > +       raw_spin_lock(&state_lock);
> > +
> > +       period = const_clock_freq / HZ;
> > +       timer_config = period & CSR_TCFG_VAL;
> > +       timer_config |= (CSR_TCFG_PERIOD | CSR_TCFG_EN);
> > +       csr_writeq(timer_config, LOONGARCH_CSR_TCFG);
> > +
> > +       raw_spin_unlock(&state_lock);
>
> I see this pattern in a couple of places, using a spinlock or raw_spinlock
> to guard MMIO access, but on many architectures a register write is
> not serialized by the following spin_unlock, unless you insert another
> read from the same address in there. E.g. on PCIe, writes are always
> posted and it would not work.
>
> Can you confirm that it works correctly on CSR registers in loongarch?
CSR on LoongArch doesn't need any barrier or flush operations,
spinlock here is used to protect the whole "read, modify and write".

Huacai
>
>          Arnd