Re: [PATCHv2 2/2] drm/amdgpu: Register MCE notifier for Aldebaran RAS

[Borislav Petkov <bp@alien8.de>]

On Sun, Sep 12, 2021 at 10:13:11PM -0400, Mukul Joshi wrote:
> On Aldebaran, GPU driver will handle bad page retirement
> even though UMC is host managed. As a result, register a
> bad page retirement handler on the mce notifier chain to
> retire bad pages on Aldebaran.
> 
> v1->v2:
> - Use smca_get_bank_type() to determine MCA bank.
> - Envelope the changes under #ifdef CONFIG_X86_MCE_AMD.
> - Use MCE_PRIORITY_UC instead of MCE_PRIO_ACCEL as we are
>   only handling uncorrectable errors.
> - Use macros to determine UMC instance and channel instance
>   where the uncorrectable error occured.
> - Update the headline.

Same note as for the previous patch.

> +static int amdgpu_bad_page_notifier(struct notifier_block *nb,
> +				    unsigned long val, void *data)
> +{
> +	struct mce *m = (struct mce *)data;
> +	struct amdgpu_device *adev = NULL;
> +	uint32_t gpu_id = 0;
> +	uint32_t umc_inst = 0;
> +	uint32_t ch_inst, channel_index = 0;
> +	struct ras_err_data err_data = {0, 0, 0, NULL};
> +	struct eeprom_table_record err_rec;
> +	uint64_t retired_page;
> +
> +	/*
> +	 * If the error was generated in UMC_V2, which belongs to GPU UMCs,
> +	 * and error occurred in DramECC (Extended error code = 0) then only
> +	 * process the error, else bail out.
> +	 */
> +	if (!m || !((smca_get_bank_type(m->bank) == SMCA_UMC_V2) &&
> +		    (XEC(m->status, 0x1f) == 0x0)))
> +		return NOTIFY_DONE;
> +
> +	/*
> +	 * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
> +	 */
> +	gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
> +
> +	adev = find_adev(gpu_id);
> +	if (!adev) {
> +		dev_warn(adev->dev, "%s: Unable to find adev for gpu_id: %d\n",
> +				     __func__, gpu_id);
> +		return NOTIFY_DONE;
> +	}
> +
> +	/*
> +	 * If it is correctable error, return.
> +	 */
> +	if (mce_is_correctable(m)) {
> +		return NOTIFY_OK;
> +	}

This can run before you find_adev().

> +static void amdgpu_register_bad_pages_mca_notifier(void)
> +{
> +	/*
> +	 * Register the x86 notifier only once
> +	 * with MCE subsystem.
> +	 */
> +	if (notifier_registered == false) {
> +		mce_register_decode_chain(&amdgpu_bad_page_nb);
> +		notifier_registered = true;
> +	}

I have a patchset which will get rid of the need to do this silliness -
only if I had some time to actually prepare it for submission... :-\

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette