Re: [PATCH 6/6] edac, mce_amd_inj: Inject errors on NBC for bank 4 errors

[Borislav Petkov <bp@alien8.de>]

On Wed, May 27, 2015 at 02:03:38PM -0500, Aravind Gopalakrishnan wrote:
> For bank 4 errors, MCE logging and reporting is done only on
> node base cores. Refer D18F3x44[NbMcaToMstCpuEn] field in
> Fam10h and later BKDGs.
> 
> This patch ensures that we inject the error on the node base core
> for bank 4 errors. Otherwise, triggering #MC or apic interrupts on
> a non node base core would not have any effect on the system.
> (i.e), we would not see any relevant output on kernel logs for
> the error we just injected.
> 
> Signed-off-by: Aravind Gopalakrishnan <Aravind.Gopalakrishnan@amd.com>
> ---
>  drivers/edac/mce_amd_inj.c | 76 ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 76 insertions(+)
> 
> diff --git a/drivers/edac/mce_amd_inj.c b/drivers/edac/mce_amd_inj.c
> index ca5b29f..bd44ba4 100644
> --- a/drivers/edac/mce_amd_inj.c
> +++ b/drivers/edac/mce_amd_inj.c
> @@ -17,9 +17,12 @@
>  #include <linux/cpu.h>
>  #include <linux/string.h>
>  #include <linux/uaccess.h>
> +#include <linux/pci.h>
>  #include <asm/mce.h>
> +#include <asm/amd_nb.h>
>  
>  #include "mce_amd.h"
> +#include "amd64_edac.h"
>  
>  /*
>   * Collect all the MCi_XXX settings
> @@ -205,6 +208,65 @@ static void trigger_thr_int(void *info)
>  	asm volatile("int $249");
>  }
>  
> +static u32 amd_get_num_nodes(void)
> +{
> +	u32 nodes = 1;
> +
> +	if (cpu_has_topoext) {
> +		u32 ecx;
> +
> +		ecx = cpuid_ecx(0x8000001e);
> +		nodes = ((ecx >> 8) & 7) + 1;
> +	} else if (static_cpu_has(X86_FEATURE_NODEID_MSR)) {
> +		u64 value;
> +
> +		rdmsrl(MSR_FAM10H_NODE_ID, value);
> +		nodes = ((value >> 3) & 7) + 1;
> +	}

So we already do that (and more) in amd_get_topology(). I'm thinking
you should take this function out of the CONFIG_X86_HT ifdeffery (also
in its caller amd_detect_cmp()) and you should save "nodes" in a local
static variable called nodes_per_processor and a small accessor called
amd_get_nodes_cnt() should return it. Similar to amd_get_nb_id().

Don't forget to add a comment explaning what that nodes_per_processor
means.

And then amd_mce_inj.c will simply use it instead of duplicating that
information here.

Please do that in 2 pre-patches.

> +
> +	return nodes;
> +}
> +
> +static u32 amd_get_nbc_for_node(int node_id)
> +{
> +	struct cpuinfo_x86 *c = &boot_cpu_data;
> +	u32 nodes, cores_per_node;
> +
> +	nodes = amd_get_num_nodes();
> +	cores_per_node = c->x86_max_cores / nodes;
> +
> +	return cores_per_node * node_id;
> +}
> +
> +static void toggle_nb_mca_mst_cpu(u16 nid)
> +{
> +	struct pci_dev *F3 = node_to_amd_nb(nid)->misc;
> +	u32 val;
> +	int err;
> +
> +	if (!F3)
> +		return;
> +
> +	err = pci_read_config_dword(F3, NBCFG, &val);
> +	if (err) {
> +		pr_err("%s: Error reading F%dx%03x.\n", __func__,
> +			PCI_FUNC(F3->devfn),
> +			NBCFG);
> +		return;
> +	}
> +
> +	if (!(val & BIT(27))) {
> +		pr_err("%s: BIOS not setting D18F3x44[NbMcaToMstCpuEn]."
> +		       "Doing that here\n", __func__);

WARNING: quoted string split across lines
#99: FILE: drivers/edac/mce_amd_inj.c:260:
+               pr_err("%s: BIOS not setting D18F3x44[NbMcaToMstCpuEn]."
+                      "Doing that here\n", __func__);

Do integrate checkpatch.pl into your workflow. It is sometimes right.

> +		val |= BIT(27);
> +		err = pci_write_config_dword(F3, NBCFG, val);
> +		if (err)
> +			pr_err("%s: Error writing F%dx%03x.\n", __func__,
> +				PCI_FUNC(F3->devfn),
> +				NBCFG);
> +	}
> +}

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--