Fix ECC error counting for AMD76x chipset, char/ecc.c driver

Summary:
* Patch is relevant to driver "char/ecc.c", for the AMD76x Athlon chipset.

* Write 1 bits, not 0 bits, to clear the ECC error flag status.
* Without this patch, Linux will detect ONLY the first single and multibit error. All subsequent errors are ignored by the hardware until the register is properly cleared.
* The patch also fast-paths the common polled operation, and simplifies code.
* Includes reference to AMD 761 spec sheet documenting the ECC register values.

* Note: this module is often not installed by default: do "modprobe ecc" then "cat /proc/ram" to check your ECC memory for detected soft errors.
* Patch is against Linux-2.6.13, the last kernel I could find with ecc.c
* Tabs suck.

I am an infrequent contributor, and did not find a matching entry in the ./MAINTAINERS file. Please help me to understand the proper procedure for submitting this patch. I understand that perhaps ecc.c is changing soon. So maybe it's not the most important patch, but it does fix a real bug, and is quite simple.

----------------------------------------------------------------------------------------
linux:/usr/src # diff -u -b orig/drivers/char/ecc.c linux-2.6.13-15/drivers/char/ecc.c > /tmp/ecc.diff linux:/usr/src # cat /tmp/ecc.diff

@@ -10,6 +10,7 @@
  */
 #define DEBUG  0
 
+
 #include <linux/config.h>
 #include <linux/version.h>
 #include <linux/module.h>

@@ -22,7 +23,7 @@
 #include <asm/io.h>
 #include <linux/proc_fs.h>
 
-#define        ECC_VER "0.14 (Oct 10 2001)"
+#define        ECC_VER "0.15 (Dec 1 2005)"
 #define KERN_ECC KERN_ALERT
 
 static struct timer_list ecctimer;

@@ -1102,15 +1103,20 @@
        }
 }
 
+
+// Spec source: AMD 761 System Controller/BIOS Guide, 24081D-February 2002
+//

http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/24081.pdf
void check_amd76x(void)
{

-       unsigned long eccstat = pci_dword(0x48);
+       u32 eccstat;
+       pci_read_config_dword(bridge, 0x48, &eccstat);
+       if(eccstat & 0x30)
+       {
        if(eccstat & 0x10)
        {
                /* bits 7-4 of eccstat indicate the row the MBE occurred. */
                int row = (eccstat >> 4) & 0xf;
                printk("<1>ECC: MBE Detected in DRAM row %d\n", row);
-               scrub_needed |= 2;
                bank[row].mbecount++;
        }
        if(eccstat & 0x20)
@@ -1118,22 +1124,9 @@
                /* bits 3-0 of eccstat indicate the row the SBE occurred. */
                int row = eccstat & 0xf;
                printk("<1>ECC: SBE Detected in DRAM row %d\n", row);
-               scrub_needed |= 1;
                bank[row].sbecount++;
        }
-       if (scrub_needed)
-       {
-               /*
-                * clear error flag bits that were set by writing 0 to them
-                * we hope the error was a fluke or something :)
-                */

- unsigned long value = eccstat;
- if (scrub_needed & 1)
- value &= 0xFFFFFDFF;
- if (scrub_needed & 2)
- value &= 0xFFFFFEFF;

-               pci_write_config_dword(bridge, 0x48, value);
-               scrub_needed = 0;
+               pci_write_config_dword(bridge, 0x48, eccstat);  // clear
by writing a 1
        }

}

drivers/char/ecc.c only lives in various vendor kernels.

you may want to check if your patch is relevant to the edac code present in -mm though.

Dave

Thanks.

The EDAC code appears to do this correctly, though it is much heavier code (involving multiple functions, and pci and irq locking that appears correct, but unnecessarily pedantic).