SB1 cache exception handling.
authorAndrew Isaacson <adi@broadcom.com>
Thu, 20 Oct 2005 06:57:40 +0000 (23:57 -0700)
committerRalf Baechle <ralf@linux-mips.org>
Sat, 29 Oct 2005 18:32:48 +0000 (19:32 +0100)
Expand SB1 cache error handling by adding SB1_CEX_ALWAYS_FATAL and
SB1_CEX_STALL, allowing configurable behavior on cache errors.

Signed-Off-By: Andy Isaacson <adi@broadcom.com>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
arch/mips/mm/cerr-sb1.c
arch/mips/mm/cex-sb1.S
arch/mips/sibyte/Kconfig

index 7166ffe635021a4124e8e9556791fc9cafafedb2..1cf3c6006ccd38f2ecaea4247489cdcdea413087 100644 (file)
 #include <linux/sched.h>
 #include <asm/mipsregs.h>
 #include <asm/sibyte/sb1250.h>
+#include <asm/sibyte/sb1250_regs.h>
 
-#ifndef CONFIG_SIBYTE_BUS_WATCHER
+#if !defined(CONFIG_SIBYTE_BUS_WATCHER) || defined(CONFIG_SIBYTE_BW_TRACE)
 #include <asm/io.h>
-#include <asm/sibyte/sb1250_regs.h>
 #include <asm/sibyte/sb1250_scd.h>
 #endif
 
+/*
+ * We'd like to dump the L2_ECC_TAG register on errors, but errata make
+ * that unsafe... So for now we don't.  (BCM1250/BCM112x erratum SOC-48.)
+ */
+#undef DUMP_L2_ECC_TAG_ON_ERROR
+
 /* SB1 definitions */
 
 /* XXX should come from config1 XXX */
@@ -139,12 +145,18 @@ static inline void breakout_cerrd(unsigned int val)
 static void check_bus_watcher(void)
 {
        uint32_t status, l2_err, memio_err;
+#ifdef DUMP_L2_ECC_TAG_ON_ERROR
+       uint64_t l2_tag;
+#endif
 
        /* Destructive read, clears register and interrupt */
        status = csr_in32(IOADDR(A_SCD_BUS_ERR_STATUS));
        /* Bit 31 is always on, but there's no #define for that */
        if (status & ~(1UL << 31)) {
                l2_err = csr_in32(IOADDR(A_BUS_L2_ERRORS));
+#ifdef DUMP_L2_ECC_TAG_ON_ERROR
+               l2_tag = in64(IO_SPACE_BASE | A_L2_ECC_TAG);
+#endif
                memio_err = csr_in32(IOADDR(A_BUS_MEM_IO_ERRORS));
                prom_printf("Bus watcher error counters: %08x %08x\n", l2_err, memio_err);
                prom_printf("\nLast recorded signature:\n");
@@ -153,6 +165,9 @@ static void check_bus_watcher(void)
                       (int)(G_SCD_BERR_TID(status) >> 6),
                       (int)G_SCD_BERR_RID(status),
                       (int)G_SCD_BERR_DCODE(status));
+#ifdef DUMP_L2_ECC_TAG_ON_ERROR
+               prom_printf("Last L2 tag w/ bad ECC: %016llx\n", l2_tag);
+#endif
        } else {
                prom_printf("Bus watcher indicates no error\n");
        }
@@ -166,6 +181,16 @@ asmlinkage void sb1_cache_error(void)
        uint64_t cerr_dpa;
        uint32_t errctl, cerr_i, cerr_d, dpalo, dpahi, eepc, res;
 
+#ifdef CONFIG_SIBYTE_BW_TRACE
+       /* Freeze the trace buffer now */
+#if defined(CONFIG_SIBYTE_BCM1x55) || defined(CONFIG_SIBYTE_BCM1x80)
+       csr_out32(M_BCM1480_SCD_TRACE_CFG_FREEZE, IO_SPACE_BASE | A_SCD_TRACE_CFG);
+#else
+       csr_out32(M_SCD_TRACE_CFG_FREEZE, IO_SPACE_BASE | A_SCD_TRACE_CFG);
+#endif
+       prom_printf("Trace buffer frozen\n");
+#endif
+
        prom_printf("Cache error exception on CPU %x:\n",
                    (read_c0_prid() >> 25) & 0x7);
 
@@ -229,11 +254,19 @@ asmlinkage void sb1_cache_error(void)
 
        check_bus_watcher();
 
-       while (1);
        /*
-        * This tends to make things get really ugly; let's just stall instead.
-        *    panic("Can't handle the cache error!");
+        * Calling panic() when a fatal cache error occurs scrambles the
+        * state of the system (and the cache), making it difficult to
+        * investigate after the fact.  However, if you just stall the CPU,
+        * the other CPU may keep on running, which is typically very
+        * undesirable.
         */
+#ifdef CONFIG_SB1_CERR_STALL
+       while (1)
+               ;
+#else
+       panic("unhandled cache error");
+#endif
 }
 
 
@@ -434,7 +467,8 @@ static struct dc_state dc_states[] = {
 };
 
 #define DC_TAG_VALID(state) \
-    (((state) == 0xf) || ((state) == 0x13) || ((state) == 0x19) || ((state == 0x16)) || ((state) == 0x1c))
+    (((state) == 0x0) || ((state) == 0xf) || ((state) == 0x13) || \
+     ((state) == 0x19) || ((state) == 0x16) || ((state) == 0x1c))
 
 static char *dc_state_str(unsigned char state)
 {
@@ -505,6 +539,7 @@ static uint32_t extract_dc(unsigned short addr, int data)
                        uint64_t datalo;
                        uint32_t datalohi, datalolo, datahi;
                        int offset;
+                       char bad_ecc = 0;
 
                        for (offset = 0; offset < 4; offset++) {
                                /* Index-load-data-D */
@@ -525,8 +560,7 @@ static uint32_t extract_dc(unsigned short addr, int data)
                                ecc = dc_ecc(datalo);
                                if (ecc != datahi) {
                                        int bits = 0;
-                                       prom_printf("  ** bad ECC (%02x %02x) ->",
-                                                   datahi, ecc);
+                                       bad_ecc |= 1 << (3-offset);
                                        ecc ^= datahi;
                                        while (ecc) {
                                                if (ecc & 1) bits++;
@@ -537,6 +571,10 @@ static uint32_t extract_dc(unsigned short addr, int data)
                                prom_printf("  %02X-%016llX", datahi, datalo);
                        }
                        prom_printf("\n");
+                       if (bad_ecc)
+                               prom_printf("  dwords w/ bad ECC: %d %d %d %d\n",
+                                           !!(bad_ecc & 8), !!(bad_ecc & 4),
+                                           !!(bad_ecc & 2), !!(bad_ecc & 1));
                }
        }
        return res;
index 2c3a23aa88c35ac5d30e75a7c3fbb226e9e6b861..0e71580774fff3673401f7b73e371af5dd123721 100644 (file)
@@ -64,6 +64,10 @@ LEAF(except_vec2_sb1)
        sd      k0,0x170($0)
        sd      k1,0x178($0)
 
+#if CONFIG_SB1_CEX_ALWAYS_FATAL
+       j       handle_vec2_sb1
+        nop
+#else
        /*
         * M_ERRCTL_RECOVERABLE is bit 31, which makes it easy to tell
         * if we can fast-path out of here for a h/w-recovered error.
@@ -134,6 +138,7 @@ unrecoverable:
        /* Unrecoverable Icache or Dcache error; log it and/or fail */
        j       handle_vec2_sb1
         nop
+#endif
 
 END(except_vec2_sb1)
 
index 6a5a08f5e212701c8aec453c1760c582b1b642f8..de46f62ac46291c28af5d1d05fcc6ef5d3baa1a7 100644 (file)
@@ -102,6 +102,14 @@ config SIMULATION
          Build a kernel suitable for running under the GDB simulator.
          Primarily adjusts the kernel's notion of time.
 
+config CONFIG_SB1_CEX_ALWAYS_FATAL
+       bool "All cache exceptions considered fatal (no recovery attempted)"
+       depends on SIBYTE_SB1xxx_SOC
+
+config CONFIG_SB1_CERR_STALL
+       bool "Stall (rather than panic) on fatal cache error"
+       depends on SIBYTE_SB1xxx_SOC
+
 config SIBYTE_CFE
        bool "Booting from CFE"
        depends on SIBYTE_SB1xxx_SOC