powerpc/fadump: Fix the race in crash_fadump().

There are chances that multiple CPUs can call crash_fadump() simultaneously
and would start duplicating same info to vmcoreinfo ELF note section. This
causes makedumpfile to fail during kdump capture. One example is,
triggering dumprestart from HMC which sends system reset to all the CPUs at
once.

makedumpfile --dump-dmesg /proc/vmcore
read_vmcoreinfo_basic_info: Invalid data in /tmp/vmcoreinfoyjgxlL: CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971
makedumpfile Failed.
Running makedumpfile --dump-dmesg /proc/vmcore failed (1).

makedumpfile  -d 31 -l /proc/vmcore
read_vmcoreinfo_basic_info: Invalid data in /tmp/vmcoreinfo1mmVdO: CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971CRASHTIME=1475605971
makedumpfile Failed.
Running makedumpfile  -d 31 -l /proc/vmcore failed (1).

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

diff --git a/arch/powerpc/kernel/fadump.c b/arch/powerpc/kernel/fadump.c
index 8f0c7c5d93f2de840826bc22f115d5528c20ea1c..8ff0dd4e77a7f3cd8d2459a4eb989c21dc80f3cc 100644 (file)
--- a/arch/powerpc/kernel/fadump.c
+++ b/arch/powerpc/kernel/fadump.c
@@ -406,12 +406,35 @@ static void register_fw_dump(struct fadump_mem_struct *fdm)
 void crash_fadump(struct pt_regs *regs, const char *str)
 {
        struct fadump_crash_info_header *fdh = NULL;
+       int old_cpu, this_cpu;
 
        if (!fw_dump.dump_registered || !fw_dump.fadumphdr_addr)
                return;
 
+       /*
+        * old_cpu == -1 means this is the first CPU which has come here,
+        * go ahead and trigger fadump.
+        *
+        * old_cpu != -1 means some other CPU has already on it's way
+        * to trigger fadump, just keep looping here.
+        */
+       this_cpu = smp_processor_id();
+       old_cpu = cmpxchg(&crashing_cpu, -1, this_cpu);
+
+       if (old_cpu != -1) {
+               /*
+                * We can't loop here indefinitely. Wait as long as fadump
+                * is in force. If we race with fadump un-registration this
+                * loop will break and then we go down to normal panic path
+                * and reboot. If fadump is in force the first crashing
+                * cpu will definitely trigger fadump.
+                */
+               while (fw_dump.dump_registered)
+                       cpu_relax();
+               return;
+       }
+
        fdh = __va(fw_dump.fadumphdr_addr);
-       crashing_cpu = smp_processor_id();
        fdh->crashing_cpu = crashing_cpu;
        crash_save_vmcoreinfo();