Commit f2a5e8f0 authored Oct 24, 2016 by Mahesh Salgaonkar Committed by Michael Ellerman Jan 30, 2017

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>

parent 79270e0a

arch/powerpc/kernel/fadump.c

+24 −1

Original line number	Diff line number	Diff line
		@@ -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();

Admin message