[PATCH] powerpc: Fix deadlock with multiple calls to smp_send_stop
From: Nicholas Piggin <npiggin@gmail.com>
Date: 2018-04-25 12:17:15
Subsystem:
linux for powerpc (32-bit and 64-bit), the rest · Maintainers:
Madhavan Srinivasan, Michael Ellerman, Linus Torvalds
smp_send_stop can lock up the IPI path for any subsequent calls,
because the receiving CPUs spin in their handler function. This
started becoming a problem with the addition of an smp_send_stop
call in the reboot path, because panics can reboot after doing
their own smp_send_stop.
The NMI IPI handler for a receiving CPU increments nmi_ipi_busy_count
over the handler function call, which causes the next
smp_send_nmi_ipi() caller to spin. Fix this by adding a special case
to the smp_send_stop handler to decrement the busy count, because it
will never return.
The smp_call_function path (used when !CONFIG_NMI_IPI) suffers from
a similar deadlock. This is fixed by having smp_send_stop only ever
do the smp_call_function once. This is a bit less robust, because
any other use of smp_call_function after smp_send_stop could deadlock,
but that hasn't been a problem before. Fixing that would take a bit
more code.
Fixes: f2748bdfe1573 ("powerpc/powernv: Always stop secondaries before reboot/shutdown")
Reported-by: Abdul Haleem <redacted>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
This supersedes "[PATCH v2] powerpc: Fix smp_send_stop NMI IPI
handling". I got the root cause for that wrong, and missed the
non-NMI case that is also affected.
arch/powerpc/kernel/smp.c | 49 +++++++++++++++++++++++++++++++++------
1 file changed, 42 insertions(+), 7 deletions(-)
diff --git a/arch/powerpc/kernel/smp.c b/arch/powerpc/kernel/smp.c
index e16ec7b3b427..9ca7148b5881 100644
--- a/arch/powerpc/kernel/smp.c
+++ b/arch/powerpc/kernel/smp.c@@ -566,10 +566,35 @@ void crash_send_ipi(void (*crash_ipi_callback)(struct pt_regs *)) #endif #ifdef CONFIG_NMI_IPI -static void stop_this_cpu(struct pt_regs *regs) -#else +static void nmi_stop_this_cpu(struct pt_regs *regs) +{ + /* + * This is a special case because it never returns, so the NMI IPI + * handling would never mark it as done, which makes any later + * smp_send_nmi_ipi() call spin forever. Mark it done now. + * + * IRQs are already hard disabled by the smp_handle_nmi_ipi. + */ + nmi_ipi_lock(); + nmi_ipi_busy_count--; + nmi_ipi_unlock(); + + /* Remove this CPU */ + set_cpu_online(smp_processor_id(), false); + + spin_begin(); + while (1) + spin_cpu_relax(); +} + +void smp_send_stop(void) +{ + smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, nmi_stop_this_cpu, 1000000); +} + +#else /* CONFIG_NMI_IPI */ + static void stop_this_cpu(void *dummy) -#endif { /* Remove this CPU */ set_cpu_online(smp_processor_id(), false);
@@ -582,12 +607,22 @@ static void stop_this_cpu(void *dummy) void smp_send_stop(void) { -#ifdef CONFIG_NMI_IPI - smp_send_nmi_ipi(NMI_IPI_ALL_OTHERS, stop_this_cpu, 1000000); -#else + static bool stopped = false; + + /* + * Prevent waiting on csd lock from a previous smp_send_stop. + * This is racy, but in general callers try to do the right + * thing and only fire off one smp_send_stop (e.g., see + * kernel/panic.c) + */ + if (stopped) + return; + + stopped = true; + smp_call_function(stop_this_cpu, NULL, 0); -#endif } +#endif /* CONFIG_NMI_IPI */ struct thread_info *current_set[NR_CPUS];
--
2.17.0