Blackfin: allow high priority domains to preempt schedule_tail()

ret_from_fork is always entered with hw interrupts off, which prevents
real-time domains to preempt the Linux kernel during part of the
initial context switch to the new task, which could in turn raise the
worst-case latency figures.

To avoid this, stall the root domain stage in the interrupt pipeline
to keep the scheduling tail code free from Linux-handled IRQs, then
enable hardware interrupts again.

Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>

1 files changed, 22 insertions, 2 deletions

diff --git a/arch/blackfin/kernel/entry.S b/arch/blackfin/kernel/entry.S
index a9cfba9946b5..3f8769b7db54 100644
--- a/arch/blackfin/kernel/entry.S
+++ b/arch/blackfin/kernel/entry.S
@@ -43,8 +43,28 @@
 
 ENTRY(_ret_from_fork)
 #ifdef CONFIG_IPIPE
-        [--sp] = reti;          /* IRQs on. */
-        SP += 4;
+        /*
+         * Hw IRQs are off on entry, and we don't want the scheduling tail
+         * code to starve high priority domains from interrupts while it
+         * runs. Therefore we first stall the root stage to have the
+         * virtual interrupt state reflect IMASK.
+         */
+        p0.l = ___ipipe_root_status;
+        p0.h = ___ipipe_root_status;
+        r4 = [p0];
+        bitset(r4, 0);
+        [p0] = r4;
+        /*
+         * Then we may enable hw IRQs, allowing preemption from high
+         * priority domains. schedule_tail() will do local_irq_enable()
+         * since Blackfin does not define __ARCH_WANT_UNLOCKED_CTXSW, so
+         * there is no need to unstall the root domain by ourselves
+         * afterwards.
+         */
+        p0.l = _bfin_irq_flags;
+        p0.h = _bfin_irq_flags;
+        r4 = [p0];
+        sti r4;
 #endif /* CONFIG_IPIPE */
         SP += -12;
         call _schedule_tail;