Blackfin/ipipe: upgrade to I-pipe mainline

This patch introduces Blackfin-specific bits to support the current
tip of the interrupt pipeline development, mainly:

- 2/3-level interrupt maps (sparse IRQs)
- generic virq handling
- sysinfo v2 format for ipipe_get_sysinfo()

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

1 files changed, 37 insertions, 4 deletions

diff --git a/arch/blackfin/mach-common/ints-priority.c b/arch/blackfin/mach-common/ints-priority.c
index 8e9d3cc30885..6cd52395a999 100644
--- a/arch/blackfin/mach-common/ints-priority.c
+++ b/arch/blackfin/mach-common/ints-priority.c
@@ -15,6 +15,7 @@
 #include <linux/kernel_stat.h>
 #include <linux/seq_file.h>
 #include <linux/irq.h>
+#include <linux/sched.h>
 #ifdef CONFIG_IPIPE
 #include <linux/ipipe.h>
 #endif
@@ -556,10 +557,9 @@ static void bfin_demux_mac_status_irq(unsigned int int_err_irq,
 static inline void bfin_set_irq_handler(unsigned irq, irq_flow_handler_t handle)
 {
 #ifdef CONFIG_IPIPE
-        _set_irq_handler(irq, handle_level_irq);
-#else
-        __set_irq_handler_unlocked(irq, handle);
+        handle = handle_level_irq;
 #endif
+        __set_irq_handler_unlocked(irq, handle);
 }
 
 static DECLARE_BITMAP(gpio_enabled, MAX_BLACKFIN_GPIOS);
@@ -1392,7 +1392,7 @@ asmlinkage int __ipipe_grab_irq(int vec, struct pt_regs *regs)
         struct ipipe_domain *this_domain = __ipipe_current_domain;
         struct ivgx *ivg_stop = ivg7_13[vec-IVG7].istop;
         struct ivgx *ivg = ivg7_13[vec-IVG7].ifirst;
-        int irq, s;
+        int irq, s = 0;
 
         if (likely(vec == EVT_IVTMR_P))
                 irq = IRQ_CORETMR;
@@ -1442,6 +1442,21 @@ asmlinkage int __ipipe_grab_irq(int vec, struct pt_regs *regs)
                         __raw_get_cpu_var(__ipipe_tick_regs).ipend |= 0x10;
         }
 
+        /*
+         * We don't want Linux interrupt handlers to run at the
+         * current core priority level (i.e. < EVT15), since this
+         * might delay other interrupts handled by a high priority
+         * domain. Here is what we do instead:
+         *
+         * - we raise the SYNCDEFER bit to prevent
+         * __ipipe_handle_irq() to sync the pipeline for the root
+         * stage for the incoming interrupt. Upon return, that IRQ is
+         * pending in the interrupt log.
+         *
+         * - we raise the TIF_IRQ_SYNC bit for the current thread, so
+         * that _schedule_and_signal_from_int will eventually sync the
+         * pipeline from EVT15.
+         */
         if (this_domain == ipipe_root_domain) {
                 s = __test_and_set_bit(IPIPE_SYNCDEFER_FLAG, &p->status);
                 barrier();
@@ -1451,6 +1466,24 @@ asmlinkage int __ipipe_grab_irq(int vec, struct pt_regs *regs)
         __ipipe_handle_irq(irq, regs);
         ipipe_trace_irq_exit(irq);
 
+        if (user_mode(regs) &&
+            !ipipe_test_foreign_stack() &&
+            (current->ipipe_flags & PF_EVTRET) != 0) {
+                /*
+                 * Testing for user_regs() does NOT fully eliminate
+                 * foreign stack contexts, because of the forged
+                 * interrupt returns we do through
+                 * __ipipe_call_irqtail. In that case, we might have
+                 * preempted a foreign stack context in a high
+                 * priority domain, with a single interrupt level now
+                 * pending after the irqtail unwinding is done. In
+                 * which case user_mode() is now true, and the event
+                 * gets dispatched spuriously.
+                 */
+                current->ipipe_flags &= ~PF_EVTRET;
+                __ipipe_dispatch_event(IPIPE_EVENT_RETURN, regs);
+        }
+
         if (this_domain == ipipe_root_domain) {
                 set_thread_flag(TIF_IRQ_SYNC);
                 if (!s) {