Merge branch 'cpus4096' into irq/threaded
[safe/jmp/linux-2.6] / arch / powerpc / kernel / irq.c
index fd4ddb8..1b55ffd 100644 (file)
@@ -7,7 +7,6 @@
  *    Copyright (C) 1996-2001 Cort Dougan
  *  Adapted for Power Macintosh by Paul Mackerras
  *    Copyright (C) 1996 Paul Mackerras (paulus@cs.anu.edu.au)
- *  Amiga/APUS changes by Jesper Skov (jskov@cygnus.co.uk).
  *
  * This program is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public License
@@ -53,6 +52,7 @@
 #include <linux/mutex.h>
 #include <linux/bootmem.h>
 #include <linux/pci.h>
+#include <linux/debugfs.h>
 
 #include <asm/uaccess.h>
 #include <asm/system.h>
 #include <asm/ptrace.h>
 #include <asm/machdep.h>
 #include <asm/udbg.h>
-#ifdef CONFIG_PPC_ISERIES
+#ifdef CONFIG_PPC64
 #include <asm/paca.h>
+#include <asm/firmware.h>
+#include <asm/lv1call.h>
 #endif
 
 int __irq_offset_value;
@@ -75,26 +77,94 @@ static int ppc_spurious_interrupts;
 EXPORT_SYMBOL(__irq_offset_value);
 atomic_t ppc_n_lost_interrupts;
 
-#ifndef CONFIG_PPC_MERGE
-#define NR_MASK_WORDS  ((NR_IRQS + 31) / 32)
-unsigned long ppc_cached_irq_mask[NR_MASK_WORDS];
-#endif
-
 #ifdef CONFIG_TAU_INT
 extern int tau_initialized;
 extern int tau_interrupts(int);
 #endif
 #endif /* CONFIG_PPC32 */
 
-#if defined(CONFIG_SMP) && !defined(CONFIG_PPC_MERGE)
-extern atomic_t ipi_recv;
-extern atomic_t ipi_sent;
-#endif
-
 #ifdef CONFIG_PPC64
 EXPORT_SYMBOL(irq_desc);
 
 int distribute_irqs = 1;
+
+static inline notrace unsigned long get_hard_enabled(void)
+{
+       unsigned long enabled;
+
+       __asm__ __volatile__("lbz %0,%1(13)"
+       : "=r" (enabled) : "i" (offsetof(struct paca_struct, hard_enabled)));
+
+       return enabled;
+}
+
+static inline notrace void set_soft_enabled(unsigned long enable)
+{
+       __asm__ __volatile__("stb %0,%1(13)"
+       : : "r" (enable), "i" (offsetof(struct paca_struct, soft_enabled)));
+}
+
+notrace void raw_local_irq_restore(unsigned long en)
+{
+       /*
+        * get_paca()->soft_enabled = en;
+        * Is it ever valid to use local_irq_restore(0) when soft_enabled is 1?
+        * That was allowed before, and in such a case we do need to take care
+        * that gcc will set soft_enabled directly via r13, not choose to use
+        * an intermediate register, lest we're preempted to a different cpu.
+        */
+       set_soft_enabled(en);
+       if (!en)
+               return;
+
+       if (firmware_has_feature(FW_FEATURE_ISERIES)) {
+               /*
+                * Do we need to disable preemption here?  Not really: in the
+                * unlikely event that we're preempted to a different cpu in
+                * between getting r13, loading its lppaca_ptr, and loading
+                * its any_int, we might call iseries_handle_interrupts without
+                * an interrupt pending on the new cpu, but that's no disaster,
+                * is it?  And the business of preempting us off the old cpu
+                * would itself involve a local_irq_restore which handles the
+                * interrupt to that cpu.
+                *
+                * But use "local_paca->lppaca_ptr" instead of "get_lppaca()"
+                * to avoid any preemption checking added into get_paca().
+                */
+               if (local_paca->lppaca_ptr->int_dword.any_int)
+                       iseries_handle_interrupts();
+       }
+
+       /*
+        * if (get_paca()->hard_enabled) return;
+        * But again we need to take care that gcc gets hard_enabled directly
+        * via r13, not choose to use an intermediate register, lest we're
+        * preempted to a different cpu in between the two instructions.
+        */
+       if (get_hard_enabled())
+               return;
+
+       /*
+        * Need to hard-enable interrupts here.  Since currently disabled,
+        * no need to take further asm precautions against preemption; but
+        * use local_paca instead of get_paca() to avoid preemption checking.
+        */
+       local_paca->hard_enabled = en;
+       if ((int)mfspr(SPRN_DEC) < 0)
+               mtspr(SPRN_DEC, 1);
+
+       /*
+        * Force the delivery of pending soft-disabled interrupts on PS3.
+        * Any HV call will have this side effect.
+        */
+       if (firmware_has_feature(FW_FEATURE_PS3_LV1)) {
+               u64 tmp;
+               lv1_get_version_info(&tmp);
+       }
+
+       __hard_irq_enable();
+}
+EXPORT_SYMBOL(raw_local_irq_restore);
 #endif /* CONFIG_PPC64 */
 
 int show_interrupts(struct seq_file *p, void *v)
@@ -120,7 +190,7 @@ int show_interrupts(struct seq_file *p, void *v)
                seq_printf(p, "%3d: ", i);
 #ifdef CONFIG_SMP
                for_each_online_cpu(j)
-                       seq_printf(p, "%10u ", kstat_cpu(j).irqs[i]);
+                       seq_printf(p, "%10u ", kstat_irqs_cpu(i, j));
 #else
                seq_printf(p, "%10u ", kstat_irqs(i));
 #endif /* CONFIG_SMP */
@@ -136,21 +206,14 @@ int show_interrupts(struct seq_file *p, void *v)
 skip:
                spin_unlock_irqrestore(&desc->lock, flags);
        } else if (i == NR_IRQS) {
-#ifdef CONFIG_PPC32
-#ifdef CONFIG_TAU_INT
+#if defined(CONFIG_PPC32) && defined(CONFIG_TAU_INT)
                if (tau_initialized){
                        seq_puts(p, "TAU: ");
                        for_each_online_cpu(j)
                                seq_printf(p, "%10u ", tau_interrupts(j));
                        seq_puts(p, "  PowerPC             Thermal Assist (cpu temp)\n");
                }
-#endif
-#if defined(CONFIG_SMP) && !defined(CONFIG_PPC_MERGE)
-               /* should this be per processor send/receive? */
-               seq_printf(p, "IPI (recv/sent): %10u/%u\n",
-                               atomic_read(&ipi_recv), atomic_read(&ipi_sent));
-#endif
-#endif /* CONFIG_PPC32 */
+#endif /* CONFIG_PPC32 && CONFIG_TAU_INT*/
                seq_printf(p, "BAD: %10u\n", ppc_spurious_interrupts);
        }
        return 0;
@@ -168,13 +231,13 @@ void fixup_irqs(cpumask_t map)
                if (irq_desc[irq].status & IRQ_PER_CPU)
                        continue;
 
-               cpus_and(mask, irq_desc[irq].affinity, map);
+               cpumask_and(&mask, irq_desc[irq].affinity, &map);
                if (any_online_cpu(mask) == NR_CPUS) {
                        printk("Breaking affinity for irq %i\n", irq);
                        mask = map;
                }
                if (irq_desc[irq].chip->set_affinity)
-                       irq_desc[irq].chip->set_affinity(irq, mask);
+                       irq_desc[irq].chip->set_affinity(irq, &mask);
                else if (irq_desc[irq].action && !(warned++))
                        printk("Cannot set affinity for irq %i\n", irq);
        }
@@ -187,12 +250,13 @@ void fixup_irqs(cpumask_t map)
 
 void do_IRQ(struct pt_regs *regs)
 {
+       struct pt_regs *old_regs = set_irq_regs(regs);
        unsigned int irq;
 #ifdef CONFIG_IRQSTACKS
        struct thread_info *curtp, *irqtp;
 #endif
 
-        irq_enter();
+       irq_enter();
 
 #ifdef CONFIG_DEBUG_STACKOVERFLOW
        /* Debugging check for stack overflow: is there less than 2KB free? */
@@ -211,12 +275,12 @@ void do_IRQ(struct pt_regs *regs)
 
        /*
         * Every platform is required to implement ppc_md.get_irq.
-        * This function will either return an irq number or -1 to
+        * This function will either return an irq number or NO_IRQ to
         * indicate there are no more pending.
-        * The value -2 is for buggy hardware and means that this IRQ
-        * has already been handled. -- Tom
+        * The value NO_IRQ_IGNORE is for buggy hardware and means that this
+        * IRQ has already been handled. -- Tom
         */
-       irq = ppc_md.get_irq(regs);
+       irq = ppc_md.get_irq();
 
        if (irq != NO_IRQ && irq != NO_IRQ_IGNORE) {
 #ifdef CONFIG_IRQSTACKS
@@ -226,25 +290,44 @@ void do_IRQ(struct pt_regs *regs)
                if (curtp != irqtp) {
                        struct irq_desc *desc = irq_desc + irq;
                        void *handler = desc->handle_irq;
+                       unsigned long saved_sp_limit = current->thread.ksp_limit;
                        if (handler == NULL)
                                handler = &__do_IRQ;
                        irqtp->task = curtp->task;
                        irqtp->flags = 0;
-                       call_handle_irq(irq, desc, regs, irqtp, handler);
+
+                       /* Copy the softirq bits in preempt_count so that the
+                        * softirq checks work in the hardirq context.
+                        */
+                       irqtp->preempt_count =
+                               (irqtp->preempt_count & ~SOFTIRQ_MASK) |
+                               (curtp->preempt_count & SOFTIRQ_MASK);
+
+                       current->thread.ksp_limit = (unsigned long)irqtp +
+                               _ALIGN_UP(sizeof(struct thread_info), 16);
+                       call_handle_irq(irq, desc, irqtp, handler);
+                       current->thread.ksp_limit = saved_sp_limit;
                        irqtp->task = NULL;
+
+
+                       /* Set any flag that may have been set on the
+                        * alternate stack
+                        */
                        if (irqtp->flags)
                                set_bits(irqtp->flags, &curtp->flags);
                } else
 #endif
-                       generic_handle_irq(irq, regs);
+                       generic_handle_irq(irq);
        } else if (irq != NO_IRQ_IGNORE)
                /* That's not SMP safe ... but who cares ? */
                ppc_spurious_interrupts++;
 
-        irq_exit();
+       irq_exit();
+       set_irq_regs(old_regs);
 
 #ifdef CONFIG_PPC_ISERIES
-       if (get_lppaca()->int_dword.fields.decr_int) {
+       if (firmware_has_feature(FW_FEATURE_ISERIES) &&
+                       get_lppaca()->int_dword.fields.decr_int) {
                get_lppaca()->int_dword.fields.decr_int = 0;
                /* Signal a fake decrementer interrupt */
                timer_interrupt(regs);
@@ -254,12 +337,44 @@ void do_IRQ(struct pt_regs *regs)
 
 void __init init_IRQ(void)
 {
-       ppc_md.init_IRQ();
-#ifdef CONFIG_PPC64
+       if (ppc_md.init_IRQ)
+               ppc_md.init_IRQ();
+
+       exc_lvl_ctx_init();
+
        irq_ctx_init();
-#endif
 }
 
+#if defined(CONFIG_BOOKE) || defined(CONFIG_40x)
+struct thread_info   *critirq_ctx[NR_CPUS] __read_mostly;
+struct thread_info    *dbgirq_ctx[NR_CPUS] __read_mostly;
+struct thread_info *mcheckirq_ctx[NR_CPUS] __read_mostly;
+
+void exc_lvl_ctx_init(void)
+{
+       struct thread_info *tp;
+       int i;
+
+       for_each_possible_cpu(i) {
+               memset((void *)critirq_ctx[i], 0, THREAD_SIZE);
+               tp = critirq_ctx[i];
+               tp->cpu = i;
+               tp->preempt_count = 0;
+
+#ifdef CONFIG_BOOKE
+               memset((void *)dbgirq_ctx[i], 0, THREAD_SIZE);
+               tp = dbgirq_ctx[i];
+               tp->cpu = i;
+               tp->preempt_count = 0;
+
+               memset((void *)mcheckirq_ctx[i], 0, THREAD_SIZE);
+               tp = mcheckirq_ctx[i];
+               tp->cpu = i;
+               tp->preempt_count = HARDIRQ_OFFSET;
+#endif
+       }
+}
+#endif
 
 #ifdef CONFIG_IRQSTACKS
 struct thread_info *softirq_ctx[NR_CPUS] __read_mostly;
@@ -274,7 +389,7 @@ void irq_ctx_init(void)
                memset((void *)softirq_ctx[i], 0, THREAD_SIZE);
                tp = softirq_ctx[i];
                tp->cpu = i;
-               tp->preempt_count = SOFTIRQ_OFFSET;
+               tp->preempt_count = 0;
 
                memset((void *)hardirq_ctx[i], 0, THREAD_SIZE);
                tp = hardirq_ctx[i];
@@ -286,11 +401,15 @@ void irq_ctx_init(void)
 static inline void do_softirq_onstack(void)
 {
        struct thread_info *curtp, *irqtp;
+       unsigned long saved_sp_limit = current->thread.ksp_limit;
 
        curtp = current_thread_info();
        irqtp = softirq_ctx[smp_processor_id()];
        irqtp->task = curtp->task;
+       current->thread.ksp_limit = (unsigned long)irqtp +
+                                   _ALIGN_UP(sizeof(struct thread_info), 16);
        call_do_softirq(irqtp);
+       current->thread.ksp_limit = saved_sp_limit;
        irqtp->task = NULL;
 }
 
@@ -312,23 +431,33 @@ void do_softirq(void)
 
        local_irq_restore(flags);
 }
-EXPORT_SYMBOL(do_softirq);
 
 
 /*
  * IRQ controller and virtual interrupts
  */
 
-#ifdef CONFIG_PPC_MERGE
-
 static LIST_HEAD(irq_hosts);
-static spinlock_t irq_big_lock = SPIN_LOCK_UNLOCKED;
-
+static DEFINE_SPINLOCK(irq_big_lock);
+static unsigned int revmap_trees_allocated;
+static DEFINE_MUTEX(revmap_trees_mutex);
 struct irq_map_entry irq_map[NR_IRQS];
 static unsigned int irq_virq_count = NR_IRQS;
 static struct irq_host *irq_default_host;
 
-struct irq_host *irq_alloc_host(unsigned int revmap_type,
+irq_hw_number_t virq_to_hw(unsigned int virq)
+{
+       return irq_map[virq].hwirq;
+}
+EXPORT_SYMBOL_GPL(virq_to_hw);
+
+static int default_irq_host_match(struct irq_host *h, struct device_node *np)
+{
+       return h->of_node != NULL && h->of_node == np;
+}
+
+struct irq_host *irq_alloc_host(struct device_node *of_node,
+                               unsigned int revmap_type,
                                unsigned int revmap_arg,
                                struct irq_host_ops *ops,
                                irq_hw_number_t inval_irq)
@@ -342,13 +471,7 @@ struct irq_host *irq_alloc_host(unsigned int revmap_type,
        /* Allocate structure and revmap table if using linear mapping */
        if (revmap_type == IRQ_HOST_MAP_LINEAR)
                size += revmap_arg * sizeof(unsigned int);
-       if (mem_init_done)
-               host = kzalloc(size, GFP_KERNEL);
-       else {
-               host = alloc_bootmem(size);
-               if (host)
-                       memset(host, 0, size);
-       }
+       host = zalloc_maybe_bootmem(size, GFP_KERNEL);
        if (host == NULL)
                return NULL;
 
@@ -356,6 +479,10 @@ struct irq_host *irq_alloc_host(unsigned int revmap_type,
        host->revmap_type = revmap_type;
        host->inval_irq = inval_irq;
        host->ops = ops;
+       host->of_node = of_node_get(of_node);
+
+       if (host->ops->match == NULL)
+               host->ops->match = default_irq_host_match;
 
        spin_lock_irqsave(&irq_big_lock, flags);
 
@@ -387,7 +514,7 @@ struct irq_host *irq_alloc_host(unsigned int revmap_type,
                host->inval_irq = 0;
                /* setup us as the host for all legacy interrupts */
                for (i = 1; i < NUM_ISA_INTERRUPTS; i++) {
-                       irq_map[i].hwirq = 0;
+                       irq_map[i].hwirq = i;
                        smp_wmb();
                        irq_map[i].host = host;
                        smp_wmb();
@@ -397,7 +524,7 @@ struct irq_host *irq_alloc_host(unsigned int revmap_type,
 
                        /* Legacy flags are left to default at this point,
                         * one can then use irq_create_mapping() to
-                        * explicitely change them
+                        * explicitly change them
                         */
                        ops->map(host, i, i);
                }
@@ -405,7 +532,7 @@ struct irq_host *irq_alloc_host(unsigned int revmap_type,
        case IRQ_HOST_MAP_LINEAR:
                rmap = (unsigned int *)(host + 1);
                for (i = 0; i < revmap_arg; i++)
-                       rmap[i] = IRQ_NONE;
+                       rmap[i] = NO_IRQ;
                host->revmap_data.linear.size = revmap_arg;
                smp_wmb();
                host->revmap_data.linear.revmap = rmap;
@@ -431,7 +558,7 @@ struct irq_host *irq_find_host(struct device_node *node)
         */
        spin_lock_irqsave(&irq_big_lock, flags);
        list_for_each_entry(h, &irq_hosts, link)
-               if (h->ops->match == NULL || h->ops->match(h, node)) {
+               if (h->ops->match(h, node)) {
                        found = h;
                        break;
                }
@@ -456,6 +583,50 @@ void irq_set_virq_count(unsigned int count)
                irq_virq_count = count;
 }
 
+static int irq_setup_virq(struct irq_host *host, unsigned int virq,
+                           irq_hw_number_t hwirq)
+{
+       /* Clear IRQ_NOREQUEST flag */
+       get_irq_desc(virq)->status &= ~IRQ_NOREQUEST;
+
+       /* map it */
+       smp_wmb();
+       irq_map[virq].hwirq = hwirq;
+       smp_mb();
+
+       if (host->ops->map(host, virq, hwirq)) {
+               pr_debug("irq: -> mapping failed, freeing\n");
+               irq_free_virt(virq, 1);
+               return -1;
+       }
+
+       return 0;
+}
+
+unsigned int irq_create_direct_mapping(struct irq_host *host)
+{
+       unsigned int virq;
+
+       if (host == NULL)
+               host = irq_default_host;
+
+       BUG_ON(host == NULL);
+       WARN_ON(host->revmap_type != IRQ_HOST_MAP_NOMAP);
+
+       virq = irq_alloc_virt(host, 1, 0);
+       if (virq == NO_IRQ) {
+               pr_debug("irq: create_direct virq allocation failed\n");
+               return NO_IRQ;
+       }
+
+       pr_debug("irq: create_direct obtained virq %d\n", virq);
+
+       if (irq_setup_virq(host, virq, virq))
+               return NO_IRQ;
+
+       return virq;
+}
+
 unsigned int irq_create_mapping(struct irq_host *host,
                                irq_hw_number_t hwirq)
 {
@@ -478,7 +649,9 @@ unsigned int irq_create_mapping(struct irq_host *host,
         * host->ops->map() to update the flags
         */
        virq = irq_find_mapping(host, hwirq);
-       if (virq != IRQ_NONE) {
+       if (virq != NO_IRQ) {
+               if (host->ops->remap)
+                       host->ops->remap(host, virq, hwirq);
                pr_debug("irq: -> existing mapping on virq %d\n", virq);
                return virq;
        }
@@ -501,24 +674,15 @@ unsigned int irq_create_mapping(struct irq_host *host,
        }
        pr_debug("irq: -> obtained virq %d\n", virq);
 
-       /* Clear IRQ_NOREQUEST flag */
-       get_irq_desc(virq)->status &= ~IRQ_NOREQUEST;
-
-       /* map it */
-       smp_wmb();
-       irq_map[virq].hwirq = hwirq;
-       smp_mb();
-       if (host->ops->map(host, virq, hwirq)) {
-               pr_debug("irq: -> mapping failed, freeing\n");
-               irq_free_virt(virq, 1);
+       if (irq_setup_virq(host, virq, hwirq))
                return NO_IRQ;
-       }
+
        return virq;
 }
 EXPORT_SYMBOL_GPL(irq_create_mapping);
 
-extern unsigned int irq_create_of_mapping(struct device_node *controller,
-                                         u32 *intspec, unsigned int intsize)
+unsigned int irq_create_of_mapping(struct device_node *controller,
+                                  u32 *intspec, unsigned int intsize)
 {
        struct irq_host *host;
        irq_hw_number_t hwirq;
@@ -571,10 +735,13 @@ EXPORT_SYMBOL_GPL(irq_of_parse_and_map);
 
 void irq_dispose_mapping(unsigned int virq)
 {
-       struct irq_host *host = irq_map[virq].host;
+       struct irq_host *host;
        irq_hw_number_t hwirq;
-       unsigned long flags;
 
+       if (virq == NO_IRQ)
+               return;
+
+       host = irq_map[virq].host;
        WARN_ON (host == NULL);
        if (host == NULL)
                return;
@@ -599,19 +766,19 @@ void irq_dispose_mapping(unsigned int virq)
        switch(host->revmap_type) {
        case IRQ_HOST_MAP_LINEAR:
                if (hwirq < host->revmap_data.linear.size)
-                       host->revmap_data.linear.revmap[hwirq] = IRQ_NONE;
+                       host->revmap_data.linear.revmap[hwirq] = NO_IRQ;
                break;
        case IRQ_HOST_MAP_TREE:
-               /* Check if radix tree allocated yet */
-               if (host->revmap_data.tree.gfp_mask == 0)
-                       break;
-               /* XXX radix tree not safe ! remove lock whem it becomes safe
-                * and use some RCU sync to make sure everything is ok before we
-                * can re-use that map entry
+               /*
+                * Check if radix tree allocated yet, if not then nothing to
+                * remove.
                 */
-               spin_lock_irqsave(&irq_big_lock, flags);
+               smp_rmb();
+               if (revmap_trees_allocated < 1)
+                       break;
+               mutex_lock(&revmap_trees_mutex);
                radix_tree_delete(&host->revmap_data.tree, hwirq);
-               spin_unlock_irqrestore(&irq_big_lock, flags);
+               mutex_unlock(&revmap_trees_mutex);
                break;
        }
 
@@ -660,46 +827,64 @@ unsigned int irq_find_mapping(struct irq_host *host,
 EXPORT_SYMBOL_GPL(irq_find_mapping);
 
 
-unsigned int irq_radix_revmap(struct irq_host *host,
-                             irq_hw_number_t hwirq)
+unsigned int irq_radix_revmap_lookup(struct irq_host *host,
+                                    irq_hw_number_t hwirq)
 {
-       struct radix_tree_root *tree;
        struct irq_map_entry *ptr;
        unsigned int virq;
-       unsigned long flags;
 
        WARN_ON(host->revmap_type != IRQ_HOST_MAP_TREE);
 
-       /* Check if the radix tree exist yet. We test the value of
-        * the gfp_mask for that. Sneaky but saves another int in the
-        * structure. If not, we fallback to slow mode
+       /*
+        * Check if the radix tree exists and has bee initialized.
+        * If not, we fallback to slow mode
         */
-       tree = &host->revmap_data.tree;
-       if (tree->gfp_mask == 0)
+       if (revmap_trees_allocated < 2)
                return irq_find_mapping(host, hwirq);
 
-       /* XXX Current radix trees are NOT SMP safe !!! Remove that lock
-        * when that is fixed (when Nick's patch gets in
+       /* Now try to resolve */
+       /*
+        * No rcu_read_lock(ing) needed, the ptr returned can't go under us
+        * as it's referencing an entry in the static irq_map table.
         */
-       spin_lock_irqsave(&irq_big_lock, flags);
+       ptr = radix_tree_lookup(&host->revmap_data.tree, hwirq);
 
-       /* Now try to resolve */
-       ptr = radix_tree_lookup(tree, hwirq);
-       /* Found it, return */
-       if (ptr) {
+       /*
+        * If found in radix tree, then fine.
+        * Else fallback to linear lookup - this should not happen in practice
+        * as it means that we failed to insert the node in the radix tree.
+        */
+       if (ptr)
                virq = ptr - irq_map;
-               goto bail;
-       }
+       else
+               virq = irq_find_mapping(host, hwirq);
 
-       /* If not there, try to insert it */
-       virq = irq_find_mapping(host, hwirq);
-       if (virq != NO_IRQ)
-               radix_tree_insert(tree, hwirq, &irq_map[virq]);
- bail:
-       spin_unlock_irqrestore(&irq_big_lock, flags);
        return virq;
 }
 
+void irq_radix_revmap_insert(struct irq_host *host, unsigned int virq,
+                            irq_hw_number_t hwirq)
+{
+
+       WARN_ON(host->revmap_type != IRQ_HOST_MAP_TREE);
+
+       /*
+        * Check if the radix tree exists yet.
+        * If not, then the irq will be inserted into the tree when it gets
+        * initialized.
+        */
+       smp_rmb();
+       if (revmap_trees_allocated < 1)
+               return;
+
+       if (virq != NO_IRQ) {
+               mutex_lock(&revmap_trees_mutex);
+               radix_tree_insert(&host->revmap_data.tree, hwirq,
+                                 &irq_map[virq]);
+               mutex_unlock(&revmap_trees_mutex);
+       }
+}
+
 unsigned int irq_linear_revmap(struct irq_host *host,
                               irq_hw_number_t hwirq)
 {
@@ -729,7 +914,6 @@ unsigned int irq_alloc_virt(struct irq_host *host,
 {
        unsigned long flags;
        unsigned int i, j, found = NO_IRQ;
-       unsigned int limit = irq_virq_count - count;
 
        if (count == 0 || count > (irq_virq_count - NUM_ISA_INTERRUPTS))
                return NO_IRQ;
@@ -746,14 +930,16 @@ unsigned int irq_alloc_virt(struct irq_host *host,
        /* Look for count consecutive numbers in the allocatable
         * (non-legacy) space
         */
-       for (i = NUM_ISA_INTERRUPTS; i <= limit; ) {
-               for (j = i; j < (i + count); j++)
-                       if (irq_map[j].host != NULL) {
-                               i = j + 1;
-                               continue;
-                       }
-               found = i;
-               break;
+       for (i = NUM_ISA_INTERRUPTS, j = 0; i < irq_virq_count; i++) {
+               if (irq_map[i].host != NULL)
+                       j = 0;
+               else
+                       j++;
+
+               if (j == count) {
+                       found = i - count + 1;
+                       break;
+               }
        }
        if (found == NO_IRQ) {
                spin_unlock_irqrestore(&irq_big_lock, flags);
@@ -805,48 +991,110 @@ void irq_early_init(void)
 static int irq_late_init(void)
 {
        struct irq_host *h;
-       unsigned long flags;
+       unsigned int i;
 
-       spin_lock_irqsave(&irq_big_lock, flags);
+       /*
+        * No mutual exclusion with respect to accessors of the tree is needed
+        * here as the synchronization is done via the state variable
+        * revmap_trees_allocated.
+        */
        list_for_each_entry(h, &irq_hosts, link) {
                if (h->revmap_type == IRQ_HOST_MAP_TREE)
-                       INIT_RADIX_TREE(&h->revmap_data.tree, GFP_ATOMIC);
+                       INIT_RADIX_TREE(&h->revmap_data.tree, GFP_KERNEL);
        }
-       spin_unlock_irqrestore(&irq_big_lock, flags);
+
+       /*
+        * Make sure the radix trees inits are visible before setting
+        * the flag
+        */
+       smp_wmb();
+       revmap_trees_allocated = 1;
+
+       /*
+        * Insert the reverse mapping for those interrupts already present
+        * in irq_map[].
+        */
+       mutex_lock(&revmap_trees_mutex);
+       for (i = 0; i < irq_virq_count; i++) {
+               if (irq_map[i].host &&
+                   (irq_map[i].host->revmap_type == IRQ_HOST_MAP_TREE))
+                       radix_tree_insert(&irq_map[i].host->revmap_data.tree,
+                                         irq_map[i].hwirq, &irq_map[i]);
+       }
+       mutex_unlock(&revmap_trees_mutex);
+
+       /*
+        * Make sure the radix trees insertions are visible before setting
+        * the flag
+        */
+       smp_wmb();
+       revmap_trees_allocated = 2;
 
        return 0;
 }
 arch_initcall(irq_late_init);
 
-#endif /* CONFIG_PPC_MERGE */
-
-#ifdef CONFIG_PCI_MSI
-int pci_enable_msi(struct pci_dev * pdev)
+#ifdef CONFIG_VIRQ_DEBUG
+static int virq_debug_show(struct seq_file *m, void *private)
 {
-       if (ppc_md.enable_msi)
-               return ppc_md.enable_msi(pdev);
-       else
-               return -1;
+       unsigned long flags;
+       irq_desc_t *desc;
+       const char *p;
+       char none[] = "none";
+       int i;
+
+       seq_printf(m, "%-5s  %-7s  %-15s  %s\n", "virq", "hwirq",
+                     "chip name", "host name");
+
+       for (i = 1; i < NR_IRQS; i++) {
+               desc = get_irq_desc(i);
+               spin_lock_irqsave(&desc->lock, flags);
+
+               if (desc->action && desc->action->handler) {
+                       seq_printf(m, "%5d  ", i);
+                       seq_printf(m, "0x%05lx  ", virq_to_hw(i));
+
+                       if (desc->chip && desc->chip->typename)
+                               p = desc->chip->typename;
+                       else
+                               p = none;
+                       seq_printf(m, "%-15s  ", p);
+
+                       if (irq_map[i].host && irq_map[i].host->of_node)
+                               p = irq_map[i].host->of_node->full_name;
+                       else
+                               p = none;
+                       seq_printf(m, "%s\n", p);
+               }
+
+               spin_unlock_irqrestore(&desc->lock, flags);
+       }
+
+       return 0;
 }
-EXPORT_SYMBOL(pci_enable_msi);
 
-void pci_disable_msi(struct pci_dev * pdev)
+static int virq_debug_open(struct inode *inode, struct file *file)
 {
-       if (ppc_md.disable_msi)
-               ppc_md.disable_msi(pdev);
+       return single_open(file, virq_debug_show, inode->i_private);
 }
-EXPORT_SYMBOL(pci_disable_msi);
 
-void pci_scan_msi_device(struct pci_dev *dev) {}
-int pci_enable_msix(struct pci_dev* dev, struct msix_entry *entries, int nvec) {return -1;}
-void pci_disable_msix(struct pci_dev *dev) {}
-void msi_remove_pci_irq_vectors(struct pci_dev *dev) {}
-void disable_msi_mode(struct pci_dev *dev, int pos, int type) {}
-void pci_no_msi(void) {}
-EXPORT_SYMBOL(pci_enable_msix);
-EXPORT_SYMBOL(pci_disable_msix);
+static const struct file_operations virq_debug_fops = {
+       .open = virq_debug_open,
+       .read = seq_read,
+       .llseek = seq_lseek,
+       .release = single_release,
+};
 
-#endif
+static int __init irq_debugfs_init(void)
+{
+       if (debugfs_create_file("virq_mapping", S_IRUGO, powerpc_debugfs_root,
+                                NULL, &virq_debug_fops) == NULL)
+               return -ENOMEM;
+
+       return 0;
+}
+__initcall(irq_debugfs_init);
+#endif /* CONFIG_VIRQ_DEBUG */
 
 #ifdef CONFIG_PPC64
 static int __init setup_noirqdistrib(char *str)