diff --git a/arch/sparc/include/asm/hugetlb.h b/arch/sparc/include/asm/hugetlb.h
index 9661e9b..7eb57d2 100644
--- a/arch/sparc/include/asm/hugetlb.h
+++ b/arch/sparc/include/asm/hugetlb.h
@@ -12,7 +12,6 @@
 
 static inline void hugetlb_prefault_arch_hook(struct mm_struct *mm)
 {
-	hugetlb_setup(mm);
 }
 
 static inline int is_hugepage_only_range(struct mm_struct *mm,
diff --git a/arch/sparc/include/asm/page_64.h b/arch/sparc/include/asm/page_64.h
index 4b39f74..e155388 100644
--- a/arch/sparc/include/asm/page_64.h
+++ b/arch/sparc/include/asm/page_64.h
@@ -27,8 +27,8 @@
 #ifndef __ASSEMBLY__
 
 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
-struct mm_struct;
-extern void hugetlb_setup(struct mm_struct *mm);
+struct pt_regs;
+extern void hugetlb_setup(struct pt_regs *regs);
 #endif
 
 #define WANT_PAGE_VIRTUAL
diff --git a/arch/sparc/include/asm/tsb.h b/arch/sparc/include/asm/tsb.h
index b4c258d..e696432 100644
--- a/arch/sparc/include/asm/tsb.h
+++ b/arch/sparc/include/asm/tsb.h
@@ -157,17 +157,26 @@
 	andn		REG2, 0x7, REG2; \
 	add		REG1, REG2, REG1;
 
-	/* This macro exists only to make the PMD translator below easier
-	 * to read.  It hides the ELF section switch for the sun4v code
-	 * patching.
+	/* These macros exists only to make the PMD translator below
+	 * easier to read.  It hides the ELF section switch for the
+	 * sun4v code patching.
 	 */
-#define OR_PTE_BIT(REG, NAME)				\
+#define OR_PTE_BIT_1INSN(REG, NAME)			\
 661:	or		REG, _PAGE_##NAME##_4U, REG;	\
 	.section	.sun4v_1insn_patch, "ax";	\
 	.word		661b;				\
 	or		REG, _PAGE_##NAME##_4V, REG;	\
 	.previous;
 
+#define OR_PTE_BIT_2INSN(REG, TMP, NAME)		\
+661:	sethi		%hi(_PAGE_##NAME##_4U), TMP;	\
+	or		REG, TMP, REG;			\
+	.section	.sun4v_2insn_patch, "ax";	\
+	.word		661b;				\
+	mov		-1, TMP;			\
+	or		REG, _PAGE_##NAME##_4V, REG;	\
+	.previous;
+
 	/* Load into REG the PTE value for VALID, CACHE, and SZHUGE.  */
 #define BUILD_PTE_VALID_SZHUGE_CACHE(REG)				   \
 661:	sethi		%uhi(_PAGE_VALID|_PAGE_SZHUGE_4U), REG;		   \
@@ -214,12 +223,13 @@
 	 andn		REG1, PMD_HUGE_PROTBITS, REG2;			      \
 	sllx		REG2, PMD_PADDR_SHIFT, REG2;			      \
 	/* REG2 now holds PFN << PAGE_SHIFT */				      \
-	andcc		REG1, PMD_HUGE_EXEC, %g0;			      \
+	andcc		REG1, PMD_HUGE_WRITE, %g0;			      \
 	bne,a,pt	%xcc, 1f;					      \
-	 OR_PTE_BIT(REG2, EXEC);					      \
-1:	andcc		REG1, PMD_HUGE_WRITE, %g0;			      \
-	bne,a,pt	%xcc, 1f;					      \
-	 OR_PTE_BIT(REG2, W);						      \
+	 OR_PTE_BIT_1INSN(REG2, W);					      \
+1:	andcc		REG1, PMD_HUGE_EXEC, %g0;			      \
+	be,pt		%xcc, 1f;					      \
+	 nop;								      \
+	OR_PTE_BIT_2INSN(REG2, REG1, EXEC);				      \
 	/* REG1 can now be clobbered, build final PTE */		      \
 1:	BUILD_PTE_VALID_SZHUGE_CACHE(REG1);				      \
 	ba,pt		%xcc, PTE_LABEL;				      \
diff --git a/arch/sparc/kernel/kernel.h b/arch/sparc/kernel/kernel.h
index 291bb5d..a702d9a 100644
--- a/arch/sparc/kernel/kernel.h
+++ b/arch/sparc/kernel/kernel.h
@@ -48,6 +48,10 @@
 extern void sun4m_unmask_profile_irq(void);
 extern void sun4m_clear_profile_irq(int cpu);
 
+/* sun4m_smp.c */
+void sun4m_cpu_pre_starting(void *arg);
+void sun4m_cpu_pre_online(void *arg);
+
 /* sun4d_irq.c */
 extern spinlock_t sun4d_imsk_lock;
 
@@ -60,6 +64,14 @@
 extern void sun4d_distribute_irqs(void);
 extern void sun4d_free_irq(unsigned int irq, void *dev_id);
 
+/* sun4d_smp.c */
+void sun4d_cpu_pre_starting(void *arg);
+void sun4d_cpu_pre_online(void *arg);
+
+/* leon_smp.c */
+void leon_cpu_pre_starting(void *arg);
+void leon_cpu_pre_online(void *arg);
+
 /* head_32.S */
 extern unsigned int t_nmi[];
 extern unsigned int linux_trap_ipi15_sun4d[];
diff --git a/arch/sparc/kernel/leon_smp.c b/arch/sparc/kernel/leon_smp.c
index 0f3fb6d..9b40c9c 100644
--- a/arch/sparc/kernel/leon_smp.c
+++ b/arch/sparc/kernel/leon_smp.c
@@ -69,31 +69,19 @@
 	return val;
 }
 
-void __cpuinit leon_callin(void)
+void __cpuinit leon_cpu_pre_starting(void *arg)
+{
+	leon_configure_cache_smp();
+}
+
+void __cpuinit leon_cpu_pre_online(void *arg)
 {
 	int cpuid = hard_smp_processor_id();
 
-	local_ops->cache_all();
-	local_ops->tlb_all();
-	leon_configure_cache_smp();
-
-	notify_cpu_starting(cpuid);
-
-	/* Get our local ticker going. */
-	register_percpu_ce(cpuid);
-
-	calibrate_delay();
-	smp_store_cpu_info(cpuid);
-
-	local_ops->cache_all();
-	local_ops->tlb_all();
-
-	/*
-	 * Unblock the master CPU _only_ when the scheduler state
-	 * of all secondary CPUs will be up-to-date, so after
-	 * the SMP initialization the master will be just allowed
-	 * to call the scheduler code.
-	 * Allow master to continue.
+	/* Allow master to continue. The master will then give us the
+	 * go-ahead by setting the smp_commenced_mask and will wait without
+	 * timeouts until our setup is completed fully (signified by
+	 * our bit being set in the cpu_online_mask).
 	 */
 	do_swap(&cpu_callin_map[cpuid], 1);
 
@@ -110,9 +98,6 @@
 
 	while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 		mb();
-
-	local_irq_enable();
-	set_cpu_online(cpuid, true);
 }
 
 /*
diff --git a/arch/sparc/kernel/smp_32.c b/arch/sparc/kernel/smp_32.c
index 79db45e..9e7e6d7 100644
--- a/arch/sparc/kernel/smp_32.c
+++ b/arch/sparc/kernel/smp_32.c
@@ -20,6 +20,7 @@
 #include <linux/seq_file.h>
 #include <linux/cache.h>
 #include <linux/delay.h>
+#include <linux/cpu.h>
 
 #include <asm/ptrace.h>
 #include <linux/atomic.h>
@@ -32,8 +33,10 @@
 #include <asm/cacheflush.h>
 #include <asm/tlbflush.h>
 #include <asm/cpudata.h>
+#include <asm/timer.h>
 #include <asm/leon.h>
 
+#include "kernel.h"
 #include "irq.h"
 
 volatile unsigned long cpu_callin_map[NR_CPUS] __cpuinitdata = {0,};
@@ -294,6 +297,89 @@
 	return ret;
 }
 
+void __cpuinit arch_cpu_pre_starting(void *arg)
+{
+	local_ops->cache_all();
+	local_ops->tlb_all();
+
+	switch(sparc_cpu_model) {
+	case sun4m:
+		sun4m_cpu_pre_starting(arg);
+		break;
+	case sun4d:
+		sun4d_cpu_pre_starting(arg);
+		break;
+	case sparc_leon:
+		leon_cpu_pre_starting(arg);
+		break;
+	default:
+		BUG();
+	}
+}
+
+void __cpuinit arch_cpu_pre_online(void *arg)
+{
+	unsigned int cpuid = hard_smp_processor_id();
+
+	register_percpu_ce(cpuid);
+
+	calibrate_delay();
+	smp_store_cpu_info(cpuid);
+
+	local_ops->cache_all();
+	local_ops->tlb_all();
+
+	switch(sparc_cpu_model) {
+	case sun4m:
+		sun4m_cpu_pre_online(arg);
+		break;
+	case sun4d:
+		sun4d_cpu_pre_online(arg);
+		break;
+	case sparc_leon:
+		leon_cpu_pre_online(arg);
+		break;
+	default:
+		BUG();
+	}
+}
+
+void __cpuinit sparc_start_secondary(void *arg)
+{
+	unsigned int cpu;
+
+	/*
+	 * SMP booting is extremely fragile in some architectures. So run
+	 * the cpu initialization code first before anything else.
+	 */
+	arch_cpu_pre_starting(arg);
+
+	preempt_disable();
+	cpu = smp_processor_id();
+
+	/* Invoke the CPU_STARTING notifier callbacks */
+	notify_cpu_starting(cpu);
+
+	arch_cpu_pre_online(arg);
+
+	/* Set the CPU in the cpu_online_mask */
+	set_cpu_online(cpu, true);
+
+	/* Enable local interrupts now */
+	local_irq_enable();
+
+	wmb();
+	cpu_idle();
+
+	/* We should never reach here! */
+	BUG();
+}
+
+void __cpuinit smp_callin(void)
+{
+	sparc_start_secondary(NULL);
+}
+
 void smp_bogo(struct seq_file *m)
 {
 	int i;
diff --git a/arch/sparc/kernel/sun4d_smp.c b/arch/sparc/kernel/sun4d_smp.c
index ddaea31..c9eb82f 100644
--- a/arch/sparc/kernel/sun4d_smp.c
+++ b/arch/sparc/kernel/sun4d_smp.c
@@ -50,10 +50,9 @@
 			      "i" (ASI_M_CTL));
 }
 
-void __cpuinit smp4d_callin(void)
+void __cpuinit sun4d_cpu_pre_starting(void *arg)
 {
 	int cpuid = hard_smp_processor_id();
-	unsigned long flags;
 
 	/* Show we are alive */
 	cpu_leds[cpuid] = 0x6;
@@ -61,26 +60,20 @@
 
 	/* Enable level15 interrupt, disable level14 interrupt for now */
 	cc_set_imsk((cc_get_imsk() & ~0x8000) | 0x4000);
+}
 
-	local_ops->cache_all();
-	local_ops->tlb_all();
+void __cpuinit sun4d_cpu_pre_online(void *arg)
+{
+	unsigned long flags;
+	int cpuid;
 
-	notify_cpu_starting(cpuid);
-	/*
-	 * Unblock the master CPU _only_ when the scheduler state
+	cpuid = hard_smp_processor_id();
+
+	/* Unblock the master CPU _only_ when the scheduler state
 	 * of all secondary CPUs will be up-to-date, so after
 	 * the SMP initialization the master will be just allowed
 	 * to call the scheduler code.
 	 */
-	/* Get our local ticker going. */
-	register_percpu_ce(cpuid);
-
-	calibrate_delay();
-	smp_store_cpu_info(cpuid);
-	local_ops->cache_all();
-	local_ops->tlb_all();
-
-	/* Allow master to continue. */
 	sun4d_swap((unsigned long *)&cpu_callin_map[cpuid], 1);
 	local_ops->cache_all();
 	local_ops->tlb_all();
@@ -106,16 +99,12 @@
 	local_ops->cache_all();
 	local_ops->tlb_all();
 
-	local_irq_enable();	/* We don't allow PIL 14 yet */
-
 	while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 		barrier();
 
 	spin_lock_irqsave(&sun4d_imsk_lock, flags);
 	cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 	spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
-	set_cpu_online(cpuid, true);
-
 }
 
 /*
diff --git a/arch/sparc/kernel/sun4m_smp.c b/arch/sparc/kernel/sun4m_smp.c
index 128af73..8a65f15 100644
--- a/arch/sparc/kernel/sun4m_smp.c
+++ b/arch/sparc/kernel/sun4m_smp.c
@@ -34,30 +34,19 @@
 	return val;
 }
 
-void __cpuinit smp4m_callin(void)
+void __cpuinit sun4m_cpu_pre_starting(void *arg)
+{
+}
+
+void __cpuinit sun4m_cpu_pre_online(void *arg)
 {
 	int cpuid = hard_smp_processor_id();
 
-	local_ops->cache_all();
-	local_ops->tlb_all();
-
-	notify_cpu_starting(cpuid);
-
-	register_percpu_ce(cpuid);
-
-	calibrate_delay();
-	smp_store_cpu_info(cpuid);
-
-	local_ops->cache_all();
-	local_ops->tlb_all();
-
-	/*
-	 * Unblock the master CPU _only_ when the scheduler state
-	 * of all secondary CPUs will be up-to-date, so after
-	 * the SMP initialization the master will be just allowed
-	 * to call the scheduler code.
+	/* Allow master to continue. The master will then give us the
+	 * go-ahead by setting the smp_commenced_mask and will wait without
+	 * timeouts until our setup is completed fully (signified by
+	 * our bit being set in the cpu_online_mask).
 	 */
-	/* Allow master to continue. */
 	swap_ulong(&cpu_callin_map[cpuid], 1);
 
 	/* XXX: What's up with all the flushes? */
@@ -75,10 +64,6 @@
 
 	while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 		mb();
-
-	local_irq_enable();
-
-	set_cpu_online(cpuid, true);
 }
 
 /*
diff --git a/arch/sparc/kernel/trampoline_32.S b/arch/sparc/kernel/trampoline_32.S
index af27aca..6cdb08c 100644
--- a/arch/sparc/kernel/trampoline_32.S
+++ b/arch/sparc/kernel/trampoline_32.S
@@ -79,18 +79,15 @@
 	 nop
 
 	/* Start this processor. */
-	call	smp4m_callin
+	call	smp_callin
 	 nop
 
-	b,a	smp_do_cpu_idle
+	b,a	smp_panic
 
 	.text
 	.align	4
 
-smp_do_cpu_idle:
-	call	cpu_idle
-	 mov	0, %o0
-
+smp_panic:
 	call	cpu_panic
 	 nop
 
@@ -144,10 +141,10 @@
 	 nop
 
 	/* Start this processor. */
-	call	smp4d_callin
+	call	smp_callin
 	 nop
 
-	b,a	smp_do_cpu_idle
+	b,a	smp_panic
 
 	__CPUINIT
 	.align	4
@@ -201,7 +198,7 @@
 	 nop
 
 	/* Start this processor. */
-	call	leon_callin
+	call	smp_callin
 	 nop
 
-	b,a	smp_do_cpu_idle
+	b,a	smp_panic
diff --git a/arch/sparc/kernel/tsb.S b/arch/sparc/kernel/tsb.S
index d4bdc7a..a313e4a 100644
--- a/arch/sparc/kernel/tsb.S
+++ b/arch/sparc/kernel/tsb.S
@@ -136,12 +136,43 @@
 	 nop
 
 	/* It is a huge page, use huge page TSB entry address we
-	 * calculated above.
+	 * calculated above.  If the huge page TSB has not been
+	 * allocated, setup a trap stack and call hugetlb_setup()
+	 * to do so, then return from the trap to replay the TLB
+	 * miss.
+	 *
+	 * This is necessary to handle the case of transparent huge
+	 * pages where we don't really have a non-atomic context
+	 * in which to allocate the hugepage TSB hash table.  When
+	 * the 'mm' faults in the hugepage for the first time, we
+	 * thus handle it here.  This also makes sure that we can
+	 * allocate the TSB hash table on the correct NUMA node.
 	 */
 	TRAP_LOAD_TRAP_BLOCK(%g7, %g2)
-	ldx		[%g7 + TRAP_PER_CPU_TSB_HUGE_TEMP], %g2
-	cmp		%g2, -1
-	movne		%xcc, %g2, %g1
+	ldx		[%g7 + TRAP_PER_CPU_TSB_HUGE_TEMP], %g1
+	cmp		%g1, -1
+	bne,pt		%xcc, 60f
+	 nop
+
+661:	rdpr		%pstate, %g5
+	wrpr		%g5, PSTATE_AG | PSTATE_MG, %pstate
+	.section	.sun4v_2insn_patch, "ax"
+	.word		661b
+	SET_GL(1)
+	nop
+	.previous
+
+	rdpr	%tl, %g3
+	cmp	%g3, 1
+	bne,pn	%xcc, winfix_trampoline
+	 nop
+	ba,pt	%xcc, etrap
+	 rd	%pc, %g7
+	call	hugetlb_setup
+	 add	%sp, PTREGS_OFF, %o0
+	ba,pt	%xcc, rtrap
+	 nop
+
 60:
 #endif
 
diff --git a/arch/sparc/mm/fault_64.c b/arch/sparc/mm/fault_64.c
index 097aee7..5062ff3 100644
--- a/arch/sparc/mm/fault_64.c
+++ b/arch/sparc/mm/fault_64.c
@@ -472,8 +472,13 @@
 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 	mm_rss = mm->context.huge_pte_count;
 	if (unlikely(mm_rss >
-		     mm->context.tsb_block[MM_TSB_HUGE].tsb_rss_limit))
-		tsb_grow(mm, MM_TSB_HUGE, mm_rss);
+		     mm->context.tsb_block[MM_TSB_HUGE].tsb_rss_limit)) {
+		if (mm->context.tsb_block[MM_TSB_HUGE].tsb)
+			tsb_grow(mm, MM_TSB_HUGE, mm_rss);
+		else
+			hugetlb_setup(regs);
+
+	}
 #endif
 	return;
 
diff --git a/arch/sparc/mm/init_64.c b/arch/sparc/mm/init_64.c
index c3b7242..82bbf04 100644
--- a/arch/sparc/mm/init_64.c
+++ b/arch/sparc/mm/init_64.c
@@ -314,16 +314,31 @@
 	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
 	unsigned long tag;
 
+	if (unlikely(!tsb))
+		return;
+
 	tsb += ((address >> tsb_hash_shift) &
 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 	tag = (address >> 22UL);
 	tsb_insert(tsb, tag, tte);
 }
 
+#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
+static inline bool is_hugetlb_pte(pte_t pte)
+{
+	if ((tlb_type == hypervisor &&
+	     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
+	    (tlb_type != hypervisor &&
+	     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
+		return true;
+	return false;
+}
+#endif
+
 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 {
-	unsigned long tsb_index, tsb_hash_shift, flags;
 	struct mm_struct *mm;
+	unsigned long flags;
 	pte_t pte = *ptep;
 
 	if (tlb_type != hypervisor) {
@@ -335,25 +350,16 @@
 
 	mm = vma->vm_mm;
 
-	tsb_index = MM_TSB_BASE;
-	tsb_hash_shift = PAGE_SHIFT;
-
 	spin_lock_irqsave(&mm->context.lock, flags);
 
 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
-	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
-		if ((tlb_type == hypervisor &&
-		     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
-		    (tlb_type != hypervisor &&
-		     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
-			tsb_index = MM_TSB_HUGE;
-			tsb_hash_shift = HPAGE_SHIFT;
-		}
-	}
+	if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
+		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, HPAGE_SHIFT,
+					address, pte_val(pte));
+	else
 #endif
-
-	__update_mmu_tsb_insert(mm, tsb_index, tsb_hash_shift,
-				address, pte_val(pte));
+		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
+					address, pte_val(pte));
 
 	spin_unlock_irqrestore(&mm->context.lock, flags);
 }
@@ -2712,14 +2718,28 @@
 		load_secondary_context(mm);
 }
 
-void hugetlb_setup(struct mm_struct *mm)
+void hugetlb_setup(struct pt_regs *regs)
 {
-	struct tsb_config *tp = &mm->context.tsb_block[MM_TSB_HUGE];
+	struct mm_struct *mm = current->mm;
+	struct tsb_config *tp;
 
-	if (likely(tp->tsb != NULL))
-		return;
+	if (in_atomic() || !mm) {
+		const struct exception_table_entry *entry;
 
-	tsb_grow(mm, MM_TSB_HUGE, 0);
+		entry = search_exception_tables(regs->tpc);
+		if (entry) {
+			regs->tpc = entry->fixup;
+			regs->tnpc = regs->tpc + 4;
+			return;
+		}
+		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
+		die_if_kernel("HugeTSB in atomic", regs);
+	}
+
+	tp = &mm->context.tsb_block[MM_TSB_HUGE];
+	if (likely(tp->tsb == NULL))
+		tsb_grow(mm, MM_TSB_HUGE, 0);
+
 	tsb_context_switch(mm);
 	smp_tsb_sync(mm);
 
diff --git a/arch/sparc/mm/tlb.c b/arch/sparc/mm/tlb.c
index 3e8fec3..ba6ae7f 100644
--- a/arch/sparc/mm/tlb.c
+++ b/arch/sparc/mm/tlb.c
@@ -135,8 +135,15 @@
 			mm->context.huge_pte_count++;
 		else
 			mm->context.huge_pte_count--;
-		if (mm->context.huge_pte_count == 1)
-			hugetlb_setup(mm);
+
+		/* Do not try to allocate the TSB hash table if we
+		 * don't have one already.  We have various locks held
+		 * and thus we'll end up doing a GFP_KERNEL allocation
+		 * in an atomic context.
+		 *
+		 * Instead, we let the first TLB miss on a hugepage
+		 * take care of this.
+		 */
 	}
 
 	if (!pmd_none(orig)) {
diff --git a/arch/sparc/mm/tsb.c b/arch/sparc/mm/tsb.c
index 7f64743..428982b 100644
--- a/arch/sparc/mm/tsb.c
+++ b/arch/sparc/mm/tsb.c
@@ -314,7 +314,7 @@
 retry_tsb_alloc:
 	gfp_flags = GFP_KERNEL;
 	if (new_size > (PAGE_SIZE * 2))
-		gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
+		gfp_flags |= __GFP_NOWARN | __GFP_NORETRY;
 
 	new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
 					gfp_flags, numa_node_id());