[linux-2.6-block.git] / arch / powerpc / mm / mmu_context_nohash.c

/*
 * This file contains the routines for handling the MMU on those
 * PowerPC implementations where the MMU is not using the hash
 * table, such as 8xx, 4xx, BookE's etc...
 *
 * Copyright 2008 Ben Herrenschmidt <benh@kernel.crashing.org>
 *                IBM Corp.
 *
 *  Derived from previous arch/powerpc/mm/mmu_context.c
 *  and arch/powerpc/include/asm/mmu_context.h
 *
 *  This program is free software; you can redistribute it and/or
 *  modify it under the terms of the GNU General Public License
 *  as published by the Free Software Foundation; either version
 *  2 of the License, or (at your option) any later version.
 *
 * TODO:
 *
 *   - The global context lock will not scale very well
 *   - The maps should be dynamically allocated to allow for processors
 *     that support more PID bits at runtime
 *   - Implement flush_tlb_mm() by making the context stale and picking
 *     a new one
 *   - More aggressively clear stale map bits and maybe find some way to
 *     also clear mm->cpu_vm_mask bits when processes are migrated
 */

//#define DEBUG_MAP_CONSISTENCY
//#define DEBUG_CLAMP_LAST_CONTEXT   31
//#define DEBUG_HARDER

/* We don't use DEBUG because it tends to be compiled in always nowadays
 * and this would generate way too much output
 */
#ifdef DEBUG_HARDER
#define pr_hard(args...)	printk(KERN_DEBUG args)
#define pr_hardcont(args...)	printk(KERN_CONT args)
#else
#define pr_hard(args...)	do { } while(0)
#define pr_hardcont(args...)	do { } while(0)
#endif

#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/init.h>
#include <linux/spinlock.h>
#include <linux/bootmem.h>
#include <linux/notifier.h>
#include <linux/cpu.h>
#include <linux/slab.h>

#include <asm/mmu_context.h>
#include <asm/tlbflush.h>

static unsigned int first_context, last_context;
static unsigned int next_context, nr_free_contexts;
static unsigned long *context_map;
static unsigned long *stale_map[NR_CPUS];
static struct mm_struct **context_mm;
static DEFINE_RAW_SPINLOCK(context_lock);

#define CTX_MAP_SIZE	\
	(sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))


/* Steal a context from a task that has one at the moment.
 *
 * This is used when we are running out of available PID numbers
 * on the processors.
 *
 * This isn't an LRU system, it just frees up each context in
 * turn (sort-of pseudo-random replacement :).  This would be the
 * place to implement an LRU scheme if anyone was motivated to do it.
 *  -- paulus
 *
 * For context stealing, we use a slightly different approach for
 * SMP and UP. Basically, the UP one is simpler and doesn't use
 * the stale map as we can just flush the local CPU
 *  -- benh
 */
#ifdef CONFIG_SMP
static unsigned int steal_context_smp(unsigned int id)
{
	struct mm_struct *mm;
	unsigned int cpu, max, i;

	max = last_context - first_context;

	/* Attempt to free next_context first and then loop until we manage */
	while (max--) {
		/* Pick up the victim mm */
		mm = context_mm[id];

		/* We have a candidate victim, check if it's active, on SMP
		 * we cannot steal active contexts
		 */
		if (mm->context.active) {
			id++;
			if (id > last_context)
				id = first_context;
			continue;
		}
		pr_hardcont(" | steal %d from 0x%p", id, mm);

		/* Mark this mm has having no context anymore */
		mm->context.id = MMU_NO_CONTEXT;

		/* Mark it stale on all CPUs that used this mm. For threaded
		 * implementations, we set it on all threads on each core
		 * represented in the mask. A future implementation will use
		 * a core map instead but this will do for now.
		 */
		for_each_cpu(cpu, mm_cpumask(mm)) {
			for (i = cpu_first_thread_sibling(cpu);
			     i <= cpu_last_thread_sibling(cpu); i++) {
				if (stale_map[i])
					__set_bit(id, stale_map[i]);
			}
			cpu = i - 1;
		}
		return id;
	}

	/* This will happen if you have more CPUs than available contexts,
	 * all we can do here is wait a bit and try again
	 */
	raw_spin_unlock(&context_lock);
	cpu_relax();
	raw_spin_lock(&context_lock);

	/* This will cause the caller to try again */
	return MMU_NO_CONTEXT;
}
#endif  /* CONFIG_SMP */

/* Note that this will also be called on SMP if all other CPUs are
 * offlined, which means that it may be called for cpu != 0. For
 * this to work, we somewhat assume that CPUs that are onlined
 * come up with a fully clean TLB (or are cleaned when offlined)
 */
static unsigned int steal_context_up(unsigned int id)
{
	struct mm_struct *mm;
	int cpu = smp_processor_id();

	/* Pick up the victim mm */
	mm = context_mm[id];

	pr_hardcont(" | steal %d from 0x%p", id, mm);

	/* Flush the TLB for that context */
	local_flush_tlb_mm(mm);

	/* Mark this mm has having no context anymore */
	mm->context.id = MMU_NO_CONTEXT;

	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
	__clear_bit(id, stale_map[cpu]);

	return id;
}

#ifdef DEBUG_MAP_CONSISTENCY
static void context_check_map(void)
{
	unsigned int id, nrf, nact;

	nrf = nact = 0;
	for (id = first_context; id <= last_context; id++) {
		int used = test_bit(id, context_map);
		if (!used)
			nrf++;
		if (used != (context_mm[id] != NULL))
			pr_err("MMU: Context %d is %s and MM is %p !\n",
			       id, used ? "used" : "free", context_mm[id]);
		if (context_mm[id] != NULL)
			nact += context_mm[id]->context.active;
	}
	if (nrf != nr_free_contexts) {
		pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
		       nr_free_contexts, nrf);
		nr_free_contexts = nrf;
	}
	if (nact > num_online_cpus())
		pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
		       nact, num_online_cpus());
	if (first_context > 0 && !test_bit(0, context_map))
		pr_err("MMU: Context 0 has been freed !!!\n");
}
#else
static void context_check_map(void) { }
#endif

void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
{
	unsigned int i, id, cpu = smp_processor_id();
	unsigned long *map;

	/* No lockless fast path .. yet */
	raw_spin_lock(&context_lock);

	pr_hard("[%d] activating context for mm @%p, active=%d, id=%d",
		cpu, next, next->context.active, next->context.id);

#ifdef CONFIG_SMP
	/* Mark us active and the previous one not anymore */
	next->context.active++;
	if (prev) {
		pr_hardcont(" (old=0x%p a=%d)", prev, prev->context.active);
		WARN_ON(prev->context.active < 1);
		prev->context.active--;
	}

 again:
#endif /* CONFIG_SMP */

	/* If we already have a valid assigned context, skip all that */
	id = next->context.id;
	if (likely(id != MMU_NO_CONTEXT)) {
#ifdef DEBUG_MAP_CONSISTENCY
		if (context_mm[id] != next)
			pr_err("MMU: mm 0x%p has id %d but context_mm[%d] says 0x%p\n",
			       next, id, id, context_mm[id]);
#endif
		goto ctxt_ok;
	}

	/* We really don't have a context, let's try to acquire one */
	id = next_context;
	if (id > last_context)
		id = first_context;
	map = context_map;

	/* No more free contexts, let's try to steal one */
	if (nr_free_contexts == 0) {
#ifdef CONFIG_SMP
		if (num_online_cpus() > 1) {
			id = steal_context_smp(id);
			if (id == MMU_NO_CONTEXT)
				goto again;
			goto stolen;
		}
#endif /* CONFIG_SMP */
		id = steal_context_up(id);
		goto stolen;
	}
	nr_free_contexts--;

	/* We know there's at least one free context, try to find it */
	while (__test_and_set_bit(id, map)) {
		id = find_next_zero_bit(map, last_context+1, id);
		if (id > last_context)
			id = first_context;
	}
 stolen:
	next_context = id + 1;
	context_mm[id] = next;
	next->context.id = id;
	pr_hardcont(" | new id=%d,nrf=%d", id, nr_free_contexts);

	context_check_map();
 ctxt_ok:

	/* If that context got marked stale on this CPU, then flush the
	 * local TLB for it and unmark it before we use it
	 */
	if (test_bit(id, stale_map[cpu])) {
		pr_hardcont(" | stale flush %d [%d..%d]",
			    id, cpu_first_thread_sibling(cpu),
			    cpu_last_thread_sibling(cpu));

		local_flush_tlb_mm(next);

		/* XXX This clear should ultimately be part of local_flush_tlb_mm */
		for (i = cpu_first_thread_sibling(cpu);
		     i <= cpu_last_thread_sibling(cpu); i++) {
			if (stale_map[i])
				__clear_bit(id, stale_map[i]);
		}
	}

	/* Flick the MMU and release lock */
	pr_hardcont(" -> %d\n", id);
	set_context(id, next->pgd);
	raw_spin_unlock(&context_lock);
}

/*
 * Set up the context for a new address space.
 */
int init_new_context(struct task_struct *t, struct mm_struct *mm)
{
	pr_hard("initing context for mm @%p\n", mm);

	mm->context.id = MMU_NO_CONTEXT;
	mm->context.active = 0;

#ifdef CONFIG_PPC_MM_SLICES
	if (slice_mm_new_context(mm))
		slice_set_user_psize(mm, mmu_virtual_psize);
#endif

	return 0;
}

/*
 * We're finished using the context for an address space.
 */
void destroy_context(struct mm_struct *mm)
{
	unsigned long flags;
	unsigned int id;

	if (mm->context.id == MMU_NO_CONTEXT)
		return;

	WARN_ON(mm->context.active != 0);

	raw_spin_lock_irqsave(&context_lock, flags);
	id = mm->context.id;
	if (id != MMU_NO_CONTEXT) {
		__clear_bit(id, context_map);
		mm->context.id = MMU_NO_CONTEXT;
#ifdef DEBUG_MAP_CONSISTENCY
		mm->context.active = 0;
#endif
		context_mm[id] = NULL;
		nr_free_contexts++;
	}
	raw_spin_unlock_irqrestore(&context_lock, flags);
}

#ifdef CONFIG_SMP

static int mmu_context_cpu_notify(struct notifier_block *self,
				  unsigned long action, void *hcpu)
{
	unsigned int cpu = (unsigned int)(long)hcpu;

	/* We don't touch CPU 0 map, it's allocated at aboot and kept
	 * around forever
	 */
	if (cpu == boot_cpuid)
		return NOTIFY_OK;

	switch (action) {
	case CPU_UP_PREPARE:
	case CPU_UP_PREPARE_FROZEN:
		pr_devel("MMU: Allocating stale context map for CPU %d\n", cpu);
		stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
		break;
#ifdef CONFIG_HOTPLUG_CPU
	case CPU_UP_CANCELED:
	case CPU_UP_CANCELED_FROZEN:
	case CPU_DEAD:
	case CPU_DEAD_FROZEN:
		pr_devel("MMU: Freeing stale context map for CPU %d\n", cpu);
		kfree(stale_map[cpu]);
		stale_map[cpu] = NULL;

		/* We also clear the cpu_vm_mask bits of CPUs going away */
		clear_tasks_mm_cpumask(cpu);
	break;
#endif /* CONFIG_HOTPLUG_CPU */
	}
	return NOTIFY_OK;
}

static struct notifier_block mmu_context_cpu_nb = {
	.notifier_call	= mmu_context_cpu_notify,
};

#endif /* CONFIG_SMP */

/*
 * Initialize the context management stuff.
 */
void __init mmu_context_init(void)
{
	/* Mark init_mm as being active on all possible CPUs since
	 * we'll get called with prev == init_mm the first time
	 * we schedule on a given CPU
	 */
	init_mm.context.active = NR_CPUS;

	/*
	 *   The MPC8xx has only 16 contexts.  We rotate through them on each
	 * task switch.  A better way would be to keep track of tasks that
	 * own contexts, and implement an LRU usage.  That way very active
	 * tasks don't always have to pay the TLB reload overhead.  The
	 * kernel pages are mapped shared, so the kernel can run on behalf
	 * of any task that makes a kernel entry.  Shared does not mean they
	 * are not protected, just that the ASID comparison is not performed.
	 *      -- Dan
	 *
	 * The IBM4xx has 256 contexts, so we can just rotate through these
	 * as a way of "switching" contexts.  If the TID of the TLB is zero,
	 * the PID/TID comparison is disabled, so we can use a TID of zero
	 * to represent all kernel pages as shared among all contexts.
	 * 	-- Dan
	 *
	 * The IBM 47x core supports 16-bit PIDs, thus 65535 contexts. We
	 * should normally never have to steal though the facility is
	 * present if needed.
	 *      -- BenH
	 */
	if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
		first_context = 0;
		last_context = 15;
	} else if (mmu_has_feature(MMU_FTR_TYPE_47x)) {
		first_context = 1;
		last_context = 65535;
	} else
#ifdef CONFIG_PPC_BOOK3E_MMU
	if (mmu_has_feature(MMU_FTR_TYPE_3E)) {
		u32 mmucfg = mfspr(SPRN_MMUCFG);
		u32 pid_bits = (mmucfg & MMUCFG_PIDSIZE_MASK)
				>> MMUCFG_PIDSIZE_SHIFT;
		first_context = 1;
		last_context = (1UL << (pid_bits + 1)) - 1;
	} else
#endif
	{
		first_context = 1;
		last_context = 255;
	}

#ifdef DEBUG_CLAMP_LAST_CONTEXT
	last_context = DEBUG_CLAMP_LAST_CONTEXT;
#endif
	/*
	 * Allocate the maps used by context management
	 */
	context_map = alloc_bootmem(CTX_MAP_SIZE);
	context_mm = alloc_bootmem(sizeof(void *) * (last_context + 1));
#ifndef CONFIG_SMP
	stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
#else
	stale_map[boot_cpuid] = alloc_bootmem(CTX_MAP_SIZE);

	register_cpu_notifier(&mmu_context_cpu_nb);
#endif

	printk(KERN_INFO
	       "MMU: Allocated %zu bytes of context maps for %d contexts\n",
	       2 * CTX_MAP_SIZE + (sizeof(void *) * (last_context + 1)),
	       last_context - first_context + 1);

	/*
	 * Some processors have too few contexts to reserve one for
	 * init_mm, and require using context 0 for a normal task.
	 * Other processors reserve the use of context zero for the kernel.
	 * This code assumes first_context < 32.
	 */
	context_map[0] = (1 << first_context) - 1;
	next_context = first_context;
	nr_free_contexts = last_context - first_context + 1;
}
Commit	Line	Data
5e696617 BH	1	/*
	2	* This file contains the routines for handling the MMU on those
	3	* PowerPC implementations where the MMU is not using the hash
	4	* table, such as 8xx, 4xx, BookE's etc...
	5	*
	6	* Copyright 2008 Ben Herrenschmidt <benh@kernel.crashing.org>
	7	* IBM Corp.
	8	*
	9	* Derived from previous arch/powerpc/mm/mmu_context.c
	10	* and arch/powerpc/include/asm/mmu_context.h
	11	*
	12	* This program is free software; you can redistribute it and/or
	13	* modify it under the terms of the GNU General Public License
	14	* as published by the Free Software Foundation; either version
	15	* 2 of the License, or (at your option) any later version.
	16	*
2ca8cf73 BH	17	* TODO:
	18	*
	19	* - The global context lock will not scale very well
	20	* - The maps should be dynamically allocated to allow for processors
	21	* that support more PID bits at runtime
	22	* - Implement flush_tlb_mm() by making the context stale and picking
	23	* a new one
	24	* - More aggressively clear stale map bits and maybe find some way to
	25	* also clear mm->cpu_vm_mask bits when processes are migrated
5e696617 BH	26	*/
5e696617 BH	27
f1167fb3 BH	28	//#define DEBUG_MAP_CONSISTENCY
f1167fb3 BH	29	//#define DEBUG_CLAMP_LAST_CONTEXT 31
fcce8109 BH	30	//#define DEBUG_HARDER
	31
	32	/* We don't use DEBUG because it tends to be compiled in always nowadays
	33	* and this would generate way too much output
	34	*/
	35	#ifdef DEBUG_HARDER
	36	#define pr_hard(args...) printk(KERN_DEBUG args)
	37	#define pr_hardcont(args...) printk(KERN_CONT args)
	38	#else
	39	#define pr_hard(args...) do { } while(0)
	40	#define pr_hardcont(args...) do { } while(0)
	41	#endif
2ca8cf73 BH	42
2ca8cf73 BH	43	#include <linux/kernel.h>
5e696617 BH	44	#include <linux/mm.h>
5e696617 BH	45	#include <linux/init.h>
77520351 BH	46	#include <linux/spinlock.h>
	47	#include <linux/bootmem.h>
	48	#include <linux/notifier.h>
	49	#include <linux/cpu.h>
5a0e3ad6	50	#include <linux/slab.h>
5e696617 BH	51
	52	#include <asm/mmu_context.h>
	53	#include <asm/tlbflush.h>
5e696617	54
77520351	55	static unsigned int first_context, last_context;
2ca8cf73	56	static unsigned int next_context, nr_free_contexts;
77520351 BH	57	static unsigned long *context_map;
	58	static unsigned long *stale_map[NR_CPUS];
	59	static struct mm_struct **context_mm;
be833f33	60	static DEFINE_RAW_SPINLOCK(context_lock);
5e696617	61
77520351 BH	62	#define CTX_MAP_SIZE \
	63	(sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
	64
	65
5e696617	66	/* Steal a context from a task that has one at the moment.
2ca8cf73 BH	67	*
	68	* This is used when we are running out of available PID numbers
	69	* on the processors.
	70	*
5e696617 BH	71	* This isn't an LRU system, it just frees up each context in
	72	* turn (sort-of pseudo-random replacement :). This would be the
	73	* place to implement an LRU scheme if anyone was motivated to do it.
	74	* -- paulus
2ca8cf73 BH	75	*
	76	* For context stealing, we use a slightly different approach for
	77	* SMP and UP. Basically, the UP one is simpler and doesn't use
	78	* the stale map as we can just flush the local CPU
	79	* -- benh
5e696617	80	*/
2ca8cf73 BH	81	#ifdef CONFIG_SMP
2ca8cf73 BH	82	static unsigned int steal_context_smp(unsigned int id)
5e696617 BH	83	{
5e696617 BH	84	struct mm_struct *mm;
fcce8109	85	unsigned int cpu, max, i;
5e696617	86
77520351	87	max = last_context - first_context;
5e696617	88
2ca8cf73 BH	89	/* Attempt to free next_context first and then loop until we manage */
	90	while (max--) {
	91	/* Pick up the victim mm */
	92	mm = context_mm[id];
5e696617	93
2ca8cf73 BH	94	/* We have a candidate victim, check if it's active, on SMP
	95	* we cannot steal active contexts
	96	*/
	97	if (mm->context.active) {
	98	id++;
77520351 BH	99	if (id > last_context)
77520351 BH	100	id = first_context;
2ca8cf73 BH	101	continue;
2ca8cf73 BH	102	}
fcce8109	103	pr_hardcont(" \| steal %d from 0x%p", id, mm);
2ca8cf73 BH	104
	105	/* Mark this mm has having no context anymore */
	106	mm->context.id = MMU_NO_CONTEXT;
	107
fcce8109 BH	108	/* Mark it stale on all CPUs that used this mm. For threaded
	109	* implementations, we set it on all threads on each core
	110	* represented in the mask. A future implementation will use
	111	* a core map instead but this will do for now.
	112	*/
	113	for_each_cpu(cpu, mm_cpumask(mm)) {
99d86705	114	for (i = cpu_first_thread_sibling(cpu);
39a421ff SW	115	i <= cpu_last_thread_sibling(cpu); i++) {
	116	if (stale_map[i])
	117	__set_bit(id, stale_map[i]);
	118	}
fcce8109 BH	119	cpu = i - 1;
fcce8109 BH	120	}
2ca8cf73 BH	121	return id;
	122	}
	123
	124	/* This will happen if you have more CPUs than available contexts,
	125	* all we can do here is wait a bit and try again
	126	*/
be833f33	127	raw_spin_unlock(&context_lock);
2ca8cf73	128	cpu_relax();
be833f33	129	raw_spin_lock(&context_lock);
3035c863 BH	130
	131	/* This will cause the caller to try again */
	132	return MMU_NO_CONTEXT;
2ca8cf73 BH	133	}
	134	#endif /* CONFIG_SMP */
	135
	136	/* Note that this will also be called on SMP if all other CPUs are
	137	* offlined, which means that it may be called for cpu != 0. For
	138	* this to work, we somewhat assume that CPUs that are onlined
	139	* come up with a fully clean TLB (or are cleaned when offlined)
5e696617	140	*/
2ca8cf73	141	static unsigned int steal_context_up(unsigned int id)
5e696617	142	{
2ca8cf73 BH	143	struct mm_struct *mm;
2ca8cf73 BH	144	int cpu = smp_processor_id();
5e696617	145
2ca8cf73 BH	146	/* Pick up the victim mm */
	147	mm = context_mm[id];
	148
fcce8109	149	pr_hardcont(" \| steal %d from 0x%p", id, mm);
5e696617	150
2ca8cf73 BH	151	/* Flush the TLB for that context */
	152	local_flush_tlb_mm(mm);
	153
8e35961b HS	154	/* Mark this mm has having no context anymore */
	155	mm->context.id = MMU_NO_CONTEXT;
	156
2ca8cf73 BH	157	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
	158	__clear_bit(id, stale_map[cpu]);
	159
	160	return id;
	161	}
	162
	163	#ifdef DEBUG_MAP_CONSISTENCY
	164	static void context_check_map(void)
	165	{
	166	unsigned int id, nrf, nact;
	167
	168	nrf = nact = 0;
77520351	169	for (id = first_context; id <= last_context; id++) {
2ca8cf73 BH	170	int used = test_bit(id, context_map);
	171	if (!used)
	172	nrf++;
	173	if (used != (context_mm[id] != NULL))
	174	pr_err("MMU: Context %d is %s and MM is %p !\n",
	175	id, used ? "used" : "free", context_mm[id]);
	176	if (context_mm[id] != NULL)
	177	nact += context_mm[id]->context.active;
5e696617	178	}
2ca8cf73 BH	179	if (nrf != nr_free_contexts) {
	180	pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
	181	nr_free_contexts, nrf);
	182	nr_free_contexts = nrf;
	183	}
	184	if (nact > num_online_cpus())
	185	pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
	186	nact, num_online_cpus());
77520351 BH	187	if (first_context > 0 && !test_bit(0, context_map))
77520351 BH	188	pr_err("MMU: Context 0 has been freed !!!\n");
5e696617	189	}
2ca8cf73 BH	190	#else
	191	static void context_check_map(void) { }
	192	#endif
5e696617 BH	193
	194	void switch_mmu_context(struct mm_struct prev, struct mm_struct next)
	195	{
67050b5c	196	unsigned int i, id, cpu = smp_processor_id();
2ca8cf73	197	unsigned long *map;
5e696617	198
2ca8cf73	199	/* No lockless fast path .. yet */
be833f33	200	raw_spin_lock(&context_lock);
2ca8cf73	201
fcce8109 BH	202	pr_hard("[%d] activating context for mm @%p, active=%d, id=%d",
fcce8109 BH	203	cpu, next, next->context.active, next->context.id);
2ca8cf73 BH	204
	205	#ifdef CONFIG_SMP
	206	/* Mark us active and the previous one not anymore */
	207	next->context.active++;
	208	if (prev) {
fcce8109	209	pr_hardcont(" (old=0x%p a=%d)", prev, prev->context.active);
2ca8cf73 BH	210	WARN_ON(prev->context.active < 1);
	211	prev->context.active--;
	212	}
3035c863 BH	213
3035c863 BH	214	again:
2ca8cf73 BH	215	#endif /* CONFIG_SMP */
	216
	217	/* If we already have a valid assigned context, skip all that */
	218	id = next->context.id;
fcce8109 BH	219	if (likely(id != MMU_NO_CONTEXT)) {
	220	#ifdef DEBUG_MAP_CONSISTENCY
	221	if (context_mm[id] != next)
	222	pr_err("MMU: mm 0x%p has id %d but context_mm[%d] says 0x%p\n",
	223	next, id, id, context_mm[id]);
	224	#endif
2ca8cf73	225	goto ctxt_ok;
fcce8109	226	}
2ca8cf73 BH	227
	228	/* We really don't have a context, let's try to acquire one */
	229	id = next_context;
77520351 BH	230	if (id > last_context)
77520351 BH	231	id = first_context;
2ca8cf73 BH	232	map = context_map;
	233
	234	/* No more free contexts, let's try to steal one */
	235	if (nr_free_contexts == 0) {
	236	#ifdef CONFIG_SMP
	237	if (num_online_cpus() > 1) {
	238	id = steal_context_smp(id);
3035c863 BH	239	if (id == MMU_NO_CONTEXT)
3035c863 BH	240	goto again;
5156ddce	241	goto stolen;
2ca8cf73 BH	242	}
	243	#endif /* CONFIG_SMP */
	244	id = steal_context_up(id);
	245	goto stolen;
	246	}
	247	nr_free_contexts--;
	248
	249	/* We know there's at least one free context, try to find it */
	250	while (__test_and_set_bit(id, map)) {
77520351 BH	251	id = find_next_zero_bit(map, last_context+1, id);
	252	if (id > last_context)
	253	id = first_context;
2ca8cf73 BH	254	}
	255	stolen:
	256	next_context = id + 1;
	257	context_mm[id] = next;
	258	next->context.id = id;
fcce8109	259	pr_hardcont(" \| new id=%d,nrf=%d", id, nr_free_contexts);
2ca8cf73 BH	260
	261	context_check_map();
	262	ctxt_ok:
	263
	264	/* If that context got marked stale on this CPU, then flush the
	265	* local TLB for it and unmark it before we use it
	266	*/
	267	if (test_bit(id, stale_map[cpu])) {
fcce8109	268	pr_hardcont(" \| stale flush %d [%d..%d]",
99d86705 VS	269	id, cpu_first_thread_sibling(cpu),
99d86705 VS	270	cpu_last_thread_sibling(cpu));
fcce8109	271
2ca8cf73 BH	272	local_flush_tlb_mm(next);
	273
	274	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
99d86705 VS	275	for (i = cpu_first_thread_sibling(cpu);
99d86705 VS	276	i <= cpu_last_thread_sibling(cpu); i++) {
39a421ff SW	277	if (stale_map[i])
39a421ff SW	278	__clear_bit(id, stale_map[i]);
67050b5c	279	}
2ca8cf73 BH	280	}
	281
	282	/* Flick the MMU and release lock */
fcce8109	283	pr_hardcont(" -> %d\n", id);
2ca8cf73	284	set_context(id, next->pgd);
be833f33	285	raw_spin_unlock(&context_lock);
5e696617 BH	286	}
	287
	288	/*
	289	* Set up the context for a new address space.
	290	*/
	291	int init_new_context(struct task_struct t, struct mm_struct mm)
	292	{
fcce8109 BH	293	pr_hard("initing context for mm @%p\n", mm);
fcce8109 BH	294
2ca8cf73 BH	295	mm->context.id = MMU_NO_CONTEXT;
	296	mm->context.active = 0;
	297
41151e77 BB	298	#ifdef CONFIG_PPC_MM_SLICES
	299	if (slice_mm_new_context(mm))
	300	slice_set_user_psize(mm, mmu_virtual_psize);
	301	#endif
	302
5e696617 BH	303	return 0;
	304	}
	305
	306	/*
	307	* We're finished using the context for an address space.
	308	*/
	309	void destroy_context(struct mm_struct *mm)
	310	{
b46b6942	311	unsigned long flags;
2ca8cf73 BH	312	unsigned int id;
	313
	314	if (mm->context.id == MMU_NO_CONTEXT)
	315	return;
	316
	317	WARN_ON(mm->context.active != 0);
	318
be833f33	319	raw_spin_lock_irqsave(&context_lock, flags);
2ca8cf73 BH	320	id = mm->context.id;
	321	if (id != MMU_NO_CONTEXT) {
	322	__clear_bit(id, context_map);
	323	mm->context.id = MMU_NO_CONTEXT;
	324	#ifdef DEBUG_MAP_CONSISTENCY
	325	mm->context.active = 0;
2ca8cf73	326	#endif
3035c863	327	context_mm[id] = NULL;
2ca8cf73	328	nr_free_contexts++;
5e696617	329	}
be833f33	330	raw_spin_unlock_irqrestore(&context_lock, flags);
5e696617 BH	331	}
5e696617 BH	332
77520351 BH	333	#ifdef CONFIG_SMP
77520351 BH	334
061d19f2 PG	335	static int mmu_context_cpu_notify(struct notifier_block *self,
061d19f2 PG	336	unsigned long action, void *hcpu)
77520351 BH	337	{
77520351 BH	338	unsigned int cpu = (unsigned int)(long)hcpu;
73863ab0	339
77520351 BH	340	/* We don't touch CPU 0 map, it's allocated at aboot and kept
	341	* around forever
	342	*/
0d35e162	343	if (cpu == boot_cpuid)
77520351 BH	344	return NOTIFY_OK;
	345
	346	switch (action) {
f5be2dc0 ME	347	case CPU_UP_PREPARE:
f5be2dc0 ME	348	case CPU_UP_PREPARE_FROZEN:
a1ac38ab	349	pr_devel("MMU: Allocating stale context map for CPU %d\n", cpu);
77520351 BH	350	stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
	351	break;
	352	#ifdef CONFIG_HOTPLUG_CPU
f5be2dc0 ME	353	case CPU_UP_CANCELED:
f5be2dc0 ME	354	case CPU_UP_CANCELED_FROZEN:
77520351 BH	355	case CPU_DEAD:
77520351 BH	356	case CPU_DEAD_FROZEN:
a1ac38ab	357	pr_devel("MMU: Freeing stale context map for CPU %d\n", cpu);
77520351 BH	358	kfree(stale_map[cpu]);
77520351 BH	359	stale_map[cpu] = NULL;
fcce8109 BH	360
fcce8109 BH	361	/* We also clear the cpu_vm_mask bits of CPUs going away */
73863ab0	362	clear_tasks_mm_cpumask(cpu);
fcce8109 BH	363	break;
fcce8109 BH	364	#endif /* CONFIG_HOTPLUG_CPU */
77520351 BH	365	}
	366	return NOTIFY_OK;
	367	}
	368
061d19f2	369	static struct notifier_block mmu_context_cpu_nb = {
77520351 BH	370	.notifier_call = mmu_context_cpu_notify,
	371	};
	372
	373	#endif /* CONFIG_SMP */
5e696617 BH	374
	375	/*
	376	* Initialize the context management stuff.
	377	*/
	378	void __init mmu_context_init(void)
	379	{
2ca8cf73 BH	380	/* Mark init_mm as being active on all possible CPUs since
	381	* we'll get called with prev == init_mm the first time
	382	* we schedule on a given CPU
	383	*/
	384	init_mm.context.active = NR_CPUS;
	385
77520351 BH	386	/*
	387	* The MPC8xx has only 16 contexts. We rotate through them on each
	388	* task switch. A better way would be to keep track of tasks that
	389	* own contexts, and implement an LRU usage. That way very active
	390	* tasks don't always have to pay the TLB reload overhead. The
	391	* kernel pages are mapped shared, so the kernel can run on behalf
	392	* of any task that makes a kernel entry. Shared does not mean they
	393	* are not protected, just that the ASID comparison is not performed.
	394	* -- Dan
	395	*
	396	* The IBM4xx has 256 contexts, so we can just rotate through these
	397	* as a way of "switching" contexts. If the TID of the TLB is zero,
	398	* the PID/TID comparison is disabled, so we can use a TID of zero
	399	* to represent all kernel pages as shared among all contexts.
	400	* -- Dan
e7f75ad0 DK	401	*
	402	* The IBM 47x core supports 16-bit PIDs, thus 65535 contexts. We
	403	* should normally never have to steal though the facility is
	404	* present if needed.
	405	* -- BenH
77520351 BH	406	*/
	407	if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
	408	first_context = 0;
	409	last_context = 15;
e7f75ad0 DK	410	} else if (mmu_has_feature(MMU_FTR_TYPE_47x)) {
	411	first_context = 1;
	412	last_context = 65535;
bd491781 BH	413	} else
	414	#ifdef CONFIG_PPC_BOOK3E_MMU
	415	if (mmu_has_feature(MMU_FTR_TYPE_3E)) {
	416	u32 mmucfg = mfspr(SPRN_MMUCFG);
	417	u32 pid_bits = (mmucfg & MMUCFG_PIDSIZE_MASK)
	418	>> MMUCFG_PIDSIZE_SHIFT;
	419	first_context = 1;
	420	last_context = (1UL << (pid_bits + 1)) - 1;
	421	} else
	422	#endif
	423	{
77520351 BH	424	first_context = 1;
	425	last_context = 255;
	426	}
	427
	428	#ifdef DEBUG_CLAMP_LAST_CONTEXT
	429	last_context = DEBUG_CLAMP_LAST_CONTEXT;
	430	#endif
	431	/*
	432	* Allocate the maps used by context management
	433	*/
	434	context_map = alloc_bootmem(CTX_MAP_SIZE);
	435	context_mm = alloc_bootmem(sizeof(void ) (last_context + 1));
0d35e162	436	#ifndef CONFIG_SMP
77520351	437	stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
0d35e162 MM	438	#else
0d35e162 MM	439	stale_map[boot_cpuid] = alloc_bootmem(CTX_MAP_SIZE);
77520351	440
77520351 BH	441	register_cpu_notifier(&mmu_context_cpu_nb);
	442	#endif
	443
	444	printk(KERN_INFO
ff7c6600	445	"MMU: Allocated %zu bytes of context maps for %d contexts\n",
77520351 BH	446	2 * CTX_MAP_SIZE + (sizeof(void ) (last_context + 1)),
	447	last_context - first_context + 1);
	448
5e696617 BH	449	/*
	450	* Some processors have too few contexts to reserve one for
	451	* init_mm, and require using context 0 for a normal task.
	452	* Other processors reserve the use of context zero for the kernel.
77520351	453	* This code assumes first_context < 32.
5e696617	454	*/
77520351 BH	455	context_map[0] = (1 << first_context) - 1;
	456	next_context = first_context;
	457	nr_free_contexts = last_context - first_context + 1;
5e696617 BH	458	}
5e696617 BH	459