[deliverable/linux.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
 */

#undef DEBUG
#define DEBUG_STEAL_ONLY
#undef DEBUG_MAP_CONSISTENCY
/*#define DEBUG_CLAMP_LAST_CONTEXT   15 */

#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 <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 spinlock_t context_lock = SPIN_LOCK_UNLOCKED;

#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;

	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_debug("[%d] steal context %d from mm @%p\n",
			 smp_processor_id(), 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_each_cpu(cpu, mm_cpumask(mm))
			__set_bit(id, stale_map[cpu]);
		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
	 */
	spin_unlock(&context_lock);
	cpu_relax();
	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_debug("[%d] steal context %d from mm @%p\n", cpu, 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 id, cpu = smp_processor_id();
	unsigned long *map;

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

#ifndef DEBUG_STEAL_ONLY
	pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n",
		 cpu, next, next->context.active, next->context.id);
#endif

#ifdef CONFIG_SMP
	/* Mark us active and the previous one not anymore */
	next->context.active++;
	if (prev) {
#ifndef DEBUG_STEAL_ONLY
		pr_debug(" old context %p active was: %d\n",
			 prev, prev->context.active);
#endif
		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))
		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;
		}
#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;

#ifndef DEBUG_STEAL_ONLY
	pr_debug("[%d] picked up new id %d, nrf is now %d\n",
		 cpu, id, nr_free_contexts);
#endif

	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_debug("[%d] flushing stale context %d for mm @%p !\n",
			 cpu, id, next);
		local_flush_tlb_mm(next);

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

	/* Flick the MMU and release lock */
	set_context(id, next->pgd);
	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)
{
	mm->context.id = MMU_NO_CONTEXT;
	mm->context.active = 0;

	return 0;
}

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

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

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

	spin_lock(&context_lock);
	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++;
	}
	spin_unlock(&context_lock);
}

#ifdef CONFIG_SMP

static int __cpuinit 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 == 0)
		return NOTIFY_OK;

	switch (action) {
	case CPU_ONLINE:
	case CPU_ONLINE_FROZEN:
		pr_debug("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_DEAD:
	case CPU_DEAD_FROZEN:
		pr_debug("MMU: Freeing stale context map for CPU %d\n", cpu);
		kfree(stale_map[cpu]);
		stale_map[cpu] = NULL;
		break;
#endif
	}
	return NOTIFY_OK;
}

static struct notifier_block __cpuinitdata 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
	 */
	if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
		first_context = 0;
		last_context = 15;
	} else {
		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));
	stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);

#ifdef CONFIG_SMP
	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
2ca8cf73 BH	28	#undef DEBUG
	29	#define DEBUG_STEAL_ONLY
	30	#undef DEBUG_MAP_CONSISTENCY
77520351	31	/#define DEBUG_CLAMP_LAST_CONTEXT 15 /
2ca8cf73 BH	32
2ca8cf73 BH	33	#include <linux/kernel.h>
5e696617 BH	34	#include <linux/mm.h>
5e696617 BH	35	#include <linux/init.h>
77520351 BH	36	#include <linux/spinlock.h>
	37	#include <linux/bootmem.h>
	38	#include <linux/notifier.h>
	39	#include <linux/cpu.h>
5e696617 BH	40
	41	#include <asm/mmu_context.h>
	42	#include <asm/tlbflush.h>
5e696617	43
77520351	44	static unsigned int first_context, last_context;
2ca8cf73	45	static unsigned int next_context, nr_free_contexts;
77520351 BH	46	static unsigned long *context_map;
	47	static unsigned long *stale_map[NR_CPUS];
	48	static struct mm_struct **context_mm;
2ca8cf73	49	static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
5e696617	50
77520351 BH	51	#define CTX_MAP_SIZE \
	52	(sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
	53
	54
5e696617	55	/* Steal a context from a task that has one at the moment.
2ca8cf73 BH	56	*
	57	* This is used when we are running out of available PID numbers
	58	* on the processors.
	59	*
5e696617 BH	60	* This isn't an LRU system, it just frees up each context in
	61	* turn (sort-of pseudo-random replacement :). This would be the
	62	* place to implement an LRU scheme if anyone was motivated to do it.
	63	* -- paulus
2ca8cf73 BH	64	*
	65	* For context stealing, we use a slightly different approach for
	66	* SMP and UP. Basically, the UP one is simpler and doesn't use
	67	* the stale map as we can just flush the local CPU
	68	* -- benh
5e696617	69	*/
2ca8cf73 BH	70	#ifdef CONFIG_SMP
2ca8cf73 BH	71	static unsigned int steal_context_smp(unsigned int id)
5e696617 BH	72	{
5e696617 BH	73	struct mm_struct *mm;
2ca8cf73	74	unsigned int cpu, max;
5e696617	75
77520351	76	max = last_context - first_context;
5e696617	77
2ca8cf73 BH	78	/* Attempt to free next_context first and then loop until we manage */
	79	while (max--) {
	80	/* Pick up the victim mm */
	81	mm = context_mm[id];
5e696617	82
2ca8cf73 BH	83	/* We have a candidate victim, check if it's active, on SMP
	84	* we cannot steal active contexts
	85	*/
	86	if (mm->context.active) {
	87	id++;
77520351 BH	88	if (id > last_context)
77520351 BH	89	id = first_context;
2ca8cf73 BH	90	continue;
	91	}
	92	pr_debug("[%d] steal context %d from mm @%p\n",
	93	smp_processor_id(), id, mm);
	94
	95	/* Mark this mm has having no context anymore */
	96	mm->context.id = MMU_NO_CONTEXT;
	97
	98	/* Mark it stale on all CPUs that used this mm */
56aa4129	99	for_each_cpu(cpu, mm_cpumask(mm))
2ca8cf73 BH	100	__set_bit(id, stale_map[cpu]);
	101	return id;
	102	}
	103
	104	/* This will happen if you have more CPUs than available contexts,
	105	* all we can do here is wait a bit and try again
	106	*/
	107	spin_unlock(&context_lock);
	108	cpu_relax();
	109	spin_lock(&context_lock);
3035c863 BH	110
	111	/* This will cause the caller to try again */
	112	return MMU_NO_CONTEXT;
2ca8cf73 BH	113	}
	114	#endif /* CONFIG_SMP */
	115
	116	/* Note that this will also be called on SMP if all other CPUs are
	117	* offlined, which means that it may be called for cpu != 0. For
	118	* this to work, we somewhat assume that CPUs that are onlined
	119	* come up with a fully clean TLB (or are cleaned when offlined)
5e696617	120	*/
2ca8cf73	121	static unsigned int steal_context_up(unsigned int id)
5e696617	122	{
2ca8cf73 BH	123	struct mm_struct *mm;
2ca8cf73 BH	124	int cpu = smp_processor_id();
5e696617	125
2ca8cf73 BH	126	/* Pick up the victim mm */
	127	mm = context_mm[id];
	128
	129	pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm);
5e696617	130
2ca8cf73 BH	131	/* Flush the TLB for that context */
	132	local_flush_tlb_mm(mm);
	133
8e35961b HS	134	/* Mark this mm has having no context anymore */
	135	mm->context.id = MMU_NO_CONTEXT;
	136
2ca8cf73 BH	137	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
	138	__clear_bit(id, stale_map[cpu]);
	139
	140	return id;
	141	}
	142
	143	#ifdef DEBUG_MAP_CONSISTENCY
	144	static void context_check_map(void)
	145	{
	146	unsigned int id, nrf, nact;
	147
	148	nrf = nact = 0;
77520351	149	for (id = first_context; id <= last_context; id++) {
2ca8cf73 BH	150	int used = test_bit(id, context_map);
	151	if (!used)
	152	nrf++;
	153	if (used != (context_mm[id] != NULL))
	154	pr_err("MMU: Context %d is %s and MM is %p !\n",
	155	id, used ? "used" : "free", context_mm[id]);
	156	if (context_mm[id] != NULL)
	157	nact += context_mm[id]->context.active;
5e696617	158	}
2ca8cf73 BH	159	if (nrf != nr_free_contexts) {
	160	pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
	161	nr_free_contexts, nrf);
	162	nr_free_contexts = nrf;
	163	}
	164	if (nact > num_online_cpus())
	165	pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
	166	nact, num_online_cpus());
77520351 BH	167	if (first_context > 0 && !test_bit(0, context_map))
77520351 BH	168	pr_err("MMU: Context 0 has been freed !!!\n");
5e696617	169	}
2ca8cf73 BH	170	#else
	171	static void context_check_map(void) { }
	172	#endif
5e696617 BH	173
	174	void switch_mmu_context(struct mm_struct prev, struct mm_struct next)
	175	{
2ca8cf73 BH	176	unsigned int id, cpu = smp_processor_id();
2ca8cf73 BH	177	unsigned long *map;
5e696617	178
2ca8cf73 BH	179	/* No lockless fast path .. yet */
	180	spin_lock(&context_lock);
	181
	182	#ifndef DEBUG_STEAL_ONLY
	183	pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n",
	184	cpu, next, next->context.active, next->context.id);
	185	#endif
	186
	187	#ifdef CONFIG_SMP
	188	/* Mark us active and the previous one not anymore */
	189	next->context.active++;
	190	if (prev) {
77520351 BH	191	#ifndef DEBUG_STEAL_ONLY
	192	pr_debug(" old context %p active was: %d\n",
	193	prev, prev->context.active);
	194	#endif
2ca8cf73 BH	195	WARN_ON(prev->context.active < 1);
	196	prev->context.active--;
	197	}
3035c863 BH	198
3035c863 BH	199	again:
2ca8cf73 BH	200	#endif /* CONFIG_SMP */
	201
	202	/* If we already have a valid assigned context, skip all that */
	203	id = next->context.id;
	204	if (likely(id != MMU_NO_CONTEXT))
	205	goto ctxt_ok;
	206
	207	/* We really don't have a context, let's try to acquire one */
	208	id = next_context;
77520351 BH	209	if (id > last_context)
77520351 BH	210	id = first_context;
2ca8cf73 BH	211	map = context_map;
	212
	213	/* No more free contexts, let's try to steal one */
	214	if (nr_free_contexts == 0) {
	215	#ifdef CONFIG_SMP
	216	if (num_online_cpus() > 1) {
	217	id = steal_context_smp(id);
3035c863 BH	218	if (id == MMU_NO_CONTEXT)
3035c863 BH	219	goto again;
2ca8cf73 BH	220	}
	221	#endif /* CONFIG_SMP */
	222	id = steal_context_up(id);
	223	goto stolen;
	224	}
	225	nr_free_contexts--;
	226
	227	/* We know there's at least one free context, try to find it */
	228	while (__test_and_set_bit(id, map)) {
77520351 BH	229	id = find_next_zero_bit(map, last_context+1, id);
	230	if (id > last_context)
	231	id = first_context;
2ca8cf73 BH	232	}
	233	stolen:
	234	next_context = id + 1;
	235	context_mm[id] = next;
	236	next->context.id = id;
	237
	238	#ifndef DEBUG_STEAL_ONLY
	239	pr_debug("[%d] picked up new id %d, nrf is now %d\n",
	240	cpu, id, nr_free_contexts);
	241	#endif
	242
	243	context_check_map();
	244	ctxt_ok:
	245
	246	/* If that context got marked stale on this CPU, then flush the
	247	* local TLB for it and unmark it before we use it
	248	*/
	249	if (test_bit(id, stale_map[cpu])) {
	250	pr_debug("[%d] flushing stale context %d for mm @%p !\n",
	251	cpu, id, next);
	252	local_flush_tlb_mm(next);
	253
	254	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
	255	__clear_bit(id, stale_map[cpu]);
	256	}
	257
	258	/* Flick the MMU and release lock */
	259	set_context(id, next->pgd);
	260	spin_unlock(&context_lock);
5e696617 BH	261	}
	262
	263	/*
	264	* Set up the context for a new address space.
	265	*/
	266	int init_new_context(struct task_struct t, struct mm_struct mm)
	267	{
2ca8cf73 BH	268	mm->context.id = MMU_NO_CONTEXT;
	269	mm->context.active = 0;
	270
5e696617 BH	271	return 0;
	272	}
	273
	274	/*
	275	* We're finished using the context for an address space.
	276	*/
	277	void destroy_context(struct mm_struct *mm)
	278	{
2ca8cf73 BH	279	unsigned int id;
	280
	281	if (mm->context.id == MMU_NO_CONTEXT)
	282	return;
	283
	284	WARN_ON(mm->context.active != 0);
	285
	286	spin_lock(&context_lock);
	287	id = mm->context.id;
	288	if (id != MMU_NO_CONTEXT) {
	289	__clear_bit(id, context_map);
	290	mm->context.id = MMU_NO_CONTEXT;
	291	#ifdef DEBUG_MAP_CONSISTENCY
	292	mm->context.active = 0;
2ca8cf73	293	#endif
3035c863	294	context_mm[id] = NULL;
2ca8cf73	295	nr_free_contexts++;
5e696617	296	}
2ca8cf73	297	spin_unlock(&context_lock);
5e696617 BH	298	}
5e696617 BH	299
77520351 BH	300	#ifdef CONFIG_SMP
	301
	302	static int __cpuinit mmu_context_cpu_notify(struct notifier_block *self,
	303	unsigned long action, void *hcpu)
	304	{
	305	unsigned int cpu = (unsigned int)(long)hcpu;
	306
	307	/* We don't touch CPU 0 map, it's allocated at aboot and kept
	308	* around forever
	309	*/
	310	if (cpu == 0)
	311	return NOTIFY_OK;
	312
	313	switch (action) {
	314	case CPU_ONLINE:
	315	case CPU_ONLINE_FROZEN:
	316	pr_debug("MMU: Allocating stale context map for CPU %d\n", cpu);
	317	stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
	318	break;
	319	#ifdef CONFIG_HOTPLUG_CPU
	320	case CPU_DEAD:
	321	case CPU_DEAD_FROZEN:
	322	pr_debug("MMU: Freeing stale context map for CPU %d\n", cpu);
	323	kfree(stale_map[cpu]);
	324	stale_map[cpu] = NULL;
	325	break;
	326	#endif
	327	}
	328	return NOTIFY_OK;
	329	}
	330
	331	static struct notifier_block __cpuinitdata mmu_context_cpu_nb = {
	332	.notifier_call = mmu_context_cpu_notify,
	333	};
	334
	335	#endif /* CONFIG_SMP */
5e696617 BH	336
	337	/*
	338	* Initialize the context management stuff.
	339	*/
	340	void __init mmu_context_init(void)
	341	{
2ca8cf73 BH	342	/* Mark init_mm as being active on all possible CPUs since
	343	* we'll get called with prev == init_mm the first time
	344	* we schedule on a given CPU
	345	*/
	346	init_mm.context.active = NR_CPUS;
	347
77520351 BH	348	/*
	349	* The MPC8xx has only 16 contexts. We rotate through them on each
	350	* task switch. A better way would be to keep track of tasks that
	351	* own contexts, and implement an LRU usage. That way very active
	352	* tasks don't always have to pay the TLB reload overhead. The
	353	* kernel pages are mapped shared, so the kernel can run on behalf
	354	* of any task that makes a kernel entry. Shared does not mean they
	355	* are not protected, just that the ASID comparison is not performed.
	356	* -- Dan
	357	*
	358	* The IBM4xx has 256 contexts, so we can just rotate through these
	359	* as a way of "switching" contexts. If the TID of the TLB is zero,
	360	* the PID/TID comparison is disabled, so we can use a TID of zero
	361	* to represent all kernel pages as shared among all contexts.
	362	* -- Dan
	363	*/
	364	if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
	365	first_context = 0;
	366	last_context = 15;
	367	} else {
	368	first_context = 1;
	369	last_context = 255;
	370	}
	371
	372	#ifdef DEBUG_CLAMP_LAST_CONTEXT
	373	last_context = DEBUG_CLAMP_LAST_CONTEXT;
	374	#endif
	375	/*
	376	* Allocate the maps used by context management
	377	*/
	378	context_map = alloc_bootmem(CTX_MAP_SIZE);
	379	context_mm = alloc_bootmem(sizeof(void ) (last_context + 1));
	380	stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
	381
	382	#ifdef CONFIG_SMP
	383	register_cpu_notifier(&mmu_context_cpu_nb);
	384	#endif
	385
	386	printk(KERN_INFO
ff7c6600	387	"MMU: Allocated %zu bytes of context maps for %d contexts\n",
77520351 BH	388	2 * CTX_MAP_SIZE + (sizeof(void ) (last_context + 1)),
	389	last_context - first_context + 1);
	390
5e696617 BH	391	/*
	392	* Some processors have too few contexts to reserve one for
	393	* init_mm, and require using context 0 for a normal task.
	394	* Other processors reserve the use of context zero for the kernel.
77520351	395	* This code assumes first_context < 32.
5e696617	396	*/
77520351 BH	397	context_map[0] = (1 << first_context) - 1;
	398	next_context = first_context;
	399	nr_free_contexts = last_context - first_context + 1;
5e696617 BH	400	}
5e696617 BH	401