[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

#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/init.h>

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

/*
 *   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
 */

#ifdef CONFIG_8xx
#define LAST_CONTEXT    	15
#define FIRST_CONTEXT    	0

#elif defined(CONFIG_4xx)
#define LAST_CONTEXT    	255
#define FIRST_CONTEXT    	1

#elif defined(CONFIG_E200) || defined(CONFIG_E500)
#define LAST_CONTEXT    	255
#define FIRST_CONTEXT    	1

#else
#error Unsupported processor type
#endif

static unsigned int next_context, nr_free_contexts;
static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1];
static struct mm_struct *context_mm[LAST_CONTEXT+1];
static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;

/* 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;

 again:
	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_mask_nr(cpu, mm->cpu_vm_mask)
			__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);
	goto again;
}
#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);

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

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

	/* 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());
}
#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) {
		WARN_ON(prev->context.active < 1);
		prev->context.active--;
	}
#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);
			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;

#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;
		context_mm[id] = NULL;
#endif
		nr_free_contexts++;
	}
	spin_unlock(&context_lock);
}


/*
 * 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;

	/*
	 * 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
	31
	32	#include <linux/kernel.h>
5e696617 BH	33	#include <linux/mm.h>
	34	#include <linux/init.h>
	35
	36	#include <asm/mmu_context.h>
	37	#include <asm/tlbflush.h>
2ca8cf73	38	#include <linux/spinlock.h>
5e696617 BH	39
	40	/*
	41	* The MPC8xx has only 16 contexts. We rotate through them on each
	42	* task switch. A better way would be to keep track of tasks that
	43	* own contexts, and implement an LRU usage. That way very active
	44	* tasks don't always have to pay the TLB reload overhead. The
	45	* kernel pages are mapped shared, so the kernel can run on behalf
	46	* of any task that makes a kernel entry. Shared does not mean they
	47	* are not protected, just that the ASID comparison is not performed.
	48	* -- Dan
	49	*
	50	* The IBM4xx has 256 contexts, so we can just rotate through these
	51	* as a way of "switching" contexts. If the TID of the TLB is zero,
	52	* the PID/TID comparison is disabled, so we can use a TID of zero
	53	* to represent all kernel pages as shared among all contexts.
	54	* -- Dan
	55	*/
	56
	57	#ifdef CONFIG_8xx
5e696617 BH	58	#define LAST_CONTEXT 15
	59	#define FIRST_CONTEXT 0
	60
	61	#elif defined(CONFIG_4xx)
5e696617 BH	62	#define LAST_CONTEXT 255
	63	#define FIRST_CONTEXT 1
	64
	65	#elif defined(CONFIG_E200) \|\| defined(CONFIG_E500)
5e696617 BH	66	#define LAST_CONTEXT 255
	67	#define FIRST_CONTEXT 1
	68
	69	#else
	70	#error Unsupported processor type
	71	#endif
	72
2ca8cf73	73	static unsigned int next_context, nr_free_contexts;
5e696617	74	static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
2ca8cf73	75	static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1];
5e696617	76	static struct mm_struct *context_mm[LAST_CONTEXT+1];
2ca8cf73	77	static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
5e696617 BH	78
5e696617 BH	79	/* Steal a context from a task that has one at the moment.
2ca8cf73 BH	80	*
	81	* This is used when we are running out of available PID numbers
	82	* on the processors.
	83	*
5e696617 BH	84	* This isn't an LRU system, it just frees up each context in
	85	* turn (sort-of pseudo-random replacement :). This would be the
	86	* place to implement an LRU scheme if anyone was motivated to do it.
	87	* -- paulus
2ca8cf73 BH	88	*
	89	* For context stealing, we use a slightly different approach for
	90	* SMP and UP. Basically, the UP one is simpler and doesn't use
	91	* the stale map as we can just flush the local CPU
	92	* -- benh
5e696617	93	*/
2ca8cf73 BH	94	#ifdef CONFIG_SMP
2ca8cf73 BH	95	static unsigned int steal_context_smp(unsigned int id)
5e696617 BH	96	{
5e696617 BH	97	struct mm_struct *mm;
2ca8cf73	98	unsigned int cpu, max;
5e696617	99
2ca8cf73 BH	100	again:
2ca8cf73 BH	101	max = LAST_CONTEXT - FIRST_CONTEXT;
5e696617	102
2ca8cf73 BH	103	/* Attempt to free next_context first and then loop until we manage */
	104	while (max--) {
	105	/* Pick up the victim mm */
	106	mm = context_mm[id];
5e696617	107
2ca8cf73 BH	108	/* We have a candidate victim, check if it's active, on SMP
	109	* we cannot steal active contexts
	110	*/
	111	if (mm->context.active) {
	112	id++;
	113	if (id > LAST_CONTEXT)
	114	id = FIRST_CONTEXT;
	115	continue;
	116	}
	117	pr_debug("[%d] steal context %d from mm @%p\n",
	118	smp_processor_id(), id, mm);
	119
	120	/* Mark this mm has having no context anymore */
	121	mm->context.id = MMU_NO_CONTEXT;
	122
	123	/* Mark it stale on all CPUs that used this mm */
	124	for_each_cpu_mask_nr(cpu, mm->cpu_vm_mask)
	125	__set_bit(id, stale_map[cpu]);
	126	return id;
	127	}
	128
	129	/* This will happen if you have more CPUs than available contexts,
	130	* all we can do here is wait a bit and try again
	131	*/
	132	spin_unlock(&context_lock);
	133	cpu_relax();
	134	spin_lock(&context_lock);
	135	goto again;
	136	}
	137	#endif /* CONFIG_SMP */
	138
	139	/* Note that this will also be called on SMP if all other CPUs are
	140	* offlined, which means that it may be called for cpu != 0. For
	141	* this to work, we somewhat assume that CPUs that are onlined
	142	* come up with a fully clean TLB (or are cleaned when offlined)
5e696617	143	*/
2ca8cf73	144	static unsigned int steal_context_up(unsigned int id)
5e696617	145	{
2ca8cf73 BH	146	struct mm_struct *mm;
2ca8cf73 BH	147	int cpu = smp_processor_id();
5e696617	148
2ca8cf73 BH	149	/* Pick up the victim mm */
	150	mm = context_mm[id];
	151
	152	pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm);
5e696617	153
2ca8cf73 BH	154	/* Mark this mm has having no context anymore */
2ca8cf73 BH	155	mm->context.id = MMU_NO_CONTEXT;
5e696617	156
2ca8cf73 BH	157	/* Flush the TLB for that context */
	158	local_flush_tlb_mm(mm);
	159
	160	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
	161	__clear_bit(id, stale_map[cpu]);
	162
	163	return id;
	164	}
	165
	166	#ifdef DEBUG_MAP_CONSISTENCY
	167	static void context_check_map(void)
	168	{
	169	unsigned int id, nrf, nact;
	170
	171	nrf = nact = 0;
	172	for (id = FIRST_CONTEXT; id <= LAST_CONTEXT; id++) {
	173	int used = test_bit(id, context_map);
	174	if (!used)
	175	nrf++;
	176	if (used != (context_mm[id] != NULL))
	177	pr_err("MMU: Context %d is %s and MM is %p !\n",
	178	id, used ? "used" : "free", context_mm[id]);
	179	if (context_mm[id] != NULL)
	180	nact += context_mm[id]->context.active;
5e696617	181	}
2ca8cf73 BH	182	if (nrf != nr_free_contexts) {
	183	pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
	184	nr_free_contexts, nrf);
	185	nr_free_contexts = nrf;
	186	}
	187	if (nact > num_online_cpus())
	188	pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
	189	nact, num_online_cpus());
5e696617	190	}
2ca8cf73 BH	191	#else
	192	static void context_check_map(void) { }
	193	#endif
5e696617 BH	194
	195	void switch_mmu_context(struct mm_struct prev, struct mm_struct next)
	196	{
2ca8cf73 BH	197	unsigned int id, cpu = smp_processor_id();
2ca8cf73 BH	198	unsigned long *map;
5e696617	199
2ca8cf73 BH	200	/* No lockless fast path .. yet */
	201	spin_lock(&context_lock);
	202
	203	#ifndef DEBUG_STEAL_ONLY
	204	pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n",
	205	cpu, next, next->context.active, next->context.id);
	206	#endif
	207
	208	#ifdef CONFIG_SMP
	209	/* Mark us active and the previous one not anymore */
	210	next->context.active++;
	211	if (prev) {
	212	WARN_ON(prev->context.active < 1);
	213	prev->context.active--;
	214	}
	215	#endif /* CONFIG_SMP */
	216
	217	/* If we already have a valid assigned context, skip all that */
	218	id = next->context.id;
	219	if (likely(id != MMU_NO_CONTEXT))
	220	goto ctxt_ok;
	221
	222	/* We really don't have a context, let's try to acquire one */
	223	id = next_context;
	224	if (id > LAST_CONTEXT)
	225	id = FIRST_CONTEXT;
	226	map = context_map;
	227
	228	/* No more free contexts, let's try to steal one */
	229	if (nr_free_contexts == 0) {
	230	#ifdef CONFIG_SMP
	231	if (num_online_cpus() > 1) {
	232	id = steal_context_smp(id);
	233	goto stolen;
	234	}
	235	#endif /* CONFIG_SMP */
	236	id = steal_context_up(id);
	237	goto stolen;
	238	}
	239	nr_free_contexts--;
	240
	241	/* We know there's at least one free context, try to find it */
	242	while (__test_and_set_bit(id, map)) {
	243	id = find_next_zero_bit(map, LAST_CONTEXT+1, id);
	244	if (id > LAST_CONTEXT)
	245	id = FIRST_CONTEXT;
	246	}
	247	stolen:
	248	next_context = id + 1;
	249	context_mm[id] = next;
	250	next->context.id = id;
	251
	252	#ifndef DEBUG_STEAL_ONLY
	253	pr_debug("[%d] picked up new id %d, nrf is now %d\n",
	254	cpu, id, nr_free_contexts);
	255	#endif
	256
	257	context_check_map();
	258	ctxt_ok:
	259
	260	/* If that context got marked stale on this CPU, then flush the
	261	* local TLB for it and unmark it before we use it
	262	*/
	263	if (test_bit(id, stale_map[cpu])) {
264	pr_debug("[%d] flushing stale context %d for mm @%p !\n",
265	cpu, id, next);
266	local_flush_tlb_mm(next);
267
268	/* XXX This clear should ultimately be part of local_flush_tlb_mm */
269	__clear_bit(id, stale_map[cpu]);
270	}
271
272	/* Flick the MMU and release lock */
273	set_context(id, next->pgd);
274	spin_unlock(&context_lock);
5e696617 BH	275	}
	276
	277	/*
	278	* Set up the context for a new address space.
	279	*/
	280	int init_new_context(struct task_struct t, struct mm_struct mm)
	281	{
2ca8cf73 BH	282	mm->context.id = MMU_NO_CONTEXT;
	283	mm->context.active = 0;
	284
5e696617 BH	285	return 0;
	286	}
	287
	288	/*
	289	* We're finished using the context for an address space.
	290	*/
	291	void destroy_context(struct mm_struct *mm)
	292	{
2ca8cf73 BH	293	unsigned int id;
	294
	295	if (mm->context.id == MMU_NO_CONTEXT)
	296	return;
	297
	298	WARN_ON(mm->context.active != 0);
	299
	300	spin_lock(&context_lock);
	301	id = mm->context.id;
	302	if (id != MMU_NO_CONTEXT) {
	303	__clear_bit(id, context_map);
	304	mm->context.id = MMU_NO_CONTEXT;
	305	#ifdef DEBUG_MAP_CONSISTENCY
	306	mm->context.active = 0;
	307	context_mm[id] = NULL;
	308	#endif
	309	nr_free_contexts++;
5e696617	310	}
2ca8cf73	311	spin_unlock(&context_lock);
5e696617 BH	312	}
	313
	314
	315	/*
	316	* Initialize the context management stuff.
	317	*/
	318	void __init mmu_context_init(void)
	319	{
2ca8cf73 BH	320	/* Mark init_mm as being active on all possible CPUs since
	321	* we'll get called with prev == init_mm the first time
	322	* we schedule on a given CPU
	323	*/
	324	init_mm.context.active = NR_CPUS;
	325
5e696617 BH	326	/*
	327	* Some processors have too few contexts to reserve one for
	328	* init_mm, and require using context 0 for a normal task.
	329	* Other processors reserve the use of context zero for the kernel.
	330	* This code assumes FIRST_CONTEXT < 32.
	331	*/
	332	context_map[0] = (1 << FIRST_CONTEXT) - 1;
2ca8cf73 BH	333	next_context = FIRST_CONTEXT;
2ca8cf73 BH	334	nr_free_contexts = LAST_CONTEXT - FIRST_CONTEXT + 1;
5e696617 BH	335	}
5e696617 BH	336