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...
6 * Copyright 2008 Ben Herrenschmidt <benh@kernel.crashing.org>
9 * Derived from previous arch/powerpc/mm/mmu_context.c
10 * and arch/powerpc/include/asm/mmu_context.h
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.
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
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
29 #define DEBUG_STEAL_ONLY
30 #undef DEBUG_MAP_CONSISTENCY
31 /*#define DEBUG_CLAMP_LAST_CONTEXT 15 */
33 #include <linux/kernel.h>
35 #include <linux/init.h>
36 #include <linux/spinlock.h>
37 #include <linux/bootmem.h>
38 #include <linux/notifier.h>
39 #include <linux/cpu.h>
41 #include <asm/mmu_context.h>
42 #include <asm/tlbflush.h>
44 static unsigned int first_context, last_context;
45 static unsigned int next_context, nr_free_contexts;
46 static unsigned long *context_map;
47 static unsigned long *stale_map[NR_CPUS];
48 static struct mm_struct **context_mm;
49 static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
51 #define CTX_MAP_SIZE \
52 (sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
55 /* Steal a context from a task that has one at the moment.
57 * This is used when we are running out of available PID numbers
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.
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
71 static unsigned int steal_context_smp(unsigned int id)
74 unsigned int cpu, max;
77 max = last_context - first_context;
79 /* Attempt to free next_context first and then loop until we manage */
81 /* Pick up the victim mm */
84 /* We have a candidate victim, check if it's active, on SMP
85 * we cannot steal active contexts
87 if (mm->context.active) {
89 if (id > last_context)
93 pr_debug("[%d] steal context %d from mm @%p\n",
94 smp_processor_id(), id, mm);
96 /* Mark this mm has having no context anymore */
97 mm->context.id = MMU_NO_CONTEXT;
99 /* Mark it stale on all CPUs that used this mm */
100 for_each_cpu_mask_nr(cpu, mm->cpu_vm_mask)
101 __set_bit(id, stale_map[cpu]);
105 /* This will happen if you have more CPUs than available contexts,
106 * all we can do here is wait a bit and try again
108 spin_unlock(&context_lock);
110 spin_lock(&context_lock);
113 #endif /* CONFIG_SMP */
115 /* Note that this will also be called on SMP if all other CPUs are
116 * offlined, which means that it may be called for cpu != 0. For
117 * this to work, we somewhat assume that CPUs that are onlined
118 * come up with a fully clean TLB (or are cleaned when offlined)
120 static unsigned int steal_context_up(unsigned int id)
122 struct mm_struct *mm;
123 int cpu = smp_processor_id();
125 /* Pick up the victim mm */
128 pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm);
130 /* Mark this mm has having no context anymore */
131 mm->context.id = MMU_NO_CONTEXT;
133 /* Flush the TLB for that context */
134 local_flush_tlb_mm(mm);
136 /* XXX This clear should ultimately be part of local_flush_tlb_mm */
137 __clear_bit(id, stale_map[cpu]);
142 #ifdef DEBUG_MAP_CONSISTENCY
143 static void context_check_map(void)
145 unsigned int id, nrf, nact;
148 for (id = first_context; id <= last_context; id++) {
149 int used = test_bit(id, context_map);
152 if (used != (context_mm[id] != NULL))
153 pr_err("MMU: Context %d is %s and MM is %p !\n",
154 id, used ? "used" : "free", context_mm[id]);
155 if (context_mm[id] != NULL)
156 nact += context_mm[id]->context.active;
158 if (nrf != nr_free_contexts) {
159 pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
160 nr_free_contexts, nrf);
161 nr_free_contexts = nrf;
163 if (nact > num_online_cpus())
164 pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
165 nact, num_online_cpus());
166 if (first_context > 0 && !test_bit(0, context_map))
167 pr_err("MMU: Context 0 has been freed !!!\n");
170 static void context_check_map(void) { }
173 void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
175 unsigned int id, cpu = smp_processor_id();
178 /* No lockless fast path .. yet */
179 spin_lock(&context_lock);
181 #ifndef DEBUG_STEAL_ONLY
182 pr_debug("[%d] activating context for mm @%p, active=%d, id=%d\n",
183 cpu, next, next->context.active, next->context.id);
187 /* Mark us active and the previous one not anymore */
188 next->context.active++;
190 #ifndef DEBUG_STEAL_ONLY
191 pr_debug(" old context %p active was: %d\n",
192 prev, prev->context.active);
194 WARN_ON(prev->context.active < 1);
195 prev->context.active--;
197 #endif /* CONFIG_SMP */
199 /* If we already have a valid assigned context, skip all that */
200 id = next->context.id;
201 if (likely(id != MMU_NO_CONTEXT))
204 /* We really don't have a context, let's try to acquire one */
206 if (id > last_context)
210 /* No more free contexts, let's try to steal one */
211 if (nr_free_contexts == 0) {
213 if (num_online_cpus() > 1) {
214 id = steal_context_smp(id);
217 #endif /* CONFIG_SMP */
218 id = steal_context_up(id);
223 /* We know there's at least one free context, try to find it */
224 while (__test_and_set_bit(id, map)) {
225 id = find_next_zero_bit(map, last_context+1, id);
226 if (id > last_context)
230 next_context = id + 1;
231 context_mm[id] = next;
232 next->context.id = id;
234 #ifndef DEBUG_STEAL_ONLY
235 pr_debug("[%d] picked up new id %d, nrf is now %d\n",
236 cpu, id, nr_free_contexts);
242 /* If that context got marked stale on this CPU, then flush the
243 * local TLB for it and unmark it before we use it
245 if (test_bit(id, stale_map[cpu])) {
246 pr_debug("[%d] flushing stale context %d for mm @%p !\n",
248 local_flush_tlb_mm(next);
250 /* XXX This clear should ultimately be part of local_flush_tlb_mm */
251 __clear_bit(id, stale_map[cpu]);
254 /* Flick the MMU and release lock */
255 set_context(id, next->pgd);
256 spin_unlock(&context_lock);
260 * Set up the context for a new address space.
262 int init_new_context(struct task_struct *t, struct mm_struct *mm)
264 mm->context.id = MMU_NO_CONTEXT;
265 mm->context.active = 0;
271 * We're finished using the context for an address space.
273 void destroy_context(struct mm_struct *mm)
277 if (mm->context.id == MMU_NO_CONTEXT)
280 WARN_ON(mm->context.active != 0);
282 spin_lock(&context_lock);
284 if (id != MMU_NO_CONTEXT) {
285 __clear_bit(id, context_map);
286 mm->context.id = MMU_NO_CONTEXT;
287 #ifdef DEBUG_MAP_CONSISTENCY
288 mm->context.active = 0;
289 context_mm[id] = NULL;
293 spin_unlock(&context_lock);
298 static int __cpuinit mmu_context_cpu_notify(struct notifier_block *self,
299 unsigned long action, void *hcpu)
301 unsigned int cpu = (unsigned int)(long)hcpu;
303 /* We don't touch CPU 0 map, it's allocated at aboot and kept
311 case CPU_ONLINE_FROZEN:
312 pr_debug("MMU: Allocating stale context map for CPU %d\n", cpu);
313 stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL);
315 #ifdef CONFIG_HOTPLUG_CPU
317 case CPU_DEAD_FROZEN:
318 pr_debug("MMU: Freeing stale context map for CPU %d\n", cpu);
319 kfree(stale_map[cpu]);
320 stale_map[cpu] = NULL;
327 static struct notifier_block __cpuinitdata mmu_context_cpu_nb = {
328 .notifier_call = mmu_context_cpu_notify,
331 #endif /* CONFIG_SMP */
334 * Initialize the context management stuff.
336 void __init mmu_context_init(void)
338 /* Mark init_mm as being active on all possible CPUs since
339 * we'll get called with prev == init_mm the first time
340 * we schedule on a given CPU
342 init_mm.context.active = NR_CPUS;
345 * The MPC8xx has only 16 contexts. We rotate through them on each
346 * task switch. A better way would be to keep track of tasks that
347 * own contexts, and implement an LRU usage. That way very active
348 * tasks don't always have to pay the TLB reload overhead. The
349 * kernel pages are mapped shared, so the kernel can run on behalf
350 * of any task that makes a kernel entry. Shared does not mean they
351 * are not protected, just that the ASID comparison is not performed.
354 * The IBM4xx has 256 contexts, so we can just rotate through these
355 * as a way of "switching" contexts. If the TID of the TLB is zero,
356 * the PID/TID comparison is disabled, so we can use a TID of zero
357 * to represent all kernel pages as shared among all contexts.
360 if (mmu_has_feature(MMU_FTR_TYPE_8xx)) {
368 #ifdef DEBUG_CLAMP_LAST_CONTEXT
369 last_context = DEBUG_CLAMP_LAST_CONTEXT;
372 * Allocate the maps used by context management
374 context_map = alloc_bootmem(CTX_MAP_SIZE);
375 context_mm = alloc_bootmem(sizeof(void *) * (last_context + 1));
376 stale_map[0] = alloc_bootmem(CTX_MAP_SIZE);
379 register_cpu_notifier(&mmu_context_cpu_nb);
383 "MMU: Allocated %d bytes of context maps for %d contexts\n",
384 2 * CTX_MAP_SIZE + (sizeof(void *) * (last_context + 1)),
385 last_context - first_context + 1);
388 * Some processors have too few contexts to reserve one for
389 * init_mm, and require using context 0 for a normal task.
390 * Other processors reserve the use of context zero for the kernel.
391 * This code assumes first_context < 32.
393 context_map[0] = (1 << first_context) - 1;
394 next_context = first_context;
395 nr_free_contexts = last_context - first_context + 1;