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
28 //#define DEBUG_MAP_CONSISTENCY
29 //#define DEBUG_CLAMP_LAST_CONTEXT 31
30 //#define DEBUG_HARDER
32 /* We don't use DEBUG because it tends to be compiled in always nowadays
33 * and this would generate way too much output
36 #define pr_hard(args...) printk(KERN_DEBUG args)
37 #define pr_hardcont(args...) printk(KERN_CONT args)
39 #define pr_hard(args...) do { } while(0)
40 #define pr_hardcont(args...) do { } while(0)
43 #include <linux/kernel.h>
45 #include <linux/init.h>
46 #include <linux/spinlock.h>
47 #include <linux/bootmem.h>
48 #include <linux/notifier.h>
49 #include <linux/cpu.h>
50 #include <linux/slab.h>
52 #include <asm/mmu_context.h>
53 #include <asm/tlbflush.h>
55 static unsigned int first_context
, last_context
;
56 static unsigned int next_context
, nr_free_contexts
;
57 static unsigned long *context_map
;
58 static unsigned long *stale_map
[NR_CPUS
];
59 static struct mm_struct
**context_mm
;
60 static DEFINE_RAW_SPINLOCK(context_lock
);
62 #define CTX_MAP_SIZE \
63 (sizeof(unsigned long) * (last_context / BITS_PER_LONG + 1))
66 /* Steal a context from a task that has one at the moment.
68 * This is used when we are running out of available PID numbers
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.
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
82 static unsigned int steal_context_smp(unsigned int id
)
85 unsigned int cpu
, max
, i
;
87 max
= last_context
- first_context
;
89 /* Attempt to free next_context first and then loop until we manage */
91 /* Pick up the victim mm */
94 /* We have a candidate victim, check if it's active, on SMP
95 * we cannot steal active contexts
97 if (mm
->context
.active
) {
99 if (id
> last_context
)
103 pr_hardcont(" | steal %d from 0x%p", id
, mm
);
105 /* Mark this mm has having no context anymore */
106 mm
->context
.id
= MMU_NO_CONTEXT
;
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.
113 for_each_cpu(cpu
, mm_cpumask(mm
)) {
114 for (i
= cpu_first_thread_sibling(cpu
);
115 i
<= cpu_last_thread_sibling(cpu
); i
++) {
117 __set_bit(id
, stale_map
[i
]);
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
127 raw_spin_unlock(&context_lock
);
129 raw_spin_lock(&context_lock
);
131 /* This will cause the caller to try again */
132 return MMU_NO_CONTEXT
;
134 #endif /* CONFIG_SMP */
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)
141 static unsigned int steal_context_up(unsigned int id
)
143 struct mm_struct
*mm
;
144 int cpu
= smp_processor_id();
146 /* Pick up the victim mm */
149 pr_hardcont(" | steal %d from 0x%p", id
, mm
);
151 /* Flush the TLB for that context */
152 local_flush_tlb_mm(mm
);
154 /* Mark this mm has having no context anymore */
155 mm
->context
.id
= MMU_NO_CONTEXT
;
157 /* XXX This clear should ultimately be part of local_flush_tlb_mm */
158 __clear_bit(id
, stale_map
[cpu
]);
163 #ifdef DEBUG_MAP_CONSISTENCY
164 static void context_check_map(void)
166 unsigned int id
, nrf
, nact
;
169 for (id
= first_context
; id
<= last_context
; id
++) {
170 int used
= test_bit(id
, context_map
);
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
;
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
;
184 if (nact
> num_online_cpus())
185 pr_err("MMU: More active contexts than CPUs ! (%d vs %d)\n",
186 nact
, num_online_cpus());
187 if (first_context
> 0 && !test_bit(0, context_map
))
188 pr_err("MMU: Context 0 has been freed !!!\n");
191 static void context_check_map(void) { }
194 void switch_mmu_context(struct mm_struct
*prev
, struct mm_struct
*next
)
196 unsigned int i
, id
, cpu
= smp_processor_id();
199 /* No lockless fast path .. yet */
200 raw_spin_lock(&context_lock
);
202 pr_hard("[%d] activating context for mm @%p, active=%d, id=%d",
203 cpu
, next
, next
->context
.active
, next
->context
.id
);
206 /* Mark us active and the previous one not anymore */
207 next
->context
.active
++;
209 pr_hardcont(" (old=0x%p a=%d)", prev
, prev
->context
.active
);
210 WARN_ON(prev
->context
.active
< 1);
211 prev
->context
.active
--;
215 #endif /* CONFIG_SMP */
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 #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
]);
228 /* We really don't have a context, let's try to acquire one */
230 if (id
> last_context
)
234 /* No more free contexts, let's try to steal one */
235 if (nr_free_contexts
== 0) {
237 if (num_online_cpus() > 1) {
238 id
= steal_context_smp(id
);
239 if (id
== MMU_NO_CONTEXT
)
243 #endif /* CONFIG_SMP */
244 id
= steal_context_up(id
);
249 /* We know there's at least one free context, try to find it */
250 while (__test_and_set_bit(id
, map
)) {
251 id
= find_next_zero_bit(map
, last_context
+1, id
);
252 if (id
> last_context
)
256 next_context
= id
+ 1;
257 context_mm
[id
] = next
;
258 next
->context
.id
= id
;
259 pr_hardcont(" | new id=%d,nrf=%d", id
, nr_free_contexts
);
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
267 if (test_bit(id
, stale_map
[cpu
])) {
268 pr_hardcont(" | stale flush %d [%d..%d]",
269 id
, cpu_first_thread_sibling(cpu
),
270 cpu_last_thread_sibling(cpu
));
272 local_flush_tlb_mm(next
);
274 /* XXX This clear should ultimately be part of local_flush_tlb_mm */
275 for (i
= cpu_first_thread_sibling(cpu
);
276 i
<= cpu_last_thread_sibling(cpu
); i
++) {
278 __clear_bit(id
, stale_map
[i
]);
282 /* Flick the MMU and release lock */
283 pr_hardcont(" -> %d\n", id
);
284 set_context(id
, next
->pgd
);
285 raw_spin_unlock(&context_lock
);
289 * Set up the context for a new address space.
291 int init_new_context(struct task_struct
*t
, struct mm_struct
*mm
)
293 pr_hard("initing context for mm @%p\n", mm
);
295 mm
->context
.id
= MMU_NO_CONTEXT
;
296 mm
->context
.active
= 0;
298 #ifdef CONFIG_PPC_MM_SLICES
299 if (slice_mm_new_context(mm
))
300 slice_set_user_psize(mm
, mmu_virtual_psize
);
307 * We're finished using the context for an address space.
309 void destroy_context(struct mm_struct
*mm
)
314 if (mm
->context
.id
== MMU_NO_CONTEXT
)
317 WARN_ON(mm
->context
.active
!= 0);
319 raw_spin_lock_irqsave(&context_lock
, flags
);
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;
327 context_mm
[id
] = NULL
;
330 raw_spin_unlock_irqrestore(&context_lock
, flags
);
335 static int mmu_context_cpu_notify(struct notifier_block
*self
,
336 unsigned long action
, void *hcpu
)
338 unsigned int cpu
= (unsigned int)(long)hcpu
;
340 /* We don't touch CPU 0 map, it's allocated at aboot and kept
343 if (cpu
== boot_cpuid
)
348 case CPU_UP_PREPARE_FROZEN
:
349 pr_devel("MMU: Allocating stale context map for CPU %d\n", cpu
);
350 stale_map
[cpu
] = kzalloc(CTX_MAP_SIZE
, GFP_KERNEL
);
352 #ifdef CONFIG_HOTPLUG_CPU
353 case CPU_UP_CANCELED
:
354 case CPU_UP_CANCELED_FROZEN
:
356 case CPU_DEAD_FROZEN
:
357 pr_devel("MMU: Freeing stale context map for CPU %d\n", cpu
);
358 kfree(stale_map
[cpu
]);
359 stale_map
[cpu
] = NULL
;
361 /* We also clear the cpu_vm_mask bits of CPUs going away */
362 clear_tasks_mm_cpumask(cpu
);
364 #endif /* CONFIG_HOTPLUG_CPU */
369 static struct notifier_block mmu_context_cpu_nb
= {
370 .notifier_call
= mmu_context_cpu_notify
,
373 #endif /* CONFIG_SMP */
376 * Initialize the context management stuff.
378 void __init
mmu_context_init(void)
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
384 init_mm
.context
.active
= NR_CPUS
;
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.
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.
402 * The IBM 47x core supports 16-bit PIDs, thus 65535 contexts. We
403 * should normally never have to steal though the facility is
407 if (mmu_has_feature(MMU_FTR_TYPE_8xx
)) {
410 } else if (mmu_has_feature(MMU_FTR_TYPE_47x
)) {
412 last_context
= 65535;
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
;
420 last_context
= (1UL << (pid_bits
+ 1)) - 1;
428 #ifdef DEBUG_CLAMP_LAST_CONTEXT
429 last_context
= DEBUG_CLAMP_LAST_CONTEXT
;
432 * Allocate the maps used by context management
434 context_map
= alloc_bootmem(CTX_MAP_SIZE
);
435 context_mm
= alloc_bootmem(sizeof(void *) * (last_context
+ 1));
437 stale_map
[0] = alloc_bootmem(CTX_MAP_SIZE
);
439 stale_map
[boot_cpuid
] = alloc_bootmem(CTX_MAP_SIZE
);
441 register_cpu_notifier(&mmu_context_cpu_nb
);
445 "MMU: Allocated %zu bytes of context maps for %d contexts\n",
446 2 * CTX_MAP_SIZE
+ (sizeof(void *) * (last_context
+ 1)),
447 last_context
- first_context
+ 1);
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.
453 * This code assumes first_context < 32.
455 context_map
[0] = (1 << first_context
) - 1;
456 next_context
= first_context
;
457 nr_free_contexts
= last_context
- first_context
+ 1;