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 | */ |
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 | |
33 | #include <linux/kernel.h> | |
5e696617 BH |
34 | #include <linux/mm.h> |
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; | |
b46b6942 | 49 | static DEFINE_SPINLOCK(context_lock); |
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 |
71 | static unsigned int steal_context_smp(unsigned int id) | |
5e696617 BH |
72 | { |
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) |
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; |
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)) |
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(); |
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 | |
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) |
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) |
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 | { | |
b46b6942 | 279 | unsigned long flags; |
2ca8cf73 BH |
280 | unsigned int id; |
281 | ||
282 | if (mm->context.id == MMU_NO_CONTEXT) | |
283 | return; | |
284 | ||
285 | WARN_ON(mm->context.active != 0); | |
286 | ||
b46b6942 | 287 | spin_lock_irqsave(&context_lock, flags); |
2ca8cf73 BH |
288 | id = mm->context.id; |
289 | if (id != MMU_NO_CONTEXT) { | |
290 | __clear_bit(id, context_map); | |
291 | mm->context.id = MMU_NO_CONTEXT; | |
292 | #ifdef DEBUG_MAP_CONSISTENCY | |
293 | mm->context.active = 0; | |
2ca8cf73 | 294 | #endif |
3035c863 | 295 | context_mm[id] = NULL; |
2ca8cf73 | 296 | nr_free_contexts++; |
5e696617 | 297 | } |
b46b6942 | 298 | spin_unlock_irqrestore(&context_lock, flags); |
5e696617 BH |
299 | } |
300 | ||
77520351 BH |
301 | #ifdef CONFIG_SMP |
302 | ||
303 | static int __cpuinit mmu_context_cpu_notify(struct notifier_block *self, | |
304 | unsigned long action, void *hcpu) | |
305 | { | |
306 | unsigned int cpu = (unsigned int)(long)hcpu; | |
307 | ||
308 | /* We don't touch CPU 0 map, it's allocated at aboot and kept | |
309 | * around forever | |
310 | */ | |
311 | if (cpu == 0) | |
312 | return NOTIFY_OK; | |
313 | ||
314 | switch (action) { | |
315 | case CPU_ONLINE: | |
316 | case CPU_ONLINE_FROZEN: | |
317 | pr_debug("MMU: Allocating stale context map for CPU %d\n", cpu); | |
318 | stale_map[cpu] = kzalloc(CTX_MAP_SIZE, GFP_KERNEL); | |
319 | break; | |
320 | #ifdef CONFIG_HOTPLUG_CPU | |
321 | case CPU_DEAD: | |
322 | case CPU_DEAD_FROZEN: | |
323 | pr_debug("MMU: Freeing stale context map for CPU %d\n", cpu); | |
324 | kfree(stale_map[cpu]); | |
325 | stale_map[cpu] = NULL; | |
326 | break; | |
327 | #endif | |
328 | } | |
329 | return NOTIFY_OK; | |
330 | } | |
331 | ||
332 | static struct notifier_block __cpuinitdata mmu_context_cpu_nb = { | |
333 | .notifier_call = mmu_context_cpu_notify, | |
334 | }; | |
335 | ||
336 | #endif /* CONFIG_SMP */ | |
5e696617 BH |
337 | |
338 | /* | |
339 | * Initialize the context management stuff. | |
340 | */ | |
341 | void __init mmu_context_init(void) | |
342 | { | |
2ca8cf73 BH |
343 | /* Mark init_mm as being active on all possible CPUs since |
344 | * we'll get called with prev == init_mm the first time | |
345 | * we schedule on a given CPU | |
346 | */ | |
347 | init_mm.context.active = NR_CPUS; | |
348 | ||
77520351 BH |
349 | /* |
350 | * The MPC8xx has only 16 contexts. We rotate through them on each | |
351 | * task switch. A better way would be to keep track of tasks that | |
352 | * own contexts, and implement an LRU usage. That way very active | |
353 | * tasks don't always have to pay the TLB reload overhead. The | |
354 | * kernel pages are mapped shared, so the kernel can run on behalf | |
355 | * of any task that makes a kernel entry. Shared does not mean they | |
356 | * are not protected, just that the ASID comparison is not performed. | |
357 | * -- Dan | |
358 | * | |
359 | * The IBM4xx has 256 contexts, so we can just rotate through these | |
360 | * as a way of "switching" contexts. If the TID of the TLB is zero, | |
361 | * the PID/TID comparison is disabled, so we can use a TID of zero | |
362 | * to represent all kernel pages as shared among all contexts. | |
363 | * -- Dan | |
364 | */ | |
365 | if (mmu_has_feature(MMU_FTR_TYPE_8xx)) { | |
366 | first_context = 0; | |
367 | last_context = 15; | |
368 | } else { | |
369 | first_context = 1; | |
370 | last_context = 255; | |
371 | } | |
372 | ||
373 | #ifdef DEBUG_CLAMP_LAST_CONTEXT | |
374 | last_context = DEBUG_CLAMP_LAST_CONTEXT; | |
375 | #endif | |
376 | /* | |
377 | * Allocate the maps used by context management | |
378 | */ | |
379 | context_map = alloc_bootmem(CTX_MAP_SIZE); | |
380 | context_mm = alloc_bootmem(sizeof(void *) * (last_context + 1)); | |
381 | stale_map[0] = alloc_bootmem(CTX_MAP_SIZE); | |
382 | ||
383 | #ifdef CONFIG_SMP | |
384 | register_cpu_notifier(&mmu_context_cpu_nb); | |
385 | #endif | |
386 | ||
387 | printk(KERN_INFO | |
ff7c6600 | 388 | "MMU: Allocated %zu bytes of context maps for %d contexts\n", |
77520351 BH |
389 | 2 * CTX_MAP_SIZE + (sizeof(void *) * (last_context + 1)), |
390 | last_context - first_context + 1); | |
391 | ||
5e696617 BH |
392 | /* |
393 | * Some processors have too few contexts to reserve one for | |
394 | * init_mm, and require using context 0 for a normal task. | |
395 | * Other processors reserve the use of context zero for the kernel. | |
77520351 | 396 | * This code assumes first_context < 32. |
5e696617 | 397 | */ |
77520351 BH |
398 | context_map[0] = (1 << first_context) - 1; |
399 | next_context = first_context; | |
400 | nr_free_contexts = last_context - first_context + 1; | |
5e696617 BH |
401 | } |
402 |