aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--arch/powerpc/include/asm/mmu-40x.h5
-rw-r--r--arch/powerpc/include/asm/mmu-44x.h5
-rw-r--r--arch/powerpc/include/asm/mmu-8xx.h3
-rw-r--r--arch/powerpc/include/asm/mmu-fsl-booke.h5
-rw-r--r--arch/powerpc/include/asm/tlbflush.h2
-rw-r--r--arch/powerpc/mm/mmu_context_nohash.c268
6 files changed, 234 insertions, 54 deletions
diff --git a/arch/powerpc/include/asm/mmu-40x.h b/arch/powerpc/include/asm/mmu-40x.h
index 3d108676584c..776f415a36aa 100644
--- a/arch/powerpc/include/asm/mmu-40x.h
+++ b/arch/powerpc/include/asm/mmu-40x.h
@@ -54,8 +54,9 @@
54#ifndef __ASSEMBLY__ 54#ifndef __ASSEMBLY__
55 55
56typedef struct { 56typedef struct {
57 unsigned long id; 57 unsigned int id;
58 unsigned long vdso_base; 58 unsigned int active;
59 unsigned long vdso_base;
59} mm_context_t; 60} mm_context_t;
60 61
61#endif /* !__ASSEMBLY__ */ 62#endif /* !__ASSEMBLY__ */
diff --git a/arch/powerpc/include/asm/mmu-44x.h b/arch/powerpc/include/asm/mmu-44x.h
index a825524c981a..b21af32ac6d6 100644
--- a/arch/powerpc/include/asm/mmu-44x.h
+++ b/arch/powerpc/include/asm/mmu-44x.h
@@ -56,8 +56,9 @@
56extern unsigned int tlb_44x_hwater; 56extern unsigned int tlb_44x_hwater;
57 57
58typedef struct { 58typedef struct {
59 unsigned long id; 59 unsigned int id;
60 unsigned long vdso_base; 60 unsigned int active;
61 unsigned long vdso_base;
61} mm_context_t; 62} mm_context_t;
62 63
63#endif /* !__ASSEMBLY__ */ 64#endif /* !__ASSEMBLY__ */
diff --git a/arch/powerpc/include/asm/mmu-8xx.h b/arch/powerpc/include/asm/mmu-8xx.h
index 9db877eb88db..07865a357848 100644
--- a/arch/powerpc/include/asm/mmu-8xx.h
+++ b/arch/powerpc/include/asm/mmu-8xx.h
@@ -137,7 +137,8 @@
137 137
138#ifndef __ASSEMBLY__ 138#ifndef __ASSEMBLY__
139typedef struct { 139typedef struct {
140 unsigned long id; 140 unsigned int id;
141 unsigned int active;
141 unsigned long vdso_base; 142 unsigned long vdso_base;
142} mm_context_t; 143} mm_context_t;
143#endif /* !__ASSEMBLY__ */ 144#endif /* !__ASSEMBLY__ */
diff --git a/arch/powerpc/include/asm/mmu-fsl-booke.h b/arch/powerpc/include/asm/mmu-fsl-booke.h
index 5588a41f439c..3f941c0f7e8e 100644
--- a/arch/powerpc/include/asm/mmu-fsl-booke.h
+++ b/arch/powerpc/include/asm/mmu-fsl-booke.h
@@ -76,8 +76,9 @@
76#ifndef __ASSEMBLY__ 76#ifndef __ASSEMBLY__
77 77
78typedef struct { 78typedef struct {
79 unsigned long id; 79 unsigned int id;
80 unsigned long vdso_base; 80 unsigned int active;
81 unsigned long vdso_base;
81} mm_context_t; 82} mm_context_t;
82#endif /* !__ASSEMBLY__ */ 83#endif /* !__ASSEMBLY__ */
83 84
diff --git a/arch/powerpc/include/asm/tlbflush.h b/arch/powerpc/include/asm/tlbflush.h
index 333c24b54379..9ed363d3de44 100644
--- a/arch/powerpc/include/asm/tlbflush.h
+++ b/arch/powerpc/include/asm/tlbflush.h
@@ -29,6 +29,8 @@
29 29
30#include <linux/mm.h> 30#include <linux/mm.h>
31 31
32#define MMU_NO_CONTEXT ((unsigned int)-1)
33
32extern void _tlbie(unsigned long address, unsigned int pid); 34extern void _tlbie(unsigned long address, unsigned int pid);
33extern void _tlbil_all(void); 35extern void _tlbil_all(void);
34extern void _tlbil_pid(unsigned int pid); 36extern void _tlbil_pid(unsigned int pid);
diff --git a/arch/powerpc/mm/mmu_context_nohash.c b/arch/powerpc/mm/mmu_context_nohash.c
index 00e02150abef..8b5de52de0ad 100644
--- a/arch/powerpc/mm/mmu_context_nohash.c
+++ b/arch/powerpc/mm/mmu_context_nohash.c
@@ -14,13 +14,28 @@
14 * as published by the Free Software Foundation; either version 14 * as published by the Free Software Foundation; either version
15 * 2 of the License, or (at your option) any later version. 15 * 2 of the License, or (at your option) any later version.
16 * 16 *
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
17 */ 26 */
18 27
28#undef DEBUG
29#define DEBUG_STEAL_ONLY
30#undef DEBUG_MAP_CONSISTENCY
31
32#include <linux/kernel.h>
19#include <linux/mm.h> 33#include <linux/mm.h>
20#include <linux/init.h> 34#include <linux/init.h>
21 35
22#include <asm/mmu_context.h> 36#include <asm/mmu_context.h>
23#include <asm/tlbflush.h> 37#include <asm/tlbflush.h>
38#include <linux/spinlock.h>
24 39
25/* 40/*
26 * The MPC8xx has only 16 contexts. We rotate through them on each 41 * The MPC8xx has only 16 contexts. We rotate through them on each
@@ -40,17 +55,14 @@
40 */ 55 */
41 56
42#ifdef CONFIG_8xx 57#ifdef CONFIG_8xx
43#define NO_CONTEXT 16
44#define LAST_CONTEXT 15 58#define LAST_CONTEXT 15
45#define FIRST_CONTEXT 0 59#define FIRST_CONTEXT 0
46 60
47#elif defined(CONFIG_4xx) 61#elif defined(CONFIG_4xx)
48#define NO_CONTEXT 256
49#define LAST_CONTEXT 255 62#define LAST_CONTEXT 255
50#define FIRST_CONTEXT 1 63#define FIRST_CONTEXT 1
51 64
52#elif defined(CONFIG_E200) || defined(CONFIG_E500) 65#elif defined(CONFIG_E200) || defined(CONFIG_E500)
53#define NO_CONTEXT 256
54#define LAST_CONTEXT 255 66#define LAST_CONTEXT 255
55#define FIRST_CONTEXT 1 67#define FIRST_CONTEXT 1
56 68
@@ -58,66 +70,208 @@
58#error Unsupported processor type 70#error Unsupported processor type
59#endif 71#endif
60 72
61static unsigned long next_mmu_context; 73static unsigned int next_context, nr_free_contexts;
62static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1]; 74static unsigned long context_map[LAST_CONTEXT / BITS_PER_LONG + 1];
63static atomic_t nr_free_contexts; 75static unsigned long stale_map[NR_CPUS][LAST_CONTEXT / BITS_PER_LONG + 1];
64static struct mm_struct *context_mm[LAST_CONTEXT+1]; 76static struct mm_struct *context_mm[LAST_CONTEXT+1];
65static void steal_context(void); 77static spinlock_t context_lock = SPIN_LOCK_UNLOCKED;
66 78
67/* Steal a context from a task that has one at the moment. 79/* Steal a context from a task that has one at the moment.
68 * This is only used on 8xx and 4xx and we presently assume that 80 *
69 * they don't do SMP. If they do then this will have to check 81 * This is used when we are running out of available PID numbers
70 * whether the MM we steal is in use. 82 * on the processors.
71 * We also assume that this is only used on systems that don't 83 *
72 * use an MMU hash table - this is true for 8xx and 4xx.
73 * This isn't an LRU system, it just frees up each context in 84 * This isn't an LRU system, it just frees up each context in
74 * turn (sort-of pseudo-random replacement :). This would be the 85 * turn (sort-of pseudo-random replacement :). This would be the
75 * place to implement an LRU scheme if anyone was motivated to do it. 86 * place to implement an LRU scheme if anyone was motivated to do it.
76 * -- paulus 87 * -- paulus
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
77 */ 93 */
78static void steal_context(void) 94#ifdef CONFIG_SMP
95static unsigned int steal_context_smp(unsigned int id)
79{ 96{
80 struct mm_struct *mm; 97 struct mm_struct *mm;
98 unsigned int cpu, max;
81 99
82 /* free up context `next_mmu_context' */ 100 again:
83 /* if we shouldn't free context 0, don't... */ 101 max = LAST_CONTEXT - FIRST_CONTEXT;
84 if (next_mmu_context < FIRST_CONTEXT)
85 next_mmu_context = FIRST_CONTEXT;
86 mm = context_mm[next_mmu_context];
87 flush_tlb_mm(mm);
88 destroy_context(mm);
89}
90 102
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];
91 107
92/* 108 /* We have a candidate victim, check if it's active, on SMP
93 * Get a new mmu context for the address space described by `mm'. 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)
94 */ 143 */
95static inline void get_mmu_context(struct mm_struct *mm) 144static unsigned int steal_context_up(unsigned int id)
96{ 145{
97 unsigned long ctx; 146 struct mm_struct *mm;
147 int cpu = smp_processor_id();
98 148
99 if (mm->context.id != NO_CONTEXT) 149 /* Pick up the victim mm */
100 return; 150 mm = context_mm[id];
151
152 pr_debug("[%d] steal context %d from mm @%p\n", cpu, id, mm);
101 153
102 while (atomic_dec_if_positive(&nr_free_contexts) < 0) 154 /* Mark this mm has having no context anymore */
103 steal_context(); 155 mm->context.id = MMU_NO_CONTEXT;
104 156
105 ctx = next_mmu_context; 157 /* Flush the TLB for that context */
106 while (test_and_set_bit(ctx, context_map)) { 158 local_flush_tlb_mm(mm);
107 ctx = find_next_zero_bit(context_map, LAST_CONTEXT+1, ctx); 159
108 if (ctx > LAST_CONTEXT) 160 /* XXX This clear should ultimately be part of local_flush_tlb_mm */
109 ctx = 0; 161 __clear_bit(id, stale_map[cpu]);
162
163 return id;
164}
165
166#ifdef DEBUG_MAP_CONSISTENCY
167static 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;
110 } 181 }
111 next_mmu_context = (ctx + 1) & LAST_CONTEXT; 182 if (nrf != nr_free_contexts) {
112 mm->context.id = ctx; 183 pr_err("MMU: Free context count out of sync ! (%d vs %d)\n",
113 context_mm[ctx] = mm; 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());
114} 190}
191#else
192static void context_check_map(void) { }
193#endif
115 194
116void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next) 195void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
117{ 196{
118 get_mmu_context(next); 197 unsigned int id, cpu = smp_processor_id();
198 unsigned long *map;
119 199
120 set_context(next->context.id, next->pgd); 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);
121} 275}
122 276
123/* 277/*
@@ -125,7 +279,9 @@ void switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
125 */ 279 */
126int init_new_context(struct task_struct *t, struct mm_struct *mm) 280int init_new_context(struct task_struct *t, struct mm_struct *mm)
127{ 281{
128 mm->context.id = NO_CONTEXT; 282 mm->context.id = MMU_NO_CONTEXT;
283 mm->context.active = 0;
284
129 return 0; 285 return 0;
130} 286}
131 287
@@ -134,13 +290,25 @@ int init_new_context(struct task_struct *t, struct mm_struct *mm)
134 */ 290 */
135void destroy_context(struct mm_struct *mm) 291void destroy_context(struct mm_struct *mm)
136{ 292{
137 preempt_disable(); 293 unsigned int id;
138 if (mm->context.id != NO_CONTEXT) { 294
139 clear_bit(mm->context.id, context_map); 295 if (mm->context.id == MMU_NO_CONTEXT)
140 mm->context.id = NO_CONTEXT; 296 return;
141 atomic_inc(&nr_free_contexts); 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++;
142 } 310 }
143 preempt_enable(); 311 spin_unlock(&context_lock);
144} 312}
145 313
146 314
@@ -149,6 +317,12 @@ void destroy_context(struct mm_struct *mm)
149 */ 317 */
150void __init mmu_context_init(void) 318void __init mmu_context_init(void)
151{ 319{
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
152 /* 326 /*
153 * Some processors have too few contexts to reserve one for 327 * Some processors have too few contexts to reserve one for
154 * init_mm, and require using context 0 for a normal task. 328 * init_mm, and require using context 0 for a normal task.
@@ -156,7 +330,7 @@ void __init mmu_context_init(void)
156 * This code assumes FIRST_CONTEXT < 32. 330 * This code assumes FIRST_CONTEXT < 32.
157 */ 331 */
158 context_map[0] = (1 << FIRST_CONTEXT) - 1; 332 context_map[0] = (1 << FIRST_CONTEXT) - 1;
159 next_mmu_context = FIRST_CONTEXT; 333 next_context = FIRST_CONTEXT;
160 atomic_set(&nr_free_contexts, LAST_CONTEXT - FIRST_CONTEXT + 1); 334 nr_free_contexts = LAST_CONTEXT - FIRST_CONTEXT + 1;
161} 335}
162 336