Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * This program is free software; you can redistribute it and/or | |
3 | * modify it under the terms of the GNU General Public License | |
4 | * as published by the Free Software Foundation; either version 2 | |
5 | * of the License, or (at your option) any later version. | |
6 | * | |
7 | * This program is distributed in the hope that it will be useful, | |
8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
10 | * GNU General Public License for more details. | |
11 | * | |
12 | * You should have received a copy of the GNU General Public License | |
13 | * along with this program; if not, write to the Free Software | |
14 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
15 | * | |
16 | * Copyright (C) 2000, 2001 Kanoj Sarcar | |
17 | * Copyright (C) 2000, 2001 Ralf Baechle | |
18 | * Copyright (C) 2000, 2001 Silicon Graphics, Inc. | |
19 | * Copyright (C) 2000, 2001, 2003 Broadcom Corporation | |
20 | */ | |
21 | #include <linux/cache.h> | |
22 | #include <linux/delay.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/interrupt.h> | |
25 | #include <linux/spinlock.h> | |
26 | #include <linux/threads.h> | |
27 | #include <linux/module.h> | |
28 | #include <linux/time.h> | |
29 | #include <linux/timex.h> | |
30 | #include <linux/sched.h> | |
31 | #include <linux/cpumask.h> | |
32 | ||
33 | #include <asm/atomic.h> | |
34 | #include <asm/cpu.h> | |
35 | #include <asm/processor.h> | |
36 | #include <asm/system.h> | |
37 | #include <asm/mmu_context.h> | |
38 | #include <asm/smp.h> | |
39 | ||
40 | cpumask_t phys_cpu_present_map; /* Bitmask of available CPUs */ | |
41 | volatile cpumask_t cpu_callin_map; /* Bitmask of started secondaries */ | |
42 | cpumask_t cpu_online_map; /* Bitmask of currently online CPUs */ | |
43 | int __cpu_number_map[NR_CPUS]; /* Map physical to logical */ | |
44 | int __cpu_logical_map[NR_CPUS]; /* Map logical to physical */ | |
45 | ||
46 | EXPORT_SYMBOL(phys_cpu_present_map); | |
47 | EXPORT_SYMBOL(cpu_online_map); | |
48 | ||
49 | static void smp_tune_scheduling (void) | |
50 | { | |
51 | struct cache_desc *cd = ¤t_cpu_data.scache; | |
52 | unsigned long cachesize; /* kB */ | |
53 | unsigned long bandwidth = 350; /* MB/s */ | |
54 | unsigned long cpu_khz; | |
55 | ||
56 | /* | |
57 | * Crude estimate until we actually meassure ... | |
58 | */ | |
59 | cpu_khz = loops_per_jiffy * 2 * HZ / 1000; | |
60 | ||
61 | /* | |
62 | * Rough estimation for SMP scheduling, this is the number of | |
63 | * cycles it takes for a fully memory-limited process to flush | |
64 | * the SMP-local cache. | |
65 | * | |
66 | * (For a P5 this pretty much means we will choose another idle | |
67 | * CPU almost always at wakeup time (this is due to the small | |
68 | * L1 cache), on PIIs it's around 50-100 usecs, depending on | |
69 | * the cache size) | |
70 | */ | |
71 | if (!cpu_khz) | |
72 | return; | |
73 | ||
74 | cachesize = cd->linesz * cd->sets * cd->ways; | |
75 | } | |
76 | ||
77 | extern void __init calibrate_delay(void); | |
78 | extern ATTRIB_NORET void cpu_idle(void); | |
79 | ||
80 | /* | |
81 | * First C code run on the secondary CPUs after being started up by | |
82 | * the master. | |
83 | */ | |
84 | asmlinkage void start_secondary(void) | |
85 | { | |
86 | unsigned int cpu = smp_processor_id(); | |
87 | ||
88 | cpu_probe(); | |
89 | cpu_report(); | |
90 | per_cpu_trap_init(); | |
91 | prom_init_secondary(); | |
92 | ||
93 | /* | |
94 | * XXX parity protection should be folded in here when it's converted | |
95 | * to an option instead of something based on .cputype | |
96 | */ | |
97 | ||
98 | calibrate_delay(); | |
99 | cpu_data[cpu].udelay_val = loops_per_jiffy; | |
100 | ||
101 | prom_smp_finish(); | |
102 | ||
103 | cpu_set(cpu, cpu_callin_map); | |
104 | ||
105 | cpu_idle(); | |
106 | } | |
107 | ||
108 | DEFINE_SPINLOCK(smp_call_lock); | |
109 | ||
110 | struct call_data_struct *call_data; | |
111 | ||
112 | /* | |
113 | * Run a function on all other CPUs. | |
114 | * <func> The function to run. This must be fast and non-blocking. | |
115 | * <info> An arbitrary pointer to pass to the function. | |
116 | * <retry> If true, keep retrying until ready. | |
117 | * <wait> If true, wait until function has completed on other CPUs. | |
118 | * [RETURNS] 0 on success, else a negative status code. | |
119 | * | |
120 | * Does not return until remote CPUs are nearly ready to execute <func> | |
121 | * or are or have executed. | |
122 | * | |
123 | * You must not call this function with disabled interrupts or from a | |
124 | * hardware interrupt handler or from a bottom half handler. | |
125 | */ | |
126 | int smp_call_function (void (*func) (void *info), void *info, int retry, | |
127 | int wait) | |
128 | { | |
129 | struct call_data_struct data; | |
130 | int i, cpus = num_online_cpus() - 1; | |
131 | int cpu = smp_processor_id(); | |
132 | ||
133 | if (!cpus) | |
134 | return 0; | |
135 | ||
136 | /* Can deadlock when called with interrupts disabled */ | |
137 | WARN_ON(irqs_disabled()); | |
138 | ||
139 | data.func = func; | |
140 | data.info = info; | |
141 | atomic_set(&data.started, 0); | |
142 | data.wait = wait; | |
143 | if (wait) | |
144 | atomic_set(&data.finished, 0); | |
145 | ||
146 | spin_lock(&smp_call_lock); | |
147 | call_data = &data; | |
148 | mb(); | |
149 | ||
150 | /* Send a message to all other CPUs and wait for them to respond */ | |
151 | for (i = 0; i < NR_CPUS; i++) | |
152 | if (cpu_online(i) && i != cpu) | |
153 | core_send_ipi(i, SMP_CALL_FUNCTION); | |
154 | ||
155 | /* Wait for response */ | |
156 | /* FIXME: lock-up detection, backtrace on lock-up */ | |
157 | while (atomic_read(&data.started) != cpus) | |
158 | barrier(); | |
159 | ||
160 | if (wait) | |
161 | while (atomic_read(&data.finished) != cpus) | |
162 | barrier(); | |
163 | spin_unlock(&smp_call_lock); | |
164 | ||
165 | return 0; | |
166 | } | |
167 | ||
168 | void smp_call_function_interrupt(void) | |
169 | { | |
170 | void (*func) (void *info) = call_data->func; | |
171 | void *info = call_data->info; | |
172 | int wait = call_data->wait; | |
173 | ||
174 | /* | |
175 | * Notify initiating CPU that I've grabbed the data and am | |
176 | * about to execute the function. | |
177 | */ | |
178 | mb(); | |
179 | atomic_inc(&call_data->started); | |
180 | ||
181 | /* | |
182 | * At this point the info structure may be out of scope unless wait==1. | |
183 | */ | |
184 | irq_enter(); | |
185 | (*func)(info); | |
186 | irq_exit(); | |
187 | ||
188 | if (wait) { | |
189 | mb(); | |
190 | atomic_inc(&call_data->finished); | |
191 | } | |
192 | } | |
193 | ||
194 | static void stop_this_cpu(void *dummy) | |
195 | { | |
196 | /* | |
197 | * Remove this CPU: | |
198 | */ | |
199 | cpu_clear(smp_processor_id(), cpu_online_map); | |
200 | local_irq_enable(); /* May need to service _machine_restart IPI */ | |
201 | for (;;); /* Wait if available. */ | |
202 | } | |
203 | ||
204 | void smp_send_stop(void) | |
205 | { | |
206 | smp_call_function(stop_this_cpu, NULL, 1, 0); | |
207 | } | |
208 | ||
209 | void __init smp_cpus_done(unsigned int max_cpus) | |
210 | { | |
211 | prom_cpus_done(); | |
212 | } | |
213 | ||
214 | /* called from main before smp_init() */ | |
215 | void __init smp_prepare_cpus(unsigned int max_cpus) | |
216 | { | |
217 | cpu_data[0].udelay_val = loops_per_jiffy; | |
218 | init_new_context(current, &init_mm); | |
219 | current_thread_info()->cpu = 0; | |
220 | smp_tune_scheduling(); | |
221 | prom_prepare_cpus(max_cpus); | |
222 | } | |
223 | ||
224 | /* preload SMP state for boot cpu */ | |
225 | void __devinit smp_prepare_boot_cpu(void) | |
226 | { | |
227 | /* | |
228 | * This assumes that bootup is always handled by the processor | |
229 | * with the logic and physical number 0. | |
230 | */ | |
231 | __cpu_number_map[0] = 0; | |
232 | __cpu_logical_map[0] = 0; | |
233 | cpu_set(0, phys_cpu_present_map); | |
234 | cpu_set(0, cpu_online_map); | |
235 | cpu_set(0, cpu_callin_map); | |
236 | } | |
237 | ||
238 | /* | |
239 | * Startup the CPU with this logical number | |
240 | */ | |
241 | static int __init do_boot_cpu(int cpu) | |
242 | { | |
243 | struct task_struct *idle; | |
244 | ||
245 | /* | |
246 | * The following code is purely to make sure | |
247 | * Linux can schedule processes on this slave. | |
248 | */ | |
249 | idle = fork_idle(cpu); | |
250 | if (IS_ERR(idle)) | |
251 | panic("failed fork for CPU %d\n", cpu); | |
252 | ||
253 | prom_boot_secondary(cpu, idle); | |
254 | ||
255 | /* XXXKW timeout */ | |
256 | while (!cpu_isset(cpu, cpu_callin_map)) | |
257 | udelay(100); | |
258 | ||
259 | cpu_set(cpu, cpu_online_map); | |
260 | ||
261 | return 0; | |
262 | } | |
263 | ||
264 | /* | |
265 | * Called once for each "cpu_possible(cpu)". Needs to spin up the cpu | |
266 | * and keep control until "cpu_online(cpu)" is set. Note: cpu is | |
267 | * physical, not logical. | |
268 | */ | |
269 | int __devinit __cpu_up(unsigned int cpu) | |
270 | { | |
271 | int ret; | |
272 | ||
273 | /* Processor goes to start_secondary(), sets online flag */ | |
274 | ret = do_boot_cpu(cpu); | |
275 | if (ret < 0) | |
276 | return ret; | |
277 | ||
278 | return 0; | |
279 | } | |
280 | ||
281 | /* Not really SMP stuff ... */ | |
282 | int setup_profiling_timer(unsigned int multiplier) | |
283 | { | |
284 | return 0; | |
285 | } | |
286 | ||
287 | static void flush_tlb_all_ipi(void *info) | |
288 | { | |
289 | local_flush_tlb_all(); | |
290 | } | |
291 | ||
292 | void flush_tlb_all(void) | |
293 | { | |
294 | on_each_cpu(flush_tlb_all_ipi, 0, 1, 1); | |
295 | } | |
296 | ||
297 | static void flush_tlb_mm_ipi(void *mm) | |
298 | { | |
299 | local_flush_tlb_mm((struct mm_struct *)mm); | |
300 | } | |
301 | ||
302 | /* | |
303 | * The following tlb flush calls are invoked when old translations are | |
304 | * being torn down, or pte attributes are changing. For single threaded | |
305 | * address spaces, a new context is obtained on the current cpu, and tlb | |
306 | * context on other cpus are invalidated to force a new context allocation | |
307 | * at switch_mm time, should the mm ever be used on other cpus. For | |
308 | * multithreaded address spaces, intercpu interrupts have to be sent. | |
309 | * Another case where intercpu interrupts are required is when the target | |
310 | * mm might be active on another cpu (eg debuggers doing the flushes on | |
311 | * behalf of debugees, kswapd stealing pages from another process etc). | |
312 | * Kanoj 07/00. | |
313 | */ | |
314 | ||
315 | void flush_tlb_mm(struct mm_struct *mm) | |
316 | { | |
317 | preempt_disable(); | |
318 | ||
319 | if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) { | |
320 | smp_call_function(flush_tlb_mm_ipi, (void *)mm, 1, 1); | |
321 | } else { | |
322 | int i; | |
323 | for (i = 0; i < num_online_cpus(); i++) | |
324 | if (smp_processor_id() != i) | |
325 | cpu_context(i, mm) = 0; | |
326 | } | |
327 | local_flush_tlb_mm(mm); | |
328 | ||
329 | preempt_enable(); | |
330 | } | |
331 | ||
332 | struct flush_tlb_data { | |
333 | struct vm_area_struct *vma; | |
334 | unsigned long addr1; | |
335 | unsigned long addr2; | |
336 | }; | |
337 | ||
338 | static void flush_tlb_range_ipi(void *info) | |
339 | { | |
340 | struct flush_tlb_data *fd = (struct flush_tlb_data *)info; | |
341 | ||
342 | local_flush_tlb_range(fd->vma, fd->addr1, fd->addr2); | |
343 | } | |
344 | ||
345 | void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end) | |
346 | { | |
347 | struct mm_struct *mm = vma->vm_mm; | |
348 | ||
349 | preempt_disable(); | |
350 | if ((atomic_read(&mm->mm_users) != 1) || (current->mm != mm)) { | |
351 | struct flush_tlb_data fd; | |
352 | ||
353 | fd.vma = vma; | |
354 | fd.addr1 = start; | |
355 | fd.addr2 = end; | |
356 | smp_call_function(flush_tlb_range_ipi, (void *)&fd, 1, 1); | |
357 | } else { | |
358 | int i; | |
359 | for (i = 0; i < num_online_cpus(); i++) | |
360 | if (smp_processor_id() != i) | |
361 | cpu_context(i, mm) = 0; | |
362 | } | |
363 | local_flush_tlb_range(vma, start, end); | |
364 | preempt_enable(); | |
365 | } | |
366 | ||
367 | static void flush_tlb_kernel_range_ipi(void *info) | |
368 | { | |
369 | struct flush_tlb_data *fd = (struct flush_tlb_data *)info; | |
370 | ||
371 | local_flush_tlb_kernel_range(fd->addr1, fd->addr2); | |
372 | } | |
373 | ||
374 | void flush_tlb_kernel_range(unsigned long start, unsigned long end) | |
375 | { | |
376 | struct flush_tlb_data fd; | |
377 | ||
378 | fd.addr1 = start; | |
379 | fd.addr2 = end; | |
380 | on_each_cpu(flush_tlb_kernel_range_ipi, (void *)&fd, 1, 1); | |
381 | } | |
382 | ||
383 | static void flush_tlb_page_ipi(void *info) | |
384 | { | |
385 | struct flush_tlb_data *fd = (struct flush_tlb_data *)info; | |
386 | ||
387 | local_flush_tlb_page(fd->vma, fd->addr1); | |
388 | } | |
389 | ||
390 | void flush_tlb_page(struct vm_area_struct *vma, unsigned long page) | |
391 | { | |
392 | preempt_disable(); | |
393 | if ((atomic_read(&vma->vm_mm->mm_users) != 1) || (current->mm != vma->vm_mm)) { | |
394 | struct flush_tlb_data fd; | |
395 | ||
396 | fd.vma = vma; | |
397 | fd.addr1 = page; | |
398 | smp_call_function(flush_tlb_page_ipi, (void *)&fd, 1, 1); | |
399 | } else { | |
400 | int i; | |
401 | for (i = 0; i < num_online_cpus(); i++) | |
402 | if (smp_processor_id() != i) | |
403 | cpu_context(i, vma->vm_mm) = 0; | |
404 | } | |
405 | local_flush_tlb_page(vma, page); | |
406 | preempt_enable(); | |
407 | } | |
408 | ||
409 | static void flush_tlb_one_ipi(void *info) | |
410 | { | |
411 | unsigned long vaddr = (unsigned long) info; | |
412 | ||
413 | local_flush_tlb_one(vaddr); | |
414 | } | |
415 | ||
416 | void flush_tlb_one(unsigned long vaddr) | |
417 | { | |
418 | smp_call_function(flush_tlb_one_ipi, (void *) vaddr, 1, 1); | |
419 | local_flush_tlb_one(vaddr); | |
420 | } | |
421 | ||
422 | EXPORT_SYMBOL(flush_tlb_page); | |
423 | EXPORT_SYMBOL(flush_tlb_one); | |
424 | EXPORT_SYMBOL(cpu_data); | |
425 | EXPORT_SYMBOL(synchronize_irq); |