]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/arch/alpha/kernel/smp.c | |
3 | * | |
4 | * 2001-07-09 Phil Ezolt (Phillip.Ezolt@compaq.com) | |
5 | * Renamed modified smp_call_function to smp_call_function_on_cpu() | |
6 | * Created an function that conforms to the old calling convention | |
7 | * of smp_call_function(). | |
8 | * | |
9 | * This is helpful for DCPI. | |
10 | * | |
11 | */ | |
12 | ||
13 | #include <linux/errno.h> | |
14 | #include <linux/kernel.h> | |
15 | #include <linux/kernel_stat.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/sched.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/threads.h> | |
20 | #include <linux/smp.h> | |
21 | #include <linux/smp_lock.h> | |
22 | #include <linux/interrupt.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/delay.h> | |
25 | #include <linux/spinlock.h> | |
26 | #include <linux/irq.h> | |
27 | #include <linux/cache.h> | |
28 | #include <linux/profile.h> | |
29 | #include <linux/bitops.h> | |
30 | ||
31 | #include <asm/hwrpb.h> | |
32 | #include <asm/ptrace.h> | |
33 | #include <asm/atomic.h> | |
34 | ||
35 | #include <asm/io.h> | |
36 | #include <asm/irq.h> | |
37 | #include <asm/pgtable.h> | |
38 | #include <asm/pgalloc.h> | |
39 | #include <asm/mmu_context.h> | |
40 | #include <asm/tlbflush.h> | |
41 | ||
42 | #include "proto.h" | |
43 | #include "irq_impl.h" | |
44 | ||
45 | ||
46 | #define DEBUG_SMP 0 | |
47 | #if DEBUG_SMP | |
48 | #define DBGS(args) printk args | |
49 | #else | |
50 | #define DBGS(args) | |
51 | #endif | |
52 | ||
53 | /* A collection of per-processor data. */ | |
54 | struct cpuinfo_alpha cpu_data[NR_CPUS]; | |
55 | ||
56 | /* A collection of single bit ipi messages. */ | |
57 | static struct { | |
58 | unsigned long bits ____cacheline_aligned; | |
59 | } ipi_data[NR_CPUS] __cacheline_aligned; | |
60 | ||
61 | enum ipi_message_type { | |
62 | IPI_RESCHEDULE, | |
63 | IPI_CALL_FUNC, | |
64 | IPI_CPU_STOP, | |
65 | }; | |
66 | ||
67 | /* Set to a secondary's cpuid when it comes online. */ | |
68 | static int smp_secondary_alive __initdata = 0; | |
69 | ||
70 | /* Which cpus ids came online. */ | |
71 | cpumask_t cpu_present_mask; | |
72 | cpumask_t cpu_online_map; | |
73 | ||
74 | EXPORT_SYMBOL(cpu_online_map); | |
75 | ||
76 | /* cpus reported in the hwrpb */ | |
77 | static unsigned long hwrpb_cpu_present_mask __initdata = 0; | |
78 | ||
79 | int smp_num_probed; /* Internal processor count */ | |
80 | int smp_num_cpus = 1; /* Number that came online. */ | |
81 | ||
82 | extern void calibrate_delay(void); | |
83 | ||
84 | \f | |
85 | ||
86 | /* | |
87 | * Called by both boot and secondaries to move global data into | |
88 | * per-processor storage. | |
89 | */ | |
90 | static inline void __init | |
91 | smp_store_cpu_info(int cpuid) | |
92 | { | |
93 | cpu_data[cpuid].loops_per_jiffy = loops_per_jiffy; | |
94 | cpu_data[cpuid].last_asn = ASN_FIRST_VERSION; | |
95 | cpu_data[cpuid].need_new_asn = 0; | |
96 | cpu_data[cpuid].asn_lock = 0; | |
97 | } | |
98 | ||
99 | /* | |
100 | * Ideally sets up per-cpu profiling hooks. Doesn't do much now... | |
101 | */ | |
102 | static inline void __init | |
103 | smp_setup_percpu_timer(int cpuid) | |
104 | { | |
105 | cpu_data[cpuid].prof_counter = 1; | |
106 | cpu_data[cpuid].prof_multiplier = 1; | |
107 | } | |
108 | ||
109 | static void __init | |
110 | wait_boot_cpu_to_stop(int cpuid) | |
111 | { | |
112 | unsigned long stop = jiffies + 10*HZ; | |
113 | ||
114 | while (time_before(jiffies, stop)) { | |
115 | if (!smp_secondary_alive) | |
116 | return; | |
117 | barrier(); | |
118 | } | |
119 | ||
120 | printk("wait_boot_cpu_to_stop: FAILED on CPU %d, hanging now\n", cpuid); | |
121 | for (;;) | |
122 | barrier(); | |
123 | } | |
124 | ||
125 | /* | |
126 | * Where secondaries begin a life of C. | |
127 | */ | |
128 | void __init | |
129 | smp_callin(void) | |
130 | { | |
131 | int cpuid = hard_smp_processor_id(); | |
132 | ||
133 | if (cpu_test_and_set(cpuid, cpu_online_map)) { | |
134 | printk("??, cpu 0x%x already present??\n", cpuid); | |
135 | BUG(); | |
136 | } | |
137 | ||
138 | /* Turn on machine checks. */ | |
139 | wrmces(7); | |
140 | ||
141 | /* Set trap vectors. */ | |
142 | trap_init(); | |
143 | ||
144 | /* Set interrupt vector. */ | |
145 | wrent(entInt, 0); | |
146 | ||
147 | /* Get our local ticker going. */ | |
148 | smp_setup_percpu_timer(cpuid); | |
149 | ||
150 | /* Call platform-specific callin, if specified */ | |
151 | if (alpha_mv.smp_callin) alpha_mv.smp_callin(); | |
152 | ||
153 | /* All kernel threads share the same mm context. */ | |
154 | atomic_inc(&init_mm.mm_count); | |
155 | current->active_mm = &init_mm; | |
156 | ||
157 | /* Must have completely accurate bogos. */ | |
158 | local_irq_enable(); | |
159 | ||
160 | /* Wait boot CPU to stop with irq enabled before running | |
161 | calibrate_delay. */ | |
162 | wait_boot_cpu_to_stop(cpuid); | |
163 | mb(); | |
164 | calibrate_delay(); | |
165 | ||
166 | smp_store_cpu_info(cpuid); | |
167 | /* Allow master to continue only after we written loops_per_jiffy. */ | |
168 | wmb(); | |
169 | smp_secondary_alive = 1; | |
170 | ||
171 | DBGS(("smp_callin: commencing CPU %d current %p active_mm %p\n", | |
172 | cpuid, current, current->active_mm)); | |
173 | ||
174 | /* Do nothing. */ | |
175 | cpu_idle(); | |
176 | } | |
177 | ||
178 | /* Wait until hwrpb->txrdy is clear for cpu. Return -1 on timeout. */ | |
179 | static int __init | |
180 | wait_for_txrdy (unsigned long cpumask) | |
181 | { | |
182 | unsigned long timeout; | |
183 | ||
184 | if (!(hwrpb->txrdy & cpumask)) | |
185 | return 0; | |
186 | ||
187 | timeout = jiffies + 10*HZ; | |
188 | while (time_before(jiffies, timeout)) { | |
189 | if (!(hwrpb->txrdy & cpumask)) | |
190 | return 0; | |
191 | udelay(10); | |
192 | barrier(); | |
193 | } | |
194 | ||
195 | return -1; | |
196 | } | |
197 | ||
198 | /* | |
199 | * Send a message to a secondary's console. "START" is one such | |
200 | * interesting message. ;-) | |
201 | */ | |
202 | static void __init | |
203 | send_secondary_console_msg(char *str, int cpuid) | |
204 | { | |
205 | struct percpu_struct *cpu; | |
206 | register char *cp1, *cp2; | |
207 | unsigned long cpumask; | |
208 | size_t len; | |
209 | ||
210 | cpu = (struct percpu_struct *) | |
211 | ((char*)hwrpb | |
212 | + hwrpb->processor_offset | |
213 | + cpuid * hwrpb->processor_size); | |
214 | ||
215 | cpumask = (1UL << cpuid); | |
216 | if (wait_for_txrdy(cpumask)) | |
217 | goto timeout; | |
218 | ||
219 | cp2 = str; | |
220 | len = strlen(cp2); | |
221 | *(unsigned int *)&cpu->ipc_buffer[0] = len; | |
222 | cp1 = (char *) &cpu->ipc_buffer[1]; | |
223 | memcpy(cp1, cp2, len); | |
224 | ||
225 | /* atomic test and set */ | |
226 | wmb(); | |
227 | set_bit(cpuid, &hwrpb->rxrdy); | |
228 | ||
229 | if (wait_for_txrdy(cpumask)) | |
230 | goto timeout; | |
231 | return; | |
232 | ||
233 | timeout: | |
234 | printk("Processor %x not ready\n", cpuid); | |
235 | } | |
236 | ||
237 | /* | |
238 | * A secondary console wants to send a message. Receive it. | |
239 | */ | |
240 | static void | |
241 | recv_secondary_console_msg(void) | |
242 | { | |
243 | int mycpu, i, cnt; | |
244 | unsigned long txrdy = hwrpb->txrdy; | |
245 | char *cp1, *cp2, buf[80]; | |
246 | struct percpu_struct *cpu; | |
247 | ||
248 | DBGS(("recv_secondary_console_msg: TXRDY 0x%lx.\n", txrdy)); | |
249 | ||
250 | mycpu = hard_smp_processor_id(); | |
251 | ||
252 | for (i = 0; i < NR_CPUS; i++) { | |
253 | if (!(txrdy & (1UL << i))) | |
254 | continue; | |
255 | ||
256 | DBGS(("recv_secondary_console_msg: " | |
257 | "TXRDY contains CPU %d.\n", i)); | |
258 | ||
259 | cpu = (struct percpu_struct *) | |
260 | ((char*)hwrpb | |
261 | + hwrpb->processor_offset | |
262 | + i * hwrpb->processor_size); | |
263 | ||
264 | DBGS(("recv_secondary_console_msg: on %d from %d" | |
265 | " HALT_REASON 0x%lx FLAGS 0x%lx\n", | |
266 | mycpu, i, cpu->halt_reason, cpu->flags)); | |
267 | ||
268 | cnt = cpu->ipc_buffer[0] >> 32; | |
269 | if (cnt <= 0 || cnt >= 80) | |
270 | strcpy(buf, "<<< BOGUS MSG >>>"); | |
271 | else { | |
272 | cp1 = (char *) &cpu->ipc_buffer[11]; | |
273 | cp2 = buf; | |
274 | strcpy(cp2, cp1); | |
275 | ||
276 | while ((cp2 = strchr(cp2, '\r')) != 0) { | |
277 | *cp2 = ' '; | |
278 | if (cp2[1] == '\n') | |
279 | cp2[1] = ' '; | |
280 | } | |
281 | } | |
282 | ||
283 | DBGS((KERN_INFO "recv_secondary_console_msg: on %d " | |
284 | "message is '%s'\n", mycpu, buf)); | |
285 | } | |
286 | ||
287 | hwrpb->txrdy = 0; | |
288 | } | |
289 | ||
290 | /* | |
291 | * Convince the console to have a secondary cpu begin execution. | |
292 | */ | |
293 | static int __init | |
294 | secondary_cpu_start(int cpuid, struct task_struct *idle) | |
295 | { | |
296 | struct percpu_struct *cpu; | |
297 | struct pcb_struct *hwpcb, *ipcb; | |
298 | unsigned long timeout; | |
299 | ||
300 | cpu = (struct percpu_struct *) | |
301 | ((char*)hwrpb | |
302 | + hwrpb->processor_offset | |
303 | + cpuid * hwrpb->processor_size); | |
304 | hwpcb = (struct pcb_struct *) cpu->hwpcb; | |
305 | ipcb = &idle->thread_info->pcb; | |
306 | ||
307 | /* Initialize the CPU's HWPCB to something just good enough for | |
308 | us to get started. Immediately after starting, we'll swpctx | |
309 | to the target idle task's pcb. Reuse the stack in the mean | |
310 | time. Precalculate the target PCBB. */ | |
311 | hwpcb->ksp = (unsigned long)ipcb + sizeof(union thread_union) - 16; | |
312 | hwpcb->usp = 0; | |
313 | hwpcb->ptbr = ipcb->ptbr; | |
314 | hwpcb->pcc = 0; | |
315 | hwpcb->asn = 0; | |
316 | hwpcb->unique = virt_to_phys(ipcb); | |
317 | hwpcb->flags = ipcb->flags; | |
318 | hwpcb->res1 = hwpcb->res2 = 0; | |
319 | ||
320 | #if 0 | |
321 | DBGS(("KSP 0x%lx PTBR 0x%lx VPTBR 0x%lx UNIQUE 0x%lx\n", | |
322 | hwpcb->ksp, hwpcb->ptbr, hwrpb->vptb, hwpcb->unique)); | |
323 | #endif | |
324 | DBGS(("Starting secondary cpu %d: state 0x%lx pal_flags 0x%lx\n", | |
325 | cpuid, idle->state, ipcb->flags)); | |
326 | ||
327 | /* Setup HWRPB fields that SRM uses to activate secondary CPU */ | |
328 | hwrpb->CPU_restart = __smp_callin; | |
329 | hwrpb->CPU_restart_data = (unsigned long) __smp_callin; | |
330 | ||
331 | /* Recalculate and update the HWRPB checksum */ | |
332 | hwrpb_update_checksum(hwrpb); | |
333 | ||
334 | /* | |
335 | * Send a "start" command to the specified processor. | |
336 | */ | |
337 | ||
338 | /* SRM III 3.4.1.3 */ | |
339 | cpu->flags |= 0x22; /* turn on Context Valid and Restart Capable */ | |
340 | cpu->flags &= ~1; /* turn off Bootstrap In Progress */ | |
341 | wmb(); | |
342 | ||
343 | send_secondary_console_msg("START\r\n", cpuid); | |
344 | ||
345 | /* Wait 10 seconds for an ACK from the console. */ | |
346 | timeout = jiffies + 10*HZ; | |
347 | while (time_before(jiffies, timeout)) { | |
348 | if (cpu->flags & 1) | |
349 | goto started; | |
350 | udelay(10); | |
351 | barrier(); | |
352 | } | |
353 | printk(KERN_ERR "SMP: Processor %d failed to start.\n", cpuid); | |
354 | return -1; | |
355 | ||
356 | started: | |
357 | DBGS(("secondary_cpu_start: SUCCESS for CPU %d!!!\n", cpuid)); | |
358 | return 0; | |
359 | } | |
360 | ||
361 | /* | |
362 | * Bring one cpu online. | |
363 | */ | |
364 | static int __init | |
365 | smp_boot_one_cpu(int cpuid) | |
366 | { | |
367 | struct task_struct *idle; | |
368 | unsigned long timeout; | |
369 | ||
370 | /* Cook up an idler for this guy. Note that the address we | |
371 | give to kernel_thread is irrelevant -- it's going to start | |
372 | where HWRPB.CPU_restart says to start. But this gets all | |
373 | the other task-y sort of data structures set up like we | |
374 | wish. We can't use kernel_thread since we must avoid | |
375 | rescheduling the child. */ | |
376 | idle = fork_idle(cpuid); | |
377 | if (IS_ERR(idle)) | |
378 | panic("failed fork for CPU %d", cpuid); | |
379 | ||
380 | DBGS(("smp_boot_one_cpu: CPU %d state 0x%lx flags 0x%lx\n", | |
381 | cpuid, idle->state, idle->flags)); | |
382 | ||
383 | /* Signal the secondary to wait a moment. */ | |
384 | smp_secondary_alive = -1; | |
385 | ||
386 | /* Whirrr, whirrr, whirrrrrrrrr... */ | |
387 | if (secondary_cpu_start(cpuid, idle)) | |
388 | return -1; | |
389 | ||
390 | /* Notify the secondary CPU it can run calibrate_delay. */ | |
391 | mb(); | |
392 | smp_secondary_alive = 0; | |
393 | ||
394 | /* We've been acked by the console; wait one second for | |
395 | the task to start up for real. */ | |
396 | timeout = jiffies + 1*HZ; | |
397 | while (time_before(jiffies, timeout)) { | |
398 | if (smp_secondary_alive == 1) | |
399 | goto alive; | |
400 | udelay(10); | |
401 | barrier(); | |
402 | } | |
403 | ||
404 | /* We failed to boot the CPU. */ | |
405 | ||
406 | printk(KERN_ERR "SMP: Processor %d is stuck.\n", cpuid); | |
407 | return -1; | |
408 | ||
409 | alive: | |
410 | /* Another "Red Snapper". */ | |
411 | return 0; | |
412 | } | |
413 | ||
414 | /* | |
415 | * Called from setup_arch. Detect an SMP system and which processors | |
416 | * are present. | |
417 | */ | |
418 | void __init | |
419 | setup_smp(void) | |
420 | { | |
421 | struct percpu_struct *cpubase, *cpu; | |
422 | unsigned long i; | |
423 | ||
424 | if (boot_cpuid != 0) { | |
425 | printk(KERN_WARNING "SMP: Booting off cpu %d instead of 0?\n", | |
426 | boot_cpuid); | |
427 | } | |
428 | ||
429 | if (hwrpb->nr_processors > 1) { | |
430 | int boot_cpu_palrev; | |
431 | ||
432 | DBGS(("setup_smp: nr_processors %ld\n", | |
433 | hwrpb->nr_processors)); | |
434 | ||
435 | cpubase = (struct percpu_struct *) | |
436 | ((char*)hwrpb + hwrpb->processor_offset); | |
437 | boot_cpu_palrev = cpubase->pal_revision; | |
438 | ||
439 | for (i = 0; i < hwrpb->nr_processors; i++) { | |
440 | cpu = (struct percpu_struct *) | |
441 | ((char *)cpubase + i*hwrpb->processor_size); | |
442 | if ((cpu->flags & 0x1cc) == 0x1cc) { | |
443 | smp_num_probed++; | |
444 | /* Assume here that "whami" == index */ | |
445 | hwrpb_cpu_present_mask |= (1UL << i); | |
446 | cpu->pal_revision = boot_cpu_palrev; | |
447 | } | |
448 | ||
449 | DBGS(("setup_smp: CPU %d: flags 0x%lx type 0x%lx\n", | |
450 | i, cpu->flags, cpu->type)); | |
451 | DBGS(("setup_smp: CPU %d: PAL rev 0x%lx\n", | |
452 | i, cpu->pal_revision)); | |
453 | } | |
454 | } else { | |
455 | smp_num_probed = 1; | |
456 | hwrpb_cpu_present_mask = (1UL << boot_cpuid); | |
457 | } | |
458 | cpu_present_mask = cpumask_of_cpu(boot_cpuid); | |
459 | ||
460 | printk(KERN_INFO "SMP: %d CPUs probed -- cpu_present_mask = %lx\n", | |
461 | smp_num_probed, hwrpb_cpu_present_mask); | |
462 | } | |
463 | ||
464 | /* | |
465 | * Called by smp_init prepare the secondaries | |
466 | */ | |
467 | void __init | |
468 | smp_prepare_cpus(unsigned int max_cpus) | |
469 | { | |
470 | int cpu_count, i; | |
471 | ||
472 | /* Take care of some initial bookkeeping. */ | |
473 | memset(ipi_data, 0, sizeof(ipi_data)); | |
474 | ||
475 | current_thread_info()->cpu = boot_cpuid; | |
476 | ||
477 | smp_store_cpu_info(boot_cpuid); | |
478 | smp_setup_percpu_timer(boot_cpuid); | |
479 | ||
480 | /* Nothing to do on a UP box, or when told not to. */ | |
481 | if (smp_num_probed == 1 || max_cpus == 0) { | |
482 | cpu_present_mask = cpumask_of_cpu(boot_cpuid); | |
483 | printk(KERN_INFO "SMP mode deactivated.\n"); | |
484 | return; | |
485 | } | |
486 | ||
487 | printk(KERN_INFO "SMP starting up secondaries.\n"); | |
488 | ||
489 | cpu_count = 1; | |
490 | for (i = 0; (i < NR_CPUS) && (cpu_count < max_cpus); i++) { | |
491 | if (i == boot_cpuid) | |
492 | continue; | |
493 | ||
494 | if (((hwrpb_cpu_present_mask >> i) & 1) == 0) | |
495 | continue; | |
496 | ||
497 | cpu_set(i, cpu_possible_map); | |
498 | cpu_count++; | |
499 | } | |
500 | ||
501 | smp_num_cpus = cpu_count; | |
502 | } | |
503 | ||
504 | void __devinit | |
505 | smp_prepare_boot_cpu(void) | |
506 | { | |
507 | /* | |
508 | * Mark the boot cpu (current cpu) as both present and online | |
509 | */ | |
510 | cpu_set(smp_processor_id(), cpu_present_mask); | |
511 | cpu_set(smp_processor_id(), cpu_online_map); | |
512 | } | |
513 | ||
514 | int __devinit | |
515 | __cpu_up(unsigned int cpu) | |
516 | { | |
517 | smp_boot_one_cpu(cpu); | |
518 | ||
519 | return cpu_online(cpu) ? 0 : -ENOSYS; | |
520 | } | |
521 | ||
522 | void __init | |
523 | smp_cpus_done(unsigned int max_cpus) | |
524 | { | |
525 | int cpu; | |
526 | unsigned long bogosum = 0; | |
527 | ||
528 | for(cpu = 0; cpu < NR_CPUS; cpu++) | |
529 | if (cpu_online(cpu)) | |
530 | bogosum += cpu_data[cpu].loops_per_jiffy; | |
531 | ||
532 | printk(KERN_INFO "SMP: Total of %d processors activated " | |
533 | "(%lu.%02lu BogoMIPS).\n", | |
534 | num_online_cpus(), | |
535 | (bogosum + 2500) / (500000/HZ), | |
536 | ((bogosum + 2500) / (5000/HZ)) % 100); | |
537 | } | |
538 | ||
539 | \f | |
540 | void | |
541 | smp_percpu_timer_interrupt(struct pt_regs *regs) | |
542 | { | |
543 | int cpu = smp_processor_id(); | |
544 | unsigned long user = user_mode(regs); | |
545 | struct cpuinfo_alpha *data = &cpu_data[cpu]; | |
546 | ||
547 | /* Record kernel PC. */ | |
548 | profile_tick(CPU_PROFILING, regs); | |
549 | ||
550 | if (!--data->prof_counter) { | |
551 | /* We need to make like a normal interrupt -- otherwise | |
552 | timer interrupts ignore the global interrupt lock, | |
553 | which would be a Bad Thing. */ | |
554 | irq_enter(); | |
555 | ||
556 | update_process_times(user); | |
557 | ||
558 | data->prof_counter = data->prof_multiplier; | |
559 | ||
560 | irq_exit(); | |
561 | } | |
562 | } | |
563 | ||
564 | int __init | |
565 | setup_profiling_timer(unsigned int multiplier) | |
566 | { | |
567 | return -EINVAL; | |
568 | } | |
569 | ||
570 | \f | |
571 | static void | |
572 | send_ipi_message(cpumask_t to_whom, enum ipi_message_type operation) | |
573 | { | |
574 | int i; | |
575 | ||
576 | mb(); | |
577 | for_each_cpu_mask(i, to_whom) | |
578 | set_bit(operation, &ipi_data[i].bits); | |
579 | ||
580 | mb(); | |
581 | for_each_cpu_mask(i, to_whom) | |
582 | wripir(i); | |
583 | } | |
584 | ||
585 | /* Structure and data for smp_call_function. This is designed to | |
586 | minimize static memory requirements. Plus it looks cleaner. */ | |
587 | ||
588 | struct smp_call_struct { | |
589 | void (*func) (void *info); | |
590 | void *info; | |
591 | long wait; | |
592 | atomic_t unstarted_count; | |
593 | atomic_t unfinished_count; | |
594 | }; | |
595 | ||
596 | static struct smp_call_struct *smp_call_function_data; | |
597 | ||
598 | /* Atomicly drop data into a shared pointer. The pointer is free if | |
599 | it is initially locked. If retry, spin until free. */ | |
600 | ||
601 | static int | |
602 | pointer_lock (void *lock, void *data, int retry) | |
603 | { | |
604 | void *old, *tmp; | |
605 | ||
606 | mb(); | |
607 | again: | |
608 | /* Compare and swap with zero. */ | |
609 | asm volatile ( | |
610 | "1: ldq_l %0,%1\n" | |
611 | " mov %3,%2\n" | |
612 | " bne %0,2f\n" | |
613 | " stq_c %2,%1\n" | |
614 | " beq %2,1b\n" | |
615 | "2:" | |
616 | : "=&r"(old), "=m"(*(void **)lock), "=&r"(tmp) | |
617 | : "r"(data) | |
618 | : "memory"); | |
619 | ||
620 | if (old == 0) | |
621 | return 0; | |
622 | if (! retry) | |
623 | return -EBUSY; | |
624 | ||
625 | while (*(void **)lock) | |
626 | barrier(); | |
627 | goto again; | |
628 | } | |
629 | ||
630 | void | |
631 | handle_ipi(struct pt_regs *regs) | |
632 | { | |
633 | int this_cpu = smp_processor_id(); | |
634 | unsigned long *pending_ipis = &ipi_data[this_cpu].bits; | |
635 | unsigned long ops; | |
636 | ||
637 | #if 0 | |
638 | DBGS(("handle_ipi: on CPU %d ops 0x%lx PC 0x%lx\n", | |
639 | this_cpu, *pending_ipis, regs->pc)); | |
640 | #endif | |
641 | ||
642 | mb(); /* Order interrupt and bit testing. */ | |
643 | while ((ops = xchg(pending_ipis, 0)) != 0) { | |
644 | mb(); /* Order bit clearing and data access. */ | |
645 | do { | |
646 | unsigned long which; | |
647 | ||
648 | which = ops & -ops; | |
649 | ops &= ~which; | |
650 | which = __ffs(which); | |
651 | ||
652 | switch (which) { | |
653 | case IPI_RESCHEDULE: | |
654 | /* Reschedule callback. Everything to be done | |
655 | is done by the interrupt return path. */ | |
656 | break; | |
657 | ||
658 | case IPI_CALL_FUNC: | |
659 | { | |
660 | struct smp_call_struct *data; | |
661 | void (*func)(void *info); | |
662 | void *info; | |
663 | int wait; | |
664 | ||
665 | data = smp_call_function_data; | |
666 | func = data->func; | |
667 | info = data->info; | |
668 | wait = data->wait; | |
669 | ||
670 | /* Notify the sending CPU that the data has been | |
671 | received, and execution is about to begin. */ | |
672 | mb(); | |
673 | atomic_dec (&data->unstarted_count); | |
674 | ||
675 | /* At this point the structure may be gone unless | |
676 | wait is true. */ | |
677 | (*func)(info); | |
678 | ||
679 | /* Notify the sending CPU that the task is done. */ | |
680 | mb(); | |
681 | if (wait) atomic_dec (&data->unfinished_count); | |
682 | break; | |
683 | } | |
684 | ||
685 | case IPI_CPU_STOP: | |
686 | halt(); | |
687 | ||
688 | default: | |
689 | printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n", | |
690 | this_cpu, which); | |
691 | break; | |
692 | } | |
693 | } while (ops); | |
694 | ||
695 | mb(); /* Order data access and bit testing. */ | |
696 | } | |
697 | ||
698 | cpu_data[this_cpu].ipi_count++; | |
699 | ||
700 | if (hwrpb->txrdy) | |
701 | recv_secondary_console_msg(); | |
702 | } | |
703 | ||
704 | void | |
705 | smp_send_reschedule(int cpu) | |
706 | { | |
707 | #ifdef DEBUG_IPI_MSG | |
708 | if (cpu == hard_smp_processor_id()) | |
709 | printk(KERN_WARNING | |
710 | "smp_send_reschedule: Sending IPI to self.\n"); | |
711 | #endif | |
712 | send_ipi_message(cpumask_of_cpu(cpu), IPI_RESCHEDULE); | |
713 | } | |
714 | ||
715 | void | |
716 | smp_send_stop(void) | |
717 | { | |
718 | cpumask_t to_whom = cpu_possible_map; | |
719 | cpu_clear(smp_processor_id(), to_whom); | |
720 | #ifdef DEBUG_IPI_MSG | |
721 | if (hard_smp_processor_id() != boot_cpu_id) | |
722 | printk(KERN_WARNING "smp_send_stop: Not on boot cpu.\n"); | |
723 | #endif | |
724 | send_ipi_message(to_whom, IPI_CPU_STOP); | |
725 | } | |
726 | ||
727 | /* | |
728 | * Run a function on all other CPUs. | |
729 | * <func> The function to run. This must be fast and non-blocking. | |
730 | * <info> An arbitrary pointer to pass to the function. | |
731 | * <retry> If true, keep retrying until ready. | |
732 | * <wait> If true, wait until function has completed on other CPUs. | |
733 | * [RETURNS] 0 on success, else a negative status code. | |
734 | * | |
735 | * Does not return until remote CPUs are nearly ready to execute <func> | |
736 | * or are or have executed. | |
737 | * You must not call this function with disabled interrupts or from a | |
738 | * hardware interrupt handler or from a bottom half handler. | |
739 | */ | |
740 | ||
741 | int | |
742 | smp_call_function_on_cpu (void (*func) (void *info), void *info, int retry, | |
743 | int wait, cpumask_t to_whom) | |
744 | { | |
745 | struct smp_call_struct data; | |
746 | unsigned long timeout; | |
747 | int num_cpus_to_call; | |
748 | ||
749 | /* Can deadlock when called with interrupts disabled */ | |
750 | WARN_ON(irqs_disabled()); | |
751 | ||
752 | data.func = func; | |
753 | data.info = info; | |
754 | data.wait = wait; | |
755 | ||
756 | cpu_clear(smp_processor_id(), to_whom); | |
757 | num_cpus_to_call = cpus_weight(to_whom); | |
758 | ||
759 | atomic_set(&data.unstarted_count, num_cpus_to_call); | |
760 | atomic_set(&data.unfinished_count, num_cpus_to_call); | |
761 | ||
762 | /* Acquire the smp_call_function_data mutex. */ | |
763 | if (pointer_lock(&smp_call_function_data, &data, retry)) | |
764 | return -EBUSY; | |
765 | ||
766 | /* Send a message to the requested CPUs. */ | |
767 | send_ipi_message(to_whom, IPI_CALL_FUNC); | |
768 | ||
769 | /* Wait for a minimal response. */ | |
770 | timeout = jiffies + HZ; | |
771 | while (atomic_read (&data.unstarted_count) > 0 | |
772 | && time_before (jiffies, timeout)) | |
773 | barrier(); | |
774 | ||
775 | /* If there's no response yet, log a message but allow a longer | |
776 | * timeout period -- if we get a response this time, log | |
777 | * a message saying when we got it.. | |
778 | */ | |
779 | if (atomic_read(&data.unstarted_count) > 0) { | |
780 | long start_time = jiffies; | |
781 | printk(KERN_ERR "%s: initial timeout -- trying long wait\n", | |
782 | __FUNCTION__); | |
783 | timeout = jiffies + 30 * HZ; | |
784 | while (atomic_read(&data.unstarted_count) > 0 | |
785 | && time_before(jiffies, timeout)) | |
786 | barrier(); | |
787 | if (atomic_read(&data.unstarted_count) <= 0) { | |
788 | long delta = jiffies - start_time; | |
789 | printk(KERN_ERR | |
790 | "%s: response %ld.%ld seconds into long wait\n", | |
791 | __FUNCTION__, delta / HZ, | |
792 | (100 * (delta - ((delta / HZ) * HZ))) / HZ); | |
793 | } | |
794 | } | |
795 | ||
796 | /* We either got one or timed out -- clear the lock. */ | |
797 | mb(); | |
798 | smp_call_function_data = NULL; | |
799 | ||
800 | /* | |
801 | * If after both the initial and long timeout periods we still don't | |
802 | * have a response, something is very wrong... | |
803 | */ | |
804 | BUG_ON(atomic_read (&data.unstarted_count) > 0); | |
805 | ||
806 | /* Wait for a complete response, if needed. */ | |
807 | if (wait) { | |
808 | while (atomic_read (&data.unfinished_count) > 0) | |
809 | barrier(); | |
810 | } | |
811 | ||
812 | return 0; | |
813 | } | |
814 | ||
815 | int | |
816 | smp_call_function (void (*func) (void *info), void *info, int retry, int wait) | |
817 | { | |
818 | return smp_call_function_on_cpu (func, info, retry, wait, | |
819 | cpu_online_map); | |
820 | } | |
821 | ||
822 | static void | |
823 | ipi_imb(void *ignored) | |
824 | { | |
825 | imb(); | |
826 | } | |
827 | ||
828 | void | |
829 | smp_imb(void) | |
830 | { | |
831 | /* Must wait other processors to flush their icache before continue. */ | |
832 | if (on_each_cpu(ipi_imb, NULL, 1, 1)) | |
833 | printk(KERN_CRIT "smp_imb: timed out\n"); | |
834 | } | |
835 | ||
836 | static void | |
837 | ipi_flush_tlb_all(void *ignored) | |
838 | { | |
839 | tbia(); | |
840 | } | |
841 | ||
842 | void | |
843 | flush_tlb_all(void) | |
844 | { | |
845 | /* Although we don't have any data to pass, we do want to | |
846 | synchronize with the other processors. */ | |
847 | if (on_each_cpu(ipi_flush_tlb_all, NULL, 1, 1)) { | |
848 | printk(KERN_CRIT "flush_tlb_all: timed out\n"); | |
849 | } | |
850 | } | |
851 | ||
852 | #define asn_locked() (cpu_data[smp_processor_id()].asn_lock) | |
853 | ||
854 | static void | |
855 | ipi_flush_tlb_mm(void *x) | |
856 | { | |
857 | struct mm_struct *mm = (struct mm_struct *) x; | |
858 | if (mm == current->active_mm && !asn_locked()) | |
859 | flush_tlb_current(mm); | |
860 | else | |
861 | flush_tlb_other(mm); | |
862 | } | |
863 | ||
864 | void | |
865 | flush_tlb_mm(struct mm_struct *mm) | |
866 | { | |
867 | preempt_disable(); | |
868 | ||
869 | if (mm == current->active_mm) { | |
870 | flush_tlb_current(mm); | |
871 | if (atomic_read(&mm->mm_users) <= 1) { | |
872 | int cpu, this_cpu = smp_processor_id(); | |
873 | for (cpu = 0; cpu < NR_CPUS; cpu++) { | |
874 | if (!cpu_online(cpu) || cpu == this_cpu) | |
875 | continue; | |
876 | if (mm->context[cpu]) | |
877 | mm->context[cpu] = 0; | |
878 | } | |
879 | preempt_enable(); | |
880 | return; | |
881 | } | |
882 | } | |
883 | ||
884 | if (smp_call_function(ipi_flush_tlb_mm, mm, 1, 1)) { | |
885 | printk(KERN_CRIT "flush_tlb_mm: timed out\n"); | |
886 | } | |
887 | ||
888 | preempt_enable(); | |
889 | } | |
890 | ||
891 | struct flush_tlb_page_struct { | |
892 | struct vm_area_struct *vma; | |
893 | struct mm_struct *mm; | |
894 | unsigned long addr; | |
895 | }; | |
896 | ||
897 | static void | |
898 | ipi_flush_tlb_page(void *x) | |
899 | { | |
900 | struct flush_tlb_page_struct *data = (struct flush_tlb_page_struct *)x; | |
901 | struct mm_struct * mm = data->mm; | |
902 | ||
903 | if (mm == current->active_mm && !asn_locked()) | |
904 | flush_tlb_current_page(mm, data->vma, data->addr); | |
905 | else | |
906 | flush_tlb_other(mm); | |
907 | } | |
908 | ||
909 | void | |
910 | flush_tlb_page(struct vm_area_struct *vma, unsigned long addr) | |
911 | { | |
912 | struct flush_tlb_page_struct data; | |
913 | struct mm_struct *mm = vma->vm_mm; | |
914 | ||
915 | preempt_disable(); | |
916 | ||
917 | if (mm == current->active_mm) { | |
918 | flush_tlb_current_page(mm, vma, addr); | |
919 | if (atomic_read(&mm->mm_users) <= 1) { | |
920 | int cpu, this_cpu = smp_processor_id(); | |
921 | for (cpu = 0; cpu < NR_CPUS; cpu++) { | |
922 | if (!cpu_online(cpu) || cpu == this_cpu) | |
923 | continue; | |
924 | if (mm->context[cpu]) | |
925 | mm->context[cpu] = 0; | |
926 | } | |
927 | preempt_enable(); | |
928 | return; | |
929 | } | |
930 | } | |
931 | ||
932 | data.vma = vma; | |
933 | data.mm = mm; | |
934 | data.addr = addr; | |
935 | ||
936 | if (smp_call_function(ipi_flush_tlb_page, &data, 1, 1)) { | |
937 | printk(KERN_CRIT "flush_tlb_page: timed out\n"); | |
938 | } | |
939 | ||
940 | preempt_enable(); | |
941 | } | |
942 | ||
943 | void | |
944 | flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end) | |
945 | { | |
946 | /* On the Alpha we always flush the whole user tlb. */ | |
947 | flush_tlb_mm(vma->vm_mm); | |
948 | } | |
949 | ||
950 | static void | |
951 | ipi_flush_icache_page(void *x) | |
952 | { | |
953 | struct mm_struct *mm = (struct mm_struct *) x; | |
954 | if (mm == current->active_mm && !asn_locked()) | |
955 | __load_new_mm_context(mm); | |
956 | else | |
957 | flush_tlb_other(mm); | |
958 | } | |
959 | ||
960 | void | |
961 | flush_icache_user_range(struct vm_area_struct *vma, struct page *page, | |
962 | unsigned long addr, int len) | |
963 | { | |
964 | struct mm_struct *mm = vma->vm_mm; | |
965 | ||
966 | if ((vma->vm_flags & VM_EXEC) == 0) | |
967 | return; | |
968 | ||
969 | preempt_disable(); | |
970 | ||
971 | if (mm == current->active_mm) { | |
972 | __load_new_mm_context(mm); | |
973 | if (atomic_read(&mm->mm_users) <= 1) { | |
974 | int cpu, this_cpu = smp_processor_id(); | |
975 | for (cpu = 0; cpu < NR_CPUS; cpu++) { | |
976 | if (!cpu_online(cpu) || cpu == this_cpu) | |
977 | continue; | |
978 | if (mm->context[cpu]) | |
979 | mm->context[cpu] = 0; | |
980 | } | |
981 | preempt_enable(); | |
982 | return; | |
983 | } | |
984 | } | |
985 | ||
986 | if (smp_call_function(ipi_flush_icache_page, mm, 1, 1)) { | |
987 | printk(KERN_CRIT "flush_icache_page: timed out\n"); | |
988 | } | |
989 | ||
990 | preempt_enable(); | |
991 | } | |
992 | \f | |
993 | #ifdef CONFIG_DEBUG_SPINLOCK | |
994 | void | |
995 | _raw_spin_unlock(spinlock_t * lock) | |
996 | { | |
997 | mb(); | |
998 | lock->lock = 0; | |
999 | ||
1000 | lock->on_cpu = -1; | |
1001 | lock->previous = NULL; | |
1002 | lock->task = NULL; | |
1003 | lock->base_file = "none"; | |
1004 | lock->line_no = 0; | |
1005 | } | |
1006 | ||
1007 | void | |
1008 | debug_spin_lock(spinlock_t * lock, const char *base_file, int line_no) | |
1009 | { | |
1010 | long tmp; | |
1011 | long stuck; | |
1012 | void *inline_pc = __builtin_return_address(0); | |
1013 | unsigned long started = jiffies; | |
1014 | int printed = 0; | |
1015 | int cpu = smp_processor_id(); | |
1016 | ||
1017 | stuck = 1L << 30; | |
1018 | try_again: | |
1019 | ||
1020 | /* Use sub-sections to put the actual loop at the end | |
1021 | of this object file's text section so as to perfect | |
1022 | branch prediction. */ | |
1023 | __asm__ __volatile__( | |
1024 | "1: ldl_l %0,%1\n" | |
1025 | " subq %2,1,%2\n" | |
1026 | " blbs %0,2f\n" | |
1027 | " or %0,1,%0\n" | |
1028 | " stl_c %0,%1\n" | |
1029 | " beq %0,3f\n" | |
1030 | "4: mb\n" | |
1031 | ".subsection 2\n" | |
1032 | "2: ldl %0,%1\n" | |
1033 | " subq %2,1,%2\n" | |
1034 | "3: blt %2,4b\n" | |
1035 | " blbs %0,2b\n" | |
1036 | " br 1b\n" | |
1037 | ".previous" | |
1038 | : "=r" (tmp), "=m" (lock->lock), "=r" (stuck) | |
1039 | : "1" (lock->lock), "2" (stuck) : "memory"); | |
1040 | ||
1041 | if (stuck < 0) { | |
1042 | printk(KERN_WARNING | |
1043 | "%s:%d spinlock stuck in %s at %p(%d)" | |
1044 | " owner %s at %p(%d) %s:%d\n", | |
1045 | base_file, line_no, | |
1046 | current->comm, inline_pc, cpu, | |
1047 | lock->task->comm, lock->previous, | |
1048 | lock->on_cpu, lock->base_file, lock->line_no); | |
1049 | stuck = 1L << 36; | |
1050 | printed = 1; | |
1051 | goto try_again; | |
1052 | } | |
1053 | ||
1054 | /* Exiting. Got the lock. */ | |
1055 | lock->on_cpu = cpu; | |
1056 | lock->previous = inline_pc; | |
1057 | lock->task = current; | |
1058 | lock->base_file = base_file; | |
1059 | lock->line_no = line_no; | |
1060 | ||
1061 | if (printed) { | |
1062 | printk(KERN_WARNING | |
1063 | "%s:%d spinlock grabbed in %s at %p(%d) %ld ticks\n", | |
1064 | base_file, line_no, current->comm, inline_pc, | |
1065 | cpu, jiffies - started); | |
1066 | } | |
1067 | } | |
1068 | ||
1069 | int | |
1070 | debug_spin_trylock(spinlock_t * lock, const char *base_file, int line_no) | |
1071 | { | |
1072 | int ret; | |
1073 | if ((ret = !test_and_set_bit(0, lock))) { | |
1074 | lock->on_cpu = smp_processor_id(); | |
1075 | lock->previous = __builtin_return_address(0); | |
1076 | lock->task = current; | |
1077 | } else { | |
1078 | lock->base_file = base_file; | |
1079 | lock->line_no = line_no; | |
1080 | } | |
1081 | return ret; | |
1082 | } | |
1083 | #endif /* CONFIG_DEBUG_SPINLOCK */ | |
1084 | \f | |
1085 | #ifdef CONFIG_DEBUG_RWLOCK | |
1086 | void _raw_write_lock(rwlock_t * lock) | |
1087 | { | |
1088 | long regx, regy; | |
1089 | int stuck_lock, stuck_reader; | |
1090 | void *inline_pc = __builtin_return_address(0); | |
1091 | ||
1092 | try_again: | |
1093 | ||
1094 | stuck_lock = 1<<30; | |
1095 | stuck_reader = 1<<30; | |
1096 | ||
1097 | __asm__ __volatile__( | |
1098 | "1: ldl_l %1,%0\n" | |
1099 | " blbs %1,6f\n" | |
1100 | " blt %1,8f\n" | |
1101 | " mov 1,%1\n" | |
1102 | " stl_c %1,%0\n" | |
1103 | " beq %1,6f\n" | |
1104 | "4: mb\n" | |
1105 | ".subsection 2\n" | |
1106 | "6: blt %3,4b # debug\n" | |
1107 | " subl %3,1,%3 # debug\n" | |
1108 | " ldl %1,%0\n" | |
1109 | " blbs %1,6b\n" | |
1110 | "8: blt %4,4b # debug\n" | |
1111 | " subl %4,1,%4 # debug\n" | |
1112 | " ldl %1,%0\n" | |
1113 | " blt %1,8b\n" | |
1114 | " br 1b\n" | |
1115 | ".previous" | |
1116 | : "=m" (*(volatile int *)lock), "=&r" (regx), "=&r" (regy), | |
1117 | "=&r" (stuck_lock), "=&r" (stuck_reader) | |
1118 | : "0" (*(volatile int *)lock), "3" (stuck_lock), "4" (stuck_reader) : "memory"); | |
1119 | ||
1120 | if (stuck_lock < 0) { | |
1121 | printk(KERN_WARNING "write_lock stuck at %p\n", inline_pc); | |
1122 | goto try_again; | |
1123 | } | |
1124 | if (stuck_reader < 0) { | |
1125 | printk(KERN_WARNING "write_lock stuck on readers at %p\n", | |
1126 | inline_pc); | |
1127 | goto try_again; | |
1128 | } | |
1129 | } | |
1130 | ||
1131 | void _raw_read_lock(rwlock_t * lock) | |
1132 | { | |
1133 | long regx; | |
1134 | int stuck_lock; | |
1135 | void *inline_pc = __builtin_return_address(0); | |
1136 | ||
1137 | try_again: | |
1138 | ||
1139 | stuck_lock = 1<<30; | |
1140 | ||
1141 | __asm__ __volatile__( | |
1142 | "1: ldl_l %1,%0;" | |
1143 | " blbs %1,6f;" | |
1144 | " subl %1,2,%1;" | |
1145 | " stl_c %1,%0;" | |
1146 | " beq %1,6f;" | |
1147 | "4: mb\n" | |
1148 | ".subsection 2\n" | |
1149 | "6: ldl %1,%0;" | |
1150 | " blt %2,4b # debug\n" | |
1151 | " subl %2,1,%2 # debug\n" | |
1152 | " blbs %1,6b;" | |
1153 | " br 1b\n" | |
1154 | ".previous" | |
1155 | : "=m" (*(volatile int *)lock), "=&r" (regx), "=&r" (stuck_lock) | |
1156 | : "0" (*(volatile int *)lock), "2" (stuck_lock) : "memory"); | |
1157 | ||
1158 | if (stuck_lock < 0) { | |
1159 | printk(KERN_WARNING "read_lock stuck at %p\n", inline_pc); | |
1160 | goto try_again; | |
1161 | } | |
1162 | } | |
1163 | #endif /* CONFIG_DEBUG_RWLOCK */ |