2 * linux/arch/alpha/kernel/smp.c
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().
9 * This is helpful for DCPI.
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>
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>
31 #include <asm/hwrpb.h>
32 #include <asm/ptrace.h>
33 #include <asm/atomic.h>
37 #include <asm/pgtable.h>
38 #include <asm/pgalloc.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
48 #define DBGS(args) printk args
53 /* A collection of per-processor data. */
54 struct cpuinfo_alpha cpu_data[NR_CPUS];
56 /* A collection of single bit ipi messages. */
58 unsigned long bits ____cacheline_aligned;
59 } ipi_data[NR_CPUS] __cacheline_aligned;
61 enum ipi_message_type {
67 /* Set to a secondary's cpuid when it comes online. */
68 static int smp_secondary_alive __initdata = 0;
70 /* Which cpus ids came online. */
71 cpumask_t cpu_present_mask;
72 cpumask_t cpu_online_map;
74 EXPORT_SYMBOL(cpu_online_map);
76 int smp_num_probed; /* Internal processor count */
77 int smp_num_cpus = 1; /* Number that came online. */
79 extern void calibrate_delay(void);
84 * Called by both boot and secondaries to move global data into
85 * per-processor storage.
87 static inline void __init
88 smp_store_cpu_info(int cpuid)
90 cpu_data[cpuid].loops_per_jiffy = loops_per_jiffy;
91 cpu_data[cpuid].last_asn = ASN_FIRST_VERSION;
92 cpu_data[cpuid].need_new_asn = 0;
93 cpu_data[cpuid].asn_lock = 0;
97 * Ideally sets up per-cpu profiling hooks. Doesn't do much now...
99 static inline void __init
100 smp_setup_percpu_timer(int cpuid)
102 cpu_data[cpuid].prof_counter = 1;
103 cpu_data[cpuid].prof_multiplier = 1;
107 wait_boot_cpu_to_stop(int cpuid)
109 unsigned long stop = jiffies + 10*HZ;
111 while (time_before(jiffies, stop)) {
112 if (!smp_secondary_alive)
117 printk("wait_boot_cpu_to_stop: FAILED on CPU %d, hanging now\n", cpuid);
123 * Where secondaries begin a life of C.
128 int cpuid = hard_smp_processor_id();
130 if (cpu_test_and_set(cpuid, cpu_online_map)) {
131 printk("??, cpu 0x%x already present??\n", cpuid);
135 /* Turn on machine checks. */
138 /* Set trap vectors. */
141 /* Set interrupt vector. */
144 /* Get our local ticker going. */
145 smp_setup_percpu_timer(cpuid);
147 /* Call platform-specific callin, if specified */
148 if (alpha_mv.smp_callin) alpha_mv.smp_callin();
150 /* All kernel threads share the same mm context. */
151 atomic_inc(&init_mm.mm_count);
152 current->active_mm = &init_mm;
154 /* Must have completely accurate bogos. */
157 /* Wait boot CPU to stop with irq enabled before running
159 wait_boot_cpu_to_stop(cpuid);
163 smp_store_cpu_info(cpuid);
164 /* Allow master to continue only after we written loops_per_jiffy. */
166 smp_secondary_alive = 1;
168 DBGS(("smp_callin: commencing CPU %d current %p active_mm %p\n",
169 cpuid, current, current->active_mm));
175 /* Wait until hwrpb->txrdy is clear for cpu. Return -1 on timeout. */
177 wait_for_txrdy (unsigned long cpumask)
179 unsigned long timeout;
181 if (!(hwrpb->txrdy & cpumask))
184 timeout = jiffies + 10*HZ;
185 while (time_before(jiffies, timeout)) {
186 if (!(hwrpb->txrdy & cpumask))
196 * Send a message to a secondary's console. "START" is one such
197 * interesting message. ;-)
200 send_secondary_console_msg(char *str, int cpuid)
202 struct percpu_struct *cpu;
203 register char *cp1, *cp2;
204 unsigned long cpumask;
207 cpu = (struct percpu_struct *)
209 + hwrpb->processor_offset
210 + cpuid * hwrpb->processor_size);
212 cpumask = (1UL << cpuid);
213 if (wait_for_txrdy(cpumask))
218 *(unsigned int *)&cpu->ipc_buffer[0] = len;
219 cp1 = (char *) &cpu->ipc_buffer[1];
220 memcpy(cp1, cp2, len);
222 /* atomic test and set */
224 set_bit(cpuid, &hwrpb->rxrdy);
226 if (wait_for_txrdy(cpumask))
231 printk("Processor %x not ready\n", cpuid);
235 * A secondary console wants to send a message. Receive it.
238 recv_secondary_console_msg(void)
241 unsigned long txrdy = hwrpb->txrdy;
242 char *cp1, *cp2, buf[80];
243 struct percpu_struct *cpu;
245 DBGS(("recv_secondary_console_msg: TXRDY 0x%lx.\n", txrdy));
247 mycpu = hard_smp_processor_id();
249 for (i = 0; i < NR_CPUS; i++) {
250 if (!(txrdy & (1UL << i)))
253 DBGS(("recv_secondary_console_msg: "
254 "TXRDY contains CPU %d.\n", i));
256 cpu = (struct percpu_struct *)
258 + hwrpb->processor_offset
259 + i * hwrpb->processor_size);
261 DBGS(("recv_secondary_console_msg: on %d from %d"
262 " HALT_REASON 0x%lx FLAGS 0x%lx\n",
263 mycpu, i, cpu->halt_reason, cpu->flags));
265 cnt = cpu->ipc_buffer[0] >> 32;
266 if (cnt <= 0 || cnt >= 80)
267 strcpy(buf, "<<< BOGUS MSG >>>");
269 cp1 = (char *) &cpu->ipc_buffer[11];
273 while ((cp2 = strchr(cp2, '\r')) != 0) {
280 DBGS((KERN_INFO "recv_secondary_console_msg: on %d "
281 "message is '%s'\n", mycpu, buf));
288 * Convince the console to have a secondary cpu begin execution.
291 secondary_cpu_start(int cpuid, struct task_struct *idle)
293 struct percpu_struct *cpu;
294 struct pcb_struct *hwpcb, *ipcb;
295 unsigned long timeout;
297 cpu = (struct percpu_struct *)
299 + hwrpb->processor_offset
300 + cpuid * hwrpb->processor_size);
301 hwpcb = (struct pcb_struct *) cpu->hwpcb;
302 ipcb = &task_thread_info(idle)->pcb;
304 /* Initialize the CPU's HWPCB to something just good enough for
305 us to get started. Immediately after starting, we'll swpctx
306 to the target idle task's pcb. Reuse the stack in the mean
307 time. Precalculate the target PCBB. */
308 hwpcb->ksp = (unsigned long)ipcb + sizeof(union thread_union) - 16;
310 hwpcb->ptbr = ipcb->ptbr;
313 hwpcb->unique = virt_to_phys(ipcb);
314 hwpcb->flags = ipcb->flags;
315 hwpcb->res1 = hwpcb->res2 = 0;
318 DBGS(("KSP 0x%lx PTBR 0x%lx VPTBR 0x%lx UNIQUE 0x%lx\n",
319 hwpcb->ksp, hwpcb->ptbr, hwrpb->vptb, hwpcb->unique));
321 DBGS(("Starting secondary cpu %d: state 0x%lx pal_flags 0x%lx\n",
322 cpuid, idle->state, ipcb->flags));
324 /* Setup HWRPB fields that SRM uses to activate secondary CPU */
325 hwrpb->CPU_restart = __smp_callin;
326 hwrpb->CPU_restart_data = (unsigned long) __smp_callin;
328 /* Recalculate and update the HWRPB checksum */
329 hwrpb_update_checksum(hwrpb);
332 * Send a "start" command to the specified processor.
335 /* SRM III 3.4.1.3 */
336 cpu->flags |= 0x22; /* turn on Context Valid and Restart Capable */
337 cpu->flags &= ~1; /* turn off Bootstrap In Progress */
340 send_secondary_console_msg("START\r\n", cpuid);
342 /* Wait 10 seconds for an ACK from the console. */
343 timeout = jiffies + 10*HZ;
344 while (time_before(jiffies, timeout)) {
350 printk(KERN_ERR "SMP: Processor %d failed to start.\n", cpuid);
354 DBGS(("secondary_cpu_start: SUCCESS for CPU %d!!!\n", cpuid));
359 * Bring one cpu online.
362 smp_boot_one_cpu(int cpuid)
364 struct task_struct *idle;
365 unsigned long timeout;
367 /* Cook up an idler for this guy. Note that the address we
368 give to kernel_thread is irrelevant -- it's going to start
369 where HWRPB.CPU_restart says to start. But this gets all
370 the other task-y sort of data structures set up like we
371 wish. We can't use kernel_thread since we must avoid
372 rescheduling the child. */
373 idle = fork_idle(cpuid);
375 panic("failed fork for CPU %d", cpuid);
377 DBGS(("smp_boot_one_cpu: CPU %d state 0x%lx flags 0x%lx\n",
378 cpuid, idle->state, idle->flags));
380 /* Signal the secondary to wait a moment. */
381 smp_secondary_alive = -1;
383 /* Whirrr, whirrr, whirrrrrrrrr... */
384 if (secondary_cpu_start(cpuid, idle))
387 /* Notify the secondary CPU it can run calibrate_delay. */
389 smp_secondary_alive = 0;
391 /* We've been acked by the console; wait one second for
392 the task to start up for real. */
393 timeout = jiffies + 1*HZ;
394 while (time_before(jiffies, timeout)) {
395 if (smp_secondary_alive == 1)
401 /* We failed to boot the CPU. */
403 printk(KERN_ERR "SMP: Processor %d is stuck.\n", cpuid);
407 /* Another "Red Snapper". */
412 * Called from setup_arch. Detect an SMP system and which processors
418 struct percpu_struct *cpubase, *cpu;
421 if (boot_cpuid != 0) {
422 printk(KERN_WARNING "SMP: Booting off cpu %d instead of 0?\n",
426 if (hwrpb->nr_processors > 1) {
429 DBGS(("setup_smp: nr_processors %ld\n",
430 hwrpb->nr_processors));
432 cpubase = (struct percpu_struct *)
433 ((char*)hwrpb + hwrpb->processor_offset);
434 boot_cpu_palrev = cpubase->pal_revision;
436 for (i = 0; i < hwrpb->nr_processors; i++) {
437 cpu = (struct percpu_struct *)
438 ((char *)cpubase + i*hwrpb->processor_size);
439 if ((cpu->flags & 0x1cc) == 0x1cc) {
441 /* Assume here that "whami" == index */
442 cpu_set(i, cpu_present_mask);
443 cpu->pal_revision = boot_cpu_palrev;
446 DBGS(("setup_smp: CPU %d: flags 0x%lx type 0x%lx\n",
447 i, cpu->flags, cpu->type));
448 DBGS(("setup_smp: CPU %d: PAL rev 0x%lx\n",
449 i, cpu->pal_revision));
453 cpu_set(boot_cpuid, cpu_present_mask);
456 printk(KERN_INFO "SMP: %d CPUs probed -- cpu_present_mask = %lx\n",
457 smp_num_probed, cpu_possible_map.bits[0]);
461 * Called by smp_init prepare the secondaries
464 smp_prepare_cpus(unsigned int max_cpus)
466 /* Take care of some initial bookkeeping. */
467 memset(ipi_data, 0, sizeof(ipi_data));
469 current_thread_info()->cpu = boot_cpuid;
471 smp_store_cpu_info(boot_cpuid);
472 smp_setup_percpu_timer(boot_cpuid);
474 /* Nothing to do on a UP box, or when told not to. */
475 if (smp_num_probed == 1 || max_cpus == 0) {
476 cpu_present_mask = cpumask_of_cpu(boot_cpuid);
477 printk(KERN_INFO "SMP mode deactivated.\n");
481 printk(KERN_INFO "SMP starting up secondaries.\n");
483 smp_num_cpus = smp_num_probed;
487 smp_prepare_boot_cpu(void)
490 * Mark the boot cpu (current cpu) as online
492 cpu_set(smp_processor_id(), cpu_online_map);
496 __cpu_up(unsigned int cpu)
498 smp_boot_one_cpu(cpu);
500 return cpu_online(cpu) ? 0 : -ENOSYS;
504 smp_cpus_done(unsigned int max_cpus)
507 unsigned long bogosum = 0;
509 for(cpu = 0; cpu < NR_CPUS; cpu++)
511 bogosum += cpu_data[cpu].loops_per_jiffy;
513 printk(KERN_INFO "SMP: Total of %d processors activated "
514 "(%lu.%02lu BogoMIPS).\n",
516 (bogosum + 2500) / (500000/HZ),
517 ((bogosum + 2500) / (5000/HZ)) % 100);
522 smp_percpu_timer_interrupt(struct pt_regs *regs)
524 int cpu = smp_processor_id();
525 unsigned long user = user_mode(regs);
526 struct cpuinfo_alpha *data = &cpu_data[cpu];
528 /* Record kernel PC. */
529 profile_tick(CPU_PROFILING, regs);
531 if (!--data->prof_counter) {
532 /* We need to make like a normal interrupt -- otherwise
533 timer interrupts ignore the global interrupt lock,
534 which would be a Bad Thing. */
537 update_process_times(user);
539 data->prof_counter = data->prof_multiplier;
546 setup_profiling_timer(unsigned int multiplier)
553 send_ipi_message(cpumask_t to_whom, enum ipi_message_type operation)
558 for_each_cpu_mask(i, to_whom)
559 set_bit(operation, &ipi_data[i].bits);
562 for_each_cpu_mask(i, to_whom)
566 /* Structure and data for smp_call_function. This is designed to
567 minimize static memory requirements. Plus it looks cleaner. */
569 struct smp_call_struct {
570 void (*func) (void *info);
573 atomic_t unstarted_count;
574 atomic_t unfinished_count;
577 static struct smp_call_struct *smp_call_function_data;
579 /* Atomicly drop data into a shared pointer. The pointer is free if
580 it is initially locked. If retry, spin until free. */
583 pointer_lock (void *lock, void *data, int retry)
589 /* Compare and swap with zero. */
597 : "=&r"(old), "=m"(*(void **)lock), "=&r"(tmp)
606 while (*(void **)lock)
612 handle_ipi(struct pt_regs *regs)
614 int this_cpu = smp_processor_id();
615 unsigned long *pending_ipis = &ipi_data[this_cpu].bits;
619 DBGS(("handle_ipi: on CPU %d ops 0x%lx PC 0x%lx\n",
620 this_cpu, *pending_ipis, regs->pc));
623 mb(); /* Order interrupt and bit testing. */
624 while ((ops = xchg(pending_ipis, 0)) != 0) {
625 mb(); /* Order bit clearing and data access. */
631 which = __ffs(which);
635 /* Reschedule callback. Everything to be done
636 is done by the interrupt return path. */
641 struct smp_call_struct *data;
642 void (*func)(void *info);
646 data = smp_call_function_data;
651 /* Notify the sending CPU that the data has been
652 received, and execution is about to begin. */
654 atomic_dec (&data->unstarted_count);
656 /* At this point the structure may be gone unless
660 /* Notify the sending CPU that the task is done. */
662 if (wait) atomic_dec (&data->unfinished_count);
670 printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n",
676 mb(); /* Order data access and bit testing. */
679 cpu_data[this_cpu].ipi_count++;
682 recv_secondary_console_msg();
686 smp_send_reschedule(int cpu)
689 if (cpu == hard_smp_processor_id())
691 "smp_send_reschedule: Sending IPI to self.\n");
693 send_ipi_message(cpumask_of_cpu(cpu), IPI_RESCHEDULE);
699 cpumask_t to_whom = cpu_possible_map;
700 cpu_clear(smp_processor_id(), to_whom);
702 if (hard_smp_processor_id() != boot_cpu_id)
703 printk(KERN_WARNING "smp_send_stop: Not on boot cpu.\n");
705 send_ipi_message(to_whom, IPI_CPU_STOP);
709 * Run a function on all other CPUs.
710 * <func> The function to run. This must be fast and non-blocking.
711 * <info> An arbitrary pointer to pass to the function.
712 * <retry> If true, keep retrying until ready.
713 * <wait> If true, wait until function has completed on other CPUs.
714 * [RETURNS] 0 on success, else a negative status code.
716 * Does not return until remote CPUs are nearly ready to execute <func>
717 * or are or have executed.
718 * You must not call this function with disabled interrupts or from a
719 * hardware interrupt handler or from a bottom half handler.
723 smp_call_function_on_cpu (void (*func) (void *info), void *info, int retry,
724 int wait, cpumask_t to_whom)
726 struct smp_call_struct data;
727 unsigned long timeout;
728 int num_cpus_to_call;
730 /* Can deadlock when called with interrupts disabled */
731 WARN_ON(irqs_disabled());
737 cpu_clear(smp_processor_id(), to_whom);
738 num_cpus_to_call = cpus_weight(to_whom);
740 atomic_set(&data.unstarted_count, num_cpus_to_call);
741 atomic_set(&data.unfinished_count, num_cpus_to_call);
743 /* Acquire the smp_call_function_data mutex. */
744 if (pointer_lock(&smp_call_function_data, &data, retry))
747 /* Send a message to the requested CPUs. */
748 send_ipi_message(to_whom, IPI_CALL_FUNC);
750 /* Wait for a minimal response. */
751 timeout = jiffies + HZ;
752 while (atomic_read (&data.unstarted_count) > 0
753 && time_before (jiffies, timeout))
756 /* If there's no response yet, log a message but allow a longer
757 * timeout period -- if we get a response this time, log
758 * a message saying when we got it..
760 if (atomic_read(&data.unstarted_count) > 0) {
761 long start_time = jiffies;
762 printk(KERN_ERR "%s: initial timeout -- trying long wait\n",
764 timeout = jiffies + 30 * HZ;
765 while (atomic_read(&data.unstarted_count) > 0
766 && time_before(jiffies, timeout))
768 if (atomic_read(&data.unstarted_count) <= 0) {
769 long delta = jiffies - start_time;
771 "%s: response %ld.%ld seconds into long wait\n",
772 __FUNCTION__, delta / HZ,
773 (100 * (delta - ((delta / HZ) * HZ))) / HZ);
777 /* We either got one or timed out -- clear the lock. */
779 smp_call_function_data = NULL;
782 * If after both the initial and long timeout periods we still don't
783 * have a response, something is very wrong...
785 BUG_ON(atomic_read (&data.unstarted_count) > 0);
787 /* Wait for a complete response, if needed. */
789 while (atomic_read (&data.unfinished_count) > 0)
797 smp_call_function (void (*func) (void *info), void *info, int retry, int wait)
799 return smp_call_function_on_cpu (func, info, retry, wait,
804 ipi_imb(void *ignored)
812 /* Must wait other processors to flush their icache before continue. */
813 if (on_each_cpu(ipi_imb, NULL, 1, 1))
814 printk(KERN_CRIT "smp_imb: timed out\n");
818 ipi_flush_tlb_all(void *ignored)
826 /* Although we don't have any data to pass, we do want to
827 synchronize with the other processors. */
828 if (on_each_cpu(ipi_flush_tlb_all, NULL, 1, 1)) {
829 printk(KERN_CRIT "flush_tlb_all: timed out\n");
833 #define asn_locked() (cpu_data[smp_processor_id()].asn_lock)
836 ipi_flush_tlb_mm(void *x)
838 struct mm_struct *mm = (struct mm_struct *) x;
839 if (mm == current->active_mm && !asn_locked())
840 flush_tlb_current(mm);
846 flush_tlb_mm(struct mm_struct *mm)
850 if (mm == current->active_mm) {
851 flush_tlb_current(mm);
852 if (atomic_read(&mm->mm_users) <= 1) {
853 int cpu, this_cpu = smp_processor_id();
854 for (cpu = 0; cpu < NR_CPUS; cpu++) {
855 if (!cpu_online(cpu) || cpu == this_cpu)
857 if (mm->context[cpu])
858 mm->context[cpu] = 0;
865 if (smp_call_function(ipi_flush_tlb_mm, mm, 1, 1)) {
866 printk(KERN_CRIT "flush_tlb_mm: timed out\n");
872 struct flush_tlb_page_struct {
873 struct vm_area_struct *vma;
874 struct mm_struct *mm;
879 ipi_flush_tlb_page(void *x)
881 struct flush_tlb_page_struct *data = (struct flush_tlb_page_struct *)x;
882 struct mm_struct * mm = data->mm;
884 if (mm == current->active_mm && !asn_locked())
885 flush_tlb_current_page(mm, data->vma, data->addr);
891 flush_tlb_page(struct vm_area_struct *vma, unsigned long addr)
893 struct flush_tlb_page_struct data;
894 struct mm_struct *mm = vma->vm_mm;
898 if (mm == current->active_mm) {
899 flush_tlb_current_page(mm, vma, addr);
900 if (atomic_read(&mm->mm_users) <= 1) {
901 int cpu, this_cpu = smp_processor_id();
902 for (cpu = 0; cpu < NR_CPUS; cpu++) {
903 if (!cpu_online(cpu) || cpu == this_cpu)
905 if (mm->context[cpu])
906 mm->context[cpu] = 0;
917 if (smp_call_function(ipi_flush_tlb_page, &data, 1, 1)) {
918 printk(KERN_CRIT "flush_tlb_page: timed out\n");
925 flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
927 /* On the Alpha we always flush the whole user tlb. */
928 flush_tlb_mm(vma->vm_mm);
932 ipi_flush_icache_page(void *x)
934 struct mm_struct *mm = (struct mm_struct *) x;
935 if (mm == current->active_mm && !asn_locked())
936 __load_new_mm_context(mm);
942 flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
943 unsigned long addr, int len)
945 struct mm_struct *mm = vma->vm_mm;
947 if ((vma->vm_flags & VM_EXEC) == 0)
952 if (mm == current->active_mm) {
953 __load_new_mm_context(mm);
954 if (atomic_read(&mm->mm_users) <= 1) {
955 int cpu, this_cpu = smp_processor_id();
956 for (cpu = 0; cpu < NR_CPUS; cpu++) {
957 if (!cpu_online(cpu) || cpu == this_cpu)
959 if (mm->context[cpu])
960 mm->context[cpu] = 0;
967 if (smp_call_function(ipi_flush_icache_page, mm, 1, 1)) {
968 printk(KERN_CRIT "flush_icache_page: timed out\n");