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_possible_map);
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_possible_map);
455 cpu_present_mask = cpumask_of_cpu(boot_cpuid);
457 printk(KERN_INFO "SMP: %d CPUs probed -- cpu_present_mask = %lx\n",
458 smp_num_probed, cpu_possible_map.bits[0]);
462 * Called by smp_init prepare the secondaries
465 smp_prepare_cpus(unsigned int max_cpus)
467 /* Take care of some initial bookkeeping. */
468 memset(ipi_data, 0, sizeof(ipi_data));
470 current_thread_info()->cpu = boot_cpuid;
472 smp_store_cpu_info(boot_cpuid);
473 smp_setup_percpu_timer(boot_cpuid);
475 /* Nothing to do on a UP box, or when told not to. */
476 if (smp_num_probed == 1 || max_cpus == 0) {
477 cpu_present_mask = cpumask_of_cpu(boot_cpuid);
478 printk(KERN_INFO "SMP mode deactivated.\n");
482 printk(KERN_INFO "SMP starting up secondaries.\n");
484 smp_num_cpus = smp_num_probed;
488 smp_prepare_boot_cpu(void)
491 * Mark the boot cpu (current cpu) as both present and online
493 cpu_set(smp_processor_id(), cpu_present_mask);
494 cpu_set(smp_processor_id(), cpu_online_map);
498 __cpu_up(unsigned int cpu)
500 smp_boot_one_cpu(cpu);
502 return cpu_online(cpu) ? 0 : -ENOSYS;
506 smp_cpus_done(unsigned int max_cpus)
509 unsigned long bogosum = 0;
511 for(cpu = 0; cpu < NR_CPUS; cpu++)
513 bogosum += cpu_data[cpu].loops_per_jiffy;
515 printk(KERN_INFO "SMP: Total of %d processors activated "
516 "(%lu.%02lu BogoMIPS).\n",
518 (bogosum + 2500) / (500000/HZ),
519 ((bogosum + 2500) / (5000/HZ)) % 100);
524 smp_percpu_timer_interrupt(struct pt_regs *regs)
526 int cpu = smp_processor_id();
527 unsigned long user = user_mode(regs);
528 struct cpuinfo_alpha *data = &cpu_data[cpu];
530 /* Record kernel PC. */
531 profile_tick(CPU_PROFILING, regs);
533 if (!--data->prof_counter) {
534 /* We need to make like a normal interrupt -- otherwise
535 timer interrupts ignore the global interrupt lock,
536 which would be a Bad Thing. */
539 update_process_times(user);
541 data->prof_counter = data->prof_multiplier;
548 setup_profiling_timer(unsigned int multiplier)
555 send_ipi_message(cpumask_t to_whom, enum ipi_message_type operation)
560 for_each_cpu_mask(i, to_whom)
561 set_bit(operation, &ipi_data[i].bits);
564 for_each_cpu_mask(i, to_whom)
568 /* Structure and data for smp_call_function. This is designed to
569 minimize static memory requirements. Plus it looks cleaner. */
571 struct smp_call_struct {
572 void (*func) (void *info);
575 atomic_t unstarted_count;
576 atomic_t unfinished_count;
579 static struct smp_call_struct *smp_call_function_data;
581 /* Atomicly drop data into a shared pointer. The pointer is free if
582 it is initially locked. If retry, spin until free. */
585 pointer_lock (void *lock, void *data, int retry)
591 /* Compare and swap with zero. */
599 : "=&r"(old), "=m"(*(void **)lock), "=&r"(tmp)
608 while (*(void **)lock)
614 handle_ipi(struct pt_regs *regs)
616 int this_cpu = smp_processor_id();
617 unsigned long *pending_ipis = &ipi_data[this_cpu].bits;
621 DBGS(("handle_ipi: on CPU %d ops 0x%lx PC 0x%lx\n",
622 this_cpu, *pending_ipis, regs->pc));
625 mb(); /* Order interrupt and bit testing. */
626 while ((ops = xchg(pending_ipis, 0)) != 0) {
627 mb(); /* Order bit clearing and data access. */
633 which = __ffs(which);
637 /* Reschedule callback. Everything to be done
638 is done by the interrupt return path. */
643 struct smp_call_struct *data;
644 void (*func)(void *info);
648 data = smp_call_function_data;
653 /* Notify the sending CPU that the data has been
654 received, and execution is about to begin. */
656 atomic_dec (&data->unstarted_count);
658 /* At this point the structure may be gone unless
662 /* Notify the sending CPU that the task is done. */
664 if (wait) atomic_dec (&data->unfinished_count);
672 printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n",
678 mb(); /* Order data access and bit testing. */
681 cpu_data[this_cpu].ipi_count++;
684 recv_secondary_console_msg();
688 smp_send_reschedule(int cpu)
691 if (cpu == hard_smp_processor_id())
693 "smp_send_reschedule: Sending IPI to self.\n");
695 send_ipi_message(cpumask_of_cpu(cpu), IPI_RESCHEDULE);
701 cpumask_t to_whom = cpu_possible_map;
702 cpu_clear(smp_processor_id(), to_whom);
704 if (hard_smp_processor_id() != boot_cpu_id)
705 printk(KERN_WARNING "smp_send_stop: Not on boot cpu.\n");
707 send_ipi_message(to_whom, IPI_CPU_STOP);
711 * Run a function on all other CPUs.
712 * <func> The function to run. This must be fast and non-blocking.
713 * <info> An arbitrary pointer to pass to the function.
714 * <retry> If true, keep retrying until ready.
715 * <wait> If true, wait until function has completed on other CPUs.
716 * [RETURNS] 0 on success, else a negative status code.
718 * Does not return until remote CPUs are nearly ready to execute <func>
719 * or are or have executed.
720 * You must not call this function with disabled interrupts or from a
721 * hardware interrupt handler or from a bottom half handler.
725 smp_call_function_on_cpu (void (*func) (void *info), void *info, int retry,
726 int wait, cpumask_t to_whom)
728 struct smp_call_struct data;
729 unsigned long timeout;
730 int num_cpus_to_call;
732 /* Can deadlock when called with interrupts disabled */
733 WARN_ON(irqs_disabled());
739 cpu_clear(smp_processor_id(), to_whom);
740 num_cpus_to_call = cpus_weight(to_whom);
742 atomic_set(&data.unstarted_count, num_cpus_to_call);
743 atomic_set(&data.unfinished_count, num_cpus_to_call);
745 /* Acquire the smp_call_function_data mutex. */
746 if (pointer_lock(&smp_call_function_data, &data, retry))
749 /* Send a message to the requested CPUs. */
750 send_ipi_message(to_whom, IPI_CALL_FUNC);
752 /* Wait for a minimal response. */
753 timeout = jiffies + HZ;
754 while (atomic_read (&data.unstarted_count) > 0
755 && time_before (jiffies, timeout))
758 /* If there's no response yet, log a message but allow a longer
759 * timeout period -- if we get a response this time, log
760 * a message saying when we got it..
762 if (atomic_read(&data.unstarted_count) > 0) {
763 long start_time = jiffies;
764 printk(KERN_ERR "%s: initial timeout -- trying long wait\n",
766 timeout = jiffies + 30 * HZ;
767 while (atomic_read(&data.unstarted_count) > 0
768 && time_before(jiffies, timeout))
770 if (atomic_read(&data.unstarted_count) <= 0) {
771 long delta = jiffies - start_time;
773 "%s: response %ld.%ld seconds into long wait\n",
774 __FUNCTION__, delta / HZ,
775 (100 * (delta - ((delta / HZ) * HZ))) / HZ);
779 /* We either got one or timed out -- clear the lock. */
781 smp_call_function_data = NULL;
784 * If after both the initial and long timeout periods we still don't
785 * have a response, something is very wrong...
787 BUG_ON(atomic_read (&data.unstarted_count) > 0);
789 /* Wait for a complete response, if needed. */
791 while (atomic_read (&data.unfinished_count) > 0)
799 smp_call_function (void (*func) (void *info), void *info, int retry, int wait)
801 return smp_call_function_on_cpu (func, info, retry, wait,
806 ipi_imb(void *ignored)
814 /* Must wait other processors to flush their icache before continue. */
815 if (on_each_cpu(ipi_imb, NULL, 1, 1))
816 printk(KERN_CRIT "smp_imb: timed out\n");
820 ipi_flush_tlb_all(void *ignored)
828 /* Although we don't have any data to pass, we do want to
829 synchronize with the other processors. */
830 if (on_each_cpu(ipi_flush_tlb_all, NULL, 1, 1)) {
831 printk(KERN_CRIT "flush_tlb_all: timed out\n");
835 #define asn_locked() (cpu_data[smp_processor_id()].asn_lock)
838 ipi_flush_tlb_mm(void *x)
840 struct mm_struct *mm = (struct mm_struct *) x;
841 if (mm == current->active_mm && !asn_locked())
842 flush_tlb_current(mm);
848 flush_tlb_mm(struct mm_struct *mm)
852 if (mm == current->active_mm) {
853 flush_tlb_current(mm);
854 if (atomic_read(&mm->mm_users) <= 1) {
855 int cpu, this_cpu = smp_processor_id();
856 for (cpu = 0; cpu < NR_CPUS; cpu++) {
857 if (!cpu_online(cpu) || cpu == this_cpu)
859 if (mm->context[cpu])
860 mm->context[cpu] = 0;
867 if (smp_call_function(ipi_flush_tlb_mm, mm, 1, 1)) {
868 printk(KERN_CRIT "flush_tlb_mm: timed out\n");
874 struct flush_tlb_page_struct {
875 struct vm_area_struct *vma;
876 struct mm_struct *mm;
881 ipi_flush_tlb_page(void *x)
883 struct flush_tlb_page_struct *data = (struct flush_tlb_page_struct *)x;
884 struct mm_struct * mm = data->mm;
886 if (mm == current->active_mm && !asn_locked())
887 flush_tlb_current_page(mm, data->vma, data->addr);
893 flush_tlb_page(struct vm_area_struct *vma, unsigned long addr)
895 struct flush_tlb_page_struct data;
896 struct mm_struct *mm = vma->vm_mm;
900 if (mm == current->active_mm) {
901 flush_tlb_current_page(mm, vma, addr);
902 if (atomic_read(&mm->mm_users) <= 1) {
903 int cpu, this_cpu = smp_processor_id();
904 for (cpu = 0; cpu < NR_CPUS; cpu++) {
905 if (!cpu_online(cpu) || cpu == this_cpu)
907 if (mm->context[cpu])
908 mm->context[cpu] = 0;
919 if (smp_call_function(ipi_flush_tlb_page, &data, 1, 1)) {
920 printk(KERN_CRIT "flush_tlb_page: timed out\n");
927 flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
929 /* On the Alpha we always flush the whole user tlb. */
930 flush_tlb_mm(vma->vm_mm);
934 ipi_flush_icache_page(void *x)
936 struct mm_struct *mm = (struct mm_struct *) x;
937 if (mm == current->active_mm && !asn_locked())
938 __load_new_mm_context(mm);
944 flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
945 unsigned long addr, int len)
947 struct mm_struct *mm = vma->vm_mm;
949 if ((vma->vm_flags & VM_EXEC) == 0)
954 if (mm == current->active_mm) {
955 __load_new_mm_context(mm);
956 if (atomic_read(&mm->mm_users) <= 1) {
957 int cpu, this_cpu = smp_processor_id();
958 for (cpu = 0; cpu < NR_CPUS; cpu++) {
959 if (!cpu_online(cpu) || cpu == this_cpu)
961 if (mm->context[cpu])
962 mm->context[cpu] = 0;
969 if (smp_call_function(ipi_flush_icache_page, mm, 1, 1)) {
970 printk(KERN_CRIT "flush_icache_page: timed out\n");