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];
55 EXPORT_SYMBOL(cpu_data);
57 /* A collection of single bit ipi messages. */
59 unsigned long bits ____cacheline_aligned;
60 } ipi_data[NR_CPUS] __cacheline_aligned;
62 enum ipi_message_type {
68 /* Set to a secondary's cpuid when it comes online. */
69 static int smp_secondary_alive __initdata = 0;
71 /* Which cpus ids came online. */
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. */
78 EXPORT_SYMBOL(smp_num_cpus);
80 extern void calibrate_delay(void);
85 * Called by both boot and secondaries to move global data into
86 * per-processor storage.
88 static inline void __init
89 smp_store_cpu_info(int cpuid)
91 cpu_data[cpuid].loops_per_jiffy = loops_per_jiffy;
92 cpu_data[cpuid].last_asn = ASN_FIRST_VERSION;
93 cpu_data[cpuid].need_new_asn = 0;
94 cpu_data[cpuid].asn_lock = 0;
98 * Ideally sets up per-cpu profiling hooks. Doesn't do much now...
100 static inline void __init
101 smp_setup_percpu_timer(int cpuid)
103 cpu_data[cpuid].prof_counter = 1;
104 cpu_data[cpuid].prof_multiplier = 1;
108 wait_boot_cpu_to_stop(int cpuid)
110 unsigned long stop = jiffies + 10*HZ;
112 while (time_before(jiffies, stop)) {
113 if (!smp_secondary_alive)
118 printk("wait_boot_cpu_to_stop: FAILED on CPU %d, hanging now\n", cpuid);
124 * Where secondaries begin a life of C.
129 int cpuid = hard_smp_processor_id();
131 if (cpu_test_and_set(cpuid, cpu_online_map)) {
132 printk("??, cpu 0x%x already present??\n", cpuid);
136 /* Turn on machine checks. */
139 /* Set trap vectors. */
142 /* Set interrupt vector. */
145 /* Get our local ticker going. */
146 smp_setup_percpu_timer(cpuid);
148 /* Call platform-specific callin, if specified */
149 if (alpha_mv.smp_callin) alpha_mv.smp_callin();
151 /* All kernel threads share the same mm context. */
152 atomic_inc(&init_mm.mm_count);
153 current->active_mm = &init_mm;
155 /* Must have completely accurate bogos. */
158 /* Wait boot CPU to stop with irq enabled before running
160 wait_boot_cpu_to_stop(cpuid);
164 smp_store_cpu_info(cpuid);
165 /* Allow master to continue only after we written loops_per_jiffy. */
167 smp_secondary_alive = 1;
169 DBGS(("smp_callin: commencing CPU %d current %p active_mm %p\n",
170 cpuid, current, current->active_mm));
176 /* Wait until hwrpb->txrdy is clear for cpu. Return -1 on timeout. */
178 wait_for_txrdy (unsigned long cpumask)
180 unsigned long timeout;
182 if (!(hwrpb->txrdy & cpumask))
185 timeout = jiffies + 10*HZ;
186 while (time_before(jiffies, timeout)) {
187 if (!(hwrpb->txrdy & cpumask))
197 * Send a message to a secondary's console. "START" is one such
198 * interesting message. ;-)
201 send_secondary_console_msg(char *str, int cpuid)
203 struct percpu_struct *cpu;
204 register char *cp1, *cp2;
205 unsigned long cpumask;
208 cpu = (struct percpu_struct *)
210 + hwrpb->processor_offset
211 + cpuid * hwrpb->processor_size);
213 cpumask = (1UL << cpuid);
214 if (wait_for_txrdy(cpumask))
219 *(unsigned int *)&cpu->ipc_buffer[0] = len;
220 cp1 = (char *) &cpu->ipc_buffer[1];
221 memcpy(cp1, cp2, len);
223 /* atomic test and set */
225 set_bit(cpuid, &hwrpb->rxrdy);
227 if (wait_for_txrdy(cpumask))
232 printk("Processor %x not ready\n", cpuid);
236 * A secondary console wants to send a message. Receive it.
239 recv_secondary_console_msg(void)
242 unsigned long txrdy = hwrpb->txrdy;
243 char *cp1, *cp2, buf[80];
244 struct percpu_struct *cpu;
246 DBGS(("recv_secondary_console_msg: TXRDY 0x%lx.\n", txrdy));
248 mycpu = hard_smp_processor_id();
250 for (i = 0; i < NR_CPUS; i++) {
251 if (!(txrdy & (1UL << i)))
254 DBGS(("recv_secondary_console_msg: "
255 "TXRDY contains CPU %d.\n", i));
257 cpu = (struct percpu_struct *)
259 + hwrpb->processor_offset
260 + i * hwrpb->processor_size);
262 DBGS(("recv_secondary_console_msg: on %d from %d"
263 " HALT_REASON 0x%lx FLAGS 0x%lx\n",
264 mycpu, i, cpu->halt_reason, cpu->flags));
266 cnt = cpu->ipc_buffer[0] >> 32;
267 if (cnt <= 0 || cnt >= 80)
268 strcpy(buf, "<<< BOGUS MSG >>>");
270 cp1 = (char *) &cpu->ipc_buffer[11];
274 while ((cp2 = strchr(cp2, '\r')) != 0) {
281 DBGS((KERN_INFO "recv_secondary_console_msg: on %d "
282 "message is '%s'\n", mycpu, buf));
289 * Convince the console to have a secondary cpu begin execution.
292 secondary_cpu_start(int cpuid, struct task_struct *idle)
294 struct percpu_struct *cpu;
295 struct pcb_struct *hwpcb, *ipcb;
296 unsigned long timeout;
298 cpu = (struct percpu_struct *)
300 + hwrpb->processor_offset
301 + cpuid * hwrpb->processor_size);
302 hwpcb = (struct pcb_struct *) cpu->hwpcb;
303 ipcb = &task_thread_info(idle)->pcb;
305 /* Initialize the CPU's HWPCB to something just good enough for
306 us to get started. Immediately after starting, we'll swpctx
307 to the target idle task's pcb. Reuse the stack in the mean
308 time. Precalculate the target PCBB. */
309 hwpcb->ksp = (unsigned long)ipcb + sizeof(union thread_union) - 16;
311 hwpcb->ptbr = ipcb->ptbr;
314 hwpcb->unique = virt_to_phys(ipcb);
315 hwpcb->flags = ipcb->flags;
316 hwpcb->res1 = hwpcb->res2 = 0;
319 DBGS(("KSP 0x%lx PTBR 0x%lx VPTBR 0x%lx UNIQUE 0x%lx\n",
320 hwpcb->ksp, hwpcb->ptbr, hwrpb->vptb, hwpcb->unique));
322 DBGS(("Starting secondary cpu %d: state 0x%lx pal_flags 0x%lx\n",
323 cpuid, idle->state, ipcb->flags));
325 /* Setup HWRPB fields that SRM uses to activate secondary CPU */
326 hwrpb->CPU_restart = __smp_callin;
327 hwrpb->CPU_restart_data = (unsigned long) __smp_callin;
329 /* Recalculate and update the HWRPB checksum */
330 hwrpb_update_checksum(hwrpb);
333 * Send a "start" command to the specified processor.
336 /* SRM III 3.4.1.3 */
337 cpu->flags |= 0x22; /* turn on Context Valid and Restart Capable */
338 cpu->flags &= ~1; /* turn off Bootstrap In Progress */
341 send_secondary_console_msg("START\r\n", cpuid);
343 /* Wait 10 seconds for an ACK from the console. */
344 timeout = jiffies + 10*HZ;
345 while (time_before(jiffies, timeout)) {
351 printk(KERN_ERR "SMP: Processor %d failed to start.\n", cpuid);
355 DBGS(("secondary_cpu_start: SUCCESS for CPU %d!!!\n", cpuid));
360 * Bring one cpu online.
363 smp_boot_one_cpu(int cpuid)
365 struct task_struct *idle;
366 unsigned long timeout;
368 /* Cook up an idler for this guy. Note that the address we
369 give to kernel_thread is irrelevant -- it's going to start
370 where HWRPB.CPU_restart says to start. But this gets all
371 the other task-y sort of data structures set up like we
372 wish. We can't use kernel_thread since we must avoid
373 rescheduling the child. */
374 idle = fork_idle(cpuid);
376 panic("failed fork for CPU %d", cpuid);
378 DBGS(("smp_boot_one_cpu: CPU %d state 0x%lx flags 0x%lx\n",
379 cpuid, idle->state, idle->flags));
381 /* Signal the secondary to wait a moment. */
382 smp_secondary_alive = -1;
384 /* Whirrr, whirrr, whirrrrrrrrr... */
385 if (secondary_cpu_start(cpuid, idle))
388 /* Notify the secondary CPU it can run calibrate_delay. */
390 smp_secondary_alive = 0;
392 /* We've been acked by the console; wait one second for
393 the task to start up for real. */
394 timeout = jiffies + 1*HZ;
395 while (time_before(jiffies, timeout)) {
396 if (smp_secondary_alive == 1)
402 /* We failed to boot the CPU. */
404 printk(KERN_ERR "SMP: Processor %d is stuck.\n", cpuid);
408 /* Another "Red Snapper". */
413 * Called from setup_arch. Detect an SMP system and which processors
419 struct percpu_struct *cpubase, *cpu;
422 if (boot_cpuid != 0) {
423 printk(KERN_WARNING "SMP: Booting off cpu %d instead of 0?\n",
427 if (hwrpb->nr_processors > 1) {
430 DBGS(("setup_smp: nr_processors %ld\n",
431 hwrpb->nr_processors));
433 cpubase = (struct percpu_struct *)
434 ((char*)hwrpb + hwrpb->processor_offset);
435 boot_cpu_palrev = cpubase->pal_revision;
437 for (i = 0; i < hwrpb->nr_processors; i++) {
438 cpu = (struct percpu_struct *)
439 ((char *)cpubase + i*hwrpb->processor_size);
440 if ((cpu->flags & 0x1cc) == 0x1cc) {
442 /* Assume here that "whami" == index */
443 cpu_set(i, cpu_present_map);
444 cpu->pal_revision = boot_cpu_palrev;
447 DBGS(("setup_smp: CPU %d: flags 0x%lx type 0x%lx\n",
448 i, cpu->flags, cpu->type));
449 DBGS(("setup_smp: CPU %d: PAL rev 0x%lx\n",
450 i, cpu->pal_revision));
456 printk(KERN_INFO "SMP: %d CPUs probed -- cpu_present_map = %lx\n",
457 smp_num_probed, cpu_present_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_map = 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)
492 __cpu_up(unsigned int cpu)
494 smp_boot_one_cpu(cpu);
496 return cpu_online(cpu) ? 0 : -ENOSYS;
500 smp_cpus_done(unsigned int max_cpus)
503 unsigned long bogosum = 0;
505 for(cpu = 0; cpu < NR_CPUS; cpu++)
507 bogosum += cpu_data[cpu].loops_per_jiffy;
509 printk(KERN_INFO "SMP: Total of %d processors activated "
510 "(%lu.%02lu BogoMIPS).\n",
512 (bogosum + 2500) / (500000/HZ),
513 ((bogosum + 2500) / (5000/HZ)) % 100);
518 smp_percpu_timer_interrupt(struct pt_regs *regs)
520 struct pt_regs *old_regs;
521 int cpu = smp_processor_id();
522 unsigned long user = user_mode(regs);
523 struct cpuinfo_alpha *data = &cpu_data[cpu];
525 old_regs = set_irq_regs(regs);
527 /* Record kernel PC. */
528 profile_tick(CPU_PROFILING);
530 if (!--data->prof_counter) {
531 /* We need to make like a normal interrupt -- otherwise
532 timer interrupts ignore the global interrupt lock,
533 which would be a Bad Thing. */
536 update_process_times(user);
538 data->prof_counter = data->prof_multiplier;
542 set_irq_regs(old_regs);
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)
795 EXPORT_SYMBOL(smp_call_function_on_cpu);
798 smp_call_function (void (*func) (void *info), void *info, int retry, int wait)
800 return smp_call_function_on_cpu (func, info, retry, wait,
803 EXPORT_SYMBOL(smp_call_function);
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");
818 EXPORT_SYMBOL(smp_imb);
821 ipi_flush_tlb_all(void *ignored)
829 /* Although we don't have any data to pass, we do want to
830 synchronize with the other processors. */
831 if (on_each_cpu(ipi_flush_tlb_all, NULL, 1, 1)) {
832 printk(KERN_CRIT "flush_tlb_all: timed out\n");
836 #define asn_locked() (cpu_data[smp_processor_id()].asn_lock)
839 ipi_flush_tlb_mm(void *x)
841 struct mm_struct *mm = (struct mm_struct *) x;
842 if (mm == current->active_mm && !asn_locked())
843 flush_tlb_current(mm);
849 flush_tlb_mm(struct mm_struct *mm)
853 if (mm == current->active_mm) {
854 flush_tlb_current(mm);
855 if (atomic_read(&mm->mm_users) <= 1) {
856 int cpu, this_cpu = smp_processor_id();
857 for (cpu = 0; cpu < NR_CPUS; cpu++) {
858 if (!cpu_online(cpu) || cpu == this_cpu)
860 if (mm->context[cpu])
861 mm->context[cpu] = 0;
868 if (smp_call_function(ipi_flush_tlb_mm, mm, 1, 1)) {
869 printk(KERN_CRIT "flush_tlb_mm: timed out\n");
874 EXPORT_SYMBOL(flush_tlb_mm);
876 struct flush_tlb_page_struct {
877 struct vm_area_struct *vma;
878 struct mm_struct *mm;
883 ipi_flush_tlb_page(void *x)
885 struct flush_tlb_page_struct *data = (struct flush_tlb_page_struct *)x;
886 struct mm_struct * mm = data->mm;
888 if (mm == current->active_mm && !asn_locked())
889 flush_tlb_current_page(mm, data->vma, data->addr);
895 flush_tlb_page(struct vm_area_struct *vma, unsigned long addr)
897 struct flush_tlb_page_struct data;
898 struct mm_struct *mm = vma->vm_mm;
902 if (mm == current->active_mm) {
903 flush_tlb_current_page(mm, vma, addr);
904 if (atomic_read(&mm->mm_users) <= 1) {
905 int cpu, this_cpu = smp_processor_id();
906 for (cpu = 0; cpu < NR_CPUS; cpu++) {
907 if (!cpu_online(cpu) || cpu == this_cpu)
909 if (mm->context[cpu])
910 mm->context[cpu] = 0;
921 if (smp_call_function(ipi_flush_tlb_page, &data, 1, 1)) {
922 printk(KERN_CRIT "flush_tlb_page: timed out\n");
927 EXPORT_SYMBOL(flush_tlb_page);
930 flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
932 /* On the Alpha we always flush the whole user tlb. */
933 flush_tlb_mm(vma->vm_mm);
935 EXPORT_SYMBOL(flush_tlb_range);
938 ipi_flush_icache_page(void *x)
940 struct mm_struct *mm = (struct mm_struct *) x;
941 if (mm == current->active_mm && !asn_locked())
942 __load_new_mm_context(mm);
948 flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
949 unsigned long addr, int len)
951 struct mm_struct *mm = vma->vm_mm;
953 if ((vma->vm_flags & VM_EXEC) == 0)
958 if (mm == current->active_mm) {
959 __load_new_mm_context(mm);
960 if (atomic_read(&mm->mm_users) <= 1) {
961 int cpu, this_cpu = smp_processor_id();
962 for (cpu = 0; cpu < NR_CPUS; cpu++) {
963 if (!cpu_online(cpu) || cpu == this_cpu)
965 if (mm->context[cpu])
966 mm->context[cpu] = 0;
973 if (smp_call_function(ipi_flush_icache_page, mm, 1, 1)) {
974 printk(KERN_CRIT "flush_icache_page: timed out\n");