]> err.no Git - linux-2.6/blobdiff - arch/s390/kernel/smp.c
Pull virt-cpu-accounting into release branch
[linux-2.6] / arch / s390 / kernel / smp.c
index 66fe28930d821fafcd187c7a1a86623c2be7cf3a..8f894d380a626316c68aeb073043f027abdcfee6 100644 (file)
@@ -54,7 +54,7 @@ EXPORT_SYMBOL(lowcore_ptr);
 cpumask_t cpu_online_map = CPU_MASK_NONE;
 EXPORT_SYMBOL(cpu_online_map);
 
-cpumask_t cpu_possible_map = CPU_MASK_NONE;
+cpumask_t cpu_possible_map = CPU_MASK_ALL;
 EXPORT_SYMBOL(cpu_possible_map);
 
 static struct task_struct *current_set[NR_CPUS];
@@ -210,6 +210,32 @@ int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 }
 EXPORT_SYMBOL(smp_call_function_single);
 
+/**
+ * smp_call_function_mask(): Run a function on a set of other CPUs.
+ * @mask: The set of cpus to run on.  Must not include the current cpu.
+ * @func: The function to run. This must be fast and non-blocking.
+ * @info: An arbitrary pointer to pass to the function.
+ * @wait: If true, wait (atomically) until function has completed on other CPUs.
+ *
+ * Returns 0 on success, else a negative status code.
+ *
+ * If @wait is true, then returns once @func has returned; otherwise
+ * it returns just before the target cpu calls @func.
+ *
+ * You must not call this function with disabled interrupts or from a
+ * hardware interrupt handler or from a bottom half handler.
+ */
+int smp_call_function_mask(cpumask_t mask, void (*func)(void *), void *info,
+                          int wait)
+{
+       preempt_disable();
+       cpu_clear(smp_processor_id(), mask);
+       __smp_call_function_map(func, info, 0, wait, mask);
+       preempt_enable();
+       return 0;
+}
+EXPORT_SYMBOL(smp_call_function_mask);
+
 void smp_send_stop(void)
 {
        int cpu, rc;
@@ -233,33 +259,6 @@ void smp_send_stop(void)
        }
 }
 
-/*
- * Reboot, halt and power_off routines for SMP.
- */
-void machine_restart_smp(char *__unused)
-{
-       smp_send_stop();
-       do_reipl();
-}
-
-void machine_halt_smp(void)
-{
-       smp_send_stop();
-       if (MACHINE_IS_VM && strlen(vmhalt_cmd) > 0)
-               __cpcmd(vmhalt_cmd, NULL, 0, NULL);
-       signal_processor(smp_processor_id(), sigp_stop_and_store_status);
-       for (;;);
-}
-
-void machine_power_off_smp(void)
-{
-       smp_send_stop();
-       if (MACHINE_IS_VM && strlen(vmpoff_cmd) > 0)
-               __cpcmd(vmpoff_cmd, NULL, 0, NULL);
-       signal_processor(smp_processor_id(), sigp_stop_and_store_status);
-       for (;;);
-}
-
 /*
  * This is the main routine where commands issued by other
  * cpus are handled.
@@ -399,7 +398,7 @@ static void __init smp_get_save_area(unsigned int cpu, unsigned int phy_cpu)
                       "kernel was compiled with NR_CPUS=%i\n", cpu, NR_CPUS);
                return;
        }
-       zfcpdump_save_areas[cpu] = alloc_bootmem(sizeof(union save_area));
+       zfcpdump_save_areas[cpu] = kmalloc(sizeof(union save_area), GFP_KERNEL);
        __cpu_logical_map[CPU_INIT_NO] = (__u16) phy_cpu;
        while (signal_processor(CPU_INIT_NO, sigp_stop_and_store_status) ==
               sigp_busy)
@@ -435,67 +434,6 @@ static int cpu_stopped(int cpu)
        return 0;
 }
 
-/*
- * Lets check how many CPUs we have.
- */
-static void __init smp_count_cpus(unsigned int *configured_cpus,
-                                 unsigned int *standby_cpus)
-{
-       unsigned int cpu;
-       struct sclp_cpu_info *info;
-       u16 boot_cpu_addr, cpu_addr;
-
-       boot_cpu_addr = S390_lowcore.cpu_data.cpu_addr;
-       current_thread_info()->cpu = 0;
-       *configured_cpus = 1;
-       *standby_cpus = 0;
-
-       info = alloc_bootmem_pages(sizeof(*info));
-       if (!info)
-               disabled_wait((unsigned long) __builtin_return_address(0));
-
-       /* Use sigp detection algorithm if sclp doesn't work. */
-       if (sclp_get_cpu_info(info)) {
-               smp_use_sigp_detection = 1;
-               for (cpu = 0; cpu <= 65535; cpu++) {
-                       if (cpu == boot_cpu_addr)
-                               continue;
-                       __cpu_logical_map[CPU_INIT_NO] = cpu;
-                       if (cpu_stopped(CPU_INIT_NO))
-                               (*configured_cpus)++;
-               }
-               goto out;
-       }
-
-       if (info->has_cpu_type) {
-               for (cpu = 0; cpu < info->combined; cpu++) {
-                       if (info->cpu[cpu].address == boot_cpu_addr) {
-                               smp_cpu_type = info->cpu[cpu].type;
-                               break;
-                       }
-               }
-       }
-       /* Count cpus. */
-       for (cpu = 0; cpu < info->combined; cpu++) {
-               if (info->has_cpu_type && info->cpu[cpu].type != smp_cpu_type)
-                       continue;
-               cpu_addr = info->cpu[cpu].address;
-               if (cpu_addr == boot_cpu_addr)
-                       continue;
-               __cpu_logical_map[CPU_INIT_NO] = cpu_addr;
-               if (!cpu_stopped(CPU_INIT_NO)) {
-                       (*standby_cpus)++;
-                       continue;
-               }
-               smp_get_save_area(*configured_cpus, cpu_addr);
-               (*configured_cpus)++;
-       }
-out:
-       printk(KERN_INFO "CPUs: %d configured, %d standby\n",
-              *configured_cpus, *standby_cpus);
-       free_bootmem((unsigned long) info, sizeof(*info));
-}
-
 static int cpu_known(int cpu_id)
 {
        int cpu;
@@ -529,7 +467,7 @@ static int smp_rescan_cpus_sigp(cpumask_t avail)
        return 0;
 }
 
-static int __init_refok smp_rescan_cpus_sclp(cpumask_t avail)
+static int smp_rescan_cpus_sclp(cpumask_t avail)
 {
        struct sclp_cpu_info *info;
        int cpu_id, logical_cpu, cpu;
@@ -538,10 +476,7 @@ static int __init_refok smp_rescan_cpus_sclp(cpumask_t avail)
        logical_cpu = first_cpu(avail);
        if (logical_cpu == NR_CPUS)
                return 0;
-       if (slab_is_available())
-               info = kmalloc(sizeof(*info), GFP_KERNEL);
-       else
-               info = alloc_bootmem(sizeof(*info));
+       info = kmalloc(sizeof(*info), GFP_KERNEL);
        if (!info)
                return -ENOMEM;
        rc = sclp_get_cpu_info(info);
@@ -564,10 +499,7 @@ static int __init_refok smp_rescan_cpus_sclp(cpumask_t avail)
                        break;
        }
 out:
-       if (slab_is_available())
-               kfree(info);
-       else
-               free_bootmem((unsigned long) info, sizeof(*info));
+       kfree(info);
        return rc;
 }
 
@@ -575,15 +507,71 @@ static int smp_rescan_cpus(void)
 {
        cpumask_t avail;
 
-       cpus_setall(avail);
-       cpus_and(avail, avail, cpu_possible_map);
-       cpus_andnot(avail, avail, cpu_present_map);
+       cpus_xor(avail, cpu_possible_map, cpu_present_map);
        if (smp_use_sigp_detection)
                return smp_rescan_cpus_sigp(avail);
        else
                return smp_rescan_cpus_sclp(avail);
 }
 
+static void __init smp_detect_cpus(void)
+{
+       unsigned int cpu, c_cpus, s_cpus;
+       struct sclp_cpu_info *info;
+       u16 boot_cpu_addr, cpu_addr;
+
+       c_cpus = 1;
+       s_cpus = 0;
+       boot_cpu_addr = S390_lowcore.cpu_data.cpu_addr;
+       info = kmalloc(sizeof(*info), GFP_KERNEL);
+       if (!info)
+               panic("smp_detect_cpus failed to allocate memory\n");
+       /* Use sigp detection algorithm if sclp doesn't work. */
+       if (sclp_get_cpu_info(info)) {
+               smp_use_sigp_detection = 1;
+               for (cpu = 0; cpu <= 65535; cpu++) {
+                       if (cpu == boot_cpu_addr)
+                               continue;
+                       __cpu_logical_map[CPU_INIT_NO] = cpu;
+                       if (!cpu_stopped(CPU_INIT_NO))
+                               continue;
+                       smp_get_save_area(c_cpus, cpu);
+                       c_cpus++;
+               }
+               goto out;
+       }
+
+       if (info->has_cpu_type) {
+               for (cpu = 0; cpu < info->combined; cpu++) {
+                       if (info->cpu[cpu].address == boot_cpu_addr) {
+                               smp_cpu_type = info->cpu[cpu].type;
+                               break;
+                       }
+               }
+       }
+
+       for (cpu = 0; cpu < info->combined; cpu++) {
+               if (info->has_cpu_type && info->cpu[cpu].type != smp_cpu_type)
+                       continue;
+               cpu_addr = info->cpu[cpu].address;
+               if (cpu_addr == boot_cpu_addr)
+                       continue;
+               __cpu_logical_map[CPU_INIT_NO] = cpu_addr;
+               if (!cpu_stopped(CPU_INIT_NO)) {
+                       s_cpus++;
+                       continue;
+               }
+               smp_get_save_area(c_cpus, cpu_addr);
+               c_cpus++;
+       }
+out:
+       kfree(info);
+       printk(KERN_INFO "CPUs: %d configured, %d standby\n", c_cpus, s_cpus);
+       get_online_cpus();
+       smp_rescan_cpus();
+       put_online_cpus();
+}
+
 /*
  *     Activate a secondary processor.
  */
@@ -627,8 +615,69 @@ static void __init smp_create_idle(unsigned int cpu)
        spin_lock_init(&(&per_cpu(s390_idle, cpu))->lock);
 }
 
+static int __cpuinit smp_alloc_lowcore(int cpu)
+{
+       unsigned long async_stack, panic_stack;
+       struct _lowcore *lowcore;
+       int lc_order;
+
+       lc_order = sizeof(long) == 8 ? 1 : 0;
+       lowcore = (void *) __get_free_pages(GFP_KERNEL | GFP_DMA, lc_order);
+       if (!lowcore)
+               return -ENOMEM;
+       async_stack = __get_free_pages(GFP_KERNEL, ASYNC_ORDER);
+       panic_stack = __get_free_page(GFP_KERNEL);
+       if (!panic_stack || !async_stack)
+               goto out;
+       memcpy(lowcore, &S390_lowcore, 512);
+       memset((char *)lowcore + 512, 0, sizeof(*lowcore) - 512);
+       lowcore->async_stack = async_stack + ASYNC_SIZE;
+       lowcore->panic_stack = panic_stack + PAGE_SIZE;
+
+#ifndef CONFIG_64BIT
+       if (MACHINE_HAS_IEEE) {
+               unsigned long save_area;
+
+               save_area = get_zeroed_page(GFP_KERNEL);
+               if (!save_area)
+                       goto out_save_area;
+               lowcore->extended_save_area_addr = (u32) save_area;
+       }
+#endif
+       lowcore_ptr[cpu] = lowcore;
+       return 0;
+
+#ifndef CONFIG_64BIT
+out_save_area:
+       free_page(panic_stack);
+#endif
+out:
+       free_pages(async_stack, ASYNC_ORDER);
+       free_pages((unsigned long) lowcore, lc_order);
+       return -ENOMEM;
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+static void smp_free_lowcore(int cpu)
+{
+       struct _lowcore *lowcore;
+       int lc_order;
+
+       lc_order = sizeof(long) == 8 ? 1 : 0;
+       lowcore = lowcore_ptr[cpu];
+#ifndef CONFIG_64BIT
+       if (MACHINE_HAS_IEEE)
+               free_page((unsigned long) lowcore->extended_save_area_addr);
+#endif
+       free_page(lowcore->panic_stack - PAGE_SIZE);
+       free_pages(lowcore->async_stack - ASYNC_SIZE, ASYNC_ORDER);
+       free_pages((unsigned long) lowcore, lc_order);
+       lowcore_ptr[cpu] = NULL;
+}
+#endif /* CONFIG_HOTPLUG_CPU */
+
 /* Upping and downing of CPUs */
-int __cpu_up(unsigned int cpu)
+int __cpuinit __cpu_up(unsigned int cpu)
 {
        struct task_struct *idle;
        struct _lowcore *cpu_lowcore;
@@ -637,6 +686,8 @@ int __cpu_up(unsigned int cpu)
 
        if (smp_cpu_state[cpu] != CPU_STATE_CONFIGURED)
                return -EIO;
+       if (smp_alloc_lowcore(cpu))
+               return -ENOMEM;
 
        ccode = signal_processor_p((__u32)(unsigned long)(lowcore_ptr[cpu]),
                                   cpu, sigp_set_prefix);
@@ -651,6 +702,7 @@ int __cpu_up(unsigned int cpu)
        cpu_lowcore = lowcore_ptr[cpu];
        cpu_lowcore->kernel_stack = (unsigned long)
                task_stack_page(idle) + THREAD_SIZE;
+       cpu_lowcore->thread_info = (unsigned long) task_thread_info(idle);
        sf = (struct stack_frame *) (cpu_lowcore->kernel_stack
                                     - sizeof(struct pt_regs)
                                     - sizeof(struct stack_frame));
@@ -664,6 +716,8 @@ int __cpu_up(unsigned int cpu)
        cpu_lowcore->percpu_offset = __per_cpu_offset[cpu];
        cpu_lowcore->current_task = (unsigned long) idle;
        cpu_lowcore->cpu_data.cpu_nr = cpu;
+       cpu_lowcore->kernel_asce = S390_lowcore.kernel_asce;
+       cpu_lowcore->ipl_device = S390_lowcore.ipl_device;
        eieio();
 
        while (signal_processor(cpu, sigp_restart) == sigp_busy)
@@ -674,41 +728,20 @@ int __cpu_up(unsigned int cpu)
        return 0;
 }
 
-static unsigned int __initdata additional_cpus;
-static unsigned int __initdata possible_cpus;
-
-void __init smp_setup_cpu_possible_map(void)
-{
-       unsigned int pos_cpus, cpu;
-       unsigned int configured_cpus, standby_cpus;
-
-       smp_count_cpus(&configured_cpus, &standby_cpus);
-       pos_cpus = min(configured_cpus + standby_cpus + additional_cpus,
-                      (unsigned int) NR_CPUS);
-       if (possible_cpus)
-               pos_cpus = min(possible_cpus, (unsigned int) NR_CPUS);
-       for (cpu = 0; cpu < pos_cpus; cpu++)
-               cpu_set(cpu, cpu_possible_map);
-       cpu_present_map = cpumask_of_cpu(0);
-       smp_rescan_cpus();
-}
-
-#ifdef CONFIG_HOTPLUG_CPU
-
-static int __init setup_additional_cpus(char *s)
-{
-       additional_cpus = simple_strtoul(s, NULL, 0);
-       return 0;
-}
-early_param("additional_cpus", setup_additional_cpus);
-
 static int __init setup_possible_cpus(char *s)
 {
-       possible_cpus = simple_strtoul(s, NULL, 0);
+       int pcpus, cpu;
+
+       pcpus = simple_strtoul(s, NULL, 0);
+       cpu_possible_map = cpumask_of_cpu(0);
+       for (cpu = 1; cpu < pcpus && cpu < NR_CPUS; cpu++)
+               cpu_set(cpu, cpu_possible_map);
        return 0;
 }
 early_param("possible_cpus", setup_possible_cpus);
 
+#ifdef CONFIG_HOTPLUG_CPU
+
 int __cpu_disable(void)
 {
        struct ec_creg_mask_parms cr_parms;
@@ -745,6 +778,7 @@ void __cpu_die(unsigned int cpu)
        /* Wait until target cpu is down */
        while (!smp_cpu_not_running(cpu))
                cpu_relax();
+       smp_free_lowcore(cpu);
        printk(KERN_INFO "Processor %d spun down\n", cpu);
 }
 
@@ -758,55 +792,45 @@ void cpu_die(void)
 
 #endif /* CONFIG_HOTPLUG_CPU */
 
-/*
- *     Cycle through the processors and setup structures.
- */
-
 void __init smp_prepare_cpus(unsigned int max_cpus)
 {
-       unsigned long stack;
+#ifndef CONFIG_64BIT
+       unsigned long save_area = 0;
+#endif
+       unsigned long async_stack, panic_stack;
+       struct _lowcore *lowcore;
        unsigned int cpu;
-       int i;
+       int lc_order;
+
+       smp_detect_cpus();
 
        /* request the 0x1201 emergency signal external interrupt */
        if (register_external_interrupt(0x1201, do_ext_call_interrupt) != 0)
                panic("Couldn't request external interrupt 0x1201");
-       memset(lowcore_ptr, 0, sizeof(lowcore_ptr));
-       /*
-        *  Initialize prefix pages and stacks for all possible cpus
-        */
        print_cpu_info(&S390_lowcore.cpu_data);
 
-       for_each_possible_cpu(i) {
-               lowcore_ptr[i] = (struct _lowcore *)
-                       __get_free_pages(GFP_KERNEL | GFP_DMA,
-                                        sizeof(void*) == 8 ? 1 : 0);
-               stack = __get_free_pages(GFP_KERNEL, ASYNC_ORDER);
-               if (!lowcore_ptr[i] || !stack)
-                       panic("smp_boot_cpus failed to allocate memory\n");
-
-               *(lowcore_ptr[i]) = S390_lowcore;
-               lowcore_ptr[i]->async_stack = stack + ASYNC_SIZE;
-               stack = __get_free_pages(GFP_KERNEL, 0);
-               if (!stack)
-                       panic("smp_boot_cpus failed to allocate memory\n");
-               lowcore_ptr[i]->panic_stack = stack + PAGE_SIZE;
+       /* Reallocate current lowcore, but keep its contents. */
+       lc_order = sizeof(long) == 8 ? 1 : 0;
+       lowcore = (void *) __get_free_pages(GFP_KERNEL | GFP_DMA, lc_order);
+       panic_stack = __get_free_page(GFP_KERNEL);
+       async_stack = __get_free_pages(GFP_KERNEL, ASYNC_ORDER);
 #ifndef CONFIG_64BIT
-               if (MACHINE_HAS_IEEE) {
-                       lowcore_ptr[i]->extended_save_area_addr =
-                               (__u32) __get_free_pages(GFP_KERNEL, 0);
-                       if (!lowcore_ptr[i]->extended_save_area_addr)
-                               panic("smp_boot_cpus failed to "
-                                     "allocate memory\n");
-               }
+       if (MACHINE_HAS_IEEE)
+               save_area = get_zeroed_page(GFP_KERNEL);
 #endif
-       }
+       local_irq_disable();
+       local_mcck_disable();
+       lowcore_ptr[smp_processor_id()] = lowcore;
+       *lowcore = S390_lowcore;
+       lowcore->panic_stack = panic_stack + PAGE_SIZE;
+       lowcore->async_stack = async_stack + ASYNC_SIZE;
 #ifndef CONFIG_64BIT
        if (MACHINE_HAS_IEEE)
-               ctl_set_bit(14, 29); /* enable extended save area */
+               lowcore->extended_save_area_addr = (u32) save_area;
 #endif
-       set_prefix((u32)(unsigned long) lowcore_ptr[smp_processor_id()]);
-
+       set_prefix((u32)(unsigned long) lowcore);
+       local_mcck_enable();
+       local_irq_enable();
        for_each_possible_cpu(cpu)
                if (cpu != smp_processor_id())
                        smp_create_idle(cpu);
@@ -816,6 +840,8 @@ void __init smp_prepare_boot_cpu(void)
 {
        BUG_ON(smp_processor_id() != 0);
 
+       current_thread_info()->cpu = 0;
+       cpu_set(0, cpu_present_map);
        cpu_set(0, cpu_online_map);
        S390_lowcore.percpu_offset = __per_cpu_offset[0];
        current_set[0] = current;
@@ -862,7 +888,7 @@ static ssize_t cpu_configure_store(struct sys_device *dev, const char *buf,
                return -EINVAL;
 
        mutex_lock(&smp_cpu_state_mutex);
-       lock_cpu_hotplug();
+       get_online_cpus();
        rc = -EBUSY;
        if (cpu_online(cpu))
                goto out;
@@ -886,7 +912,7 @@ static ssize_t cpu_configure_store(struct sys_device *dev, const char *buf,
                break;
        }
 out:
-       unlock_cpu_hotplug();
+       put_online_cpus();
        mutex_unlock(&smp_cpu_state_mutex);
        return rc ? rc : count;
 }
@@ -998,7 +1024,7 @@ static struct notifier_block __cpuinitdata smp_cpu_nb = {
        .notifier_call = smp_cpu_notify,
 };
 
-static int smp_add_present_cpu(int cpu)
+static int __devinit smp_add_present_cpu(int cpu)
 {
        struct cpu *c = &per_cpu(cpu_devices, cpu);
        struct sys_device *s = &c->sysdev;
@@ -1026,15 +1052,15 @@ out:
 }
 
 #ifdef CONFIG_HOTPLUG_CPU
-static ssize_t rescan_store(struct sys_device *dev, const char *buf,
-                           size_t count)
+static ssize_t __ref rescan_store(struct sys_device *dev,
+                                 const char *buf, size_t count)
 {
        cpumask_t newcpus;
        int cpu;
        int rc;
 
        mutex_lock(&smp_cpu_state_mutex);
-       lock_cpu_hotplug();
+       get_online_cpus();
        newcpus = cpu_present_map;
        rc = smp_rescan_cpus();
        if (rc)
@@ -1047,7 +1073,7 @@ static ssize_t rescan_store(struct sys_device *dev, const char *buf,
        }
        rc = 0;
 out:
-       unlock_cpu_hotplug();
+       put_online_cpus();
        mutex_unlock(&smp_cpu_state_mutex);
        return rc ? rc : count;
 }