arch/powerpc/platforms/powernv/smp.c

   1 /*
   2  * SMP support for PowerNV machines.
   3  *
   4  * Copyright 2011 IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/sched.h>
  15 #include <linux/smp.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/delay.h>
  18 #include <linux/init.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/cpu.h>
  21
  22 #include <asm/irq.h>
  23 #include <asm/smp.h>
  24 #include <asm/paca.h>
  25 #include <asm/machdep.h>
  26 #include <asm/cputable.h>
  27 #include <asm/firmware.h>
  28 #include <asm/vdso_datapage.h>
  29 #include <asm/cputhreads.h>
  30 #include <asm/xics.h>
  31 #include <asm/opal.h>
  32 #include <asm/runlatch.h>
  33 #include <asm/code-patching.h>
  34 #include <asm/dbell.h>
  35 #include <asm/kvm_ppc.h>
  36 #include <asm/ppc-opcode.h>
  37
  38 #include "powernv.h"
  39
  40 #ifdef DEBUG
  41 #include <asm/udbg.h>
  42 #define DBG(fmt...) udbg_printf(fmt)
  43 #else
  44 #define DBG(fmt...)
  45 #endif
  46
  47 static void pnv_smp_setup_cpu(int cpu)
  48 {
  49         if (cpu != boot_cpuid)
  50                 xics_setup_cpu();
  51
  52 #ifdef CONFIG_PPC_DOORBELL
  53         if (cpu_has_feature(CPU_FTR_DBELL))
  54                 doorbell_setup_this_cpu();
  55 #endif
  56 }
  57
  58 static int pnv_smp_kick_cpu(int nr)
  59 {
  60         unsigned int pcpu = get_hard_smp_processor_id(nr);
  61         unsigned long start_here =
  62                         __pa(ppc_function_entry(generic_secondary_smp_init));
  63         long rc;
  64         uint8_t status;
  65
  66         BUG_ON(nr < 0 || nr >= NR_CPUS);
  67
  68         /*
  69          * If we already started or OPAL is not supported, we just
  70          * kick the CPU via the PACA
  71          */
  72         if (paca[nr].cpu_start || !firmware_has_feature(FW_FEATURE_OPAL))
  73                 goto kick;
  74
  75         /*
  76          * At this point, the CPU can either be spinning on the way in
  77          * from kexec or be inside OPAL waiting to be started for the
  78          * first time. OPAL v3 allows us to query OPAL to know if it
  79          * has the CPUs, so we do that
  80          */
  81         rc = opal_query_cpu_status(pcpu, &status);
  82         if (rc != OPAL_SUCCESS) {
  83                 pr_warn("OPAL Error %ld querying CPU %d state\n", rc, nr);
  84                 return -ENODEV;
  85         }
  86
  87         /*
  88          * Already started, just kick it, probably coming from
  89          * kexec and spinning
  90          */
  91         if (status == OPAL_THREAD_STARTED)
  92                 goto kick;
  93
  94         /*
  95          * Available/inactive, let's kick it
  96          */
  97         if (status == OPAL_THREAD_INACTIVE) {
  98                 pr_devel("OPAL: Starting CPU %d (HW 0x%x)...\n", nr, pcpu);
  99                 rc = opal_start_cpu(pcpu, start_here);
 100                 if (rc != OPAL_SUCCESS) {
 101                         pr_warn("OPAL Error %ld starting CPU %d\n", rc, nr);
 102                         return -ENODEV;
 103                 }
 104         } else {
 105                 /*
 106                  * An unavailable CPU (or any other unknown status)
 107                  * shouldn't be started. It should also
 108                  * not be in the possible map but currently it can
 109                  * happen
 110                  */
 111                 pr_devel("OPAL: CPU %d (HW 0x%x) is unavailable"
 112                          " (status %d)...\n", nr, pcpu, status);
 113                 return -ENODEV;
 114         }
 115
 116 kick:
 117         return smp_generic_kick_cpu(nr);
 118 }
 119
 120 #ifdef CONFIG_HOTPLUG_CPU
 121
 122 static int pnv_smp_cpu_disable(void)
 123 {
 124         int cpu = smp_processor_id();
 125
 126         /* This is identical to pSeries... might consolidate by
 127          * moving migrate_irqs_away to a ppc_md with default to
 128          * the generic fixup_irqs. --BenH.
 129          */
 130         set_cpu_online(cpu, false);
 131         vdso_data->processorCount--;
 132         if (cpu == boot_cpuid)
 133                 boot_cpuid = cpumask_any(cpu_online_mask);
 134         xics_migrate_irqs_away();
 135         return 0;
 136 }
 137
 138 static void pnv_smp_cpu_kill_self(void)
 139 {
 140         unsigned int cpu;
 141         unsigned long srr1, wmask;
 142         u32 idle_states;
 143
 144         /* Standard hot unplug procedure */
 145         local_irq_disable();
 146         idle_task_exit();
 147         current->active_mm = NULL; /* for sanity */
 148         cpu = smp_processor_id();
 149         DBG("CPU%d offline\n", cpu);
 150         generic_set_cpu_dead(cpu);
 151         smp_wmb();
 152
 153         wmask = SRR1_WAKEMASK;
 154         if (cpu_has_feature(CPU_FTR_ARCH_207S))
 155                 wmask = SRR1_WAKEMASK_P8;
 156
 157         idle_states = pnv_get_supported_cpuidle_states();
 158
 159         /* We don't want to take decrementer interrupts while we are offline,
 160          * so clear LPCR:PECE1. We keep PECE2 (and LPCR_PECE_HVEE on P9)
 161          * enabled as to let IPIs in.
 162          */
 163         mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) & ~(u64)LPCR_PECE1);
 164
 165         /*
 166          * Hard-disable interrupts, and then clear irq_happened flags
 167          * that we can safely ignore while off-line, since they
 168          * are for things for which we do no processing when off-line
 169          * (or in the case of HMI, all the processing we need to do
 170          * is done in lower-level real-mode code).
 171          */
 172         hard_irq_disable();
 173         local_paca->irq_happened &= ~(PACA_IRQ_DEC | PACA_IRQ_HMI);
 174
 175         while (!generic_check_cpu_restart(cpu)) {
 176                 /*
 177                  * Clear IPI flag, since we don't handle IPIs while
 178                  * offline, except for those when changing micro-threading
 179                  * mode, which are handled explicitly below, and those
 180                  * for coming online, which are handled via
 181                  * generic_check_cpu_restart() calls.
 182                  */
 183                 kvmppc_set_host_ipi(cpu, 0);
 184
 185                 ppc64_runlatch_off();
 186
 187                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 188                         srr1 = power9_idle_stop(pnv_deepest_stop_psscr_val,
 189                                                 pnv_deepest_stop_psscr_mask);
 190                 } else if (idle_states & OPAL_PM_WINKLE_ENABLED) {
 191                         srr1 = power7_winkle();
 192                 } else if ((idle_states & OPAL_PM_SLEEP_ENABLED) ||
 193                            (idle_states & OPAL_PM_SLEEP_ENABLED_ER1)) {
 194                         srr1 = power7_sleep();
 195                 } else {
 196                         srr1 = power7_nap(1);
 197                 }
 198
 199                 ppc64_runlatch_on();
 200
 201                 /*
 202                  * If the SRR1 value indicates that we woke up due to
 203                  * an external interrupt, then clear the interrupt.
 204                  * We clear the interrupt before checking for the
 205                  * reason, so as to avoid a race where we wake up for
 206                  * some other reason, find nothing and clear the interrupt
 207                  * just as some other cpu is sending us an interrupt.
 208                  * If we returned from power7_nap as a result of
 209                  * having finished executing in a KVM guest, then srr1
 210                  * contains 0.
 211                  */
 212                 if (((srr1 & wmask) == SRR1_WAKEEE) ||
 213                     ((srr1 & wmask) == SRR1_WAKEHVI) ||
 214                     (local_paca->irq_happened & PACA_IRQ_EE)) {
 215                         if (cpu_has_feature(CPU_FTR_ARCH_300))
 216                                 icp_opal_flush_interrupt();
 217                         else
 218                                 icp_native_flush_interrupt();
 219                 } else if ((srr1 & wmask) == SRR1_WAKEHDBELL) {
 220                         unsigned long msg = PPC_DBELL_TYPE(PPC_DBELL_SERVER);
 221                         asm volatile(PPC_MSGCLR(%0) : : "r" (msg));
 222                 }
 223                 local_paca->irq_happened &= ~(PACA_IRQ_EE | PACA_IRQ_DBELL);
 224                 smp_mb();
 225
 226                 if (cpu_core_split_required())
 227                         continue;
 228
 229                 if (srr1 && !generic_check_cpu_restart(cpu))
 230                         DBG("CPU%d Unexpected exit while offline !\n", cpu);
 231         }
 232
 233         /* Re-enable decrementer interrupts */
 234         mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) | LPCR_PECE1);
 235         DBG("CPU%d coming online...\n", cpu);
 236 }
 237
 238 #endif /* CONFIG_HOTPLUG_CPU */
 239
 240 static int pnv_cpu_bootable(unsigned int nr)
 241 {
 242         /*
 243          * Starting with POWER8, the subcore logic relies on all threads of a
 244          * core being booted so that they can participate in split mode
 245          * switches. So on those machines we ignore the smt_enabled_at_boot
 246          * setting (smt-enabled on the kernel command line).
 247          */
 248         if (cpu_has_feature(CPU_FTR_ARCH_207S))
 249                 return 1;
 250
 251         return smp_generic_cpu_bootable(nr);
 252 }
 253
 254 static struct smp_ops_t pnv_smp_ops = {
 255         .message_pass   = smp_muxed_ipi_message_pass,
 256         .cause_ipi      = NULL, /* Filled at runtime by xics_smp_probe() */
 257         .probe          = xics_smp_probe,
 258         .kick_cpu       = pnv_smp_kick_cpu,
 259         .setup_cpu      = pnv_smp_setup_cpu,
 260         .cpu_bootable   = pnv_cpu_bootable,
 261 #ifdef CONFIG_HOTPLUG_CPU
 262         .cpu_disable    = pnv_smp_cpu_disable,
 263         .cpu_die        = generic_cpu_die,
 264 #endif /* CONFIG_HOTPLUG_CPU */
 265 };
 266
 267 /* This is called very early during platform setup_arch */
 268 void __init pnv_smp_init(void)
 269 {
 270         smp_ops = &pnv_smp_ops;
 271
 272 #ifdef CONFIG_HOTPLUG_CPU
 273         ppc_md.cpu_die  = pnv_smp_cpu_kill_self;
 274 #endif
 275 }