arch/powerpc/platforms/powernv/setup.c

   1 /*
   2  * PowerNV setup code.
   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 #undef DEBUG
  13
  14 #include <linux/cpu.h>
  15 #include <linux/errno.h>
  16 #include <linux/sched.h>
  17 #include <linux/kernel.h>
  18 #include <linux/tty.h>
  19 #include <linux/reboot.h>
  20 #include <linux/init.h>
  21 #include <linux/console.h>
  22 #include <linux/delay.h>
  23 #include <linux/irq.h>
  24 #include <linux/seq_file.h>
  25 #include <linux/of.h>
  26 #include <linux/of_fdt.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/bug.h>
  29 #include <linux/pci.h>
  30 #include <linux/cpufreq.h>
  31
  32 #include <asm/machdep.h>
  33 #include <asm/firmware.h>
  34 #include <asm/xics.h>
  35 #include <asm/rtas.h>
  36 #include <asm/opal.h>
  37 #include <asm/kexec.h>
  38 #include <asm/smp.h>
  39
  40 #include "powernv.h"
  41
  42 static void __init pnv_setup_arch(void)
  43 {
  44         set_arch_panic_timeout(10, ARCH_PANIC_TIMEOUT);
  45
  46         /* Initialize SMP */
  47         pnv_smp_init();
  48
  49         /* Setup PCI */
  50         pnv_pci_init();
  51
  52         /* Setup RTC and NVRAM callbacks */
  53         if (firmware_has_feature(FW_FEATURE_OPAL))
  54                 opal_nvram_init();
  55
  56         /* Enable NAP mode */
  57         powersave_nap = 1;
  58
  59         /* XXX PMCS */
  60 }
  61
  62 static void __init pnv_init_early(void)
  63 {
  64         /*
  65          * Initialize the LPC bus now so that legacy serial
  66          * ports can be found on it
  67          */
  68         opal_lpc_init();
  69
  70 #ifdef CONFIG_HVC_OPAL
  71         if (firmware_has_feature(FW_FEATURE_OPAL))
  72                 hvc_opal_init_early();
  73         else
  74 #endif
  75                 add_preferred_console("hvc", 0, NULL);
  76 }
  77
  78 static void __init pnv_init_IRQ(void)
  79 {
  80         xics_init();
  81
  82         WARN_ON(!ppc_md.get_irq);
  83 }
  84
  85 static void pnv_show_cpuinfo(struct seq_file *m)
  86 {
  87         struct device_node *root;
  88         const char *model = "";
  89
  90         root = of_find_node_by_path("/");
  91         if (root)
  92                 model = of_get_property(root, "model", NULL);
  93         seq_printf(m, "machine\t\t: PowerNV %s\n", model);
  94         if (firmware_has_feature(FW_FEATURE_OPALv3))
  95                 seq_printf(m, "firmware\t: OPAL v3\n");
  96         else if (firmware_has_feature(FW_FEATURE_OPALv2))
  97                 seq_printf(m, "firmware\t: OPAL v2\n");
  98         else if (firmware_has_feature(FW_FEATURE_OPAL))
  99                 seq_printf(m, "firmware\t: OPAL v1\n");
 100         else
 101                 seq_printf(m, "firmware\t: BML\n");
 102         of_node_put(root);
 103 }
 104
 105 static void pnv_prepare_going_down(void)
 106 {
 107         /*
 108          * Disable all notifiers from OPAL, we can't
 109          * service interrupts anymore anyway
 110          */
 111         opal_notifier_disable();
 112
 113         /* Soft disable interrupts */
 114         local_irq_disable();
 115
 116         /*
 117          * Return secondary CPUs to firwmare if a flash update
 118          * is pending otherwise we will get all sort of error
 119          * messages about CPU being stuck etc.. This will also
 120          * have the side effect of hard disabling interrupts so
 121          * past this point, the kernel is effectively dead.
 122          */
 123         opal_flash_term_callback();
 124 }
 125
 126 static void  __noreturn pnv_restart(char *cmd)
 127 {
 128         long rc = OPAL_BUSY;
 129
 130         pnv_prepare_going_down();
 131
 132         while (rc == OPAL_BUSY || rc == OPAL_BUSY_EVENT) {
 133                 rc = opal_cec_reboot();
 134                 if (rc == OPAL_BUSY_EVENT)
 135                         opal_poll_events(NULL);
 136                 else
 137                         mdelay(10);
 138         }
 139         for (;;)
 140                 opal_poll_events(NULL);
 141 }
 142
 143 static void __noreturn pnv_power_off(void)
 144 {
 145         long rc = OPAL_BUSY;
 146
 147         pnv_prepare_going_down();
 148
 149         while (rc == OPAL_BUSY || rc == OPAL_BUSY_EVENT) {
 150                 rc = opal_cec_power_down(0);
 151                 if (rc == OPAL_BUSY_EVENT)
 152                         opal_poll_events(NULL);
 153                 else
 154                         mdelay(10);
 155         }
 156         for (;;)
 157                 opal_poll_events(NULL);
 158 }
 159
 160 static void __noreturn pnv_halt(void)
 161 {
 162         pnv_power_off();
 163 }
 164
 165 static void pnv_progress(char *s, unsigned short hex)
 166 {
 167 }
 168
 169 static int pnv_dma_set_mask(struct device *dev, u64 dma_mask)
 170 {
 171         if (dev_is_pci(dev))
 172                 return pnv_pci_dma_set_mask(to_pci_dev(dev), dma_mask);
 173         return __dma_set_mask(dev, dma_mask);
 174 }
 175
 176 static u64 pnv_dma_get_required_mask(struct device *dev)
 177 {
 178         if (dev_is_pci(dev))
 179                 return pnv_pci_dma_get_required_mask(to_pci_dev(dev));
 180
 181         return __dma_get_required_mask(dev);
 182 }
 183
 184 static void pnv_shutdown(void)
 185 {
 186         /* Let the PCI code clear up IODA tables */
 187         pnv_pci_shutdown();
 188
 189         /*
 190          * Stop OPAL activity: Unregister all OPAL interrupts so they
 191          * don't fire up while we kexec and make sure all potentially
 192          * DMA'ing ops are complete (such as dump retrieval).
 193          */
 194         opal_shutdown();
 195 }
 196
 197 #ifdef CONFIG_KEXEC
 198 static void pnv_kexec_wait_secondaries_down(void)
 199 {
 200         int my_cpu, i, notified = -1;
 201
 202         my_cpu = get_cpu();
 203
 204         for_each_online_cpu(i) {
 205                 uint8_t status;
 206                 int64_t rc;
 207
 208                 if (i == my_cpu)
 209                         continue;
 210
 211                 for (;;) {
 212                         rc = opal_query_cpu_status(get_hard_smp_processor_id(i),
 213                                                    &status);
 214                         if (rc != OPAL_SUCCESS || status != OPAL_THREAD_STARTED)
 215                                 break;
 216                         barrier();
 217                         if (i != notified) {
 218                                 printk(KERN_INFO "kexec: waiting for cpu %d "
 219                                        "(physical %d) to enter OPAL\n",
 220                                        i, paca[i].hw_cpu_id);
 221                                 notified = i;
 222                         }
 223                 }
 224         }
 225 }
 226
 227 static void pnv_kexec_cpu_down(int crash_shutdown, int secondary)
 228 {
 229         xics_kexec_teardown_cpu(secondary);
 230
 231         /* On OPAL v3, we return all CPUs to firmware */
 232
 233         if (!firmware_has_feature(FW_FEATURE_OPALv3))
 234                 return;
 235
 236         if (secondary) {
 237                 /* Return secondary CPUs to firmware on OPAL v3 */
 238                 mb();
 239                 get_paca()->kexec_state = KEXEC_STATE_REAL_MODE;
 240                 mb();
 241
 242                 /* Return the CPU to OPAL */
 243                 opal_return_cpu();
 244         } else if (crash_shutdown) {
 245                 /*
 246                  * On crash, we don't wait for secondaries to go
 247                  * down as they might be unreachable or hung, so
 248                  * instead we just wait a bit and move on.
 249                  */
 250                 mdelay(1);
 251         } else {
 252                 /* Primary waits for the secondaries to have reached OPAL */
 253                 pnv_kexec_wait_secondaries_down();
 254         }
 255 }
 256 #endif /* CONFIG_KEXEC */
 257
 258 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 259 static unsigned long pnv_memory_block_size(void)
 260 {
 261         return 256UL * 1024 * 1024;
 262 }
 263 #endif
 264
 265 static void __init pnv_setup_machdep_opal(void)
 266 {
 267         ppc_md.get_boot_time = opal_get_boot_time;
 268         ppc_md.restart = pnv_restart;
 269         pm_power_off = pnv_power_off;
 270         ppc_md.halt = pnv_halt;
 271         ppc_md.machine_check_exception = opal_machine_check;
 272         ppc_md.mce_check_early_recovery = opal_mce_check_early_recovery;
 273         ppc_md.hmi_exception_early = opal_hmi_exception_early;
 274         ppc_md.handle_hmi_exception = opal_handle_hmi_exception;
 275 }
 276
 277 #ifdef CONFIG_PPC_POWERNV_RTAS
 278 static void __init pnv_setup_machdep_rtas(void)
 279 {
 280         if (rtas_token("get-time-of-day") != RTAS_UNKNOWN_SERVICE) {
 281                 ppc_md.get_boot_time = rtas_get_boot_time;
 282                 ppc_md.get_rtc_time = rtas_get_rtc_time;
 283                 ppc_md.set_rtc_time = rtas_set_rtc_time;
 284         }
 285         ppc_md.restart = rtas_restart;
 286         pm_power_off = rtas_power_off;
 287         ppc_md.halt = rtas_halt;
 288 }
 289 #endif /* CONFIG_PPC_POWERNV_RTAS */
 290
 291 static int __init pnv_probe(void)
 292 {
 293         unsigned long root = of_get_flat_dt_root();
 294
 295         if (!of_flat_dt_is_compatible(root, "ibm,powernv"))
 296                 return 0;
 297
 298         hpte_init_native();
 299
 300         if (firmware_has_feature(FW_FEATURE_OPAL))
 301                 pnv_setup_machdep_opal();
 302 #ifdef CONFIG_PPC_POWERNV_RTAS
 303         else if (rtas.base)
 304                 pnv_setup_machdep_rtas();
 305 #endif /* CONFIG_PPC_POWERNV_RTAS */
 306
 307         pr_debug("PowerNV detected !\n");
 308
 309         return 1;
 310 }
 311
 312 /*
 313  * Returns the cpu frequency for 'cpu' in Hz. This is used by
 314  * /proc/cpuinfo
 315  */
 316 static unsigned long pnv_get_proc_freq(unsigned int cpu)
 317 {
 318         unsigned long ret_freq;
 319
 320         ret_freq = cpufreq_quick_get(cpu) * 1000ul;
 321
 322         /*
 323          * If the backend cpufreq driver does not exist,
 324          * then fallback to old way of reporting the clockrate.
 325          */
 326         if (!ret_freq)
 327                 ret_freq = ppc_proc_freq;
 328         return ret_freq;
 329 }
 330
 331 define_machine(powernv) {
 332         .name                   = "PowerNV",
 333         .probe                  = pnv_probe,
 334         .init_early             = pnv_init_early,
 335         .setup_arch             = pnv_setup_arch,
 336         .init_IRQ               = pnv_init_IRQ,
 337         .show_cpuinfo           = pnv_show_cpuinfo,
 338         .get_proc_freq          = pnv_get_proc_freq,
 339         .progress               = pnv_progress,
 340         .machine_shutdown       = pnv_shutdown,
 341         .power_save             = power7_idle,
 342         .calibrate_decr         = generic_calibrate_decr,
 343         .dma_set_mask           = pnv_dma_set_mask,
 344         .dma_get_required_mask  = pnv_dma_get_required_mask,
 345 #ifdef CONFIG_KEXEC
 346         .kexec_cpu_down         = pnv_kexec_cpu_down,
 347 #endif
 348 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 349         .memory_block_size      = pnv_memory_block_size,
 350 #endif
 351 };