arch/powerpc/platforms/pseries/setup.c

   1 /*
   2  *  64-bit pSeries and RS/6000 setup code.
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Adapted from 'alpha' version by Gary Thomas
   6  *  Modified by Cort Dougan (cort@cs.nmt.edu)
   7  *  Modified by PPC64 Team, IBM Corp
   8  *
   9  * This program is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU General Public License
  11  * as published by the Free Software Foundation; either version
  12  * 2 of the License, or (at your option) any later version.
  13  */
  14
  15 /*
  16  * bootup setup stuff..
  17  */
  18
  19 #include <linux/cpu.h>
  20 #include <linux/errno.h>
  21 #include <linux/sched.h>
  22 #include <linux/kernel.h>
  23 #include <linux/mm.h>
  24 #include <linux/stddef.h>
  25 #include <linux/unistd.h>
  26 #include <linux/user.h>
  27 #include <linux/tty.h>
  28 #include <linux/major.h>
  29 #include <linux/interrupt.h>
  30 #include <linux/reboot.h>
  31 #include <linux/init.h>
  32 #include <linux/ioport.h>
  33 #include <linux/console.h>
  34 #include <linux/pci.h>
  35 #include <linux/utsname.h>
  36 #include <linux/adb.h>
  37 #include <linux/export.h>
  38 #include <linux/delay.h>
  39 #include <linux/irq.h>
  40 #include <linux/seq_file.h>
  41 #include <linux/root_dev.h>
  42 #include <linux/of.h>
  43 #include <linux/kexec.h>
  44
  45 #include <asm/mmu.h>
  46 #include <asm/processor.h>
  47 #include <asm/io.h>
  48 #include <asm/pgtable.h>
  49 #include <asm/prom.h>
  50 #include <asm/rtas.h>
  51 #include <asm/pci-bridge.h>
  52 #include <asm/iommu.h>
  53 #include <asm/dma.h>
  54 #include <asm/machdep.h>
  55 #include <asm/irq.h>
  56 #include <asm/time.h>
  57 #include <asm/nvram.h>
  58 #include <asm/pmc.h>
  59 #include <asm/mpic.h>
  60 #include <asm/xics.h>
  61 #include <asm/ppc-pci.h>
  62 #include <asm/i8259.h>
  63 #include <asm/udbg.h>
  64 #include <asm/smp.h>
  65 #include <asm/firmware.h>
  66 #include <asm/eeh.h>
  67 #include <asm/reg.h>
  68 #include <asm/plpar_wrappers.h>
  69
  70 #include "pseries.h"
  71
  72 int CMO_PrPSP = -1;
  73 int CMO_SecPSP = -1;
  74 unsigned long CMO_PageSize = (ASM_CONST(1) << IOMMU_PAGE_SHIFT_4K);
  75 EXPORT_SYMBOL(CMO_PageSize);
  76
  77 int fwnmi_active;  /* TRUE if an FWNMI handler is present */
  78
  79 static struct device_node *pSeries_mpic_node;
  80
  81 static void pSeries_show_cpuinfo(struct seq_file *m)
  82 {
  83         struct device_node *root;
  84         const char *model = "";
  85
  86         root = of_find_node_by_path("/");
  87         if (root)
  88                 model = of_get_property(root, "model", NULL);
  89         seq_printf(m, "machine\t\t: CHRP %s\n", model);
  90         of_node_put(root);
  91 }
  92
  93 /* Initialize firmware assisted non-maskable interrupts if
  94  * the firmware supports this feature.
  95  */
  96 static void __init fwnmi_init(void)
  97 {
  98         unsigned long system_reset_addr, machine_check_addr;
  99
 100         int ibm_nmi_register = rtas_token("ibm,nmi-register");
 101         if (ibm_nmi_register == RTAS_UNKNOWN_SERVICE)
 102                 return;
 103
 104         /* If the kernel's not linked at zero we point the firmware at low
 105          * addresses anyway, and use a trampoline to get to the real code. */
 106         system_reset_addr  = __pa(system_reset_fwnmi) - PHYSICAL_START;
 107         machine_check_addr = __pa(machine_check_fwnmi) - PHYSICAL_START;
 108
 109         if (0 == rtas_call(ibm_nmi_register, 2, 1, NULL, system_reset_addr,
 110                                 machine_check_addr))
 111                 fwnmi_active = 1;
 112 }
 113
 114 static void pseries_8259_cascade(unsigned int irq, struct irq_desc *desc)
 115 {
 116         struct irq_chip *chip = irq_desc_get_chip(desc);
 117         unsigned int cascade_irq = i8259_irq();
 118
 119         if (cascade_irq != NO_IRQ)
 120                 generic_handle_irq(cascade_irq);
 121
 122         chip->irq_eoi(&desc->irq_data);
 123 }
 124
 125 static void __init pseries_setup_i8259_cascade(void)
 126 {
 127         struct device_node *np, *old, *found = NULL;
 128         unsigned int cascade;
 129         const u32 *addrp;
 130         unsigned long intack = 0;
 131         int naddr;
 132
 133         for_each_node_by_type(np, "interrupt-controller") {
 134                 if (of_device_is_compatible(np, "chrp,iic")) {
 135                         found = np;
 136                         break;
 137                 }
 138         }
 139
 140         if (found == NULL) {
 141                 printk(KERN_DEBUG "pic: no ISA interrupt controller\n");
 142                 return;
 143         }
 144
 145         cascade = irq_of_parse_and_map(found, 0);
 146         if (cascade == NO_IRQ) {
 147                 printk(KERN_ERR "pic: failed to map cascade interrupt");
 148                 return;
 149         }
 150         pr_debug("pic: cascade mapped to irq %d\n", cascade);
 151
 152         for (old = of_node_get(found); old != NULL ; old = np) {
 153                 np = of_get_parent(old);
 154                 of_node_put(old);
 155                 if (np == NULL)
 156                         break;
 157                 if (strcmp(np->name, "pci") != 0)
 158                         continue;
 159                 addrp = of_get_property(np, "8259-interrupt-acknowledge", NULL);
 160                 if (addrp == NULL)
 161                         continue;
 162                 naddr = of_n_addr_cells(np);
 163                 intack = addrp[naddr-1];
 164                 if (naddr > 1)
 165                         intack |= ((unsigned long)addrp[naddr-2]) << 32;
 166         }
 167         if (intack)
 168                 printk(KERN_DEBUG "pic: PCI 8259 intack at 0x%016lx\n", intack);
 169         i8259_init(found, intack);
 170         of_node_put(found);
 171         irq_set_chained_handler(cascade, pseries_8259_cascade);
 172 }
 173
 174 static void __init pseries_mpic_init_IRQ(void)
 175 {
 176         struct device_node *np;
 177         const unsigned int *opprop;
 178         unsigned long openpic_addr = 0;
 179         int naddr, n, i, opplen;
 180         struct mpic *mpic;
 181
 182         np = of_find_node_by_path("/");
 183         naddr = of_n_addr_cells(np);
 184         opprop = of_get_property(np, "platform-open-pic", &opplen);
 185         if (opprop != NULL) {
 186                 openpic_addr = of_read_number(opprop, naddr);
 187                 printk(KERN_DEBUG "OpenPIC addr: %lx\n", openpic_addr);
 188         }
 189         of_node_put(np);
 190
 191         BUG_ON(openpic_addr == 0);
 192
 193         /* Setup the openpic driver */
 194         mpic = mpic_alloc(pSeries_mpic_node, openpic_addr,
 195                         MPIC_NO_RESET, 16, 0, " MPIC     ");
 196         BUG_ON(mpic == NULL);
 197
 198         /* Add ISUs */
 199         opplen /= sizeof(u32);
 200         for (n = 0, i = naddr; i < opplen; i += naddr, n++) {
 201                 unsigned long isuaddr = of_read_number(opprop + i, naddr);
 202                 mpic_assign_isu(mpic, n, isuaddr);
 203         }
 204
 205         /* Setup top-level get_irq */
 206         ppc_md.get_irq = mpic_get_irq;
 207
 208         /* All ISUs are setup, complete initialization */
 209         mpic_init(mpic);
 210
 211         /* Look for cascade */
 212         pseries_setup_i8259_cascade();
 213 }
 214
 215 static void __init pseries_xics_init_IRQ(void)
 216 {
 217         xics_init();
 218         pseries_setup_i8259_cascade();
 219 }
 220
 221 static void pseries_lpar_enable_pmcs(void)
 222 {
 223         unsigned long set, reset;
 224
 225         set = 1UL << 63;
 226         reset = 0;
 227         plpar_hcall_norets(H_PERFMON, set, reset);
 228 }
 229
 230 static void __init pseries_discover_pic(void)
 231 {
 232         struct device_node *np;
 233         const char *typep;
 234
 235         for_each_node_by_name(np, "interrupt-controller") {
 236                 typep = of_get_property(np, "compatible", NULL);
 237                 if (strstr(typep, "open-pic")) {
 238                         pSeries_mpic_node = of_node_get(np);
 239                         ppc_md.init_IRQ       = pseries_mpic_init_IRQ;
 240                         setup_kexec_cpu_down_mpic();
 241                         smp_init_pseries_mpic();
 242                         return;
 243                 } else if (strstr(typep, "ppc-xicp")) {
 244                         ppc_md.init_IRQ       = pseries_xics_init_IRQ;
 245                         setup_kexec_cpu_down_xics();
 246                         smp_init_pseries_xics();
 247                         return;
 248                 }
 249         }
 250         printk(KERN_ERR "pSeries_discover_pic: failed to recognize"
 251                " interrupt-controller\n");
 252 }
 253
 254 static int pci_dn_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *node)
 255 {
 256         struct device_node *np = node;
 257         struct pci_dn *pci = NULL;
 258         int err = NOTIFY_OK;
 259
 260         switch (action) {
 261         case OF_RECONFIG_ATTACH_NODE:
 262                 pci = np->parent->data;
 263                 if (pci) {
 264                         update_dn_pci_info(np, pci->phb);
 265
 266                         /* Create EEH device for the OF node */
 267                         eeh_dev_init(np, pci->phb);
 268                 }
 269                 break;
 270         default:
 271                 err = NOTIFY_DONE;
 272                 break;
 273         }
 274         return err;
 275 }
 276
 277 static struct notifier_block pci_dn_reconfig_nb = {
 278         .notifier_call = pci_dn_reconfig_notifier,
 279 };
 280
 281 struct kmem_cache *dtl_cache;
 282
 283 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 284 /*
 285  * Allocate space for the dispatch trace log for all possible cpus
 286  * and register the buffers with the hypervisor.  This is used for
 287  * computing time stolen by the hypervisor.
 288  */
 289 static int alloc_dispatch_logs(void)
 290 {
 291         int cpu, ret;
 292         struct paca_struct *pp;
 293         struct dtl_entry *dtl;
 294
 295         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
 296                 return 0;
 297
 298         if (!dtl_cache)
 299                 return 0;
 300
 301         for_each_possible_cpu(cpu) {
 302                 pp = &paca[cpu];
 303                 dtl = kmem_cache_alloc(dtl_cache, GFP_KERNEL);
 304                 if (!dtl) {
 305                         pr_warn("Failed to allocate dispatch trace log for cpu %d\n",
 306                                 cpu);
 307                         pr_warn("Stolen time statistics will be unreliable\n");
 308                         break;
 309                 }
 310
 311                 pp->dtl_ridx = 0;
 312                 pp->dispatch_log = dtl;
 313                 pp->dispatch_log_end = dtl + N_DISPATCH_LOG;
 314                 pp->dtl_curr = dtl;
 315         }
 316
 317         /* Register the DTL for the current (boot) cpu */
 318         dtl = get_paca()->dispatch_log;
 319         get_paca()->dtl_ridx = 0;
 320         get_paca()->dtl_curr = dtl;
 321         get_paca()->lppaca_ptr->dtl_idx = 0;
 322
 323         /* hypervisor reads buffer length from this field */
 324         dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
 325         ret = register_dtl(hard_smp_processor_id(), __pa(dtl));
 326         if (ret)
 327                 pr_err("WARNING: DTL registration of cpu %d (hw %d) failed "
 328                        "with %d\n", smp_processor_id(),
 329                        hard_smp_processor_id(), ret);
 330         get_paca()->lppaca_ptr->dtl_enable_mask = 2;
 331
 332         return 0;
 333 }
 334 #else /* !CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 335 static inline int alloc_dispatch_logs(void)
 336 {
 337         return 0;
 338 }
 339 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE */
 340
 341 static int alloc_dispatch_log_kmem_cache(void)
 342 {
 343         dtl_cache = kmem_cache_create("dtl", DISPATCH_LOG_BYTES,
 344                                                 DISPATCH_LOG_BYTES, 0, NULL);
 345         if (!dtl_cache) {
 346                 pr_warn("Failed to create dispatch trace log buffer cache\n");
 347                 pr_warn("Stolen time statistics will be unreliable\n");
 348                 return 0;
 349         }
 350
 351         return alloc_dispatch_logs();
 352 }
 353 machine_early_initcall(pseries, alloc_dispatch_log_kmem_cache);
 354
 355 static void pseries_lpar_idle(void)
 356 {
 357         /*
 358          * Default handler to go into low thread priority and possibly
 359          * low power mode by cedeing processor to hypervisor
 360          */
 361
 362         /* Indicate to hypervisor that we are idle. */
 363         get_lppaca()->idle = 1;
 364
 365         /*
 366          * Yield the processor to the hypervisor.  We return if
 367          * an external interrupt occurs (which are driven prior
 368          * to returning here) or if a prod occurs from another
 369          * processor. When returning here, external interrupts
 370          * are enabled.
 371          */
 372         cede_processor();
 373
 374         get_lppaca()->idle = 0;
 375 }
 376
 377 /*
 378  * Enable relocation on during exceptions. This has partition wide scope and
 379  * may take a while to complete, if it takes longer than one second we will
 380  * just give up rather than wasting any more time on this - if that turns out
 381  * to ever be a problem in practice we can move this into a kernel thread to
 382  * finish off the process later in boot.
 383  */
 384 long pSeries_enable_reloc_on_exc(void)
 385 {
 386         long rc;
 387         unsigned int delay, total_delay = 0;
 388
 389         while (1) {
 390                 rc = enable_reloc_on_exceptions();
 391                 if (!H_IS_LONG_BUSY(rc))
 392                         return rc;
 393
 394                 delay = get_longbusy_msecs(rc);
 395                 total_delay += delay;
 396                 if (total_delay > 1000) {
 397                         pr_warn("Warning: Giving up waiting to enable "
 398                                 "relocation on exceptions (%u msec)!\n",
 399                                 total_delay);
 400                         return rc;
 401                 }
 402
 403                 mdelay(delay);
 404         }
 405 }
 406 EXPORT_SYMBOL(pSeries_enable_reloc_on_exc);
 407
 408 long pSeries_disable_reloc_on_exc(void)
 409 {
 410         long rc;
 411
 412         while (1) {
 413                 rc = disable_reloc_on_exceptions();
 414                 if (!H_IS_LONG_BUSY(rc))
 415                         return rc;
 416                 mdelay(get_longbusy_msecs(rc));
 417         }
 418 }
 419 EXPORT_SYMBOL(pSeries_disable_reloc_on_exc);
 420
 421 #ifdef CONFIG_KEXEC
 422 static void pSeries_machine_kexec(struct kimage *image)
 423 {
 424         long rc;
 425
 426         if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
 427                 rc = pSeries_disable_reloc_on_exc();
 428                 if (rc != H_SUCCESS)
 429                         pr_warning("Warning: Failed to disable relocation on "
 430                                    "exceptions: %ld\n", rc);
 431         }
 432
 433         default_machine_kexec(image);
 434 }
 435 #endif
 436
 437 #ifdef __LITTLE_ENDIAN__
 438 long pseries_big_endian_exceptions(void)
 439 {
 440         long rc;
 441
 442         while (1) {
 443                 rc = enable_big_endian_exceptions();
 444                 if (!H_IS_LONG_BUSY(rc))
 445                         return rc;
 446                 mdelay(get_longbusy_msecs(rc));
 447         }
 448 }
 449
 450 static long pseries_little_endian_exceptions(void)
 451 {
 452         long rc;
 453
 454         while (1) {
 455                 rc = enable_little_endian_exceptions();
 456                 if (!H_IS_LONG_BUSY(rc))
 457                         return rc;
 458                 mdelay(get_longbusy_msecs(rc));
 459         }
 460 }
 461 #endif
 462
 463 static void __init pSeries_setup_arch(void)
 464 {
 465         set_arch_panic_timeout(10, ARCH_PANIC_TIMEOUT);
 466
 467         /* Discover PIC type and setup ppc_md accordingly */
 468         pseries_discover_pic();
 469
 470         /* openpic global configuration register (64-bit format). */
 471         /* openpic Interrupt Source Unit pointer (64-bit format). */
 472         /* python0 facility area (mmio) (64-bit format) REAL address. */
 473
 474         /* init to some ~sane value until calibrate_delay() runs */
 475         loops_per_jiffy = 50000000;
 476
 477         fwnmi_init();
 478
 479         /* By default, only probe PCI (can be overriden by rtas_pci) */
 480         pci_add_flags(PCI_PROBE_ONLY);
 481
 482         /* Find and initialize PCI host bridges */
 483         init_pci_config_tokens();
 484         find_and_init_phbs();
 485         of_reconfig_notifier_register(&pci_dn_reconfig_nb);
 486
 487         pSeries_nvram_init();
 488
 489         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 490                 vpa_init(boot_cpuid);
 491                 ppc_md.power_save = pseries_lpar_idle;
 492                 ppc_md.enable_pmcs = pseries_lpar_enable_pmcs;
 493         } else {
 494                 /* No special idle routine */
 495                 ppc_md.enable_pmcs = power4_enable_pmcs;
 496         }
 497
 498         ppc_md.pcibios_root_bridge_prepare = pseries_root_bridge_prepare;
 499
 500         if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
 501                 long rc;
 502                 if ((rc = pSeries_enable_reloc_on_exc()) != H_SUCCESS) {
 503                         pr_warn("Unable to enable relocation on exceptions: "
 504                                 "%ld\n", rc);
 505                 }
 506         }
 507 }
 508
 509 static int __init pSeries_init_panel(void)
 510 {
 511         /* Manually leave the kernel version on the panel. */
 512 #ifdef __BIG_ENDIAN__
 513         ppc_md.progress("Linux ppc64\n", 0);
 514 #else
 515         ppc_md.progress("Linux ppc64le\n", 0);
 516 #endif
 517         ppc_md.progress(init_utsname()->version, 0);
 518
 519         return 0;
 520 }
 521 machine_arch_initcall(pseries, pSeries_init_panel);
 522
 523 static int pseries_set_dabr(unsigned long dabr, unsigned long dabrx)
 524 {
 525         return plpar_hcall_norets(H_SET_DABR, dabr);
 526 }
 527
 528 static int pseries_set_xdabr(unsigned long dabr, unsigned long dabrx)
 529 {
 530         /* Have to set at least one bit in the DABRX according to PAPR */
 531         if (dabrx == 0 && dabr == 0)
 532                 dabrx = DABRX_USER;
 533         /* PAPR says we can only set kernel and user bits */
 534         dabrx &= DABRX_KERNEL | DABRX_USER;
 535
 536         return plpar_hcall_norets(H_SET_XDABR, dabr, dabrx);
 537 }
 538
 539 static int pseries_set_dawr(unsigned long dawr, unsigned long dawrx)
 540 {
 541         /* PAPR says we can't set HYP */
 542         dawrx &= ~DAWRX_HYP;
 543
 544         return  plapr_set_watchpoint0(dawr, dawrx);
 545 }
 546
 547 #define CMO_CHARACTERISTICS_TOKEN 44
 548 #define CMO_MAXLENGTH 1026
 549
 550 void pSeries_coalesce_init(void)
 551 {
 552         struct hvcall_mpp_x_data mpp_x_data;
 553
 554         if (firmware_has_feature(FW_FEATURE_CMO) && !h_get_mpp_x(&mpp_x_data))
 555                 powerpc_firmware_features |= FW_FEATURE_XCMO;
 556         else
 557                 powerpc_firmware_features &= ~FW_FEATURE_XCMO;
 558 }
 559
 560 /**
 561  * fw_cmo_feature_init - FW_FEATURE_CMO is not stored in ibm,hypertas-functions,
 562  * handle that here. (Stolen from parse_system_parameter_string)
 563  */
 564 void pSeries_cmo_feature_init(void)
 565 {
 566         char *ptr, *key, *value, *end;
 567         int call_status;
 568         int page_order = IOMMU_PAGE_SHIFT_4K;
 569
 570         pr_debug(" -> fw_cmo_feature_init()\n");
 571         spin_lock(&rtas_data_buf_lock);
 572         memset(rtas_data_buf, 0, RTAS_DATA_BUF_SIZE);
 573         call_status = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
 574                                 NULL,
 575                                 CMO_CHARACTERISTICS_TOKEN,
 576                                 __pa(rtas_data_buf),
 577                                 RTAS_DATA_BUF_SIZE);
 578
 579         if (call_status != 0) {
 580                 spin_unlock(&rtas_data_buf_lock);
 581                 pr_debug("CMO not available\n");
 582                 pr_debug(" <- fw_cmo_feature_init()\n");
 583                 return;
 584         }
 585
 586         end = rtas_data_buf + CMO_MAXLENGTH - 2;
 587         ptr = rtas_data_buf + 2;        /* step over strlen value */
 588         key = value = ptr;
 589
 590         while (*ptr && (ptr <= end)) {
 591                 /* Separate the key and value by replacing '=' with '\0' and
 592                  * point the value at the string after the '='
 593                  */
 594                 if (ptr[0] == '=') {
 595                         ptr[0] = '\0';
 596                         value = ptr + 1;
 597                 } else if (ptr[0] == '\0' || ptr[0] == ',') {
 598                         /* Terminate the string containing the key/value pair */
 599                         ptr[0] = '\0';
 600
 601                         if (key == value) {
 602                                 pr_debug("Malformed key/value pair\n");
 603                                 /* Never found a '=', end processing */
 604                                 break;
 605                         }
 606
 607                         if (0 == strcmp(key, "CMOPageSize"))
 608                                 page_order = simple_strtol(value, NULL, 10);
 609                         else if (0 == strcmp(key, "PrPSP"))
 610                                 CMO_PrPSP = simple_strtol(value, NULL, 10);
 611                         else if (0 == strcmp(key, "SecPSP"))
 612                                 CMO_SecPSP = simple_strtol(value, NULL, 10);
 613                         value = key = ptr + 1;
 614                 }
 615                 ptr++;
 616         }
 617
 618         /* Page size is returned as the power of 2 of the page size,
 619          * convert to the page size in bytes before returning
 620          */
 621         CMO_PageSize = 1 << page_order;
 622         pr_debug("CMO_PageSize = %lu\n", CMO_PageSize);
 623
 624         if (CMO_PrPSP != -1 || CMO_SecPSP != -1) {
 625                 pr_info("CMO enabled\n");
 626                 pr_debug("CMO enabled, PrPSP=%d, SecPSP=%d\n", CMO_PrPSP,
 627                          CMO_SecPSP);
 628                 powerpc_firmware_features |= FW_FEATURE_CMO;
 629                 pSeries_coalesce_init();
 630         } else
 631                 pr_debug("CMO not enabled, PrPSP=%d, SecPSP=%d\n", CMO_PrPSP,
 632                          CMO_SecPSP);
 633         spin_unlock(&rtas_data_buf_lock);
 634         pr_debug(" <- fw_cmo_feature_init()\n");
 635 }
 636
 637 /*
 638  * Early initialization.  Relocation is on but do not reference unbolted pages
 639  */
 640 static void __init pSeries_init_early(void)
 641 {
 642         pr_debug(" -> pSeries_init_early()\n");
 643
 644 #ifdef CONFIG_HVC_CONSOLE
 645         if (firmware_has_feature(FW_FEATURE_LPAR))
 646                 hvc_vio_init_early();
 647 #endif
 648         if (firmware_has_feature(FW_FEATURE_XDABR))
 649                 ppc_md.set_dabr = pseries_set_xdabr;
 650         else if (firmware_has_feature(FW_FEATURE_DABR))
 651                 ppc_md.set_dabr = pseries_set_dabr;
 652
 653         if (firmware_has_feature(FW_FEATURE_SET_MODE))
 654                 ppc_md.set_dawr = pseries_set_dawr;
 655
 656         pSeries_cmo_feature_init();
 657         iommu_init_early_pSeries();
 658
 659         pr_debug(" <- pSeries_init_early()\n");
 660 }
 661
 662 /*
 663  * Called very early, MMU is off, device-tree isn't unflattened
 664  */
 665
 666 static int __init pseries_probe_fw_features(unsigned long node,
 667                                             const char *uname, int depth,
 668                                             void *data)
 669 {
 670         const char *prop;
 671         int len;
 672         static int hypertas_found;
 673         static int vec5_found;
 674
 675         if (depth != 1)
 676                 return 0;
 677
 678         if (!strcmp(uname, "rtas") || !strcmp(uname, "rtas@0")) {
 679                 prop = of_get_flat_dt_prop(node, "ibm,hypertas-functions",
 680                                            &len);
 681                 if (prop) {
 682                         powerpc_firmware_features |= FW_FEATURE_LPAR;
 683                         fw_hypertas_feature_init(prop, len);
 684                 }
 685
 686                 hypertas_found = 1;
 687         }
 688
 689         if (!strcmp(uname, "chosen")) {
 690                 prop = of_get_flat_dt_prop(node, "ibm,architecture-vec-5",
 691                                            &len);
 692                 if (prop)
 693                         fw_vec5_feature_init(prop, len);
 694
 695                 vec5_found = 1;
 696         }
 697
 698         return hypertas_found && vec5_found;
 699 }
 700
 701 static int __init pSeries_probe(void)
 702 {
 703         unsigned long root = of_get_flat_dt_root();
 704         const char *dtype = of_get_flat_dt_prop(root, "device_type", NULL);
 705
 706         if (dtype == NULL)
 707                 return 0;
 708         if (strcmp(dtype, "chrp"))
 709                 return 0;
 710
 711         /* Cell blades firmware claims to be chrp while it's not. Until this
 712          * is fixed, we need to avoid those here.
 713          */
 714         if (of_flat_dt_is_compatible(root, "IBM,CPBW-1.0") ||
 715             of_flat_dt_is_compatible(root, "IBM,CBEA"))
 716                 return 0;
 717
 718         pr_debug("pSeries detected, looking for LPAR capability...\n");
 719
 720         /* Now try to figure out if we are running on LPAR */
 721         of_scan_flat_dt(pseries_probe_fw_features, NULL);
 722
 723 #ifdef __LITTLE_ENDIAN__
 724         if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
 725                 long rc;
 726                 /*
 727                  * Tell the hypervisor that we want our exceptions to
 728                  * be taken in little endian mode. If this fails we don't
 729                  * want to use BUG() because it will trigger an exception.
 730                  */
 731                 rc = pseries_little_endian_exceptions();
 732                 if (rc) {
 733                         ppc_md.progress("H_SET_MODE LE exception fail", 0);
 734                         panic("Could not enable little endian exceptions");
 735                 }
 736         }
 737 #endif
 738
 739         if (firmware_has_feature(FW_FEATURE_LPAR))
 740                 hpte_init_lpar();
 741         else
 742                 hpte_init_native();
 743
 744         pr_debug("Machine is%s LPAR !\n",
 745                  (powerpc_firmware_features & FW_FEATURE_LPAR) ? "" : " not");
 746
 747         return 1;
 748 }
 749
 750 static int pSeries_pci_probe_mode(struct pci_bus *bus)
 751 {
 752         if (firmware_has_feature(FW_FEATURE_LPAR))
 753                 return PCI_PROBE_DEVTREE;
 754         return PCI_PROBE_NORMAL;
 755 }
 756
 757 /**
 758  * pSeries_power_off - tell firmware about how to power off the system.
 759  *
 760  * This function calls either the power-off rtas token in normal cases
 761  * or the ibm,power-off-ups token (if present & requested) in case of
 762  * a power failure. If power-off token is used, power on will only be
 763  * possible with power button press. If ibm,power-off-ups token is used
 764  * it will allow auto poweron after power is restored.
 765  */
 766 static void pSeries_power_off(void)
 767 {
 768         int rc;
 769         int rtas_poweroff_ups_token = rtas_token("ibm,power-off-ups");
 770
 771         if (rtas_flash_term_hook)
 772                 rtas_flash_term_hook(SYS_POWER_OFF);
 773
 774         if (rtas_poweron_auto == 0 ||
 775                 rtas_poweroff_ups_token == RTAS_UNKNOWN_SERVICE) {
 776                 rc = rtas_call(rtas_token("power-off"), 2, 1, NULL, -1, -1);
 777                 printk(KERN_INFO "RTAS power-off returned %d\n", rc);
 778         } else {
 779                 rc = rtas_call(rtas_poweroff_ups_token, 0, 1, NULL);
 780                 printk(KERN_INFO "RTAS ibm,power-off-ups returned %d\n", rc);
 781         }
 782         for (;;);
 783 }
 784
 785 #ifndef CONFIG_PCI
 786 void pSeries_final_fixup(void) { }
 787 #endif
 788
 789 define_machine(pseries) {
 790         .name                   = "pSeries",
 791         .probe                  = pSeries_probe,
 792         .setup_arch             = pSeries_setup_arch,
 793         .init_early             = pSeries_init_early,
 794         .show_cpuinfo           = pSeries_show_cpuinfo,
 795         .log_error              = pSeries_log_error,
 796         .pcibios_fixup          = pSeries_final_fixup,
 797         .pci_probe_mode         = pSeries_pci_probe_mode,
 798         .restart                = rtas_restart,
 799         .power_off              = pSeries_power_off,
 800         .halt                   = rtas_halt,
 801         .panic                  = rtas_os_term,
 802         .get_boot_time          = rtas_get_boot_time,
 803         .get_rtc_time           = rtas_get_rtc_time,
 804         .set_rtc_time           = rtas_set_rtc_time,
 805         .calibrate_decr         = generic_calibrate_decr,
 806         .progress               = rtas_progress,
 807         .system_reset_exception = pSeries_system_reset_exception,
 808         .machine_check_exception = pSeries_machine_check_exception,
 809 #ifdef CONFIG_KEXEC
 810         .machine_kexec          = pSeries_machine_kexec,
 811 #endif
 812 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
 813         .memory_block_size      = pseries_memory_block_size,
 814 #endif
 815 };