sys/ia64/ia64/machdep.c

   1 /*-
   2  * Copyright (c) 2003,2004 Marcel Moolenaar
   3  * Copyright (c) 2000,2001 Doug Rabson
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 #include <sys/cdefs.h>
  29 __FBSDID("$FreeBSD$");
  30
  31 #include "opt_compat.h"
  32 #include "opt_ddb.h"
  33 #include "opt_kstack_pages.h"
  34 #include "opt_msgbuf.h"
  35 #include "opt_sched.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/proc.h>
  39 #include <sys/systm.h>
  40 #include <sys/bio.h>
  41 #include <sys/buf.h>
  42 #include <sys/bus.h>
  43 #include <sys/cons.h>
  44 #include <sys/cpu.h>
  45 #include <sys/eventhandler.h>
  46 #include <sys/exec.h>
  47 #include <sys/imgact.h>
  48 #include <sys/kdb.h>
  49 #include <sys/kernel.h>
  50 #include <sys/linker.h>
  51 #include <sys/lock.h>
  52 #include <sys/malloc.h>
  53 #include <sys/mbuf.h>
  54 #include <sys/msgbuf.h>
  55 #include <sys/pcpu.h>
  56 #include <sys/ptrace.h>
  57 #include <sys/random.h>
  58 #include <sys/reboot.h>
  59 #include <sys/sched.h>
  60 #include <sys/signalvar.h>
  61 #include <sys/syscall.h>
  62 #include <sys/sysctl.h>
  63 #include <sys/sysproto.h>
  64 #include <sys/ucontext.h>
  65 #include <sys/uio.h>
  66 #include <sys/uuid.h>
  67 #include <sys/vmmeter.h>
  68 #include <sys/vnode.h>
  69
  70 #include <ddb/ddb.h>
  71
  72 #include <net/netisr.h>
  73
  74 #include <vm/vm.h>
  75 #include <vm/vm_extern.h>
  76 #include <vm/vm_kern.h>
  77 #include <vm/vm_page.h>
  78 #include <vm/vm_map.h>
  79 #include <vm/vm_object.h>
  80 #include <vm/vm_pager.h>
  81
  82 #include <machine/bootinfo.h>
  83 #include <machine/clock.h>
  84 #include <machine/cpu.h>
  85 #include <machine/efi.h>
  86 #include <machine/elf.h>
  87 #include <machine/fpu.h>
  88 #include <machine/mca.h>
  89 #include <machine/md_var.h>
  90 #include <machine/mutex.h>
  91 #include <machine/pal.h>
  92 #include <machine/pcb.h>
  93 #include <machine/reg.h>
  94 #include <machine/sal.h>
  95 #include <machine/sigframe.h>
  96 #ifdef SMP
  97 #include <machine/smp.h>
  98 #endif
  99 #include <machine/unwind.h>
 100 #include <machine/vmparam.h>
 101
 102 #include <i386/include/specialreg.h>
 103
 104 u_int64_t processor_frequency;
 105 u_int64_t bus_frequency;
 106 u_int64_t itc_frequency;
 107 int cold = 1;
 108
 109 u_int64_t pa_bootinfo;
 110 struct bootinfo bootinfo;
 111
 112 struct pcpu pcpu0;
 113
 114 extern u_int64_t kernel_text[], _end[];
 115
 116 extern u_int64_t ia64_gateway_page[];
 117 extern u_int64_t break_sigtramp[];
 118 extern u_int64_t epc_sigtramp[];
 119
 120 struct fpswa_iface *fpswa_iface;
 121
 122 u_int64_t ia64_pal_base;
 123 u_int64_t ia64_port_base;
 124
 125 static int ia64_sync_icache_needed;
 126
 127 char machine[] = MACHINE;
 128 SYSCTL_STRING(_hw, HW_MACHINE, machine, CTLFLAG_RD, machine, 0, "");
 129
 130 static char cpu_model[64];
 131 SYSCTL_STRING(_hw, HW_MODEL, model, CTLFLAG_RD, cpu_model, 0,
 132     "The CPU model name");
 133
 134 static char cpu_family[64];
 135 SYSCTL_STRING(_hw, OID_AUTO, family, CTLFLAG_RD, cpu_family, 0,
 136     "The CPU family name");
 137
 138 #ifdef DDB
 139 extern vm_offset_t ksym_start, ksym_end;
 140 #endif
 141
 142 static void cpu_startup(void *);
 143 SYSINIT(cpu, SI_SUB_CPU, SI_ORDER_FIRST, cpu_startup, NULL);
 144
 145 struct msgbuf *msgbufp = NULL;
 146
 147 /* Other subsystems (e.g., ACPI) can hook this later. */
 148 void (*cpu_idle_hook)(void) = NULL;
 149
 150 long Maxmem = 0;
 151 long realmem = 0;
 152
 153 #define PHYSMAP_SIZE    (2 * VM_PHYSSEG_MAX)
 154
 155 vm_paddr_t phys_avail[PHYSMAP_SIZE + 2];
 156
 157 /* must be 2 less so 0 0 can signal end of chunks */
 158 #define PHYS_AVAIL_ARRAY_END ((sizeof(phys_avail) / sizeof(vm_offset_t)) - 2)
 159
 160 struct kva_md_info kmi;
 161
 162 #define Mhz     1000000L
 163 #define Ghz     (1000L*Mhz)
 164
 165 static void
 166 identifycpu(void)
 167 {
 168         char vendor[17];
 169         char *family_name, *model_name;
 170         u_int64_t features, tmp;
 171         int number, revision, model, family, archrev;
 172
 173         /*
 174          * Assumes little-endian.
 175          */
 176         *(u_int64_t *) &vendor[0] = ia64_get_cpuid(0);
 177         *(u_int64_t *) &vendor[8] = ia64_get_cpuid(1);
 178         vendor[16] = '\0';
 179
 180         tmp = ia64_get_cpuid(3);
 181         number = (tmp >> 0) & 0xff;
 182         revision = (tmp >> 8) & 0xff;
 183         model = (tmp >> 16) & 0xff;
 184         family = (tmp >> 24) & 0xff;
 185         archrev = (tmp >> 32) & 0xff;
 186
 187         family_name = model_name = "unknown";
 188         switch (family) {
 189         case 0x07:
 190                 family_name = "Itanium";
 191                 model_name = "Merced";
 192                 break;
 193         case 0x1f:
 194                 family_name = "Itanium 2";
 195                 switch (model) {
 196                 case 0x00:
 197                         model_name = "McKinley";
 198                         break;
 199                 case 0x01:
 200                         /*
 201                          * Deerfield is a low-voltage variant based on the
 202                          * Madison core. We need circumstantial evidence
 203                          * (i.e. the clock frequency) to identify those.
 204                          * Allow for roughly 1% error margin.
 205                          */
 206                         tmp = processor_frequency >> 7;
 207                         if ((processor_frequency - tmp) < 1*Ghz &&
 208                             (processor_frequency + tmp) >= 1*Ghz)
 209                                 model_name = "Deerfield";
 210                         else
 211                                 model_name = "Madison";
 212                         break;
 213                 case 0x02:
 214                         model_name = "Madison II";
 215                         break;
 216                 }
 217                 break;
 218         case 0x20:
 219                 ia64_sync_icache_needed = 1;
 220
 221                 family_name = "Itanium 2";
 222                 switch (model) {
 223                 case 0x00:
 224                         model_name = "Montecito";
 225                         break;
 226                 }
 227                 break;
 228         }
 229         snprintf(cpu_family, sizeof(cpu_family), "%s", family_name);
 230         snprintf(cpu_model, sizeof(cpu_model), "%s", model_name);
 231
 232         features = ia64_get_cpuid(4);
 233
 234         printf("CPU: %s (", model_name);
 235         if (processor_frequency) {
 236                 printf("%ld.%02ld-Mhz ",
 237                     (processor_frequency + 4999) / Mhz,
 238                     ((processor_frequency + 4999) / (Mhz/100)) % 100);
 239         }
 240         printf("%s)\n", family_name);
 241         printf("  Origin = \"%s\"  Revision = %d\n", vendor, revision);
 242         printf("  Features = 0x%b\n", (u_int32_t) features,
 243             "\020"
 244             "\001LB"    /* long branch (brl) instruction. */
 245             "\002SD"    /* Spontaneous deferral. */
 246             "\003AO"    /* 16-byte atomic operations (ld, st, cmpxchg). */ );
 247 }
 248
 249 static void
 250 cpu_startup(dummy)
 251         void *dummy;
 252 {
 253
 254         /*
 255          * Good {morning,afternoon,evening,night}.
 256          */
 257         identifycpu();
 258
 259 #ifdef PERFMON
 260         perfmon_init();
 261 #endif
 262         printf("real memory  = %ld (%ld MB)\n", ia64_ptob(Maxmem),
 263             ia64_ptob(Maxmem) / 1048576);
 264         realmem = Maxmem;
 265
 266         /*
 267          * Display any holes after the first chunk of extended memory.
 268          */
 269         if (bootverbose) {
 270                 int indx;
 271
 272                 printf("Physical memory chunk(s):\n");
 273                 for (indx = 0; phys_avail[indx + 1] != 0; indx += 2) {
 274                         long size1 = phys_avail[indx + 1] - phys_avail[indx];
 275
 276                         printf("0x%08lx - 0x%08lx, %ld bytes (%ld pages)\n",
 277                             phys_avail[indx], phys_avail[indx + 1] - 1, size1,
 278                             size1 >> PAGE_SHIFT);
 279                 }
 280         }
 281
 282         vm_ksubmap_init(&kmi);
 283
 284         printf("avail memory = %ld (%ld MB)\n", ptoa(cnt.v_free_count),
 285             ptoa(cnt.v_free_count) / 1048576);
 286
 287         if (fpswa_iface == NULL)
 288                 printf("Warning: no FPSWA package supplied\n");
 289         else
 290                 printf("FPSWA Revision = 0x%lx, Entry = %p\n",
 291                     (long)fpswa_iface->if_rev, (void *)fpswa_iface->if_fpswa);
 292
 293         /*
 294          * Set up buffers, so they can be used to read disk labels.
 295          */
 296         bufinit();
 297         vm_pager_bufferinit();
 298
 299         /*
 300          * Traverse the MADT to discover IOSAPIC and Local SAPIC
 301          * information.
 302          */
 303         ia64_probe_sapics();
 304         ia64_mca_init();
 305 }
 306
 307 void
 308 cpu_boot(int howto)
 309 {
 310
 311         efi_reset_system();
 312 }
 313
 314 void
 315 cpu_flush_dcache(void *ptr, size_t len)
 316 {
 317         vm_offset_t lim, va;
 318
 319         va = (uintptr_t)ptr & ~31;
 320         lim = (uintptr_t)ptr + len;
 321         while (va < lim) {
 322                 ia64_fc(va);
 323                 va += 32;
 324         }
 325
 326         ia64_srlz_d();
 327 }
 328
 329 /* Get current clock frequency for the given cpu id. */
 330 int
 331 cpu_est_clockrate(int cpu_id, uint64_t *rate)
 332 {
 333
 334         if (pcpu_find(cpu_id) == NULL || rate == NULL)
 335                 return (EINVAL);
 336         *rate = processor_frequency;
 337         return (0);
 338 }
 339
 340 void
 341 cpu_halt()
 342 {
 343
 344         efi_reset_system();
 345 }
 346
 347 void
 348 cpu_idle(int busy)
 349 {
 350         struct ia64_pal_result res;
 351
 352         if (cpu_idle_hook != NULL)
 353                 (*cpu_idle_hook)();
 354         else
 355                 res = ia64_call_pal_static(PAL_HALT_LIGHT, 0, 0, 0);
 356 }
 357
 358 int
 359 cpu_idle_wakeup(int cpu)
 360 {
 361
 362         return (0);
 363 }
 364
 365 void
 366 cpu_reset()
 367 {
 368
 369         cpu_boot(0);
 370 }
 371
 372 void
 373 cpu_switch(struct thread *old, struct thread *new, struct mtx *mtx)
 374 {
 375         struct pcb *oldpcb, *newpcb;
 376
 377         oldpcb = old->td_pcb;
 378 #ifdef COMPAT_IA32
 379         ia32_savectx(oldpcb);
 380 #endif
 381         if (PCPU_GET(fpcurthread) == old)
 382                 old->td_frame->tf_special.psr |= IA64_PSR_DFH;
 383         if (!savectx(oldpcb)) {
 384                 old->td_lock = mtx;
 385 #if defined(SCHED_ULE) && defined(SMP)
 386                 /* td_lock is volatile */
 387                 while (new->td_lock == &blocked_lock)
 388                         ;
 389 #endif
 390                 newpcb = new->td_pcb;
 391                 oldpcb->pcb_current_pmap =
 392                     pmap_switch(newpcb->pcb_current_pmap);
 393                 PCPU_SET(curthread, new);
 394 #ifdef COMPAT_IA32
 395                 ia32_restorectx(newpcb);
 396 #endif
 397                 if (PCPU_GET(fpcurthread) == new)
 398                         new->td_frame->tf_special.psr &= ~IA64_PSR_DFH;
 399                 restorectx(newpcb);
 400                 /* We should not get here. */
 401                 panic("cpu_switch: restorectx() returned");
 402                 /* NOTREACHED */
 403         }
 404 }
 405
 406 void
 407 cpu_throw(struct thread *old __unused, struct thread *new)
 408 {
 409         struct pcb *newpcb;
 410
 411         newpcb = new->td_pcb;
 412         (void)pmap_switch(newpcb->pcb_current_pmap);
 413         PCPU_SET(curthread, new);
 414 #ifdef COMPAT_IA32
 415         ia32_restorectx(newpcb);
 416 #endif
 417         restorectx(newpcb);
 418         /* We should not get here. */
 419         panic("cpu_throw: restorectx() returned");
 420         /* NOTREACHED */
 421 }
 422
 423 void
 424 cpu_pcpu_init(struct pcpu *pcpu, int cpuid, size_t size)
 425 {
 426
 427         /*
 428          * Set pc_acpi_id to "uninitialized".
 429          * See sys/dev/acpica/acpi_cpu.c
 430          */
 431         pcpu->pc_acpi_id = 0xffffffff;
 432 }
 433
 434 void
 435 spinlock_enter(void)
 436 {
 437         struct thread *td;
 438
 439         td = curthread;
 440         if (td->td_md.md_spinlock_count == 0)
 441                 td->td_md.md_saved_intr = intr_disable();
 442         td->td_md.md_spinlock_count++;
 443         critical_enter();
 444 }
 445
 446 void
 447 spinlock_exit(void)
 448 {
 449         struct thread *td;
 450
 451         td = curthread;
 452         critical_exit();
 453         td->td_md.md_spinlock_count--;
 454         if (td->td_md.md_spinlock_count == 0)
 455                 intr_restore(td->td_md.md_saved_intr);
 456 }
 457
 458 void
 459 map_vhpt(uintptr_t vhpt)
 460 {
 461         pt_entry_t pte;
 462         uint64_t psr;
 463
 464         pte = PTE_PRESENT | PTE_MA_WB | PTE_ACCESSED | PTE_DIRTY |
 465             PTE_PL_KERN | PTE_AR_RW;
 466         pte |= vhpt & PTE_PPN_MASK;
 467
 468         __asm __volatile("ptr.d %0,%1" :: "r"(vhpt),
 469             "r"(IA64_ID_PAGE_SHIFT<<2));
 470
 471         __asm __volatile("mov   %0=psr" : "=r"(psr));
 472         __asm __volatile("rsm   psr.ic|psr.i");
 473         ia64_srlz_i();
 474         ia64_set_ifa(vhpt);
 475         ia64_set_itir(IA64_ID_PAGE_SHIFT << 2);
 476         ia64_srlz_d();
 477         __asm __volatile("itr.d dtr[%0]=%1" :: "r"(2), "r"(pte));
 478         __asm __volatile("mov   psr.l=%0" :: "r" (psr));
 479         ia64_srlz_i();
 480 }
 481
 482 void
 483 map_pal_code(void)
 484 {
 485         pt_entry_t pte;
 486         uint64_t psr;
 487
 488         if (ia64_pal_base == 0)
 489                 return;
 490
 491         pte = PTE_PRESENT | PTE_MA_WB | PTE_ACCESSED | PTE_DIRTY |
 492             PTE_PL_KERN | PTE_AR_RWX;
 493         pte |= ia64_pal_base & PTE_PPN_MASK;
 494
 495         __asm __volatile("ptr.d %0,%1; ptr.i %0,%1" ::
 496             "r"(IA64_PHYS_TO_RR7(ia64_pal_base)), "r"(IA64_ID_PAGE_SHIFT<<2));
 497
 498         __asm __volatile("mov   %0=psr" : "=r"(psr));
 499         __asm __volatile("rsm   psr.ic|psr.i");
 500         ia64_srlz_i();
 501         ia64_set_ifa(IA64_PHYS_TO_RR7(ia64_pal_base));
 502         ia64_set_itir(IA64_ID_PAGE_SHIFT << 2);
 503         ia64_srlz_d();
 504         __asm __volatile("itr.d dtr[%0]=%1" :: "r"(1), "r"(pte));
 505         ia64_srlz_d();
 506         __asm __volatile("itr.i itr[%0]=%1" :: "r"(1), "r"(pte));
 507         __asm __volatile("mov   psr.l=%0" :: "r" (psr));
 508         ia64_srlz_i();
 509 }
 510
 511 void
 512 map_gateway_page(void)
 513 {
 514         pt_entry_t pte;
 515         uint64_t psr;
 516
 517         pte = PTE_PRESENT | PTE_MA_WB | PTE_ACCESSED | PTE_DIRTY |
 518             PTE_PL_KERN | PTE_AR_X_RX;
 519         pte |= (uint64_t)ia64_gateway_page & PTE_PPN_MASK;
 520
 521         __asm __volatile("ptr.d %0,%1; ptr.i %0,%1" ::
 522             "r"(VM_MAX_ADDRESS), "r"(PAGE_SHIFT << 2));
 523
 524         __asm __volatile("mov   %0=psr" : "=r"(psr));
 525         __asm __volatile("rsm   psr.ic|psr.i");
 526         ia64_srlz_i();
 527         ia64_set_ifa(VM_MAX_ADDRESS);
 528         ia64_set_itir(PAGE_SHIFT << 2);
 529         ia64_srlz_d();
 530         __asm __volatile("itr.d dtr[%0]=%1" :: "r"(3), "r"(pte));
 531         ia64_srlz_d();
 532         __asm __volatile("itr.i itr[%0]=%1" :: "r"(3), "r"(pte));
 533         __asm __volatile("mov   psr.l=%0" :: "r" (psr));
 534         ia64_srlz_i();
 535
 536         /* Expose the mapping to userland in ar.k5 */
 537         ia64_set_k5(VM_MAX_ADDRESS);
 538 }
 539
 540 static void
 541 calculate_frequencies(void)
 542 {
 543         struct ia64_sal_result sal;
 544         struct ia64_pal_result pal;
 545
 546         sal = ia64_sal_entry(SAL_FREQ_BASE, 0, 0, 0, 0, 0, 0, 0);
 547         pal = ia64_call_pal_static(PAL_FREQ_RATIOS, 0, 0, 0);
 548
 549         if (sal.sal_status == 0 && pal.pal_status == 0) {
 550                 if (bootverbose) {
 551                         printf("Platform clock frequency %ld Hz\n",
 552                                sal.sal_result[0]);
 553                         printf("Processor ratio %ld/%ld, Bus ratio %ld/%ld, "
 554                                "ITC ratio %ld/%ld\n",
 555                                pal.pal_result[0] >> 32,
 556                                pal.pal_result[0] & ((1L << 32) - 1),
 557                                pal.pal_result[1] >> 32,
 558                                pal.pal_result[1] & ((1L << 32) - 1),
 559                                pal.pal_result[2] >> 32,
 560                                pal.pal_result[2] & ((1L << 32) - 1));
 561                 }
 562                 processor_frequency =
 563                         sal.sal_result[0] * (pal.pal_result[0] >> 32)
 564                         / (pal.pal_result[0] & ((1L << 32) - 1));
 565                 bus_frequency =
 566                         sal.sal_result[0] * (pal.pal_result[1] >> 32)
 567                         / (pal.pal_result[1] & ((1L << 32) - 1));
 568                 itc_frequency =
 569                         sal.sal_result[0] * (pal.pal_result[2] >> 32)
 570                         / (pal.pal_result[2] & ((1L << 32) - 1));
 571         }
 572 }
 573
 574 struct ia64_init_return
 575 ia64_init(void)
 576 {
 577         struct ia64_init_return ret;
 578         int phys_avail_cnt;
 579         vm_offset_t kernstart, kernend;
 580         vm_offset_t kernstartpfn, kernendpfn, pfn0, pfn1;
 581         char *p;
 582         struct efi_md *md;
 583         int metadata_missing;
 584
 585         /* NO OUTPUT ALLOWED UNTIL FURTHER NOTICE */
 586
 587         /*
 588          * TODO: Disable interrupts, floating point etc.
 589          * Maybe flush cache and tlb
 590          */
 591         ia64_set_fpsr(IA64_FPSR_DEFAULT);
 592
 593         /*
 594          * TODO: Get critical system information (if possible, from the
 595          * information provided by the boot program).
 596          */
 597
 598         /*
 599          * pa_bootinfo is the physical address of the bootinfo block as
 600          * passed to us by the loader and set in locore.s.
 601          */
 602         bootinfo = *(struct bootinfo *)(IA64_PHYS_TO_RR7(pa_bootinfo));
 603
 604         if (bootinfo.bi_magic != BOOTINFO_MAGIC || bootinfo.bi_version != 1) {
 605                 bzero(&bootinfo, sizeof(bootinfo));
 606                 bootinfo.bi_kernend = (vm_offset_t) round_page(_end);
 607         }
 608
 609         /*
 610          * Look for the I/O ports first - we need them for console
 611          * probing.
 612          */
 613         for (md = efi_md_first(); md != NULL; md = efi_md_next(md)) {
 614                 switch (md->md_type) {
 615                 case EFI_MD_TYPE_IOPORT:
 616                         ia64_port_base = IA64_PHYS_TO_RR6(md->md_phys);
 617                         break;
 618                 case EFI_MD_TYPE_PALCODE:
 619                         ia64_pal_base = md->md_phys;
 620                         break;
 621                 }
 622         }
 623
 624         metadata_missing = 0;
 625         if (bootinfo.bi_modulep)
 626                 preload_metadata = (caddr_t)bootinfo.bi_modulep;
 627         else
 628                 metadata_missing = 1;
 629
 630         if (envmode == 0 && bootinfo.bi_envp)
 631                 kern_envp = (caddr_t)bootinfo.bi_envp;
 632         else
 633                 kern_envp = static_env;
 634
 635         /*
 636          * Look at arguments passed to us and compute boothowto.
 637          */
 638         boothowto = bootinfo.bi_boothowto;
 639
 640         /*
 641          * Catch case of boot_verbose set in environment.
 642          */
 643         if ((p = getenv("boot_verbose")) != NULL) {
 644                 if (strcmp(p, "yes") == 0 || strcmp(p, "YES") == 0) {
 645                         boothowto |= RB_VERBOSE;
 646                 }
 647                 freeenv(p);
 648         }
 649
 650         if (boothowto & RB_VERBOSE)
 651                 bootverbose = 1;
 652
 653         /*
 654          * Find the beginning and end of the kernel.
 655          */
 656         kernstart = trunc_page(kernel_text);
 657 #ifdef DDB
 658         ksym_start = bootinfo.bi_symtab;
 659         ksym_end = bootinfo.bi_esymtab;
 660         kernend = (vm_offset_t)round_page(ksym_end);
 661 #else
 662         kernend = (vm_offset_t)round_page(_end);
 663 #endif
 664         /* But if the bootstrap tells us otherwise, believe it! */
 665         if (bootinfo.bi_kernend)
 666                 kernend = round_page(bootinfo.bi_kernend);
 667
 668         /*
 669          * Setup the PCPU data for the bootstrap processor. It is needed
 670          * by printf(). Also, since printf() has critical sections, we
 671          * need to initialize at least pc_curthread.
 672          */
 673         pcpup = &pcpu0;
 674         ia64_set_k4((u_int64_t)pcpup);
 675         pcpu_init(pcpup, 0, sizeof(pcpu0));
 676         dpcpu_init((void *)kernend, 0);
 677         kernend += DPCPU_SIZE;
 678         PCPU_SET(curthread, &thread0);
 679
 680         /*
 681          * Initialize the console before we print anything out.
 682          */
 683         cninit();
 684
 685         /* OUTPUT NOW ALLOWED */
 686
 687         if (ia64_pal_base != 0) {
 688                 ia64_pal_base &= ~IA64_ID_PAGE_MASK;
 689                 /*
 690                  * We use a TR to map the first 256M of memory - this might
 691                  * cover the palcode too.
 692                  */
 693                 if (ia64_pal_base == 0)
 694                         printf("PAL code mapped by the kernel's TR\n");
 695         } else
 696                 printf("PAL code not found\n");
 697
 698         /*
 699          * Wire things up so we can call the firmware.
 700          */
 701         map_pal_code();
 702         efi_boot_minimal(bootinfo.bi_systab);
 703         ia64_sal_init();
 704         calculate_frequencies();
 705
 706         if (metadata_missing)
 707                 printf("WARNING: loader(8) metadata is missing!\n");
 708
 709         /* Get FPSWA interface */
 710         fpswa_iface = (bootinfo.bi_fpswa == 0) ? NULL :
 711             (struct fpswa_iface *)IA64_PHYS_TO_RR7(bootinfo.bi_fpswa);
 712
 713         /* Init basic tunables, including hz */
 714         init_param1();
 715
 716         p = getenv("kernelname");
 717         if (p) {
 718                 strncpy(kernelname, p, sizeof(kernelname) - 1);
 719                 freeenv(p);
 720         }
 721
 722         kernstartpfn = atop(IA64_RR_MASK(kernstart));
 723         kernendpfn = atop(IA64_RR_MASK(kernend));
 724
 725         /*
 726          * Size the memory regions and load phys_avail[] with the results.
 727          */
 728
 729         /*
 730          * Find out how much memory is available, by looking at
 731          * the memory descriptors.
 732          */
 733
 734 #ifdef DEBUG_MD
 735         printf("Memory descriptor count: %d\n", mdcount);
 736 #endif
 737
 738         phys_avail_cnt = 0;
 739         for (md = efi_md_first(); md != NULL; md = efi_md_next(md)) {
 740 #ifdef DEBUG_MD
 741                 printf("MD %p: type %d pa 0x%lx cnt 0x%lx\n", md,
 742                     md->md_type, md->md_phys, md->md_pages);
 743 #endif
 744
 745                 pfn0 = ia64_btop(round_page(md->md_phys));
 746                 pfn1 = ia64_btop(trunc_page(md->md_phys + md->md_pages * 4096));
 747                 if (pfn1 <= pfn0)
 748                         continue;
 749
 750                 if (md->md_type != EFI_MD_TYPE_FREE)
 751                         continue;
 752
 753                 /*
 754                  * We have a memory descriptor that describes conventional
 755                  * memory that is for general use. We must determine if the
 756                  * loader has put the kernel in this region.
 757                  */
 758                 physmem += (pfn1 - pfn0);
 759                 if (pfn0 <= kernendpfn && kernstartpfn <= pfn1) {
 760                         /*
 761                          * Must compute the location of the kernel
 762                          * within the segment.
 763                          */
 764 #ifdef DEBUG_MD
 765                         printf("Descriptor %p contains kernel\n", mp);
 766 #endif
 767                         if (pfn0 < kernstartpfn) {
 768                                 /*
 769                                  * There is a chunk before the kernel.
 770                                  */
 771 #ifdef DEBUG_MD
 772                                 printf("Loading chunk before kernel: "
 773                                        "0x%lx / 0x%lx\n", pfn0, kernstartpfn);
 774 #endif
 775                                 phys_avail[phys_avail_cnt] = ia64_ptob(pfn0);
 776                                 phys_avail[phys_avail_cnt+1] = ia64_ptob(kernstartpfn);
 777                                 phys_avail_cnt += 2;
 778                         }
 779                         if (kernendpfn < pfn1) {
 780                                 /*
 781                                  * There is a chunk after the kernel.
 782                                  */
 783 #ifdef DEBUG_MD
 784                                 printf("Loading chunk after kernel: "
 785                                        "0x%lx / 0x%lx\n", kernendpfn, pfn1);
 786 #endif
 787                                 phys_avail[phys_avail_cnt] = ia64_ptob(kernendpfn);
 788                                 phys_avail[phys_avail_cnt+1] = ia64_ptob(pfn1);
 789                                 phys_avail_cnt += 2;
 790                         }
 791                 } else {
 792                         /*
 793                          * Just load this cluster as one chunk.
 794                          */
 795 #ifdef DEBUG_MD
 796                         printf("Loading descriptor %d: 0x%lx / 0x%lx\n", i,
 797                                pfn0, pfn1);
 798 #endif
 799                         phys_avail[phys_avail_cnt] = ia64_ptob(pfn0);
 800                         phys_avail[phys_avail_cnt+1] = ia64_ptob(pfn1);
 801                         phys_avail_cnt += 2;
 802
 803                 }
 804         }
 805         phys_avail[phys_avail_cnt] = 0;
 806
 807         Maxmem = physmem;
 808         init_param2(physmem);
 809
 810         /*
 811          * Initialize error message buffer (at end of core).
 812          */
 813         msgbufp = (struct msgbuf *)pmap_steal_memory(MSGBUF_SIZE);
 814         msgbufinit(msgbufp, MSGBUF_SIZE);
 815
 816         proc_linkup0(&proc0, &thread0);
 817         /*
 818          * Init mapping for kernel stack for proc 0
 819          */
 820         thread0.td_kstack = pmap_steal_memory(KSTACK_PAGES * PAGE_SIZE);
 821         thread0.td_kstack_pages = KSTACK_PAGES;
 822
 823         mutex_init();
 824
 825         /*
 826          * Initialize the rest of proc 0's PCB.
 827          *
 828          * Set the kernel sp, reserving space for an (empty) trapframe,
 829          * and make proc0's trapframe pointer point to it for sanity.
 830          * Initialise proc0's backing store to start after u area.
 831          */
 832         cpu_thread_alloc(&thread0);
 833         thread0.td_frame->tf_flags = FRAME_SYSCALL;
 834         thread0.td_pcb->pcb_special.sp =
 835             (u_int64_t)thread0.td_frame - 16;
 836         thread0.td_pcb->pcb_special.bspstore = thread0.td_kstack;
 837
 838         /*
 839          * Initialize the virtual memory system.
 840          */
 841         pmap_bootstrap();
 842
 843         /*
 844          * Initialize debuggers, and break into them if appropriate.
 845          */
 846         kdb_init();
 847
 848 #ifdef KDB
 849         if (boothowto & RB_KDB)
 850                 kdb_enter(KDB_WHY_BOOTFLAGS,
 851                     "Boot flags requested debugger\n");
 852 #endif
 853
 854         ia64_set_tpr(0);
 855         ia64_srlz_d();
 856
 857         ret.bspstore = thread0.td_pcb->pcb_special.bspstore;
 858         ret.sp = thread0.td_pcb->pcb_special.sp;
 859         return (ret);
 860 }
 861
 862 __volatile void *
 863 ia64_ioport_address(u_int port)
 864 {
 865         uint64_t addr;
 866
 867         addr = (port > 0xffff) ? IA64_PHYS_TO_RR6((uint64_t)port) :
 868             ia64_port_base | ((port & 0xfffc) << 10) | (port & 0xFFF);
 869         return ((__volatile void *)addr);
 870 }
 871
 872 uint64_t
 873 ia64_get_hcdp(void)
 874 {
 875
 876         return (bootinfo.bi_hcdp);
 877 }
 878
 879 void
 880 bzero(void *buf, size_t len)
 881 {
 882         caddr_t p = buf;
 883
 884         while (((vm_offset_t) p & (sizeof(u_long) - 1)) && len) {
 885                 *p++ = 0;
 886                 len--;
 887         }
 888         while (len >= sizeof(u_long) * 8) {
 889                 *(u_long*) p = 0;
 890                 *((u_long*) p + 1) = 0;
 891                 *((u_long*) p + 2) = 0;
 892                 *((u_long*) p + 3) = 0;
 893                 len -= sizeof(u_long) * 8;
 894                 *((u_long*) p + 4) = 0;
 895                 *((u_long*) p + 5) = 0;
 896                 *((u_long*) p + 6) = 0;
 897                 *((u_long*) p + 7) = 0;
 898                 p += sizeof(u_long) * 8;
 899         }
 900         while (len >= sizeof(u_long)) {
 901                 *(u_long*) p = 0;
 902                 len -= sizeof(u_long);
 903                 p += sizeof(u_long);
 904         }
 905         while (len) {
 906                 *p++ = 0;
 907                 len--;
 908         }
 909 }
 910
 911 void
 912 DELAY(int n)
 913 {
 914         u_int64_t start, end, now;
 915
 916         sched_pin();
 917
 918         start = ia64_get_itc();
 919         end = start + (itc_frequency * n) / 1000000;
 920         /* printf("DELAY from 0x%lx to 0x%lx\n", start, end); */
 921         do {
 922                 now = ia64_get_itc();
 923         } while (now < end || (now > start && end < start));
 924
 925         sched_unpin();
 926 }
 927
 928 /*
 929  * Send an interrupt (signal) to a process.
 930  */
 931 void
 932 sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 933 {
 934         struct proc *p;
 935         struct thread *td;
 936         struct trapframe *tf;
 937         struct sigacts *psp;
 938         struct sigframe sf, *sfp;
 939         u_int64_t sbs, sp;
 940         int oonstack;
 941         int sig;
 942         u_long code;
 943
 944         td = curthread;
 945         p = td->td_proc;
 946         PROC_LOCK_ASSERT(p, MA_OWNED);
 947         sig = ksi->ksi_signo;
 948         code = ksi->ksi_code;
 949         psp = p->p_sigacts;
 950         mtx_assert(&psp->ps_mtx, MA_OWNED);
 951         tf = td->td_frame;
 952         sp = tf->tf_special.sp;
 953         oonstack = sigonstack(sp);
 954         sbs = 0;
 955
 956         /* save user context */
 957         bzero(&sf, sizeof(struct sigframe));
 958         sf.sf_uc.uc_sigmask = *mask;
 959         sf.sf_uc.uc_stack = td->td_sigstk;
 960         sf.sf_uc.uc_stack.ss_flags = (td->td_pflags & TDP_ALTSTACK)
 961             ? ((oonstack) ? SS_ONSTACK : 0) : SS_DISABLE;
 962
 963         /*
 964          * Allocate and validate space for the signal handler
 965          * context. Note that if the stack is in P0 space, the
 966          * call to grow() is a nop, and the useracc() check
 967          * will fail if the process has not already allocated
 968          * the space with a `brk'.
 969          */
 970         if ((td->td_pflags & TDP_ALTSTACK) != 0 && !oonstack &&
 971             SIGISMEMBER(psp->ps_sigonstack, sig)) {
 972                 sbs = (u_int64_t)td->td_sigstk.ss_sp;
 973                 sbs = (sbs + 15) & ~15;
 974                 sfp = (struct sigframe *)(sbs + td->td_sigstk.ss_size);
 975 #if defined(COMPAT_43)
 976                 td->td_sigstk.ss_flags |= SS_ONSTACK;
 977 #endif
 978         } else
 979                 sfp = (struct sigframe *)sp;
 980         sfp = (struct sigframe *)((u_int64_t)(sfp - 1) & ~15);
 981
 982         /* Fill in the siginfo structure for POSIX handlers. */
 983         if (SIGISMEMBER(psp->ps_siginfo, sig)) {
 984                 sf.sf_si = ksi->ksi_info;
 985                 sf.sf_si.si_signo = sig;
 986                 /*
 987                  * XXX this shouldn't be here after code in trap.c
 988                  * is fixed
 989                  */
 990                 sf.sf_si.si_addr = (void*)tf->tf_special.ifa;
 991                 code = (u_int64_t)&sfp->sf_si;
 992         }
 993
 994         mtx_unlock(&psp->ps_mtx);
 995         PROC_UNLOCK(p);
 996
 997         get_mcontext(td, &sf.sf_uc.uc_mcontext, 0);
 998
 999         /* Copy the frame out to userland. */
1000         if (copyout(&sf, sfp, sizeof(sf)) != 0) {
1001                 /*
1002                  * Process has trashed its stack; give it an illegal
1003                  * instruction to halt it in its tracks.
1004                  */
1005                 PROC_LOCK(p);
1006                 sigexit(td, SIGILL);
1007                 return;
1008         }
1009
1010         if ((tf->tf_flags & FRAME_SYSCALL) == 0) {
1011                 tf->tf_special.psr &= ~IA64_PSR_RI;
1012                 tf->tf_special.iip = ia64_get_k5() +
1013                     ((uint64_t)break_sigtramp - (uint64_t)ia64_gateway_page);
1014         } else
1015                 tf->tf_special.iip = ia64_get_k5() +
1016                     ((uint64_t)epc_sigtramp - (uint64_t)ia64_gateway_page);
1017
1018         /*
1019          * Setup the trapframe to return to the signal trampoline. We pass
1020          * information to the trampoline in the following registers:
1021          *
1022          *      gp      new backing store or NULL
1023          *      r8      signal number
1024          *      r9      signal code or siginfo pointer
1025          *      r10     signal handler (function descriptor)
1026          */
1027         tf->tf_special.sp = (u_int64_t)sfp - 16;
1028         tf->tf_special.gp = sbs;
1029         tf->tf_special.bspstore = sf.sf_uc.uc_mcontext.mc_special.bspstore;
1030         tf->tf_special.ndirty = 0;
1031         tf->tf_special.rnat = sf.sf_uc.uc_mcontext.mc_special.rnat;
1032         tf->tf_scratch.gr8 = sig;
1033         tf->tf_scratch.gr9 = code;
1034         tf->tf_scratch.gr10 = (u_int64_t)catcher;
1035
1036         PROC_LOCK(p);
1037         mtx_lock(&psp->ps_mtx);
1038 }
1039
1040 /*
1041  * System call to cleanup state after a signal
1042  * has been taken.  Reset signal mask and
1043  * stack state from context left by sendsig (above).
1044  * Return to previous pc and psl as specified by
1045  * context left by sendsig. Check carefully to
1046  * make sure that the user has not modified the
1047  * state to gain improper privileges.
1048  *
1049  * MPSAFE
1050  */
1051 int
1052 sigreturn(struct thread *td,
1053         struct sigreturn_args /* {
1054                 ucontext_t *sigcntxp;
1055         } */ *uap)
1056 {
1057         ucontext_t uc;
1058         struct trapframe *tf;
1059         struct pcb *pcb;
1060
1061         tf = td->td_frame;
1062         pcb = td->td_pcb;
1063
1064         /*
1065          * Fetch the entire context structure at once for speed.
1066          * We don't use a normal argument to simplify RSE handling.
1067          */
1068         if (copyin(uap->sigcntxp, (caddr_t)&uc, sizeof(uc)))
1069                 return (EFAULT);
1070
1071         set_mcontext(td, &uc.uc_mcontext);
1072
1073 #if defined(COMPAT_43)
1074         if (sigonstack(tf->tf_special.sp))
1075                 td->td_sigstk.ss_flags |= SS_ONSTACK;
1076         else
1077                 td->td_sigstk.ss_flags &= ~SS_ONSTACK;
1078 #endif
1079         kern_sigprocmask(td, SIG_SETMASK, &uc.uc_sigmask, NULL, 0);
1080
1081         return (EJUSTRETURN);
1082 }
1083
1084 #ifdef COMPAT_FREEBSD4
1085 int
1086 freebsd4_sigreturn(struct thread *td, struct freebsd4_sigreturn_args *uap)
1087 {
1088
1089         return sigreturn(td, (struct sigreturn_args *)uap);
1090 }
1091 #endif
1092
1093 /*
1094  * Construct a PCB from a trapframe. This is called from kdb_trap() where
1095  * we want to start a backtrace from the function that caused us to enter
1096  * the debugger. We have the context in the trapframe, but base the trace
1097  * on the PCB. The PCB doesn't have to be perfect, as long as it contains
1098  * enough for a backtrace.
1099  */
1100 void
1101 makectx(struct trapframe *tf, struct pcb *pcb)
1102 {
1103
1104         pcb->pcb_special = tf->tf_special;
1105         pcb->pcb_special.__spare = ~0UL;        /* XXX see unwind.c */
1106         save_callee_saved(&pcb->pcb_preserved);
1107         save_callee_saved_fp(&pcb->pcb_preserved_fp);
1108 }
1109
1110 int
1111 ia64_flush_dirty(struct thread *td, struct _special *r)
1112 {
1113         struct iovec iov;
1114         struct uio uio;
1115         uint64_t bspst, kstk, rnat;
1116         int error, locked;
1117
1118         if (r->ndirty == 0)
1119                 return (0);
1120
1121         kstk = td->td_kstack + (r->bspstore & 0x1ffUL);
1122         if (td == curthread) {
1123                 __asm __volatile("mov   ar.rsc=0;;");
1124                 __asm __volatile("mov   %0=ar.bspstore" : "=r"(bspst));
1125                 /* Make sure we have all the user registers written out. */
1126                 if (bspst - kstk < r->ndirty) {
1127                         __asm __volatile("flushrs;;");
1128                         __asm __volatile("mov   %0=ar.bspstore" : "=r"(bspst));
1129                 }
1130                 __asm __volatile("mov   %0=ar.rnat;;" : "=r"(rnat));
1131                 __asm __volatile("mov   ar.rsc=3");
1132                 error = copyout((void*)kstk, (void*)r->bspstore, r->ndirty);
1133                 kstk += r->ndirty;
1134                 r->rnat = (bspst > kstk && (bspst & 0x1ffL) < (kstk & 0x1ffL))
1135                     ? *(uint64_t*)(kstk | 0x1f8L) : rnat;
1136         } else {
1137                 locked = PROC_LOCKED(td->td_proc);
1138                 if (!locked)
1139                         PHOLD(td->td_proc);
1140                 iov.iov_base = (void*)(uintptr_t)kstk;
1141                 iov.iov_len = r->ndirty;
1142                 uio.uio_iov = &iov;
1143                 uio.uio_iovcnt = 1;
1144                 uio.uio_offset = r->bspstore;
1145                 uio.uio_resid = r->ndirty;
1146                 uio.uio_segflg = UIO_SYSSPACE;
1147                 uio.uio_rw = UIO_WRITE;
1148                 uio.uio_td = td;
1149                 error = proc_rwmem(td->td_proc, &uio);
1150                 /*
1151                  * XXX proc_rwmem() doesn't currently return ENOSPC,
1152                  * so I think it can bogusly return 0. Neither do
1153                  * we allow short writes.
1154                  */
1155                 if (uio.uio_resid != 0 && error == 0)
1156                         error = ENOSPC;
1157                 if (!locked)
1158                         PRELE(td->td_proc);
1159         }
1160
1161         r->bspstore += r->ndirty;
1162         r->ndirty = 0;
1163         return (error);
1164 }
1165
1166 int
1167 get_mcontext(struct thread *td, mcontext_t *mc, int flags)
1168 {
1169         struct trapframe *tf;
1170         int error;
1171
1172         tf = td->td_frame;
1173         bzero(mc, sizeof(*mc));
1174         mc->mc_special = tf->tf_special;
1175         error = ia64_flush_dirty(td, &mc->mc_special);
1176         if (tf->tf_flags & FRAME_SYSCALL) {
1177                 mc->mc_flags |= _MC_FLAGS_SYSCALL_CONTEXT;
1178                 mc->mc_scratch = tf->tf_scratch;
1179                 if (flags & GET_MC_CLEAR_RET) {
1180                         mc->mc_scratch.gr8 = 0;
1181                         mc->mc_scratch.gr9 = 0;
1182                         mc->mc_scratch.gr10 = 0;
1183                         mc->mc_scratch.gr11 = 0;
1184                 }
1185         } else {
1186                 mc->mc_flags |= _MC_FLAGS_ASYNC_CONTEXT;
1187                 mc->mc_scratch = tf->tf_scratch;
1188                 mc->mc_scratch_fp = tf->tf_scratch_fp;
1189                 /*
1190                  * XXX If the thread never used the high FP registers, we
1191                  * probably shouldn't waste time saving them.
1192                  */
1193                 ia64_highfp_save(td);
1194                 mc->mc_flags |= _MC_FLAGS_HIGHFP_VALID;
1195                 mc->mc_high_fp = td->td_pcb->pcb_high_fp;
1196         }
1197         save_callee_saved(&mc->mc_preserved);
1198         save_callee_saved_fp(&mc->mc_preserved_fp);
1199         return (error);
1200 }
1201
1202 int
1203 set_mcontext(struct thread *td, const mcontext_t *mc)
1204 {
1205         struct _special s;
1206         struct trapframe *tf;
1207         uint64_t psrmask;
1208
1209         tf = td->td_frame;
1210
1211         KASSERT((tf->tf_special.ndirty & ~PAGE_MASK) == 0,
1212             ("Whoa there! We have more than 8KB of dirty registers!"));
1213
1214         s = mc->mc_special;
1215         /*
1216          * Only copy the user mask and the restart instruction bit from
1217          * the new context.
1218          */
1219         psrmask = IA64_PSR_BE | IA64_PSR_UP | IA64_PSR_AC | IA64_PSR_MFL |
1220             IA64_PSR_MFH | IA64_PSR_RI;
1221         s.psr = (tf->tf_special.psr & ~psrmask) | (s.psr & psrmask);
1222         /* We don't have any dirty registers of the new context. */
1223         s.ndirty = 0;
1224         if (mc->mc_flags & _MC_FLAGS_ASYNC_CONTEXT) {
1225                 /*
1226                  * We can get an async context passed to us while we
1227                  * entered the kernel through a syscall: sigreturn(2)
1228                  * takes contexts that could previously be the result of
1229                  * a trap or interrupt.
1230                  * Hence, we cannot assert that the trapframe is not
1231                  * a syscall frame, but we can assert that it's at
1232                  * least an expected syscall.
1233                  */
1234                 if (tf->tf_flags & FRAME_SYSCALL) {
1235                         KASSERT(tf->tf_scratch.gr15 == SYS_sigreturn, ("foo"));
1236                         tf->tf_flags &= ~FRAME_SYSCALL;
1237                 }
1238                 tf->tf_scratch = mc->mc_scratch;
1239                 tf->tf_scratch_fp = mc->mc_scratch_fp;
1240                 if (mc->mc_flags & _MC_FLAGS_HIGHFP_VALID)
1241                         td->td_pcb->pcb_high_fp = mc->mc_high_fp;
1242         } else {
1243                 KASSERT((tf->tf_flags & FRAME_SYSCALL) != 0, ("foo"));
1244                 if ((mc->mc_flags & _MC_FLAGS_SYSCALL_CONTEXT) == 0) {
1245                         s.cfm = tf->tf_special.cfm;
1246                         s.iip = tf->tf_special.iip;
1247                         tf->tf_scratch.gr15 = 0;        /* Clear syscall nr. */
1248                 } else
1249                         tf->tf_scratch = mc->mc_scratch;
1250         }
1251         tf->tf_special = s;
1252         restore_callee_saved(&mc->mc_preserved);
1253         restore_callee_saved_fp(&mc->mc_preserved_fp);
1254
1255         return (0);
1256 }
1257
1258 /*
1259  * Clear registers on exec.
1260  */
1261 void
1262 exec_setregs(struct thread *td, u_long entry, u_long stack, u_long ps_strings)
1263 {
1264         struct trapframe *tf;
1265         uint64_t *ksttop, *kst;
1266
1267         tf = td->td_frame;
1268         ksttop = (uint64_t*)(td->td_kstack + tf->tf_special.ndirty +
1269             (tf->tf_special.bspstore & 0x1ffUL));
1270
1271         /*
1272          * We can ignore up to 8KB of dirty registers by masking off the
1273          * lower 13 bits in exception_restore() or epc_syscall(). This
1274          * should be enough for a couple of years, but if there are more
1275          * than 8KB of dirty registers, we lose track of the bottom of
1276          * the kernel stack. The solution is to copy the active part of
1277          * the kernel stack down 1 page (or 2, but not more than that)
1278          * so that we always have less than 8KB of dirty registers.
1279          */
1280         KASSERT((tf->tf_special.ndirty & ~PAGE_MASK) == 0,
1281             ("Whoa there! We have more than 8KB of dirty registers!"));
1282
1283         bzero(&tf->tf_special, sizeof(tf->tf_special));
1284         if ((tf->tf_flags & FRAME_SYSCALL) == 0) {      /* break syscalls. */
1285                 bzero(&tf->tf_scratch, sizeof(tf->tf_scratch));
1286                 bzero(&tf->tf_scratch_fp, sizeof(tf->tf_scratch_fp));
1287                 tf->tf_special.cfm = (1UL<<63) | (3UL<<7) | 3UL;
1288                 tf->tf_special.bspstore = IA64_BACKINGSTORE;
1289                 /*
1290                  * Copy the arguments onto the kernel register stack so that
1291                  * they get loaded by the loadrs instruction. Skip over the
1292                  * NaT collection points.
1293                  */
1294                 kst = ksttop - 1;
1295                 if (((uintptr_t)kst & 0x1ff) == 0x1f8)
1296                         *kst-- = 0;
1297                 *kst-- = 0;
1298                 if (((uintptr_t)kst & 0x1ff) == 0x1f8)
1299                         *kst-- = 0;
1300                 *kst-- = ps_strings;
1301                 if (((uintptr_t)kst & 0x1ff) == 0x1f8)
1302                         *kst-- = 0;
1303                 *kst = stack;
1304                 tf->tf_special.ndirty = (ksttop - kst) << 3;
1305         } else {                                /* epc syscalls (default). */
1306                 tf->tf_special.cfm = (3UL<<62) | (3UL<<7) | 3UL;
1307                 tf->tf_special.bspstore = IA64_BACKINGSTORE + 24;
1308                 /*
1309                  * Write values for out0, out1 and out2 to the user's backing
1310                  * store and arrange for them to be restored into the user's
1311                  * initial register frame.
1312                  * Assumes that (bspstore & 0x1f8) < 0x1e0.
1313                  */
1314                 suword((caddr_t)tf->tf_special.bspstore - 24, stack);
1315                 suword((caddr_t)tf->tf_special.bspstore - 16, ps_strings);
1316                 suword((caddr_t)tf->tf_special.bspstore -  8, 0);
1317         }
1318
1319         tf->tf_special.iip = entry;
1320         tf->tf_special.sp = (stack & ~15) - 16;
1321         tf->tf_special.rsc = 0xf;
1322         tf->tf_special.fpsr = IA64_FPSR_DEFAULT;
1323         tf->tf_special.psr = IA64_PSR_IC | IA64_PSR_I | IA64_PSR_IT |
1324             IA64_PSR_DT | IA64_PSR_RT | IA64_PSR_DFH | IA64_PSR_BN |
1325             IA64_PSR_CPL_USER;
1326 }
1327
1328 int
1329 ptrace_set_pc(struct thread *td, unsigned long addr)
1330 {
1331         uint64_t slot;
1332
1333         switch (addr & 0xFUL) {
1334         case 0:
1335                 slot = IA64_PSR_RI_0;
1336                 break;
1337         case 1:
1338                 /* XXX we need to deal with MLX bundles here */
1339                 slot = IA64_PSR_RI_1;
1340                 break;
1341         case 2:
1342                 slot = IA64_PSR_RI_2;
1343                 break;
1344         default:
1345                 return (EINVAL);
1346         }
1347
1348         td->td_frame->tf_special.iip = addr & ~0x0FULL;
1349         td->td_frame->tf_special.psr =
1350             (td->td_frame->tf_special.psr & ~IA64_PSR_RI) | slot;
1351         return (0);
1352 }
1353
1354 int
1355 ptrace_single_step(struct thread *td)
1356 {
1357         struct trapframe *tf;
1358
1359         /*
1360          * There's no way to set single stepping when we're leaving the
1361          * kernel through the EPC syscall path. The way we solve this is
1362          * by enabling the lower-privilege trap so that we re-enter the
1363          * kernel as soon as the privilege level changes. See trap.c for
1364          * how we proceed from there.
1365          */
1366         tf = td->td_frame;
1367         if (tf->tf_flags & FRAME_SYSCALL)
1368                 tf->tf_special.psr |= IA64_PSR_LP;
1369         else
1370                 tf->tf_special.psr |= IA64_PSR_SS;
1371         return (0);
1372 }
1373
1374 int
1375 ptrace_clear_single_step(struct thread *td)
1376 {
1377         struct trapframe *tf;
1378
1379         /*
1380          * Clear any and all status bits we may use to implement single
1381          * stepping.
1382          */
1383         tf = td->td_frame;
1384         tf->tf_special.psr &= ~IA64_PSR_SS;
1385         tf->tf_special.psr &= ~IA64_PSR_LP;
1386         tf->tf_special.psr &= ~IA64_PSR_TB;
1387         return (0);
1388 }
1389
1390 int
1391 fill_regs(struct thread *td, struct reg *regs)
1392 {
1393         struct trapframe *tf;
1394
1395         tf = td->td_frame;
1396         regs->r_special = tf->tf_special;
1397         regs->r_scratch = tf->tf_scratch;
1398         save_callee_saved(&regs->r_preserved);
1399         return (0);
1400 }
1401
1402 int
1403 set_regs(struct thread *td, struct reg *regs)
1404 {
1405         struct trapframe *tf;
1406         int error;
1407
1408         tf = td->td_frame;
1409         error = ia64_flush_dirty(td, &tf->tf_special);
1410         if (!error) {
1411                 tf->tf_special = regs->r_special;
1412                 tf->tf_special.bspstore += tf->tf_special.ndirty;
1413                 tf->tf_special.ndirty = 0;
1414                 tf->tf_scratch = regs->r_scratch;
1415                 restore_callee_saved(&regs->r_preserved);
1416         }
1417         return (error);
1418 }
1419
1420 int
1421 fill_dbregs(struct thread *td, struct dbreg *dbregs)
1422 {
1423
1424         return (ENOSYS);
1425 }
1426
1427 int
1428 set_dbregs(struct thread *td, struct dbreg *dbregs)
1429 {
1430
1431         return (ENOSYS);
1432 }
1433
1434 int
1435 fill_fpregs(struct thread *td, struct fpreg *fpregs)
1436 {
1437         struct trapframe *frame = td->td_frame;
1438         struct pcb *pcb = td->td_pcb;
1439
1440         /* Save the high FP registers. */
1441         ia64_highfp_save(td);
1442
1443         fpregs->fpr_scratch = frame->tf_scratch_fp;
1444         save_callee_saved_fp(&fpregs->fpr_preserved);
1445         fpregs->fpr_high = pcb->pcb_high_fp;
1446         return (0);
1447 }
1448
1449 int
1450 set_fpregs(struct thread *td, struct fpreg *fpregs)
1451 {
1452         struct trapframe *frame = td->td_frame;
1453         struct pcb *pcb = td->td_pcb;
1454
1455         /* Throw away the high FP registers (should be redundant). */
1456         ia64_highfp_drop(td);
1457
1458         frame->tf_scratch_fp = fpregs->fpr_scratch;
1459         restore_callee_saved_fp(&fpregs->fpr_preserved);
1460         pcb->pcb_high_fp = fpregs->fpr_high;
1461         return (0);
1462 }
1463
1464 void
1465 ia64_sync_icache(vm_offset_t va, vm_offset_t sz)
1466 {
1467         vm_offset_t lim;
1468
1469         if (!ia64_sync_icache_needed)
1470                 return;
1471
1472         lim = va + sz;
1473         while (va < lim) {
1474                 ia64_fc_i(va);
1475                 va += 32;       /* XXX */
1476         }
1477
1478         ia64_sync_i();
1479         ia64_srlz_i();
1480 }