1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24 */
25 /*
26 * Copyright (c) 2010, Intel Corporation.
27 * All rights reserved.
28 */
29
30 #include <sys/types.h>
31 #include <sys/t_lock.h>
32 #include <sys/param.h>
33 #include <sys/segments.h>
34 #include <sys/sysmacros.h>
35 #include <sys/signal.h>
36 #include <sys/systm.h>
37 #include <sys/user.h>
38 #include <sys/mman.h>
39 #include <sys/vm.h>
40
41 #include <sys/disp.h>
42 #include <sys/class.h>
43
44 #include <sys/proc.h>
45 #include <sys/buf.h>
46 #include <sys/kmem.h>
47
48 #include <sys/reboot.h>
49 #include <sys/uadmin.h>
50 #include <sys/callb.h>
51
52 #include <sys/cred.h>
53 #include <sys/vnode.h>
54 #include <sys/file.h>
55
56 #include <sys/procfs.h>
57 #include <sys/acct.h>
58
59 #include <sys/vfs.h>
60 #include <sys/dnlc.h>
61 #include <sys/var.h>
62 #include <sys/cmn_err.h>
63 #include <sys/utsname.h>
64 #include <sys/debug.h>
65
66 #include <sys/dumphdr.h>
67 #include <sys/bootconf.h>
68 #include <sys/varargs.h>
69 #include <sys/promif.h>
70 #include <sys/modctl.h>
71
72 #include <sys/consdev.h>
73 #include <sys/frame.h>
74
75 #include <sys/sunddi.h>
76 #include <sys/ddidmareq.h>
77 #include <sys/psw.h>
78 #include <sys/regset.h>
79 #include <sys/privregs.h>
80 #include <sys/clock.h>
81 #include <sys/tss.h>
82 #include <sys/cpu.h>
83 #include <sys/stack.h>
84 #include <sys/trap.h>
85 #include <sys/pic.h>
86 #include <vm/hat.h>
87 #include <vm/anon.h>
88 #include <vm/as.h>
89 #include <vm/page.h>
90 #include <vm/seg.h>
91 #include <vm/seg_kmem.h>
92 #include <vm/seg_map.h>
93 #include <vm/seg_vn.h>
94 #include <vm/seg_kp.h>
95 #include <vm/hat_i86.h>
96 #include <sys/swap.h>
97 #include <sys/thread.h>
98 #include <sys/sysconf.h>
99 #include <sys/vm_machparam.h>
100 #include <sys/archsystm.h>
101 #include <sys/machsystm.h>
102 #include <sys/machlock.h>
103 #include <sys/x_call.h>
104 #include <sys/instance.h>
105
106 #include <sys/time.h>
107 #include <sys/smp_impldefs.h>
108 #include <sys/psm_types.h>
109 #include <sys/atomic.h>
110 #include <sys/panic.h>
111 #include <sys/cpuvar.h>
112 #include <sys/dtrace.h>
113 #include <sys/bl.h>
114 #include <sys/nvpair.h>
115 #include <sys/x86_archext.h>
116 #include <sys/pool_pset.h>
117 #include <sys/autoconf.h>
118 #include <sys/mem.h>
119 #include <sys/dumphdr.h>
120 #include <sys/compress.h>
121 #include <sys/cpu_module.h>
122 #if defined(__xpv)
123 #include <sys/hypervisor.h>
124 #include <sys/xpv_panic.h>
125 #endif
126
127 #include <sys/fastboot.h>
128 #include <sys/machelf.h>
129 #include <sys/kobj.h>
130 #include <sys/multiboot.h>
131
132 #ifdef TRAPTRACE
133 #include <sys/traptrace.h>
134 #endif /* TRAPTRACE */
135
136 #include <c2/audit.h>
137 #include <sys/clock_impl.h>
138
139 extern void audit_enterprom(int);
140 extern void audit_exitprom(int);
141
142 /*
143 * Tunable to enable apix PSM; if set to 0, pcplusmp PSM will be used.
144 */
145 int apix_enable = 1;
146
147 int apic_nvidia_io_max = 0; /* no. of NVIDIA i/o apics */
148
149 /*
150 * Occassionally the kernel knows better whether to power-off or reboot.
151 */
152 int force_shutdown_method = AD_UNKNOWN;
153
154 /*
155 * The panicbuf array is used to record messages and state:
156 */
157 char panicbuf[PANICBUFSIZE];
158
159 /*
160 * Flags to control Dynamic Reconfiguration features.
161 */
162 uint64_t plat_dr_options;
163
164 /*
165 * Maximum physical address for memory DR operations.
166 */
167 uint64_t plat_dr_physmax;
168
169 /*
170 * maxphys - used during physio
171 * klustsize - used for klustering by swapfs and specfs
172 */
173 int maxphys = 56 * 1024; /* XXX See vm_subr.c - max b_count in physio */
174 int klustsize = 56 * 1024;
175
176 caddr_t p0_va; /* Virtual address for accessing physical page 0 */
177
178 /*
179 * defined here, though unused on x86,
180 * to make kstat_fr.c happy.
181 */
182 int vac;
183
184 void debug_enter(char *);
185
186 extern void pm_cfb_check_and_powerup(void);
187 extern void pm_cfb_rele(void);
188
189 extern fastboot_info_t newkernel;
190
191 /*
192 * Machine dependent code to reboot.
193 * "mdep" is interpreted as a character pointer; if non-null, it is a pointer
194 * to a string to be used as the argument string when rebooting.
195 *
196 * "invoke_cb" is a boolean. It is set to true when mdboot() can safely
197 * invoke CB_CL_MDBOOT callbacks before shutting the system down, i.e. when
198 * we are in a normal shutdown sequence (interrupts are not blocked, the
199 * system is not panic'ing or being suspended).
200 */
201 /*ARGSUSED*/
202 void
203 mdboot(int cmd, int fcn, char *mdep, boolean_t invoke_cb)
204 {
205 processorid_t bootcpuid = 0;
206 static int is_first_quiesce = 1;
207 static int is_first_reset = 1;
208 int reset_status = 0;
209 static char fallback_str[] = "Falling back to regular reboot.\n";
210
211 if (fcn == AD_FASTREBOOT && !newkernel.fi_valid)
212 fcn = AD_BOOT;
213
214 if (!panicstr) {
215 kpreempt_disable();
216 if (fcn == AD_FASTREBOOT) {
217 mutex_enter(&cpu_lock);
218 if (CPU_ACTIVE(cpu_get(bootcpuid))) {
219 affinity_set(bootcpuid);
220 }
221 mutex_exit(&cpu_lock);
222 } else {
223 affinity_set(CPU_CURRENT);
224 }
225 }
226
227 if (force_shutdown_method != AD_UNKNOWN)
228 fcn = force_shutdown_method;
229
230 /*
231 * XXX - rconsvp is set to NULL to ensure that output messages
232 * are sent to the underlying "hardware" device using the
233 * monitor's printf routine since we are in the process of
234 * either rebooting or halting the machine.
235 */
236 rconsvp = NULL;
237
238 /*
239 * Print the reboot message now, before pausing other cpus.
240 * There is a race condition in the printing support that
241 * can deadlock multiprocessor machines.
242 */
243 if (!(fcn == AD_HALT || fcn == AD_POWEROFF))
244 prom_printf("rebooting...\n");
245
246 if (IN_XPV_PANIC())
247 reset();
248
249 /*
250 * We can't bring up the console from above lock level, so do it now
251 */
252 pm_cfb_check_and_powerup();
253
254 /* make sure there are no more changes to the device tree */
255 devtree_freeze();
256
257 if (invoke_cb)
258 (void) callb_execute_class(CB_CL_MDBOOT, NULL);
259
260 /*
261 * Clear any unresolved UEs from memory.
262 */
263 page_retire_mdboot();
264
265 #if defined(__xpv)
266 /*
267 * XXPV Should probably think some more about how we deal
268 * with panicing before it's really safe to panic.
269 * On hypervisors, we reboot very quickly.. Perhaps panic
270 * should only attempt to recover by rebooting if,
271 * say, we were able to mount the root filesystem,
272 * or if we successfully launched init(1m).
273 */
274 if (panicstr && proc_init == NULL)
275 (void) HYPERVISOR_shutdown(SHUTDOWN_poweroff);
276 #endif
277 /*
278 * stop other cpus and raise our priority. since there is only
279 * one active cpu after this, and our priority will be too high
280 * for us to be preempted, we're essentially single threaded
281 * from here on out.
282 */
283 (void) spl6();
284 if (!panicstr) {
285 mutex_enter(&cpu_lock);
286 pause_cpus(NULL, NULL);
287 mutex_exit(&cpu_lock);
288 }
289
290 /*
291 * If the system is panicking, the preloaded kernel is valid, and
292 * fastreboot_onpanic has been set, and the system has been up for
293 * longer than fastreboot_onpanic_uptime (default to 10 minutes),
294 * choose Fast Reboot.
295 */
296 if (fcn == AD_BOOT && panicstr && newkernel.fi_valid &&
297 fastreboot_onpanic &&
298 (panic_lbolt - lbolt_at_boot) > fastreboot_onpanic_uptime) {
299 fcn = AD_FASTREBOOT;
300 }
301
302 /*
303 * Try to quiesce devices.
304 */
305 if (is_first_quiesce) {
306 /*
307 * Clear is_first_quiesce before calling quiesce_devices()
308 * so that if quiesce_devices() causes panics, it will not
309 * be invoked again.
310 */
311 is_first_quiesce = 0;
312
313 quiesce_active = 1;
314 quiesce_devices(ddi_root_node(), &reset_status);
315 if (reset_status == -1) {
316 if (fcn == AD_FASTREBOOT && !force_fastreboot) {
317 prom_printf("Driver(s) not capable of fast "
318 "reboot.\n");
319 prom_printf(fallback_str);
320 fastreboot_capable = 0;
321 fcn = AD_BOOT;
322 } else if (fcn != AD_FASTREBOOT)
323 fastreboot_capable = 0;
324 }
325 quiesce_active = 0;
326 }
327
328 /*
329 * Try to reset devices. reset_leaves() should only be called
330 * a) when there are no other threads that could be accessing devices,
331 * and
332 * b) on a system that's not capable of fast reboot (fastreboot_capable
333 * being 0), or on a system where quiesce_devices() failed to
334 * complete (quiesce_active being 1).
335 */
336 if (is_first_reset && (!fastreboot_capable || quiesce_active)) {
337 /*
338 * Clear is_first_reset before calling reset_devices()
339 * so that if reset_devices() causes panics, it will not
340 * be invoked again.
341 */
342 is_first_reset = 0;
343 reset_leaves();
344 }
345
346 /* Verify newkernel checksum */
347 if (fastreboot_capable && fcn == AD_FASTREBOOT &&
348 fastboot_cksum_verify(&newkernel) != 0) {
349 fastreboot_capable = 0;
350 prom_printf("Fast reboot: checksum failed for the new "
351 "kernel.\n");
352 prom_printf(fallback_str);
353 }
354
355 (void) spl8();
356
357 if (fastreboot_capable && fcn == AD_FASTREBOOT) {
358 /*
359 * psm_shutdown is called within fast_reboot()
360 */
361 fast_reboot();
362 } else {
363 (*psm_shutdownf)(cmd, fcn);
364
365 if (fcn == AD_HALT || fcn == AD_POWEROFF)
366 halt((char *)NULL);
367 else
368 prom_reboot("");
369 }
370 /*NOTREACHED*/
371 }
372
373 /* mdpreboot - may be called prior to mdboot while root fs still mounted */
374 /*ARGSUSED*/
375 void
376 mdpreboot(int cmd, int fcn, char *mdep)
377 {
378 if (fcn == AD_FASTREBOOT && !fastreboot_capable) {
379 fcn = AD_BOOT;
380 #ifdef __xpv
381 cmn_err(CE_WARN, "Fast reboot is not supported on xVM");
382 #else
383 cmn_err(CE_WARN,
384 "Fast reboot is not supported on this platform%s",
385 fastreboot_nosup_message());
386 #endif
387 }
388
389 if (fcn == AD_FASTREBOOT) {
390 fastboot_load_kernel(mdep);
391 if (!newkernel.fi_valid)
392 fcn = AD_BOOT;
393 }
394
395 (*psm_preshutdownf)(cmd, fcn);
396 }
397
398 static void
399 stop_other_cpus(void)
400 {
401 ulong_t s = clear_int_flag(); /* fast way to keep CPU from changing */
402 cpuset_t xcset;
403
404 CPUSET_ALL_BUT(xcset, CPU->cpu_id);
405 xc_priority(0, 0, 0, CPUSET2BV(xcset), (xc_func_t)mach_cpu_halt);
406 restore_int_flag(s);
407 }
408
409 /*
410 * Machine dependent abort sequence handling
411 */
412 void
413 abort_sequence_enter(char *msg)
414 {
415 if (abort_enable == 0) {
416 if (AU_ZONE_AUDITING(GET_KCTX_GZ))
417 audit_enterprom(0);
418 return;
419 }
420 if (AU_ZONE_AUDITING(GET_KCTX_GZ))
421 audit_enterprom(1);
422 debug_enter(msg);
423 if (AU_ZONE_AUDITING(GET_KCTX_GZ))
424 audit_exitprom(1);
425 }
426
427 /*
428 * Enter debugger. Called when the user types ctrl-alt-d or whenever
429 * code wants to enter the debugger and possibly resume later.
430 */
431 void
432 debug_enter(
433 char *msg) /* message to print, possibly NULL */
434 {
435 if (dtrace_debugger_init != NULL)
436 (*dtrace_debugger_init)();
437
438 if (msg)
439 prom_printf("%s\n", msg);
440
441 if (boothowto & RB_DEBUG)
442 kmdb_enter();
443
444 if (dtrace_debugger_fini != NULL)
445 (*dtrace_debugger_fini)();
446 }
447
448 void
449 reset(void)
450 {
451 extern void acpi_reset_system();
452 #if !defined(__xpv)
453 ushort_t *bios_memchk;
454
455 /*
456 * Can't use psm_map_phys or acpi_reset_system before the hat is
457 * initialized.
458 */
459 if (khat_running) {
460 bios_memchk = (ushort_t *)psm_map_phys(0x472,
461 sizeof (ushort_t), PROT_READ | PROT_WRITE);
462 if (bios_memchk)
463 *bios_memchk = 0x1234; /* bios memory check disable */
464
465 if (options_dip != NULL &&
466 ddi_prop_exists(DDI_DEV_T_ANY, ddi_root_node(), 0,
467 "efi-systab")) {
468 efi_reset();
469 }
470
471 /*
472 * The problem with using stubs is that we can call
473 * acpi_reset_system only after the kernel is up and running.
474 *
475 * We should create a global state to keep track of how far
476 * up the kernel is but for the time being we will depend on
477 * bootops. bootops cleared in startup_end().
478 */
479 if (bootops == NULL)
480 acpi_reset_system();
481 }
482
483 pc_reset();
484 #else
485 if (IN_XPV_PANIC()) {
486 if (khat_running && bootops == NULL) {
487 acpi_reset_system();
488 }
489
490 pc_reset();
491 }
492
493 (void) HYPERVISOR_shutdown(SHUTDOWN_reboot);
494 panic("HYPERVISOR_shutdown() failed");
495 #endif
496 /*NOTREACHED*/
497 }
498
499 /*
500 * Halt the machine and return to the monitor
501 */
502 void
503 halt(char *s)
504 {
505 stop_other_cpus(); /* send stop signal to other CPUs */
506 if (s)
507 prom_printf("(%s) \n", s);
508 prom_exit_to_mon();
509 /*NOTREACHED*/
510 }
511
512 /*
513 * Initiate interrupt redistribution.
514 */
515 void
516 i_ddi_intr_redist_all_cpus()
517 {
518 }
519
520 /*
521 * XXX These probably ought to live somewhere else
522 * XXX They are called from mem.c
523 */
524
525 /*
526 * Convert page frame number to an OBMEM page frame number
527 * (i.e. put in the type bits -- zero for this implementation)
528 */
529 pfn_t
530 impl_obmem_pfnum(pfn_t pf)
531 {
532 return (pf);
533 }
534
535 #ifdef NM_DEBUG
536 int nmi_test = 0; /* checked in intentry.s during clock int */
537 int nmtest = -1;
538 nmfunc1(arg, rp)
539 int arg;
540 struct regs *rp;
541 {
542 printf("nmi called with arg = %x, regs = %x\n", arg, rp);
543 nmtest += 50;
544 if (arg == nmtest) {
545 printf("ip = %x\n", rp->r_pc);
546 return (1);
547 }
548 return (0);
549 }
550
551 #endif
552
553 #include <sys/bootsvcs.h>
554
555 /* Hacked up initialization for initial kernel check out is HERE. */
556 /* The basic steps are: */
557 /* kernel bootfuncs definition/initialization for KADB */
558 /* kadb bootfuncs pointer initialization */
559 /* putchar/getchar (interrupts disabled) */
560
561 /* kadb bootfuncs pointer initialization */
562
563 int
564 sysp_getchar()
565 {
566 int i;
567 ulong_t s;
568
569 if (cons_polledio == NULL) {
570 /* Uh oh */
571 prom_printf("getchar called with no console\n");
572 for (;;)
573 /* LOOP FOREVER */;
574 }
575
576 s = clear_int_flag();
577 i = cons_polledio->cons_polledio_getchar(
578 cons_polledio->cons_polledio_argument);
579 restore_int_flag(s);
580 return (i);
581 }
582
583 void
584 sysp_putchar(int c)
585 {
586 ulong_t s;
587
588 /*
589 * We have no alternative but to drop the output on the floor.
590 */
591 if (cons_polledio == NULL ||
592 cons_polledio->cons_polledio_putchar == NULL)
593 return;
594
595 s = clear_int_flag();
596 cons_polledio->cons_polledio_putchar(
597 cons_polledio->cons_polledio_argument, c);
598 restore_int_flag(s);
599 }
600
601 int
602 sysp_ischar()
603 {
604 int i;
605 ulong_t s;
606
607 if (cons_polledio == NULL ||
608 cons_polledio->cons_polledio_ischar == NULL)
609 return (0);
610
611 s = clear_int_flag();
612 i = cons_polledio->cons_polledio_ischar(
613 cons_polledio->cons_polledio_argument);
614 restore_int_flag(s);
615 return (i);
616 }
617
618 int
619 goany(void)
620 {
621 prom_printf("Type any key to continue ");
622 (void) prom_getchar();
623 prom_printf("\n");
624 return (1);
625 }
626
627 static struct boot_syscalls kern_sysp = {
628 sysp_getchar, /* unchar (*getchar)(); 7 */
629 sysp_putchar, /* int (*putchar)(); 8 */
630 sysp_ischar, /* int (*ischar)(); 9 */
631 };
632
633 #if defined(__xpv)
634 int using_kern_polledio;
635 #endif
636
637 void
638 kadb_uses_kernel()
639 {
640 /*
641 * This routine is now totally misnamed, since it does not in fact
642 * control kadb's I/O; it only controls the kernel's prom_* I/O.
643 */
644 sysp = &kern_sysp;
645 #if defined(__xpv)
646 using_kern_polledio = 1;
647 #endif
648 }
649
650 /*
651 * the interface to the outside world
652 */
653
654 /*
655 * poll_port -- wait for a register to achieve a
656 * specific state. Arguments are a mask of bits we care about,
657 * and two sub-masks. To return normally, all the bits in the
658 * first sub-mask must be ON, all the bits in the second sub-
659 * mask must be OFF. If about seconds pass without the register
660 * achieving the desired bit configuration, we return 1, else
661 * 0.
662 */
663 int
664 poll_port(ushort_t port, ushort_t mask, ushort_t onbits, ushort_t offbits)
665 {
666 int i;
667 ushort_t maskval;
668
669 for (i = 500000; i; i--) {
670 maskval = inb(port) & mask;
671 if (((maskval & onbits) == onbits) &&
672 ((maskval & offbits) == 0))
673 return (0);
674 drv_usecwait(10);
675 }
676 return (1);
677 }
678
679 /*
680 * set_idle_cpu is called from idle() when a CPU becomes idle.
681 */
682 /*LINTED: static unused */
683 static uint_t last_idle_cpu;
684
685 /*ARGSUSED*/
686 void
687 set_idle_cpu(int cpun)
688 {
689 last_idle_cpu = cpun;
690 (*psm_set_idle_cpuf)(cpun);
691 }
692
693 /*
694 * unset_idle_cpu is called from idle() when a CPU is no longer idle.
695 */
696 /*ARGSUSED*/
697 void
698 unset_idle_cpu(int cpun)
699 {
700 (*psm_unset_idle_cpuf)(cpun);
701 }
702
703 /*
704 * This routine is almost correct now, but not quite. It still needs the
705 * equivalent concept of "hres_last_tick", just like on the sparc side.
706 * The idea is to take a snapshot of the hi-res timer while doing the
707 * hrestime_adj updates under hres_lock in locore, so that the small
708 * interval between interrupt assertion and interrupt processing is
709 * accounted for correctly. Once we have this, the code below should
710 * be modified to subtract off hres_last_tick rather than hrtime_base.
711 *
712 * I'd have done this myself, but I don't have source to all of the
713 * vendor-specific hi-res timer routines (grrr...). The generic hook I
714 * need is something like "gethrtime_unlocked()", which would be just like
715 * gethrtime() but would assume that you're already holding CLOCK_LOCK().
716 * This is what the GET_HRTIME() macro is for on sparc (although it also
717 * serves the function of making time available without a function call
718 * so you don't take a register window overflow while traps are disabled).
719 */
720 void
721 pc_gethrestime(timestruc_t *tp)
722 {
723 int lock_prev;
724 timestruc_t now;
725 int nslt; /* nsec since last tick */
726 int adj; /* amount of adjustment to apply */
727
728 loop:
729 lock_prev = hres_lock;
730 now = hrestime;
731 nslt = (int)(gethrtime() - hres_last_tick);
732 if (nslt < 0) {
733 /*
734 * nslt < 0 means a tick came between sampling
735 * gethrtime() and hres_last_tick; restart the loop
736 */
737
738 goto loop;
739 }
740 now.tv_nsec += nslt;
741 if (hrestime_adj != 0) {
742 if (hrestime_adj > 0) {
743 adj = (nslt >> ADJ_SHIFT);
744 if (adj > hrestime_adj)
745 adj = (int)hrestime_adj;
746 } else {
747 adj = -(nslt >> ADJ_SHIFT);
748 if (adj < hrestime_adj)
749 adj = (int)hrestime_adj;
750 }
751 now.tv_nsec += adj;
752 }
753 while ((unsigned long)now.tv_nsec >= NANOSEC) {
754
755 /*
756 * We might have a large adjustment or have been in the
757 * debugger for a long time; take care of (at most) four
758 * of those missed seconds (tv_nsec is 32 bits, so
759 * anything >4s will be wrapping around). However,
760 * anything more than 2 seconds out of sync will trigger
761 * timedelta from clock() to go correct the time anyway,
762 * so do what we can, and let the big crowbar do the
763 * rest. A similar correction while loop exists inside
764 * hres_tick(); in all cases we'd like tv_nsec to
765 * satisfy 0 <= tv_nsec < NANOSEC to avoid confusing
766 * user processes, but if tv_sec's a little behind for a
767 * little while, that's OK; time still monotonically
768 * increases.
769 */
770
771 now.tv_nsec -= NANOSEC;
772 now.tv_sec++;
773 }
774 if ((hres_lock & ~1) != lock_prev)
775 goto loop;
776
777 *tp = now;
778 }
779
780 void
781 gethrestime_lasttick(timespec_t *tp)
782 {
783 int s;
784
785 s = hr_clock_lock();
786 *tp = hrestime;
787 hr_clock_unlock(s);
788 }
789
790 time_t
791 gethrestime_sec(void)
792 {
793 timestruc_t now;
794
795 gethrestime(&now);
796 return (now.tv_sec);
797 }
798
799 /*
800 * Initialize a kernel thread's stack
801 */
802
803 caddr_t
804 thread_stk_init(caddr_t stk)
805 {
806 ASSERT(((uintptr_t)stk & (STACK_ALIGN - 1)) == 0);
807 return (stk - SA(MINFRAME));
808 }
809
810 /*
811 * Initialize lwp's kernel stack.
812 */
813
814 #ifdef TRAPTRACE
815 /*
816 * There's a tricky interdependency here between use of sysenter and
817 * TRAPTRACE which needs recording to avoid future confusion (this is
818 * about the third time I've re-figured this out ..)
819 *
820 * Here's how debugging lcall works with TRAPTRACE.
821 *
822 * 1 We're in userland with a breakpoint on the lcall instruction.
823 * 2 We execute the instruction - the instruction pushes the userland
824 * %ss, %esp, %efl, %cs, %eip on the stack and zips into the kernel
825 * via the call gate.
826 * 3 The hardware raises a debug trap in kernel mode, the hardware
827 * pushes %efl, %cs, %eip and gets to dbgtrap via the idt.
828 * 4 dbgtrap pushes the error code and trapno and calls cmntrap
829 * 5 cmntrap finishes building a trap frame
830 * 6 The TRACE_REGS macros in cmntrap copy a REGSIZE worth chunk
831 * off the stack into the traptrace buffer.
832 *
833 * This means that the traptrace buffer contains the wrong values in
834 * %esp and %ss, but everything else in there is correct.
835 *
836 * Here's how debugging sysenter works with TRAPTRACE.
837 *
838 * a We're in userland with a breakpoint on the sysenter instruction.
839 * b We execute the instruction - the instruction pushes -nothing-
840 * on the stack, but sets %cs, %eip, %ss, %esp to prearranged
841 * values to take us to sys_sysenter, at the top of the lwp's
842 * stack.
843 * c goto 3
844 *
845 * At this point, because we got into the kernel without the requisite
846 * five pushes on the stack, if we didn't make extra room, we'd
847 * end up with the TRACE_REGS macro fetching the saved %ss and %esp
848 * values from negative (unmapped) stack addresses -- which really bites.
849 * That's why we do the '-= 8' below.
850 *
851 * XXX Note that reading "up" lwp0's stack works because t0 is declared
852 * right next to t0stack in locore.s
853 */
854 #endif
855
856 caddr_t
857 lwp_stk_init(klwp_t *lwp, caddr_t stk)
858 {
859 caddr_t oldstk;
860 struct pcb *pcb = &lwp->lwp_pcb;
861
862 oldstk = stk;
863 stk -= SA(sizeof (struct regs) + SA(MINFRAME));
864 #ifdef TRAPTRACE
865 stk -= 2 * sizeof (greg_t); /* space for phony %ss:%sp (see above) */
866 #endif
867 stk = (caddr_t)((uintptr_t)stk & ~(STACK_ALIGN - 1ul));
868 bzero(stk, oldstk - stk);
869 lwp->lwp_regs = (void *)(stk + SA(MINFRAME));
870
871 /*
872 * Arrange that the virtualized %fs and %gs GDT descriptors
873 * have a well-defined initial state (present, ring 3
874 * and of type data).
875 */
876 #if defined(__amd64)
877 if (lwp_getdatamodel(lwp) == DATAMODEL_NATIVE)
878 pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
879 else
880 pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_u32desc;
881 #elif defined(__i386)
882 pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
883 #endif /* __i386 */
884 lwp_installctx(lwp);
885 return (stk);
886 }
887
888 /*ARGSUSED*/
889 void
890 lwp_stk_fini(klwp_t *lwp)
891 {}
892
893 /*
894 * If we're not the panic CPU, we wait in panic_idle for reboot.
895 */
896 void
897 panic_idle(void)
898 {
899 splx(ipltospl(CLOCK_LEVEL));
900 (void) setjmp(&curthread->t_pcb);
901
902 dumpsys_helper();
903
904 #ifndef __xpv
905 for (;;)
906 i86_halt();
907 #else
908 for (;;)
909 ;
910 #endif
911 }
912
913 /*
914 * Stop the other CPUs by cross-calling them and forcing them to enter
915 * the panic_idle() loop above.
916 */
917 /*ARGSUSED*/
918 void
919 panic_stopcpus(cpu_t *cp, kthread_t *t, int spl)
920 {
921 processorid_t i;
922 cpuset_t xcset;
923
924 /*
925 * In the case of a Xen panic, the hypervisor has already stopped
926 * all of the CPUs.
927 */
928 if (!IN_XPV_PANIC()) {
929 (void) splzs();
930
931 CPUSET_ALL_BUT(xcset, cp->cpu_id);
932 xc_priority(0, 0, 0, CPUSET2BV(xcset), (xc_func_t)panic_idle);
933 }
934
935 for (i = 0; i < NCPU; i++) {
936 if (i != cp->cpu_id && cpu[i] != NULL &&
937 (cpu[i]->cpu_flags & CPU_EXISTS))
938 cpu[i]->cpu_flags |= CPU_QUIESCED;
939 }
940 }
941
942 /*
943 * Platform callback following each entry to panicsys().
944 */
945 /*ARGSUSED*/
946 void
947 panic_enter_hw(int spl)
948 {
949 /* Nothing to do here */
950 }
951
952 /*
953 * Platform-specific code to execute after panicstr is set: we invoke
954 * the PSM entry point to indicate that a panic has occurred.
955 */
956 /*ARGSUSED*/
957 void
958 panic_quiesce_hw(panic_data_t *pdp)
959 {
960 psm_notifyf(PSM_PANIC_ENTER);
961
962 cmi_panic_callback();
963
964 #ifdef TRAPTRACE
965 /*
966 * Turn off TRAPTRACE
967 */
968 TRAPTRACE_FREEZE;
969 #endif /* TRAPTRACE */
970 }
971
972 /*
973 * Platform callback prior to writing crash dump.
974 */
975 /*ARGSUSED*/
976 void
977 panic_dump_hw(int spl)
978 {
979 /* Nothing to do here */
980 }
981
982 void *
983 plat_traceback(void *fpreg)
984 {
985 #ifdef __xpv
986 if (IN_XPV_PANIC())
987 return (xpv_traceback(fpreg));
988 #endif
989 return (fpreg);
990 }
991
992 /*ARGSUSED*/
993 void
994 plat_tod_fault(enum tod_fault_type tod_bad)
995 {}
996
997 /*ARGSUSED*/
998 int
999 blacklist(int cmd, const char *scheme, nvlist_t *fmri, const char *class)
1000 {
1001 return (ENOTSUP);
1002 }
1003
1004 /*
1005 * The underlying console output routines are protected by raising IPL in case
1006 * we are still calling into the early boot services. Once we start calling
1007 * the kernel console emulator, it will disable interrupts completely during
1008 * character rendering (see sysp_putchar, for example). Refer to the comments
1009 * and code in common/os/console.c for more information on these callbacks.
1010 */
1011 /*ARGSUSED*/
1012 int
1013 console_enter(int busy)
1014 {
1015 return (splzs());
1016 }
1017
1018 /*ARGSUSED*/
1019 void
1020 console_exit(int busy, int spl)
1021 {
1022 splx(spl);
1023 }
1024
1025 /*
1026 * Allocate a region of virtual address space, unmapped.
1027 * Stubbed out except on sparc, at least for now.
1028 */
1029 /*ARGSUSED*/
1030 void *
1031 boot_virt_alloc(void *addr, size_t size)
1032 {
1033 return (addr);
1034 }
1035
1036 volatile unsigned long tenmicrodata;
1037
1038 void
1039 tenmicrosec(void)
1040 {
1041 extern int gethrtime_hires;
1042
1043 if (gethrtime_hires) {
1044 hrtime_t start, end;
1045 start = end = gethrtime();
1046 while ((end - start) < (10 * (NANOSEC / MICROSEC))) {
1047 SMT_PAUSE();
1048 end = gethrtime();
1049 }
1050 } else {
1051 #if defined(__xpv)
1052 hrtime_t newtime;
1053
1054 newtime = xpv_gethrtime() + 10000; /* now + 10 us */
1055 while (xpv_gethrtime() < newtime)
1056 SMT_PAUSE();
1057 #else /* __xpv */
1058 int i;
1059
1060 /*
1061 * Artificial loop to induce delay.
1062 */
1063 for (i = 0; i < microdata; i++)
1064 tenmicrodata = microdata;
1065 #endif /* __xpv */
1066 }
1067 }
1068
1069 /*
1070 * get_cpu_mstate() is passed an array of timestamps, NCMSTATES
1071 * long, and it fills in the array with the time spent on cpu in
1072 * each of the mstates, where time is returned in nsec.
1073 *
1074 * No guarantee is made that the returned values in times[] will
1075 * monotonically increase on sequential calls, although this will
1076 * be true in the long run. Any such guarantee must be handled by
1077 * the caller, if needed. This can happen if we fail to account
1078 * for elapsed time due to a generation counter conflict, yet we
1079 * did account for it on a prior call (see below).
1080 *
1081 * The complication is that the cpu in question may be updating
1082 * its microstate at the same time that we are reading it.
1083 * Because the microstate is only updated when the CPU's state
1084 * changes, the values in cpu_intracct[] can be indefinitely out
1085 * of date. To determine true current values, it is necessary to
1086 * compare the current time with cpu_mstate_start, and add the
1087 * difference to times[cpu_mstate].
1088 *
1089 * This can be a problem if those values are changing out from
1090 * under us. Because the code path in new_cpu_mstate() is
1091 * performance critical, we have not added a lock to it. Instead,
1092 * we have added a generation counter. Before beginning
1093 * modifications, the counter is set to 0. After modifications,
1094 * it is set to the old value plus one.
1095 *
1096 * get_cpu_mstate() will not consider the values of cpu_mstate
1097 * and cpu_mstate_start to be usable unless the value of
1098 * cpu_mstate_gen is both non-zero and unchanged, both before and
1099 * after reading the mstate information. Note that we must
1100 * protect against out-of-order loads around accesses to the
1101 * generation counter. Also, this is a best effort approach in
1102 * that we do not retry should the counter be found to have
1103 * changed.
1104 *
1105 * cpu_intracct[] is used to identify time spent in each CPU
1106 * mstate while handling interrupts. Such time should be reported
1107 * against system time, and so is subtracted out from its
1108 * corresponding cpu_acct[] time and added to
1109 * cpu_acct[CMS_SYSTEM].
1110 */
1111
1112 void
1113 get_cpu_mstate(cpu_t *cpu, hrtime_t *times)
1114 {
1115 int i;
1116 hrtime_t now, start;
1117 uint16_t gen;
1118 uint16_t state;
1119 hrtime_t intracct[NCMSTATES];
1120
1121 /*
1122 * Load all volatile state under the protection of membar.
1123 * cpu_acct[cpu_mstate] must be loaded to avoid double counting
1124 * of (now - cpu_mstate_start) by a change in CPU mstate that
1125 * arrives after we make our last check of cpu_mstate_gen.
1126 */
1127
1128 now = gethrtime_unscaled();
1129 gen = cpu->cpu_mstate_gen;
1130
1131 membar_consumer(); /* guarantee load ordering */
1132 start = cpu->cpu_mstate_start;
1133 state = cpu->cpu_mstate;
1134 for (i = 0; i < NCMSTATES; i++) {
1135 intracct[i] = cpu->cpu_intracct[i];
1136 times[i] = cpu->cpu_acct[i];
1137 }
1138 membar_consumer(); /* guarantee load ordering */
1139
1140 if (gen != 0 && gen == cpu->cpu_mstate_gen && now > start)
1141 times[state] += now - start;
1142
1143 for (i = 0; i < NCMSTATES; i++) {
1144 if (i == CMS_SYSTEM)
1145 continue;
1146 times[i] -= intracct[i];
1147 if (times[i] < 0) {
1148 intracct[i] += times[i];
1149 times[i] = 0;
1150 }
1151 times[CMS_SYSTEM] += intracct[i];
1152 scalehrtime(×[i]);
1153 }
1154 scalehrtime(×[CMS_SYSTEM]);
1155 }
1156
1157 /*
1158 * This is a version of the rdmsr instruction that allows
1159 * an error code to be returned in the case of failure.
1160 */
1161 int
1162 checked_rdmsr(uint_t msr, uint64_t *value)
1163 {
1164 if (!is_x86_feature(x86_featureset, X86FSET_MSR))
1165 return (ENOTSUP);
1166 *value = rdmsr(msr);
1167 return (0);
1168 }
1169
1170 /*
1171 * This is a version of the wrmsr instruction that allows
1172 * an error code to be returned in the case of failure.
1173 */
1174 int
1175 checked_wrmsr(uint_t msr, uint64_t value)
1176 {
1177 if (!is_x86_feature(x86_featureset, X86FSET_MSR))
1178 return (ENOTSUP);
1179 wrmsr(msr, value);
1180 return (0);
1181 }
1182
1183 /*
1184 * The mem driver's usual method of using hat_devload() to establish a
1185 * temporary mapping will not work for foreign pages mapped into this
1186 * domain or for the special hypervisor-provided pages. For the foreign
1187 * pages, we often don't know which domain owns them, so we can't ask the
1188 * hypervisor to set up a new mapping. For the other pages, we don't have
1189 * a pfn, so we can't create a new PTE. For these special cases, we do a
1190 * direct uiomove() from the existing kernel virtual address.
1191 */
1192 /*ARGSUSED*/
1193 int
1194 plat_mem_do_mmio(struct uio *uio, enum uio_rw rw)
1195 {
1196 #if defined(__xpv)
1197 void *va = (void *)(uintptr_t)uio->uio_loffset;
1198 off_t pageoff = uio->uio_loffset & PAGEOFFSET;
1199 size_t nbytes = MIN((size_t)(PAGESIZE - pageoff),
1200 (size_t)uio->uio_iov->iov_len);
1201
1202 if ((rw == UIO_READ &&
1203 (va == HYPERVISOR_shared_info || va == xen_info)) ||
1204 (pfn_is_foreign(hat_getpfnum(kas.a_hat, va))))
1205 return (uiomove(va, nbytes, rw, uio));
1206 #endif
1207 return (ENOTSUP);
1208 }
1209
1210 pgcnt_t
1211 num_phys_pages()
1212 {
1213 pgcnt_t npages = 0;
1214 struct memlist *mp;
1215
1216 #if defined(__xpv)
1217 if (DOMAIN_IS_INITDOMAIN(xen_info))
1218 return (xpv_nr_phys_pages());
1219 #endif /* __xpv */
1220
1221 for (mp = phys_install; mp != NULL; mp = mp->ml_next)
1222 npages += mp->ml_size >> PAGESHIFT;
1223
1224 return (npages);
1225 }
1226
1227 /* cpu threshold for compressed dumps */
1228 #ifdef _LP64
1229 uint_t dump_plat_mincpu_default = DUMP_PLAT_X86_64_MINCPU;
1230 #else
1231 uint_t dump_plat_mincpu_default = DUMP_PLAT_X86_32_MINCPU;
1232 #endif
1233
1234 int
1235 dump_plat_addr()
1236 {
1237 #ifdef __xpv
1238 pfn_t pfn = mmu_btop(xen_info->shared_info) | PFN_IS_FOREIGN_MFN;
1239 mem_vtop_t mem_vtop;
1240 int cnt;
1241
1242 /*
1243 * On the hypervisor, we want to dump the page with shared_info on it.
1244 */
1245 if (!IN_XPV_PANIC()) {
1246 mem_vtop.m_as = &kas;
1247 mem_vtop.m_va = HYPERVISOR_shared_info;
1248 mem_vtop.m_pfn = pfn;
1249 dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
1250 cnt = 1;
1251 } else {
1252 cnt = dump_xpv_addr();
1253 }
1254 return (cnt);
1255 #else
1256 return (0);
1257 #endif
1258 }
1259
1260 void
1261 dump_plat_pfn()
1262 {
1263 #ifdef __xpv
1264 pfn_t pfn = mmu_btop(xen_info->shared_info) | PFN_IS_FOREIGN_MFN;
1265
1266 if (!IN_XPV_PANIC())
1267 dumpvp_write(&pfn, sizeof (pfn));
1268 else
1269 dump_xpv_pfn();
1270 #endif
1271 }
1272
1273 /*ARGSUSED*/
1274 int
1275 dump_plat_data(void *dump_cbuf)
1276 {
1277 #ifdef __xpv
1278 uint32_t csize;
1279 int cnt;
1280
1281 if (!IN_XPV_PANIC()) {
1282 csize = (uint32_t)compress(HYPERVISOR_shared_info, dump_cbuf,
1283 PAGESIZE);
1284 dumpvp_write(&csize, sizeof (uint32_t));
1285 dumpvp_write(dump_cbuf, csize);
1286 cnt = 1;
1287 } else {
1288 cnt = dump_xpv_data(dump_cbuf);
1289 }
1290 return (cnt);
1291 #else
1292 return (0);
1293 #endif
1294 }
1295
1296 /*
1297 * Calculates a linear address, given the CS selector and PC values
1298 * by looking up the %cs selector process's LDT or the CPU's GDT.
1299 * proc->p_ldtlock must be held across this call.
1300 */
1301 int
1302 linear_pc(struct regs *rp, proc_t *p, caddr_t *linearp)
1303 {
1304 user_desc_t *descrp;
1305 caddr_t baseaddr;
1306 uint16_t idx = SELTOIDX(rp->r_cs);
1307
1308 ASSERT(rp->r_cs <= 0xFFFF);
1309 ASSERT(MUTEX_HELD(&p->p_ldtlock));
1310
1311 if (SELISLDT(rp->r_cs)) {
1312 /*
1313 * Currently 64 bit processes cannot have private LDTs.
1314 */
1315 ASSERT(p->p_model != DATAMODEL_LP64);
1316
1317 if (p->p_ldt == NULL)
1318 return (-1);
1319
1320 descrp = &p->p_ldt[idx];
1321 baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1322
1323 /*
1324 * Calculate the linear address (wraparound is not only ok,
1325 * it's expected behavior). The cast to uint32_t is because
1326 * LDT selectors are only allowed in 32-bit processes.
1327 */
1328 *linearp = (caddr_t)(uintptr_t)(uint32_t)((uintptr_t)baseaddr +
1329 rp->r_pc);
1330 } else {
1331 #ifdef DEBUG
1332 descrp = &CPU->cpu_gdt[idx];
1333 baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1334 /* GDT-based descriptors' base addresses should always be 0 */
1335 ASSERT(baseaddr == 0);
1336 #endif
1337 *linearp = (caddr_t)(uintptr_t)rp->r_pc;
1338 }
1339
1340 return (0);
1341 }
1342
1343 /*
1344 * The implementation of dtrace_linear_pc is similar to the that of
1345 * linear_pc, above, but here we acquire p_ldtlock before accessing
1346 * p_ldt. This implementation is used by the pid provider; we prefix
1347 * it with "dtrace_" to avoid inducing spurious tracing events.
1348 */
1349 int
1350 dtrace_linear_pc(struct regs *rp, proc_t *p, caddr_t *linearp)
1351 {
1352 user_desc_t *descrp;
1353 caddr_t baseaddr;
1354 uint16_t idx = SELTOIDX(rp->r_cs);
1355
1356 ASSERT(rp->r_cs <= 0xFFFF);
1357
1358 if (SELISLDT(rp->r_cs)) {
1359 /*
1360 * Currently 64 bit processes cannot have private LDTs.
1361 */
1362 ASSERT(p->p_model != DATAMODEL_LP64);
1363
1364 mutex_enter(&p->p_ldtlock);
1365 if (p->p_ldt == NULL) {
1366 mutex_exit(&p->p_ldtlock);
1367 return (-1);
1368 }
1369 descrp = &p->p_ldt[idx];
1370 baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1371 mutex_exit(&p->p_ldtlock);
1372
1373 /*
1374 * Calculate the linear address (wraparound is not only ok,
1375 * it's expected behavior). The cast to uint32_t is because
1376 * LDT selectors are only allowed in 32-bit processes.
1377 */
1378 *linearp = (caddr_t)(uintptr_t)(uint32_t)((uintptr_t)baseaddr +
1379 rp->r_pc);
1380 } else {
1381 #ifdef DEBUG
1382 descrp = &CPU->cpu_gdt[idx];
1383 baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1384 /* GDT-based descriptors' base addresses should always be 0 */
1385 ASSERT(baseaddr == 0);
1386 #endif
1387 *linearp = (caddr_t)(uintptr_t)rp->r_pc;
1388 }
1389
1390 return (0);
1391 }
1392
1393 /*
1394 * We need to post a soft interrupt to reprogram the lbolt cyclic when
1395 * switching from event to cyclic driven lbolt. The following code adds
1396 * and posts the softint for x86.
1397 */
1398 static ddi_softint_hdl_impl_t lbolt_softint_hdl =
1399 {NULL, 0, {{NULL}}, NULL, NULL, NULL, NULL, NULL};
1400
1401 void
1402 lbolt_softint_add(void)
1403 {
1404 (void) add_avsoftintr((void *)&lbolt_softint_hdl, LOCK_LEVEL,
1405 (avfunc)lbolt_ev_to_cyclic, "lbolt_ev_to_cyclic", NULL, NULL);
1406 }
1407
1408 void
1409 lbolt_softint_post(void)
1410 {
1411 (*setsoftint)(CBE_LOCK_PIL, lbolt_softint_hdl.ih_pending);
1412 }
1413
1414 boolean_t
1415 plat_dr_check_capability(uint64_t features)
1416 {
1417 return ((plat_dr_options & features) == features);
1418 }
1419
1420 boolean_t
1421 plat_dr_support_cpu(void)
1422 {
1423 return (plat_dr_options & PLAT_DR_FEATURE_CPU);
1424 }
1425
1426 boolean_t
1427 plat_dr_support_memory(void)
1428 {
1429 return (plat_dr_options & PLAT_DR_FEATURE_MEMORY);
1430 }
1431
1432 void
1433 plat_dr_enable_capability(uint64_t features)
1434 {
1435 atomic_or_64(&plat_dr_options, features);
1436 }
1437
1438 void
1439 plat_dr_disable_capability(uint64_t features)
1440 {
1441 atomic_and_64(&plat_dr_options, ~features);
1442 }