xref: /titanic_41/usr/src/uts/i86xpv/os/xpv_panic.c (revision cd37da7426f0c49c14ad9a8a07638ca971477566)
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 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/clock.h>
31 #include <sys/psm.h>
32 #include <sys/archsystm.h>
33 #include <sys/machsystm.h>
34 #include <sys/compress.h>
35 #include <sys/modctl.h>
36 #include <sys/trap.h>
37 #include <sys/panic.h>
38 #include <sys/regset.h>
39 #include <sys/frame.h>
40 #include <sys/kobj.h>
41 #include <sys/apic.h>
42 #include <sys/dumphdr.h>
43 #include <sys/mem.h>
44 #include <sys/x86_archext.h>
45 #include <sys/xpv_panic.h>
46 #include <sys/boot_console.h>
47 #include <sys/bootsvcs.h>
48 #include <sys/consdev.h>
49 #include <vm/hat_pte.h>
50 #include <vm/hat_i86.h>
51 
52 /* XXX: need to add a PAE version too, if we ever support both PAE and non */
53 #if defined(__i386)
54 #define	XPV_FILENAME	"/boot/xen-syms"
55 #else
56 #define	XPV_FILENAME	"/boot/amd64/xen-syms"
57 #endif
58 #define	XPV_MODNAME	"xpv"
59 
60 int xpv_panicking = 0;
61 
62 struct module *xpv_module;
63 struct modctl *xpv_modctl;
64 
65 #define	ALIGN(x, a)	((a) == 0 ? (uintptr_t)(x) : \
66 	(((uintptr_t)(x) + (uintptr_t)(a) - 1l) & ~((uintptr_t)(a) - 1l)))
67 
68 /* Pointer to the xpv_panic_info structure handed to us by Xen.  */
69 static struct panic_info *xpv_panic_info = NULL;
70 
71 /* Timer support */
72 #define	NSEC_SHIFT 5
73 #define	T_XPV_TIMER	0xd1
74 #define	XPV_TIMER_INTERVAL	1000	/* 1000 microseconds */
75 static uint32_t *xpv_apicadr = NULL;
76 static uint_t	nsec_scale;
77 
78 /* IDT support */
79 #pragma	align	16(xpv_panic_idt)
80 static gate_desc_t	xpv_panic_idt[NIDT];	/* interrupt descriptor table */
81 
82 /* Xen pagetables mapped into our HAT's ptable windows */
83 static pfn_t ptable_pfn[MAX_NUM_LEVEL];
84 
85 /* Number of MMU_PAGESIZE pages we're adding to the Solaris dump */
86 static int xpv_dump_pages;
87 
88 /*
89  * Some commonly used values that we don't want to recompute over and over.
90  */
91 static int xpv_panic_nptes[MAX_NUM_LEVEL];
92 static ulong_t xpv_panic_cr3;
93 static uintptr_t xpv_end;
94 
95 static void xpv_panic_console_print(const char *fmt, ...);
96 static void (*xpv_panic_printf)(const char *, ...) = xpv_panic_console_print;
97 
98 #define	CONSOLE_BUF_SIZE	256
99 static char console_buffer[CONSOLE_BUF_SIZE];
100 static boolean_t use_polledio;
101 
102 static void
103 xpv_panic_putc(int m)
104 {
105 	struct cons_polledio *c = cons_polledio;
106 
107 	/* This really shouldn't happen */
108 	if (console == CONS_HYPERVISOR)
109 		return;
110 
111 	if (use_polledio == B_TRUE)
112 		c->cons_polledio_putchar(c->cons_polledio_argument, m);
113 	else
114 		bcons_putchar(m);
115 }
116 
117 static void
118 xpv_panic_puts(char *msg)
119 {
120 	char *m;
121 
122 	dump_timeleft = dump_timeout;
123 	for (m = msg; *m; m++)
124 		xpv_panic_putc((int)*m);
125 }
126 
127 static void
128 xpv_panic_console_print(const char *fmt, ...)
129 {
130 	va_list ap;
131 
132 	va_start(ap, fmt);
133 	(void) vsnprintf(console_buffer, sizeof (console_buffer), fmt, ap);
134 	va_end(ap);
135 
136 	xpv_panic_puts(console_buffer);
137 }
138 
139 static void
140 xpv_panic_map(int level, pfn_t pfn)
141 {
142 	x86pte_t pte, *pteptr;
143 
144 	/*
145 	 * The provided pfn represents a level 'level' page table.  Map it
146 	 * into the 'level' slot in the list of page table windows.
147 	 */
148 	pteptr = (x86pte_t *)PWIN_PTE_VA(level);
149 	pte = pfn_to_pa(pfn) | PT_VALID;
150 
151 	XPV_ALLOW_PAGETABLE_UPDATES();
152 	if (mmu.pae_hat)
153 		*pteptr = pte;
154 	else
155 		*(x86pte32_t *)pteptr = pte;
156 	XPV_DISALLOW_PAGETABLE_UPDATES();
157 
158 	mmu_tlbflush_entry(PWIN_VA(level));
159 }
160 
161 /*
162  * Walk the page tables to find the pfn mapped by the given va.
163  */
164 static pfn_t
165 xpv_va_walk(uintptr_t *vaddr)
166 {
167 	int l, idx;
168 	pfn_t pfn;
169 	x86pte_t pte;
170 	x86pte_t *ptep;
171 	uintptr_t va = *vaddr;
172 	uintptr_t scan_va;
173 	caddr_t ptable_window;
174 	static pfn_t toplevel_pfn;
175 	static uintptr_t lastva;
176 
177 	/*
178 	 * If we do anything other than a simple scan through memory, don't
179 	 * trust the mapped page tables.
180 	 */
181 	if (va != lastva + MMU_PAGESIZE)
182 		for (l = mmu.max_level; l >= 0; l--)
183 			ptable_pfn[l] = PFN_INVALID;
184 
185 	toplevel_pfn = mmu_btop(xpv_panic_cr3);
186 
187 	while (va < xpv_end && va >= *vaddr) {
188 		/* Find the lowest table with any entry for va */
189 		pfn = toplevel_pfn;
190 		for (l = mmu.max_level; l >= 0; l--) {
191 			if (ptable_pfn[l] != pfn) {
192 				xpv_panic_map(l, pfn);
193 				ptable_pfn[l] = pfn;
194 			}
195 
196 			/*
197 			 * Search this pagetable for any mapping to an
198 			 * address >= va.
199 			 */
200 			ptable_window = PWIN_VA(l);
201 			if (l == mmu.max_level && mmu.pae_hat)
202 				ptable_window +=
203 				    (xpv_panic_cr3 & MMU_PAGEOFFSET);
204 
205 			idx = (va >> LEVEL_SHIFT(l)) & (xpv_panic_nptes[l] - 1);
206 			scan_va = va;
207 			while (idx < xpv_panic_nptes[l] && scan_va < xpv_end &&
208 			    scan_va >= *vaddr) {
209 				ptep = (x86pte_t *)(ptable_window +
210 				    (idx << mmu.pte_size_shift));
211 				pte = GET_PTE(ptep);
212 				if (pte & PTE_VALID)
213 					break;
214 				idx++;
215 				scan_va += mmu.level_size[l];
216 			}
217 			va = scan_va;
218 
219 			/*
220 			 * See if we've hit the end of the range.
221 			 */
222 			if (scan_va >= xpv_end || scan_va < *vaddr) {
223 				va = scan_va;
224 				break;
225 			}
226 
227 			/*
228 			 * If there are no valid mappings in this table, we
229 			 * can skip to the end of the VA range it covers.
230 			 */
231 			if (idx == xpv_panic_nptes[l]) {
232 				va = NEXT_ENTRY_VA(va, l + 1);
233 				break;
234 			}
235 
236 			/*
237 			 * If this mapping is for a pagetable, we drop down
238 			 * to the next level in the hierarchy and look for
239 			 * a mapping in it.
240 			 */
241 			pfn = PTE2MFN(pte, l);
242 			if (!PTE_ISPAGE(pte, l))
243 				continue;
244 
245 			/*
246 			 * The APIC page is magic.  Nothing to see here;
247 			 * move along.
248 			 */
249 			if (((uintptr_t)xpv_apicadr & MMU_PAGEMASK) ==
250 			    (va & MMU_PAGEMASK)) {
251 				va += MMU_PAGESIZE;
252 				break;
253 			}
254 
255 			/* We also want to skip the Xen version of KPM */
256 			if (va >= (uintptr_t)xpv_panic_info->pi_ram_start &&
257 			    va < (uintptr_t)xpv_panic_info->pi_ram_end) {
258 				va = (uintptr_t)xpv_panic_info->pi_ram_end;
259 				break;
260 			}
261 
262 			/*
263 			 * The Xen panic code only handles small pages.  If
264 			 * this mapping is for a large page, we need to
265 			 * identify the consituent page that covers the
266 			 * specific VA we were looking for.
267 			 */
268 			if (l > 0) {
269 				if (l > 1)
270 					panic("Xen panic can't cope with "
271 					    "giant pages.");
272 				idx = (va >> LEVEL_SHIFT(0)) &
273 				    (xpv_panic_nptes[0] - 1);
274 				pfn += idx;
275 			}
276 
277 			*vaddr = va;
278 			lastva = va;
279 			return (pfn | PFN_IS_FOREIGN_MFN);
280 		}
281 	}
282 	return (PFN_INVALID);
283 }
284 
285 /*
286  * Walk through the Xen VA space, finding pages that are mapped in.
287  *
288  * These pages all have MFNs rather than PFNs, meaning they may be outside
289  * the physical address space the kernel knows about, or they may collide
290  * with PFNs the kernel is using.
291  *
292  * The obvious trick of just adding the PFN_IS_FOREIGN_MFN bit to the MFNs
293  * to avoid collisions doesn't work.  The pages need to be written to disk
294  * in PFN-order or savecore gets confused.  We can't allocate memory to
295  * contruct a sorted pfn->VA reverse mapping, so we have to write the pages
296  * to disk in VA order.
297  *
298  * To square this circle, we simply make up PFNs for each of Xen's pages.
299  * We assign each mapped page a fake PFN in ascending order.  These fake
300  * PFNs each have the FOREIGN bit set, ensuring that they fall outside the
301  * range of Solaris PFNs written by the kernel.
302  */
303 int
304 dump_xpv_addr()
305 {
306 	uintptr_t va;
307 	mem_vtop_t mem_vtop;
308 
309 	xpv_dump_pages = 0;
310 	va = xen_virt_start;
311 
312 	while (xpv_va_walk(&va) != PFN_INVALID) {
313 		mem_vtop.m_as = &kas;
314 		mem_vtop.m_va = (void *)va;
315 		mem_vtop.m_pfn = (pfn_t)xpv_dump_pages | PFN_IS_FOREIGN_MFN;
316 
317 		dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
318 		xpv_dump_pages++;
319 
320 		va += MMU_PAGESIZE;
321 	}
322 
323 	/*
324 	 * Add the shared_info page.  This page actually ends up in the
325 	 * dump twice: once for the Xen va and once for the Solaris va.
326 	 * This isn't ideal, but we don't know the address Xen is using for
327 	 * the page, so we can't share it.
328 	 */
329 	mem_vtop.m_as = &kas;
330 	mem_vtop.m_va = HYPERVISOR_shared_info;
331 	mem_vtop.m_pfn = (pfn_t)xpv_dump_pages | PFN_IS_FOREIGN_MFN;
332 	dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
333 	xpv_dump_pages++;
334 
335 	return (xpv_dump_pages);
336 }
337 
338 void
339 dump_xpv_pfn()
340 {
341 	pfn_t pfn;
342 	int cnt;
343 
344 	for (cnt = 0; cnt < xpv_dump_pages; cnt++) {
345 		pfn = (pfn_t)cnt | PFN_IS_FOREIGN_MFN;
346 		dumpvp_write(&pfn, sizeof (pfn));
347 	}
348 }
349 
350 int
351 dump_xpv_data(void *dump_cbuf)
352 {
353 	uintptr_t va;
354 	uint32_t csize;
355 	int cnt = 0;
356 
357 	/*
358 	 * XXX: we should probably run this data through a UE check.  The
359 	 * catch is that the UE code relies on on_trap() and getpfnum()
360 	 * working.
361 	 */
362 	va = xen_virt_start;
363 
364 	while (xpv_va_walk(&va) != PFN_INVALID) {
365 		csize = (uint32_t)compress((void *)va, dump_cbuf, PAGESIZE);
366 		dumpvp_write(&csize, sizeof (uint32_t));
367 		dumpvp_write(dump_cbuf, csize);
368 		if (dump_ioerr) {
369 			dumphdr->dump_flags &= ~DF_COMPLETE;
370 			return (cnt);
371 		}
372 		cnt++;
373 		va += MMU_PAGESIZE;
374 	}
375 
376 	/*
377 	 * Finally, dump the shared_info page
378 	 */
379 	csize = (uint32_t)compress((void *)HYPERVISOR_shared_info, dump_cbuf,
380 	    PAGESIZE);
381 	dumpvp_write(&csize, sizeof (uint32_t));
382 	dumpvp_write(dump_cbuf, csize);
383 	if (dump_ioerr)
384 		dumphdr->dump_flags &= ~DF_COMPLETE;
385 	cnt++;
386 
387 	return (cnt);
388 }
389 
390 static void *
391 showstack(void *fpreg, int xpv_only)
392 {
393 	struct frame *fpp;
394 	ulong_t off;
395 	char *sym;
396 	uintptr_t pc, fp, lastfp;
397 	uintptr_t minaddr = min(KERNELBASE, xen_virt_start);
398 
399 	fp = (uintptr_t)fpreg;
400 	if (fp < minaddr) {
401 		xpv_panic_printf("Bad frame ptr: 0x%p\n", fpreg);
402 		return (fpreg);
403 	}
404 
405 	do {
406 		fpp = (struct frame *)fp;
407 		pc = fpp->fr_savpc;
408 
409 		if ((xpv_only != 0) &&
410 		    (fp > xpv_end || fp < xen_virt_start))
411 			break;
412 		if ((sym = kobj_getsymname(pc, &off)) != NULL)
413 			xpv_panic_printf("%08lx %s:%s+%lx\n", fp,
414 			    mod_containing_pc((caddr_t)pc), sym, off);
415 		else if ((pc >= xen_virt_start) && (pc <= xpv_end))
416 			xpv_panic_printf("%08lx 0x%lx (in Xen)\n", fp, pc);
417 		else
418 			xpv_panic_printf("%08lx %lx\n", fp, pc);
419 
420 		lastfp = fp;
421 		fp = fpp->fr_savfp;
422 
423 		/*
424 		 * Xen marks an exception frame by inverting the frame
425 		 * pointer.
426 		 */
427 		if (fp < lastfp) {
428 			if ((~fp > minaddr) && ((~fp) ^ lastfp) < 0xfff)
429 				fp = ~fp;
430 		}
431 	} while (fp > lastfp);
432 	return ((void *)fp);
433 }
434 
435 void *
436 xpv_traceback(void *fpreg)
437 {
438 	return (showstack(fpreg, 1));
439 }
440 
441 #if defined(__amd64)
442 static void
443 xpv_panic_hypercall(ulong_t call)
444 {
445 	panic("Illegally issued hypercall %d during panic!\n", (int)call);
446 }
447 #endif
448 
449 void
450 xpv_die(struct regs *rp)
451 {
452 	struct panic_trap_info ti;
453 	struct cregs creg;
454 
455 	ti.trap_regs = rp;
456 	ti.trap_type = rp->r_trapno;
457 
458 	curthread->t_panic_trap = &ti;
459 	if (ti.trap_type == T_PGFLT) {
460 		getcregs(&creg);
461 		ti.trap_addr = (caddr_t)creg.cr_cr2;
462 		panic("Fatal pagefault at 0x%lx.  fault addr=0x%p  rp=0x%p",
463 		    rp->r_pc, ti.trap_addr, rp);
464 	} else {
465 		ti.trap_addr = (caddr_t)rp->r_pc;
466 		panic("Fatal trap %ld at 0x%lx.  rp=0x%p", rp->r_trapno,
467 		    rp->r_pc, rp);
468 	}
469 }
470 
471 /*
472  * Build IDT to handle a Xen panic
473  */
474 static void
475 switch_to_xpv_panic_idt()
476 {
477 	int i;
478 	desctbr_t idtr;
479 	gate_desc_t *idt = xpv_panic_idt;
480 	selector_t cs = get_cs_register();
481 
482 	for (i = 0; i < 32; i++)
483 		set_gatesegd(&idt[i], &xpv_invaltrap, cs, SDT_SYSIGT, TRP_XPL);
484 
485 	set_gatesegd(&idt[T_ZERODIV], &xpv_div0trap, cs, SDT_SYSIGT, TRP_XPL);
486 	set_gatesegd(&idt[T_SGLSTP], &xpv_dbgtrap, cs, SDT_SYSIGT, TRP_XPL);
487 	set_gatesegd(&idt[T_NMIFLT], &xpv_nmiint, cs, SDT_SYSIGT, TRP_XPL);
488 	set_gatesegd(&idt[T_BOUNDFLT], &xpv_boundstrap, cs, SDT_SYSIGT,
489 	    TRP_XPL);
490 	set_gatesegd(&idt[T_ILLINST], &xpv_invoptrap, cs, SDT_SYSIGT, TRP_XPL);
491 	set_gatesegd(&idt[T_NOEXTFLT], &xpv_ndptrap, cs, SDT_SYSIGT, TRP_XPL);
492 	set_gatesegd(&idt[T_TSSFLT], &xpv_invtsstrap, cs, SDT_SYSIGT, TRP_XPL);
493 	set_gatesegd(&idt[T_SEGFLT], &xpv_segnptrap, cs, SDT_SYSIGT, TRP_XPL);
494 	set_gatesegd(&idt[T_STKFLT], &xpv_stktrap, cs, SDT_SYSIGT, TRP_XPL);
495 	set_gatesegd(&idt[T_GPFLT], &xpv_gptrap, cs, SDT_SYSIGT, TRP_XPL);
496 	set_gatesegd(&idt[T_PGFLT], &xpv_pftrap, cs, SDT_SYSIGT, TRP_XPL);
497 	set_gatesegd(&idt[T_EXTERRFLT], &xpv_ndperr, cs, SDT_SYSIGT, TRP_XPL);
498 	set_gatesegd(&idt[T_ALIGNMENT], &xpv_achktrap, cs, SDT_SYSIGT, TRP_XPL);
499 	set_gatesegd(&idt[T_MCE], &xpv_mcetrap, cs, SDT_SYSIGT, TRP_XPL);
500 	set_gatesegd(&idt[T_SIMDFPE], &xpv_xmtrap, cs, SDT_SYSIGT, TRP_XPL);
501 
502 	/*
503 	 * We have no double fault handler.  Any single fault represents a
504 	 * catastrophic failure for us, so there is no attempt to handle
505 	 * them cleanly: we just print a message and reboot.  If we
506 	 * encounter a second fault while doing that, there is nothing
507 	 * else we can do.
508 	 */
509 
510 	/*
511 	 * Be prepared to absorb any stray device interrupts received
512 	 * while writing the core to disk.
513 	 */
514 	for (i = 33; i < NIDT; i++)
515 		set_gatesegd(&idt[i], &xpv_surprise_intr, cs, SDT_SYSIGT,
516 		    TRP_XPL);
517 
518 	/* The one interrupt we expect to get is from the APIC timer.  */
519 	set_gatesegd(&idt[T_XPV_TIMER], &xpv_timer_trap, cs, SDT_SYSIGT,
520 	    TRP_XPL);
521 
522 	idtr.dtr_base = (uintptr_t)xpv_panic_idt;
523 	idtr.dtr_limit = sizeof (xpv_panic_idt) - 1;
524 	wr_idtr(&idtr);
525 
526 #if defined(__amd64)
527 	/* Catch any hypercalls. */
528 	wrmsr(MSR_AMD_LSTAR, (uintptr_t)xpv_panic_hypercall);
529 	wrmsr(MSR_AMD_CSTAR, (uintptr_t)xpv_panic_hypercall);
530 #endif
531 }
532 
533 static void
534 xpv_apic_clkinit()
535 {
536 	uint_t		apic_ticks = 0;
537 
538 	/*
539 	 * Measure how many APIC ticks there are within a fixed time
540 	 * period.  We're going to be fairly coarse here.  This timer is
541 	 * just being used to detect a stalled panic, so as long as we have
542 	 * the right order of magnitude, everything should be fine.
543 	 */
544 	xpv_apicadr[APIC_SPUR_INT_REG] = AV_UNIT_ENABLE | APIC_SPUR_INTR;
545 	xpv_apicadr[APIC_LOCAL_TIMER] = AV_MASK;
546 	xpv_apicadr[APIC_INT_VECT0] = AV_MASK;	/* local intr reg 0 */
547 
548 	xpv_apicadr[APIC_DIVIDE_REG] = 0;
549 	xpv_apicadr[APIC_INIT_COUNT] = APIC_MAXVAL;
550 	drv_usecwait(XPV_TIMER_INTERVAL);
551 	apic_ticks = APIC_MAXVAL - xpv_apicadr[APIC_CURR_COUNT];
552 
553 	/*
554 	 * apic_ticks now represents roughly how many apic ticks comprise
555 	 * one timeout interval.  Program the timer to send us an interrupt
556 	 * every time that interval expires.
557 	 */
558 	xpv_apicadr[APIC_LOCAL_TIMER] = T_XPV_TIMER | AV_TIME;
559 	xpv_apicadr[APIC_INIT_COUNT] = apic_ticks;
560 	xpv_apicadr[APIC_EOI_REG] = 0;
561 }
562 
563 void
564 xpv_timer_tick(void)
565 {
566 	static int ticks = 0;
567 
568 	if (ticks++ >= MICROSEC / XPV_TIMER_INTERVAL) {
569 		ticks = 0;
570 		if (dump_timeleft && (--dump_timeleft == 0))
571 			panic("Xen panic timeout\n");
572 	}
573 	xpv_apicadr[APIC_EOI_REG] = 0;
574 }
575 
576 void
577 xpv_interrupt(void)
578 {
579 #ifdef	DEBUG
580 	static int cnt = 0;
581 
582 	if (cnt++ < 10)
583 		xpv_panic_printf("Unexpected interrupt received.\n");
584 	if ((cnt < 1000) && ((cnt % 100) == 0))
585 		xpv_panic_printf("%d unexpected interrupts received.\n", cnt);
586 #endif
587 
588 	xpv_apicadr[APIC_EOI_REG] = 0;
589 }
590 
591 /*
592  * Managing time in panic context is trivial.  We only have a single CPU,
593  * we never get rescheduled, we never get suspended.  We just need to
594  * convert clock ticks into nanoseconds.
595  */
596 static hrtime_t
597 xpv_panic_gethrtime(void)
598 {
599 	hrtime_t tsc, hrt;
600 	unsigned int *l = (unsigned int *)&(tsc);
601 
602 	tsc = __rdtsc_insn();
603 	hrt = (mul32(l[1], nsec_scale) << NSEC_SHIFT) +
604 	    (mul32(l[0], nsec_scale) >> (32 - NSEC_SHIFT));
605 
606 	return (hrt);
607 }
608 
609 static void
610 xpv_panic_time_init()
611 {
612 	nsec_scale =
613 	    CPU->cpu_m.mcpu_vcpu_info->time.tsc_to_system_mul >> NSEC_SHIFT;
614 
615 	gethrtimef = xpv_panic_gethrtime;
616 }
617 
618 static void
619 xpv_panicsys(struct regs *rp, char *fmt, ...)
620 {
621 	extern void panicsys(const char *, va_list, struct regs *, int);
622 	va_list alist;
623 
624 	va_start(alist, fmt);
625 	panicsys(fmt, alist, rp, 1);
626 	va_end(alist);
627 }
628 
629 void
630 xpv_do_panic(void *arg)
631 {
632 	struct panic_info *pip = (struct panic_info *)arg;
633 	int l;
634 	struct cregs creg;
635 #if defined(__amd64)
636 	extern uintptr_t postbootkernelbase;
637 #endif
638 
639 	if (xpv_panicking++ > 0)
640 		panic("multiple calls to xpv_do_panic()");
641 
642 	/*
643 	 * Indicate to the underlying panic framework that a panic has been
644 	 * initiated.  This is ordinarily done as part of vpanic().  Since
645 	 * we already have all the register state saved by the hypervisor,
646 	 * we skip that and jump straight into the panic processing code.
647 	 */
648 	(void) panic_trigger(&panic_quiesce);
649 
650 #if defined(__amd64)
651 	/*
652 	 * bzero() and bcopy() get unhappy when asked to operate on
653 	 * addresses outside of the kernel.  At this point Xen is really a
654 	 * part of the kernel, so we update the routines' notion of where
655 	 * the kernel starts.
656 	 */
657 	postbootkernelbase = xen_virt_start;
658 #endif
659 
660 #if defined(HYPERVISOR_VIRT_END)
661 	xpv_end = HYPERVISOR_VIRT_END;
662 #else
663 	xpv_end = (uintptr_t)UINTPTR_MAX - sizeof (uintptr_t);
664 #endif
665 
666 	/*
667 	 * If we were redirecting console output to the hypervisor, we have
668 	 * to stop.
669 	 */
670 	use_polledio = B_FALSE;
671 	if (console == CONS_HYPERVISOR) {
672 		bcons_device_change(CONS_HYPERVISOR);
673 	} else if (cons_polledio != NULL &&
674 	    cons_polledio->cons_polledio_putchar != NULL)  {
675 		if (cons_polledio->cons_polledio_enter != NULL)
676 			cons_polledio->cons_polledio_enter(
677 			    cons_polledio->cons_polledio_argument);
678 		use_polledio = 1;
679 	}
680 
681 	/* Make sure we handle all console output from here on. */
682 	sysp->bsvc_putchar = xpv_panic_putc;
683 
684 	/*
685 	 * If we find an unsupported panic_info structure, there's not much
686 	 * we can do other than complain, plow on, and hope for the best.
687 	 */
688 	if (pip->pi_version != PANIC_INFO_VERSION)
689 		xpv_panic_printf("Warning: Xen is using an unsupported "
690 		    "version of the panic_info structure.\n");
691 
692 	xpv_panic_info = pip;
693 
694 	/*
695 	 * Make sure we are running on the Solaris %gs.  The Xen panic code
696 	 * should already have set up the GDT properly.
697 	 */
698 	xpv_panic_resetgs();
699 #if defined(__amd64)
700 	wrmsr(MSR_AMD_GSBASE, (uint64_t)&cpus[0]);
701 #endif
702 
703 	xpv_panic_time_init();
704 
705 	/*
706 	 * Switch to our own IDT, avoiding any accidental returns to Xen
707 	 * world.
708 	 */
709 	switch_to_xpv_panic_idt();
710 
711 	/*
712 	 * Initialize the APIC timer, which is used to detect a hung dump
713 	 * attempt.
714 	 */
715 	xpv_apicadr = pip->pi_apic;
716 	xpv_apic_clkinit();
717 
718 	/*
719 	 * Set up a few values that we'll need repeatedly.
720 	 */
721 	getcregs(&creg);
722 	xpv_panic_cr3 = creg.cr_cr3;
723 	for (l = mmu.max_level; l >= 0; l--)
724 		xpv_panic_nptes[l] = mmu.ptes_per_table;
725 #ifdef __i386
726 	if (mmu.pae_hat)
727 		xpv_panic_nptes[mmu.max_level] = 4;
728 #endif
729 
730 	/* Add the fake Xen module to the module list */
731 	if (xpv_module != NULL) {
732 		extern int last_module_id;
733 
734 		xpv_modctl->mod_id = last_module_id++;
735 		xpv_modctl->mod_next = &modules;
736 		xpv_modctl->mod_prev = modules.mod_prev;
737 		modules.mod_prev->mod_next = xpv_modctl;
738 		modules.mod_prev = xpv_modctl;
739 	}
740 	xpv_panic_printf = printf;
741 	xpv_panicsys((struct regs *)pip->pi_regs, pip->pi_panicstr);
742 	xpv_panic_printf("Failed to reboot following panic.\n");
743 	for (;;)
744 		;
745 }
746 
747 /*
748  * Set up the necessary data structures to pretend that the Xen hypervisor
749  * is a loadable module, allowing mdb to find the Xen symbols in a crash
750  * dump.  Since these symbols all map to VA space Solaris doesn't normally
751  * have access to, we don't link these structures into the kernel's lists
752  * until/unless we hit a Xen panic.
753  *
754  * The observant reader will note a striking amount of overlap between this
755  * code and that found in krtld.  While it would be handy if we could just
756  * ask krtld to do this work for us, it's not that simple.  Among the
757  * complications: we're not actually loading the text here (grub did it at
758  * boot), the .text section is writable, there are no relocations to do,
759  * none of the module text/data is in readable memory, etc.  Training krtld
760  * to deal with this weird module is as complicated, and more risky, than
761  * reimplementing the necessary subset of it here.
762  */
763 static void
764 init_xen_module()
765 {
766 	struct _buf *file = NULL;
767 	struct module *mp;
768 	struct modctl *mcp;
769 	int i, shn;
770 	Shdr *shp, *ctf_shp;
771 	char *names = NULL;
772 	size_t n, namesize, text_align, data_align;
773 #if defined(__amd64)
774 	const char machine = EM_AMD64;
775 #else
776 	const char machine = EM_386;
777 #endif
778 
779 	/* Allocate and init the module structure */
780 	mp = kmem_zalloc(sizeof (*mp), KM_SLEEP);
781 	mp->filename = kobj_zalloc(strlen(XPV_FILENAME) + 1, KM_SLEEP);
782 	(void) strcpy(mp->filename, XPV_FILENAME);
783 
784 	/* Allocate and init the modctl structure */
785 	mcp = kmem_zalloc(sizeof (*mcp), KM_SLEEP);
786 	mcp->mod_modname = kobj_zalloc(strlen(XPV_MODNAME) + 1, KM_SLEEP);
787 	(void) strcpy(mcp->mod_modname, XPV_MODNAME);
788 	mcp->mod_filename = kobj_zalloc(strlen(XPV_FILENAME) + 1, KM_SLEEP);
789 	(void) strcpy(mcp->mod_filename, XPV_FILENAME);
790 	mcp->mod_inprogress_thread = (kthread_id_t)-1;
791 	mcp->mod_ref = 1;
792 	mcp->mod_loaded = 1;
793 	mcp->mod_loadcnt = 1;
794 	mcp->mod_mp = mp;
795 
796 	/*
797 	 * Try to open a Xen image that hasn't had its symbol and CTF
798 	 * information stripped off.
799 	 */
800 	file = kobj_open_file(XPV_FILENAME);
801 	if (file == (struct _buf *)-1) {
802 		file = NULL;
803 		goto err;
804 	}
805 
806 	/*
807 	 * Read the header and ensure that this is an ELF file for the
808 	 * proper ISA.  If it's not, somebody has done something very
809 	 * stupid.  Why bother?  See Mencken.
810 	 */
811 	if (kobj_read_file(file, (char *)&mp->hdr, sizeof (mp->hdr), 0) < 0)
812 		goto err;
813 	for (i = 0; i < SELFMAG; i++)
814 		if (mp->hdr.e_ident[i] != ELFMAG[i])
815 			goto err;
816 	if ((mp->hdr.e_ident[EI_DATA] != ELFDATA2LSB) ||
817 	    (mp->hdr.e_machine != machine))
818 		goto err;
819 
820 	/* Read in the section headers */
821 	n = mp->hdr.e_shentsize * mp->hdr.e_shnum;
822 	mp->shdrs = kmem_zalloc(n, KM_SLEEP);
823 	if (kobj_read_file(file, mp->shdrs, n, mp->hdr.e_shoff) < 0)
824 		goto err;
825 
826 	/* Read the section names */
827 	shp = (Shdr *)(mp->shdrs + mp->hdr.e_shstrndx * mp->hdr.e_shentsize);
828 	namesize = shp->sh_size;
829 	names = kmem_zalloc(shp->sh_size, KM_SLEEP);
830 	if (kobj_read_file(file, names, shp->sh_size, shp->sh_offset) < 0)
831 		goto err;
832 
833 	/*
834 	 * Fill in the text and data size fields.
835 	 */
836 	ctf_shp = NULL;
837 	text_align = data_align = 0;
838 	for (shn = 1; shn < mp->hdr.e_shnum; shn++) {
839 		shp = (Shdr *)(mp->shdrs + shn * mp->hdr.e_shentsize);
840 
841 		/* Sanity check the offset of the section name */
842 		if (shp->sh_name >= namesize)
843 			continue;
844 
845 		/* If we find the symtab section, remember it for later. */
846 		if (shp->sh_type == SHT_SYMTAB) {
847 			mp->symtbl_section = shn;
848 			mp->symhdr = shp;
849 			continue;
850 		}
851 
852 		/* If we find the CTF section, remember it for later. */
853 		if ((shp->sh_size != 0) &&
854 		    (strcmp(names + shp->sh_name, ".SUNW_ctf") == 0)) {
855 			ctf_shp = shp;
856 			continue;
857 		}
858 
859 		if (!(shp->sh_flags & SHF_ALLOC))
860 			continue;
861 
862 		/*
863 		 * Xen marks its text section as writable, so we need to
864 		 * look for the name - not just the flag.
865 		 */
866 		if ((strcmp(&names[shp->sh_name], ".text") != NULL) &&
867 		    (shp->sh_flags & SHF_WRITE) != 0) {
868 			if (shp->sh_addralign > data_align)
869 				data_align = shp->sh_addralign;
870 			mp->data_size = ALIGN(mp->data_size, data_align);
871 			mp->data_size += ALIGN(shp->sh_size, 8);
872 			if (mp->data == NULL || mp->data > (char *)shp->sh_addr)
873 				mp->data = (char *)shp->sh_addr;
874 		} else {
875 			if (shp->sh_addralign > text_align)
876 				text_align = shp->sh_addralign;
877 			mp->text_size = ALIGN(mp->text_size, text_align);
878 			mp->text_size += ALIGN(shp->sh_size, 8);
879 			if (mp->text == NULL || mp->text > (char *)shp->sh_addr)
880 				mp->text = (char *)shp->sh_addr;
881 		}
882 	}
883 	kmem_free(names, namesize);
884 	names = NULL;
885 	shp = NULL;
886 	mcp->mod_text = mp->text;
887 	mcp->mod_text_size = mp->text_size;
888 
889 	/*
890 	 * If we have symbol table and string table sections, read them in
891 	 * now.  If we don't, we just plow on.  We'll still get a valid
892 	 * core dump, but finding anything useful will be just a bit
893 	 * harder.
894 	 *
895 	 * Note: we don't bother with a hash table.  We'll never do a
896 	 * symbol lookup unless we crash, and then mdb creates its own.  We
897 	 * also don't try to perform any relocations.  Xen should be loaded
898 	 * exactly where the ELF file indicates, and the symbol information
899 	 * in the file should be complete and correct already.  Static
900 	 * linking ain't all bad.
901 	 */
902 	if ((mp->symhdr != NULL) && (mp->symhdr->sh_link < mp->hdr.e_shnum)) {
903 		mp->strhdr = (Shdr *)
904 		    (mp->shdrs + mp->symhdr->sh_link * mp->hdr.e_shentsize);
905 		mp->nsyms = mp->symhdr->sh_size / mp->symhdr->sh_entsize;
906 
907 		/* Allocate space for the symbol table and strings.  */
908 		mp->symsize = mp->symhdr->sh_size +
909 		    mp->nsyms * sizeof (symid_t) + mp->strhdr->sh_size;
910 		mp->symspace = kmem_zalloc(mp->symsize, KM_SLEEP);
911 		mp->symtbl = mp->symspace;
912 		mp->strings = (char *)(mp->symtbl + mp->symhdr->sh_size);
913 
914 		if ((kobj_read_file(file, mp->symtbl,
915 		    mp->symhdr->sh_size, mp->symhdr->sh_offset) < 0) ||
916 		    (kobj_read_file(file, mp->strings,
917 		    mp->strhdr->sh_size, mp->strhdr->sh_offset) < 0))
918 			goto err;
919 	}
920 
921 	/*
922 	 * Read in the CTF section
923 	 */
924 	if ((ctf_shp != NULL) && ((moddebug & MODDEBUG_NOCTF) == 0)) {
925 		mp->ctfdata = kmem_zalloc(ctf_shp->sh_size, KM_SLEEP);
926 		mp->ctfsize = ctf_shp->sh_size;
927 		if (kobj_read_file(file, mp->ctfdata, mp->ctfsize,
928 		    ctf_shp->sh_offset) < 0)
929 			goto err;
930 	}
931 
932 	kobj_close_file(file);
933 
934 	xpv_module = mp;
935 	xpv_modctl = mcp;
936 	return;
937 
938 err:
939 	cmn_err(CE_WARN, "Failed to initialize xpv module.");
940 	if (file != NULL)
941 		kobj_close_file(file);
942 
943 	kmem_free(mp->filename, strlen(XPV_FILENAME) + 1);
944 	if (mp->shdrs != NULL)
945 		kmem_free(mp->shdrs, mp->hdr.e_shentsize * mp->hdr.e_shnum);
946 	if (mp->symspace != NULL)
947 		kmem_free(mp->symspace, mp->symsize);
948 	if (mp->ctfdata != NULL)
949 		kmem_free(mp->ctfdata, mp->ctfsize);
950 	kmem_free(mp, sizeof (*mp));
951 	kmem_free(mcp->mod_filename, strlen(XPV_FILENAME) + 1);
952 	kmem_free(mcp->mod_modname, strlen(XPV_MODNAME) + 1);
953 	kmem_free(mcp, sizeof (*mcp));
954 	if (names != NULL)
955 		kmem_free(names, namesize);
956 }
957 
958 void
959 xpv_panic_init()
960 {
961 	xen_platform_op_t op;
962 	int i;
963 
964 	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
965 
966 	for (i = 0; i < mmu.num_level; i++)
967 		ptable_pfn[i] = PFN_INVALID;
968 
969 	/* Let Xen know where to jump if/when it panics. */
970 	op.cmd = XENPF_panic_init;
971 	op.interface_version = XENPF_INTERFACE_VERSION;
972 	op.u.panic_init.panic_addr = (unsigned long)xpv_panic_hdlr;
973 
974 	(void) HYPERVISOR_platform_op(&op);
975 
976 	init_xen_module();
977 }
978