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 * Copyright (c) 2012 Gary Mills
23 * Copyright 2016 PALO, Richard.
24 *
25 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
26 */
27
28 #include <sys/types.h>
29 #include <sys/clock.h>
30 #include <sys/psm.h>
31 #include <sys/archsystm.h>
32 #include <sys/machsystm.h>
33 #include <sys/compress.h>
34 #include <sys/modctl.h>
35 #include <sys/trap.h>
36 #include <sys/panic.h>
37 #include <sys/regset.h>
38 #include <sys/frame.h>
39 #include <sys/kobj.h>
40 #include <sys/apic.h>
41 #include <sys/apic_timer.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 * There are up to two large swathes of RAM that we don't want to include
90 * in the dump: those that comprise the Xen version of segkpm. On 32-bit
91 * systems there is no such region of memory. On 64-bit systems, there
92 * should be just a single contiguous region that corresponds to all of
93 * physical memory. The tricky bit is that Xen's heap sometimes lives in
94 * the middle of their segkpm, and is mapped using only kpm-like addresses.
95 * In that case, we need to skip the swathes before and after Xen's heap.
96 */
97 uintptr_t kpm1_low = 0;
98 uintptr_t kpm1_high = 0;
99 uintptr_t kpm2_low = 0;
100 uintptr_t kpm2_high = 0;
101
102 /*
103 * Some commonly used values that we don't want to recompute over and over.
104 */
105 static int xpv_panic_nptes[MAX_NUM_LEVEL];
106 static ulong_t xpv_panic_cr3;
107 static uintptr_t xpv_end;
108
109 static void xpv_panic_console_print(const char *fmt, ...);
110 static void (*xpv_panic_printf)(const char *, ...) = xpv_panic_console_print;
111
112 #define CONSOLE_BUF_SIZE 256
113 static char console_buffer[CONSOLE_BUF_SIZE];
114 static boolean_t use_polledio;
115
116 /*
117 * Pointers to machine check panic info (if any).
118 */
119 xpv_mca_panic_data_t *xpv_mca_panic_data = NULL;
120
121 static void
122 xpv_panic_putc(int m)
123 {
124 struct cons_polledio *c = cons_polledio;
125
126 /* This really shouldn't happen */
127 if (boot_console_type(NULL) == CONS_HYPERVISOR)
128 return;
129
130 if (use_polledio == B_TRUE)
131 c->cons_polledio_putchar(c->cons_polledio_argument, m);
132 else
133 bcons_putchar(m);
134 }
135
136 static void
137 xpv_panic_puts(char *msg)
138 {
139 char *m;
140
141 dump_timeleft = dump_timeout;
142 for (m = msg; *m; m++)
143 xpv_panic_putc((int)*m);
144 }
145
146 static void
147 xpv_panic_console_print(const char *fmt, ...)
148 {
149 va_list ap;
150
151 va_start(ap, fmt);
152 (void) vsnprintf(console_buffer, sizeof (console_buffer), fmt, ap);
153 va_end(ap);
154
155 xpv_panic_puts(console_buffer);
156 }
157
158 static void
159 xpv_panic_map(int level, pfn_t pfn)
160 {
161 x86pte_t pte, *pteptr;
162
163 /*
164 * The provided pfn represents a level 'level' page table. Map it
165 * into the 'level' slot in the list of page table windows.
166 */
167 pteptr = (x86pte_t *)PWIN_PTE_VA(level);
168 pte = pfn_to_pa(pfn) | PT_VALID;
169
170 XPV_ALLOW_PAGETABLE_UPDATES();
171 if (mmu.pae_hat)
172 *pteptr = pte;
173 else
174 *(x86pte32_t *)pteptr = pte;
175 XPV_DISALLOW_PAGETABLE_UPDATES();
176
177 mmu_tlbflush_entry(PWIN_VA(level));
178 }
179
180 /*
181 * Walk the page tables to find the pfn mapped by the given va.
182 */
183 static pfn_t
184 xpv_va_walk(uintptr_t *vaddr)
185 {
186 int l, idx;
187 pfn_t pfn;
188 x86pte_t pte;
189 x86pte_t *ptep;
190 uintptr_t va = *vaddr;
191 uintptr_t scan_va;
192 caddr_t ptable_window;
193 static pfn_t toplevel_pfn;
194 static uintptr_t lastva;
195
196 /*
197 * If we do anything other than a simple scan through memory, don't
198 * trust the mapped page tables.
199 */
200 if (va != lastva + MMU_PAGESIZE)
201 for (l = mmu.max_level; l >= 0; l--)
202 ptable_pfn[l] = PFN_INVALID;
203
204 toplevel_pfn = mmu_btop(xpv_panic_cr3);
205
206 while (va < xpv_end && va >= *vaddr) {
207 /* Find the lowest table with any entry for va */
208 pfn = toplevel_pfn;
209 for (l = mmu.max_level; l >= 0; l--) {
210 if (ptable_pfn[l] != pfn) {
211 xpv_panic_map(l, pfn);
212 ptable_pfn[l] = pfn;
213 }
214
215 /*
216 * Search this pagetable for any mapping to an
217 * address >= va.
218 */
219 ptable_window = PWIN_VA(l);
220 if (l == mmu.max_level && mmu.pae_hat)
221 ptable_window +=
222 (xpv_panic_cr3 & MMU_PAGEOFFSET);
223
224 idx = (va >> LEVEL_SHIFT(l)) & (xpv_panic_nptes[l] - 1);
225 scan_va = va;
226 while (idx < xpv_panic_nptes[l] && scan_va < xpv_end &&
227 scan_va >= *vaddr) {
228 ptep = (x86pte_t *)(ptable_window +
229 (idx << mmu.pte_size_shift));
230 pte = GET_PTE(ptep);
231 if (pte & PTE_VALID)
232 break;
233 idx++;
234 scan_va += mmu.level_size[l];
235 }
236
237 /*
238 * If there are no valid mappings in this table, we
239 * can skip to the end of the VA range it covers.
240 */
241 if (idx == xpv_panic_nptes[l]) {
242 va = NEXT_ENTRY_VA(va, l + 1);
243 break;
244 }
245
246 va = scan_va;
247 /*
248 * See if we've hit the end of the range.
249 */
250 if (va >= xpv_end || va < *vaddr)
251 break;
252
253 /*
254 * If this mapping is for a pagetable, we drop down
255 * to the next level in the hierarchy and look for
256 * a mapping in it.
257 */
258 pfn = PTE2MFN(pte, l);
259 if (!PTE_ISPAGE(pte, l))
260 continue;
261
262 /*
263 * The APIC page is magic. Nothing to see here;
264 * move along.
265 */
266 if (((uintptr_t)xpv_apicadr & MMU_PAGEMASK) ==
267 (va & MMU_PAGEMASK)) {
268 va += MMU_PAGESIZE;
269 break;
270 }
271
272 /*
273 * See if the address is within one of the two
274 * kpm-like regions we want to skip.
275 */
276 if (va >= kpm1_low && va < kpm1_high) {
277 va = kpm1_high;
278 break;
279 }
280 if (va >= kpm2_low && va < kpm2_high) {
281 va = kpm2_high;
282 break;
283 }
284
285 /*
286 * The Xen panic code only handles small pages. If
287 * this mapping is for a large page, we need to
288 * identify the consituent page that covers the
289 * specific VA we were looking for.
290 */
291 if (l > 0) {
292 if (l > 1)
293 panic("Xen panic can't cope with "
294 "giant pages.");
295 idx = (va >> LEVEL_SHIFT(0)) &
296 (xpv_panic_nptes[0] - 1);
297 pfn += idx;
298 }
299
300 *vaddr = va;
301 lastva = va;
302 return (pfn | PFN_IS_FOREIGN_MFN);
303 }
304 }
305 return (PFN_INVALID);
306 }
307
308 /*
309 * Walk through the Xen VA space, finding pages that are mapped in.
310 *
311 * These pages all have MFNs rather than PFNs, meaning they may be outside
312 * the physical address space the kernel knows about, or they may collide
313 * with PFNs the kernel is using.
314 *
315 * The obvious trick of just adding the PFN_IS_FOREIGN_MFN bit to the MFNs
316 * to avoid collisions doesn't work. The pages need to be written to disk
317 * in PFN-order or savecore gets confused. We can't allocate memory to
318 * contruct a sorted pfn->VA reverse mapping, so we have to write the pages
319 * to disk in VA order.
320 *
321 * To square this circle, we simply make up PFNs for each of Xen's pages.
322 * We assign each mapped page a fake PFN in ascending order. These fake
323 * PFNs each have the FOREIGN bit set, ensuring that they fall outside the
324 * range of Solaris PFNs written by the kernel.
325 */
326 int
327 dump_xpv_addr()
328 {
329 uintptr_t va;
330 mem_vtop_t mem_vtop;
331
332 xpv_dump_pages = 0;
333 va = xen_virt_start;
334
335 while (xpv_va_walk(&va) != PFN_INVALID) {
336 mem_vtop.m_as = &kas;
337 mem_vtop.m_va = (void *)va;
338 mem_vtop.m_pfn = (pfn_t)xpv_dump_pages | PFN_IS_FOREIGN_MFN;
339
340 dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
341 xpv_dump_pages++;
342
343 va += MMU_PAGESIZE;
344 }
345
346 /*
347 * Add the shared_info page. This page actually ends up in the
348 * dump twice: once for the Xen va and once for the Solaris va.
349 * This isn't ideal, but we don't know the address Xen is using for
350 * the page, so we can't share it.
351 */
352 mem_vtop.m_as = &kas;
353 mem_vtop.m_va = HYPERVISOR_shared_info;
354 mem_vtop.m_pfn = (pfn_t)xpv_dump_pages | PFN_IS_FOREIGN_MFN;
355 dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
356 xpv_dump_pages++;
357
358 return (xpv_dump_pages);
359 }
360
361 void
362 dump_xpv_pfn()
363 {
364 pfn_t pfn;
365 int cnt;
366
367 for (cnt = 0; cnt < xpv_dump_pages; cnt++) {
368 pfn = (pfn_t)cnt | PFN_IS_FOREIGN_MFN;
369 dumpvp_write(&pfn, sizeof (pfn));
370 }
371 }
372
373 int
374 dump_xpv_data(void *dump_cbuf)
375 {
376 uintptr_t va;
377 uint32_t csize;
378 int cnt = 0;
379
380 /*
381 * XXX: we should probably run this data through a UE check. The
382 * catch is that the UE code relies on on_trap() and getpfnum()
383 * working.
384 */
385 va = xen_virt_start;
386
387 while (xpv_va_walk(&va) != PFN_INVALID) {
388 csize = (uint32_t)compress((void *)va, dump_cbuf, PAGESIZE);
389 dumpvp_write(&csize, sizeof (uint32_t));
390 dumpvp_write(dump_cbuf, csize);
391 if (dump_ioerr) {
392 dumphdr->dump_flags &= ~DF_COMPLETE;
393 return (cnt);
394 }
395 cnt++;
396 va += MMU_PAGESIZE;
397 }
398
399 /*
400 * Finally, dump the shared_info page
401 */
402 csize = (uint32_t)compress((void *)HYPERVISOR_shared_info, dump_cbuf,
403 PAGESIZE);
404 dumpvp_write(&csize, sizeof (uint32_t));
405 dumpvp_write(dump_cbuf, csize);
406 if (dump_ioerr)
407 dumphdr->dump_flags &= ~DF_COMPLETE;
408 cnt++;
409
410 return (cnt);
411 }
412
413 static void *
414 showstack(void *fpreg, int xpv_only)
415 {
416 struct frame *fpp;
417 ulong_t off;
418 char *sym;
419 uintptr_t pc, fp, lastfp;
420 uintptr_t minaddr = min(KERNELBASE, xen_virt_start);
421
422 fp = (uintptr_t)fpreg;
423 if (fp < minaddr) {
424 xpv_panic_printf("Bad frame ptr: 0x%p\n", fpreg);
425 return (fpreg);
426 }
427
428 do {
429 fpp = (struct frame *)fp;
430 pc = fpp->fr_savpc;
431
432 if ((xpv_only != 0) &&
433 (fp > xpv_end || fp < xen_virt_start))
434 break;
435 if ((sym = kobj_getsymname(pc, &off)) != NULL)
436 xpv_panic_printf("%08lx %s:%s+%lx\n", fp,
437 mod_containing_pc((caddr_t)pc), sym, off);
438 else if ((pc >= xen_virt_start) && (pc <= xpv_end))
439 xpv_panic_printf("%08lx 0x%lx (in Xen)\n", fp, pc);
440 else
441 xpv_panic_printf("%08lx %lx\n", fp, pc);
442
443 lastfp = fp;
444 fp = fpp->fr_savfp;
445
446 /*
447 * Xen marks an exception frame by inverting the frame
448 * pointer.
449 */
450 if (fp < lastfp) {
451 if ((~fp > minaddr) && ((~fp) ^ lastfp) < 0xfff)
452 fp = ~fp;
453 }
454 } while (fp > lastfp);
455 return ((void *)fp);
456 }
457
458 void *
459 xpv_traceback(void *fpreg)
460 {
461 return (showstack(fpreg, 1));
462 }
463
464 #if defined(__amd64)
465 static void
466 xpv_panic_hypercall(ulong_t call)
467 {
468 panic("Illegally issued hypercall %d during panic!\n", (int)call);
469 }
470 #endif
471
472 void
473 xpv_die(struct regs *rp)
474 {
475 struct panic_trap_info ti;
476 struct cregs creg;
477
478 ti.trap_regs = rp;
479 ti.trap_type = rp->r_trapno;
480
481 curthread->t_panic_trap = &ti;
482 if (ti.trap_type == T_PGFLT) {
483 getcregs(&creg);
484 ti.trap_addr = (caddr_t)creg.cr_cr2;
485 panic("Fatal pagefault at 0x%lx. fault addr=0x%p rp=0x%p",
486 rp->r_pc, (void *)ti.trap_addr, (void *)rp);
487 } else {
488 ti.trap_addr = (caddr_t)rp->r_pc;
489 panic("Fatal trap %ld at 0x%lx. rp=0x%p", rp->r_trapno,
490 rp->r_pc, (void *)rp);
491 }
492 }
493
494 /*
495 * Build IDT to handle a Xen panic
496 */
497 static void
498 switch_to_xpv_panic_idt()
499 {
500 int i;
501 desctbr_t idtr;
502 gate_desc_t *idt = xpv_panic_idt;
503 selector_t cs = get_cs_register();
504
505 for (i = 0; i < 32; i++)
506 set_gatesegd(&idt[i], &xpv_invaltrap, cs, SDT_SYSIGT, TRP_XPL,
507 0);
508
509 set_gatesegd(&idt[T_ZERODIV], &xpv_div0trap, cs, SDT_SYSIGT, TRP_XPL,
510 0);
511 set_gatesegd(&idt[T_SGLSTP], &xpv_dbgtrap, cs, SDT_SYSIGT, TRP_XPL, 0);
512 set_gatesegd(&idt[T_NMIFLT], &xpv_nmiint, cs, SDT_SYSIGT, TRP_XPL, 0);
513 set_gatesegd(&idt[T_BOUNDFLT], &xpv_boundstrap, cs, SDT_SYSIGT,
514 TRP_XPL, 0);
515 set_gatesegd(&idt[T_ILLINST], &xpv_invoptrap, cs, SDT_SYSIGT, TRP_XPL,
516 0);
517 set_gatesegd(&idt[T_NOEXTFLT], &xpv_ndptrap, cs, SDT_SYSIGT, TRP_XPL,
518 0);
519 set_gatesegd(&idt[T_TSSFLT], &xpv_invtsstrap, cs, SDT_SYSIGT, TRP_XPL,
520 0);
521 set_gatesegd(&idt[T_SEGFLT], &xpv_segnptrap, cs, SDT_SYSIGT, TRP_XPL,
522 0);
523 set_gatesegd(&idt[T_STKFLT], &xpv_stktrap, cs, SDT_SYSIGT, TRP_XPL, 0);
524 set_gatesegd(&idt[T_GPFLT], &xpv_gptrap, cs, SDT_SYSIGT, TRP_XPL, 0);
525 set_gatesegd(&idt[T_PGFLT], &xpv_pftrap, cs, SDT_SYSIGT, TRP_XPL, 0);
526 set_gatesegd(&idt[T_EXTERRFLT], &xpv_ndperr, cs, SDT_SYSIGT, TRP_XPL,
527 0);
528 set_gatesegd(&idt[T_ALIGNMENT], &xpv_achktrap, cs, SDT_SYSIGT, TRP_XPL,
529 0);
530 set_gatesegd(&idt[T_MCE], &xpv_mcetrap, cs, SDT_SYSIGT, TRP_XPL, 0);
531 set_gatesegd(&idt[T_SIMDFPE], &xpv_xmtrap, cs, SDT_SYSIGT, TRP_XPL, 0);
532
533 /*
534 * We have no double fault handler. Any single fault represents a
535 * catastrophic failure for us, so there is no attempt to handle
536 * them cleanly: we just print a message and reboot. If we
537 * encounter a second fault while doing that, there is nothing
538 * else we can do.
539 */
540
541 /*
542 * Be prepared to absorb any stray device interrupts received
543 * while writing the core to disk.
544 */
545 for (i = 33; i < NIDT; i++)
546 set_gatesegd(&idt[i], &xpv_surprise_intr, cs, SDT_SYSIGT,
547 TRP_XPL, 0);
548
549 /* The one interrupt we expect to get is from the APIC timer. */
550 set_gatesegd(&idt[T_XPV_TIMER], &xpv_timer_trap, cs, SDT_SYSIGT,
551 TRP_XPL, 0);
552
553 idtr.dtr_base = (uintptr_t)xpv_panic_idt;
554 idtr.dtr_limit = sizeof (xpv_panic_idt) - 1;
555 wr_idtr(&idtr);
556
557 #if defined(__amd64)
558 /* Catch any hypercalls. */
559 wrmsr(MSR_AMD_LSTAR, (uintptr_t)xpv_panic_hypercall);
560 wrmsr(MSR_AMD_CSTAR, (uintptr_t)xpv_panic_hypercall);
561 #endif
562 }
563
564 static void
565 xpv_apic_clkinit()
566 {
567 uint_t apic_ticks = 0;
568
569 /*
570 * Measure how many APIC ticks there are within a fixed time
571 * period. We're going to be fairly coarse here. This timer is
572 * just being used to detect a stalled panic, so as long as we have
573 * the right order of magnitude, everything should be fine.
574 */
575 xpv_apicadr[APIC_SPUR_INT_REG] = AV_UNIT_ENABLE | APIC_SPUR_INTR;
576 xpv_apicadr[APIC_LOCAL_TIMER] = AV_MASK;
577 xpv_apicadr[APIC_INT_VECT0] = AV_MASK; /* local intr reg 0 */
578
579 xpv_apicadr[APIC_DIVIDE_REG] = 0;
580 xpv_apicadr[APIC_INIT_COUNT] = APIC_MAXVAL;
581 drv_usecwait(XPV_TIMER_INTERVAL);
582 apic_ticks = APIC_MAXVAL - xpv_apicadr[APIC_CURR_COUNT];
583
584 /*
585 * apic_ticks now represents roughly how many apic ticks comprise
586 * one timeout interval. Program the timer to send us an interrupt
587 * every time that interval expires.
588 */
589 xpv_apicadr[APIC_LOCAL_TIMER] = T_XPV_TIMER | AV_PERIODIC;
590 xpv_apicadr[APIC_INIT_COUNT] = apic_ticks;
591 xpv_apicadr[APIC_EOI_REG] = 0;
592 }
593
594 void
595 xpv_timer_tick(void)
596 {
597 static int ticks = 0;
598
599 if (ticks++ >= MICROSEC / XPV_TIMER_INTERVAL) {
600 ticks = 0;
601 if (dump_timeleft && (--dump_timeleft == 0))
602 panic("Xen panic timeout\n");
603 }
604 xpv_apicadr[APIC_EOI_REG] = 0;
605 }
606
607 void
608 xpv_interrupt(void)
609 {
610 #ifdef DEBUG
611 static int cnt = 0;
612
613 if (cnt++ < 10)
614 xpv_panic_printf("Unexpected interrupt received.\n");
615 if ((cnt < 1000) && ((cnt % 100) == 0))
616 xpv_panic_printf("%d unexpected interrupts received.\n", cnt);
617 #endif
618
619 xpv_apicadr[APIC_EOI_REG] = 0;
620 }
621
622 /*
623 * Managing time in panic context is trivial. We only have a single CPU,
624 * we never get rescheduled, we never get suspended. We just need to
625 * convert clock ticks into nanoseconds.
626 */
627 static hrtime_t
628 xpv_panic_gethrtime(void)
629 {
630 hrtime_t tsc, hrt;
631 unsigned int *l = (unsigned int *)&(tsc);
632
633 tsc = __rdtsc_insn();
634 hrt = (mul32(l[1], nsec_scale) << NSEC_SHIFT) +
635 (mul32(l[0], nsec_scale) >> (32 - NSEC_SHIFT));
636
637 return (hrt);
638 }
639
640 static void
641 xpv_panic_time_init()
642 {
643 nsec_scale =
644 CPU->cpu_m.mcpu_vcpu_info->time.tsc_to_system_mul >> NSEC_SHIFT;
645
646 gethrtimef = xpv_panic_gethrtime;
647 }
648
649 static void
650 xpv_panicsys(struct regs *rp, char *fmt, ...)
651 {
652 extern void panicsys(const char *, va_list, struct regs *, int);
653 va_list alist;
654
655 va_start(alist, fmt);
656 panicsys(fmt, alist, rp, 1);
657 va_end(alist);
658 }
659
660 void
661 xpv_do_panic(void *arg)
662 {
663 struct panic_info *pip = (struct panic_info *)arg;
664 int l;
665 struct cregs creg;
666 #if defined(__amd64)
667 extern uintptr_t postbootkernelbase;
668 #endif
669
670 if (xpv_panicking++ > 0)
671 panic("multiple calls to xpv_do_panic()");
672
673 /*
674 * Indicate to the underlying panic framework that a panic has been
675 * initiated. This is ordinarily done as part of vpanic(). Since
676 * we already have all the register state saved by the hypervisor,
677 * we skip that and jump straight into the panic processing code.
678 *
679 * XXX If another thread grabs and wins the panic_quiesce trigger
680 * then we'll have two threads in panicsys believing they are in
681 * charge of the panic attempt!
682 */
683 (void) panic_trigger(&panic_quiesce);
684
685 #if defined(__amd64)
686 /*
687 * bzero() and bcopy() get unhappy when asked to operate on
688 * addresses outside of the kernel. At this point Xen is really a
689 * part of the kernel, so we update the routines' notion of where
690 * the kernel starts.
691 */
692 postbootkernelbase = xen_virt_start;
693 #endif
694
695 #if defined(HYPERVISOR_VIRT_END)
696 xpv_end = HYPERVISOR_VIRT_END;
697 #else
698 xpv_end = (uintptr_t)UINTPTR_MAX - sizeof (uintptr_t);
699 #endif
700
701 /*
702 * If we were redirecting console output to the hypervisor, we have
703 * to stop.
704 */
705 use_polledio = B_FALSE;
706 if (boot_console_type(NULL) == CONS_HYPERVISOR) {
707 bcons_device_change(CONS_HYPERVISOR);
708 } else if (cons_polledio != NULL &&
709 cons_polledio->cons_polledio_putchar != NULL) {
710 if (cons_polledio->cons_polledio_enter != NULL)
711 cons_polledio->cons_polledio_enter(
712 cons_polledio->cons_polledio_argument);
713 use_polledio = 1;
714 }
715
716 /* Make sure we handle all console output from here on. */
717 sysp->bsvc_putchar = xpv_panic_putc;
718
719 /*
720 * If we find an unsupported panic_info structure, there's not much
721 * we can do other than complain, plow on, and hope for the best.
722 */
723 if (pip->pi_version != PANIC_INFO_VERSION)
724 xpv_panic_printf("Warning: Xen is using an unsupported "
725 "version of the panic_info structure.\n");
726
727 xpv_panic_info = pip;
728
729 #if defined(__amd64)
730 kpm1_low = (uintptr_t)xpv_panic_info->pi_ram_start;
731 if (xpv_panic_info->pi_xen_start == NULL) {
732 kpm1_high = (uintptr_t)xpv_panic_info->pi_ram_end;
733 } else {
734 kpm1_high = (uintptr_t)xpv_panic_info->pi_xen_start;
735 kpm2_low = (uintptr_t)xpv_panic_info->pi_xen_end;
736 kpm2_high = (uintptr_t)xpv_panic_info->pi_ram_end;
737 }
738 #endif
739
740 /*
741 * Make sure we are running on the Solaris %gs. The Xen panic code
742 * should already have set up the GDT properly.
743 */
744 xpv_panic_resetgs();
745 #if defined(__amd64)
746 wrmsr(MSR_AMD_GSBASE, (uint64_t)&cpus[0]);
747 #endif
748
749 xpv_panic_time_init();
750
751 /*
752 * Switch to our own IDT, avoiding any accidental returns to Xen
753 * world.
754 */
755 switch_to_xpv_panic_idt();
756
757 /*
758 * Initialize the APIC timer, which is used to detect a hung dump
759 * attempt.
760 */
761 xpv_apicadr = pip->pi_apic;
762 xpv_apic_clkinit();
763
764 /*
765 * Set up a few values that we'll need repeatedly.
766 */
767 getcregs(&creg);
768 xpv_panic_cr3 = creg.cr_cr3;
769 for (l = mmu.max_level; l >= 0; l--)
770 xpv_panic_nptes[l] = mmu.ptes_per_table;
771 #ifdef __i386
772 if (mmu.pae_hat)
773 xpv_panic_nptes[mmu.max_level] = 4;
774 #endif
775
776 /* Add the fake Xen module to the module list */
777 if (xpv_module != NULL) {
778 extern int last_module_id;
779
780 xpv_modctl->mod_id = last_module_id++;
781 xpv_modctl->mod_next = &modules;
782 xpv_modctl->mod_prev = modules.mod_prev;
783 modules.mod_prev->mod_next = xpv_modctl;
784 modules.mod_prev = xpv_modctl;
785 }
786
787 if (pip->pi_mca.mpd_magic == MCA_PANICDATA_MAGIC)
788 xpv_mca_panic_data = &pip->pi_mca;
789
790 xpv_panic_printf = printf;
791 xpv_panicsys((struct regs *)pip->pi_regs, pip->pi_panicstr);
792 xpv_panic_printf("Failed to reboot following panic.\n");
793 for (;;)
794 ;
795 }
796
797 /*
798 * Set up the necessary data structures to pretend that the Xen hypervisor
799 * is a loadable module, allowing mdb to find the Xen symbols in a crash
800 * dump. Since these symbols all map to VA space Solaris doesn't normally
801 * have access to, we don't link these structures into the kernel's lists
802 * until/unless we hit a Xen panic.
803 *
804 * The observant reader will note a striking amount of overlap between this
805 * code and that found in krtld. While it would be handy if we could just
806 * ask krtld to do this work for us, it's not that simple. Among the
807 * complications: we're not actually loading the text here (grub did it at
808 * boot), the .text section is writable, there are no relocations to do,
809 * none of the module text/data is in readable memory, etc. Training krtld
810 * to deal with this weird module is as complicated, and more risky, than
811 * reimplementing the necessary subset of it here.
812 */
813 static void
814 init_xen_module()
815 {
816 struct _buf *file = NULL;
817 struct module *mp;
818 struct modctl *mcp;
819 int i, shn;
820 Shdr *shp, *ctf_shp;
821 char *names = NULL;
822 size_t n, namesize, text_align, data_align;
823 #if defined(__amd64)
824 const char machine = EM_AMD64;
825 #else
826 const char machine = EM_386;
827 #endif
828
829 /* Allocate and init the module structure */
830 mp = kmem_zalloc(sizeof (*mp), KM_SLEEP);
831 mp->filename = kobj_zalloc(strlen(XPV_FILENAME) + 1, KM_SLEEP);
832 (void) strcpy(mp->filename, XPV_FILENAME);
833
834 /* Allocate and init the modctl structure */
835 mcp = kmem_zalloc(sizeof (*mcp), KM_SLEEP);
836 mcp->mod_modname = kobj_zalloc(strlen(XPV_MODNAME) + 1, KM_SLEEP);
837 (void) strcpy(mcp->mod_modname, XPV_MODNAME);
838 mcp->mod_filename = kobj_zalloc(strlen(XPV_FILENAME) + 1, KM_SLEEP);
839 (void) strcpy(mcp->mod_filename, XPV_FILENAME);
840 mcp->mod_inprogress_thread = (kthread_id_t)-1;
841 mcp->mod_ref = 1;
842 mcp->mod_loaded = 1;
843 mcp->mod_loadcnt = 1;
844 mcp->mod_mp = mp;
845
846 /*
847 * Try to open a Xen image that hasn't had its symbol and CTF
848 * information stripped off.
849 */
850 file = kobj_open_file(XPV_FILENAME);
851 if (file == (struct _buf *)-1) {
852 file = NULL;
853 goto err;
854 }
855
856 /*
857 * Read the header and ensure that this is an ELF file for the
858 * proper ISA. If it's not, somebody has done something very
859 * stupid. Why bother? See Mencken.
860 */
861 if (kobj_read_file(file, (char *)&mp->hdr, sizeof (mp->hdr), 0) < 0)
862 goto err;
863 for (i = 0; i < SELFMAG; i++)
864 if (mp->hdr.e_ident[i] != ELFMAG[i])
865 goto err;
866 if ((mp->hdr.e_ident[EI_DATA] != ELFDATA2LSB) ||
867 (mp->hdr.e_machine != machine))
868 goto err;
869
870 /* Read in the section headers */
871 n = mp->hdr.e_shentsize * mp->hdr.e_shnum;
872 mp->shdrs = kmem_zalloc(n, KM_SLEEP);
873 if (kobj_read_file(file, mp->shdrs, n, mp->hdr.e_shoff) < 0)
874 goto err;
875
876 /* Read the section names */
877 shp = (Shdr *)(mp->shdrs + mp->hdr.e_shstrndx * mp->hdr.e_shentsize);
878 namesize = shp->sh_size;
879 names = kmem_zalloc(shp->sh_size, KM_SLEEP);
880 if (kobj_read_file(file, names, shp->sh_size, shp->sh_offset) < 0)
881 goto err;
882
883 /*
884 * Fill in the text and data size fields.
885 */
886 ctf_shp = NULL;
887 text_align = data_align = 0;
888 for (shn = 1; shn < mp->hdr.e_shnum; shn++) {
889 shp = (Shdr *)(mp->shdrs + shn * mp->hdr.e_shentsize);
890
891 /* Sanity check the offset of the section name */
892 if (shp->sh_name >= namesize)
893 continue;
894
895 /* If we find the symtab section, remember it for later. */
896 if (shp->sh_type == SHT_SYMTAB) {
897 mp->symtbl_section = shn;
898 mp->symhdr = shp;
899 continue;
900 }
901
902 /* If we find the CTF section, remember it for later. */
903 if ((shp->sh_size != 0) &&
904 (strcmp(names + shp->sh_name, ".SUNW_ctf") == 0)) {
905 ctf_shp = shp;
906 continue;
907 }
908
909 if (!(shp->sh_flags & SHF_ALLOC))
910 continue;
911
912 /*
913 * Xen marks its text section as writable, so we need to
914 * look for the name - not just the flag.
915 */
916 if ((strcmp(&names[shp->sh_name], ".text") != 0) &&
917 (shp->sh_flags & SHF_WRITE) != 0) {
918 if (shp->sh_addralign > data_align)
919 data_align = shp->sh_addralign;
920 mp->data_size = ALIGN(mp->data_size, data_align);
921 mp->data_size += ALIGN(shp->sh_size, 8);
922 if (mp->data == NULL || mp->data > (char *)shp->sh_addr)
923 mp->data = (char *)shp->sh_addr;
924 } else {
925 if (shp->sh_addralign > text_align)
926 text_align = shp->sh_addralign;
927 mp->text_size = ALIGN(mp->text_size, text_align);
928 mp->text_size += ALIGN(shp->sh_size, 8);
929 if (mp->text == NULL || mp->text > (char *)shp->sh_addr)
930 mp->text = (char *)shp->sh_addr;
931 }
932 }
933 kmem_free(names, namesize);
934 names = NULL;
935 shp = NULL;
936 mcp->mod_text = mp->text;
937 mcp->mod_text_size = mp->text_size;
938
939 /*
940 * If we have symbol table and string table sections, read them in
941 * now. If we don't, we just plow on. We'll still get a valid
942 * core dump, but finding anything useful will be just a bit
943 * harder.
944 *
945 * Note: we don't bother with a hash table. We'll never do a
946 * symbol lookup unless we crash, and then mdb creates its own. We
947 * also don't try to perform any relocations. Xen should be loaded
948 * exactly where the ELF file indicates, and the symbol information
949 * in the file should be complete and correct already. Static
950 * linking ain't all bad.
951 */
952 if ((mp->symhdr != NULL) && (mp->symhdr->sh_link < mp->hdr.e_shnum)) {
953 mp->strhdr = (Shdr *)
954 (mp->shdrs + mp->symhdr->sh_link * mp->hdr.e_shentsize);
955 mp->nsyms = mp->symhdr->sh_size / mp->symhdr->sh_entsize;
956
957 /* Allocate space for the symbol table and strings. */
958 mp->symsize = mp->symhdr->sh_size +
959 mp->nsyms * sizeof (symid_t) + mp->strhdr->sh_size;
960 mp->symspace = kmem_zalloc(mp->symsize, KM_SLEEP);
961 mp->symtbl = mp->symspace;
962 mp->strings = (char *)(mp->symtbl + mp->symhdr->sh_size);
963
964 if ((kobj_read_file(file, mp->symtbl,
965 mp->symhdr->sh_size, mp->symhdr->sh_offset) < 0) ||
966 (kobj_read_file(file, mp->strings,
967 mp->strhdr->sh_size, mp->strhdr->sh_offset) < 0))
968 goto err;
969 }
970
971 /*
972 * Read in the CTF section
973 */
974 if ((ctf_shp != NULL) && ((moddebug & MODDEBUG_NOCTF) == 0)) {
975 mp->ctfdata = kmem_zalloc(ctf_shp->sh_size, KM_SLEEP);
976 mp->ctfsize = ctf_shp->sh_size;
977 if (kobj_read_file(file, mp->ctfdata, mp->ctfsize,
978 ctf_shp->sh_offset) < 0)
979 goto err;
980 }
981
982 kobj_close_file(file);
983
984 xpv_module = mp;
985 xpv_modctl = mcp;
986 return;
987
988 err:
989 cmn_err(CE_WARN, "Failed to initialize xpv module.");
990 if (file != NULL)
991 kobj_close_file(file);
992
993 kmem_free(mp->filename, strlen(XPV_FILENAME) + 1);
994 if (mp->shdrs != NULL)
995 kmem_free(mp->shdrs, mp->hdr.e_shentsize * mp->hdr.e_shnum);
996 if (mp->symspace != NULL)
997 kmem_free(mp->symspace, mp->symsize);
998 if (mp->ctfdata != NULL)
999 kmem_free(mp->ctfdata, mp->ctfsize);
1000 kmem_free(mp, sizeof (*mp));
1001 kmem_free(mcp->mod_filename, strlen(XPV_FILENAME) + 1);
1002 kmem_free(mcp->mod_modname, strlen(XPV_MODNAME) + 1);
1003 kmem_free(mcp, sizeof (*mcp));
1004 if (names != NULL)
1005 kmem_free(names, namesize);
1006 }
1007
1008 void
1009 xpv_panic_init()
1010 {
1011 xen_platform_op_t op;
1012 int i;
1013
1014 ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
1015
1016 for (i = 0; i < mmu.num_level; i++)
1017 ptable_pfn[i] = PFN_INVALID;
1018
1019 /* Let Xen know where to jump if/when it panics. */
1020 op.cmd = XENPF_panic_init;
1021 op.interface_version = XENPF_INTERFACE_VERSION;
1022 op.u.panic_init.panic_addr = (unsigned long)xpv_panic_hdlr;
1023
1024 (void) HYPERVISOR_platform_op(&op);
1025
1026 init_xen_module();
1027 }