Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Core of Xen paravirt_ops implementation.
4 *
5 * This file contains the xen_paravirt_ops structure itself, and the
6 * implementations for:
7 * - privileged instructions
8 * - interrupt flags
9 * - segment operations
10 * - booting and setup
11 *
12 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
13 */
14
15#include <linux/cpu.h>
16#include <linux/kernel.h>
17#include <linux/init.h>
18#include <linux/smp.h>
19#include <linux/preempt.h>
20#include <linux/hardirq.h>
21#include <linux/percpu.h>
22#include <linux/delay.h>
23#include <linux/start_kernel.h>
24#include <linux/sched.h>
25#include <linux/kprobes.h>
26#include <linux/kstrtox.h>
27#include <linux/memblock.h>
28#include <linux/export.h>
29#include <linux/mm.h>
30#include <linux/page-flags.h>
31#include <linux/pci.h>
32#include <linux/gfp.h>
33#include <linux/edd.h>
34#include <linux/reboot.h>
35#include <linux/virtio_anchor.h>
36#include <linux/stackprotector.h>
37
38#include <xen/xen.h>
39#include <xen/events.h>
40#include <xen/interface/xen.h>
41#include <xen/interface/version.h>
42#include <xen/interface/physdev.h>
43#include <xen/interface/vcpu.h>
44#include <xen/interface/memory.h>
45#include <xen/interface/nmi.h>
46#include <xen/interface/xen-mca.h>
47#include <xen/features.h>
48#include <xen/page.h>
49#include <xen/hvc-console.h>
50#include <xen/acpi.h>
51
52#include <asm/paravirt.h>
53#include <asm/apic.h>
54#include <asm/page.h>
55#include <asm/xen/pci.h>
56#include <asm/xen/hypercall.h>
57#include <asm/xen/hypervisor.h>
58#include <asm/xen/cpuid.h>
59#include <asm/fixmap.h>
60#include <asm/processor.h>
61#include <asm/proto.h>
62#include <asm/msr-index.h>
63#include <asm/traps.h>
64#include <asm/setup.h>
65#include <asm/desc.h>
66#include <asm/pgalloc.h>
67#include <asm/tlbflush.h>
68#include <asm/reboot.h>
69#include <asm/hypervisor.h>
70#include <asm/mach_traps.h>
71#include <asm/mwait.h>
72#include <asm/pci_x86.h>
73#include <asm/cpu.h>
74#ifdef CONFIG_X86_IOPL_IOPERM
75#include <asm/io_bitmap.h>
76#endif
77
78#ifdef CONFIG_ACPI
79#include <linux/acpi.h>
80#include <asm/acpi.h>
81#include <acpi/pdc_intel.h>
82#include <acpi/processor.h>
83#include <xen/interface/platform.h>
84#endif
85
86#include "xen-ops.h"
87#include "mmu.h"
88#include "smp.h"
89#include "multicalls.h"
90#include "pmu.h"
91
92#include "../kernel/cpu/cpu.h" /* get_cpu_cap() */
93
94void *xen_initial_gdt;
95
96static int xen_cpu_up_prepare_pv(unsigned int cpu);
97static int xen_cpu_dead_pv(unsigned int cpu);
98
99struct tls_descs {
100 struct desc_struct desc[3];
101};
102
103/*
104 * Updating the 3 TLS descriptors in the GDT on every task switch is
105 * surprisingly expensive so we avoid updating them if they haven't
106 * changed. Since Xen writes different descriptors than the one
107 * passed in the update_descriptor hypercall we keep shadow copies to
108 * compare against.
109 */
110static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
111
112static __read_mostly bool xen_msr_safe = IS_ENABLED(CONFIG_XEN_PV_MSR_SAFE);
113
114static int __init parse_xen_msr_safe(char *str)
115{
116 if (str)
117 return kstrtobool(str, &xen_msr_safe);
118 return -EINVAL;
119}
120early_param("xen_msr_safe", parse_xen_msr_safe);
121
122static void __init xen_pv_init_platform(void)
123{
124 /* PV guests can't operate virtio devices without grants. */
125 if (IS_ENABLED(CONFIG_XEN_VIRTIO))
126 virtio_set_mem_acc_cb(xen_virtio_restricted_mem_acc);
127
128 populate_extra_pte(fix_to_virt(FIX_PARAVIRT_BOOTMAP));
129
130 set_fixmap(FIX_PARAVIRT_BOOTMAP, xen_start_info->shared_info);
131 HYPERVISOR_shared_info = (void *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
132
133 /* xen clock uses per-cpu vcpu_info, need to init it for boot cpu */
134 xen_vcpu_info_reset(0);
135
136 /* pvclock is in shared info area */
137 xen_init_time_ops();
138}
139
140static void __init xen_pv_guest_late_init(void)
141{
142#ifndef CONFIG_SMP
143 /* Setup shared vcpu info for non-smp configurations */
144 xen_setup_vcpu_info_placement();
145#endif
146}
147
148static __read_mostly unsigned int cpuid_leaf5_ecx_val;
149static __read_mostly unsigned int cpuid_leaf5_edx_val;
150
151static void xen_cpuid(unsigned int *ax, unsigned int *bx,
152 unsigned int *cx, unsigned int *dx)
153{
154 unsigned maskebx = ~0;
155
156 /*
157 * Mask out inconvenient features, to try and disable as many
158 * unsupported kernel subsystems as possible.
159 */
160 switch (*ax) {
161 case CPUID_MWAIT_LEAF:
162 /* Synthesize the values.. */
163 *ax = 0;
164 *bx = 0;
165 *cx = cpuid_leaf5_ecx_val;
166 *dx = cpuid_leaf5_edx_val;
167 return;
168
169 case 0xb:
170 /* Suppress extended topology stuff */
171 maskebx = 0;
172 break;
173 }
174
175 asm(XEN_EMULATE_PREFIX "cpuid"
176 : "=a" (*ax),
177 "=b" (*bx),
178 "=c" (*cx),
179 "=d" (*dx)
180 : "0" (*ax), "2" (*cx));
181
182 *bx &= maskebx;
183}
184
185static bool __init xen_check_mwait(void)
186{
187#ifdef CONFIG_ACPI
188 struct xen_platform_op op = {
189 .cmd = XENPF_set_processor_pminfo,
190 .u.set_pminfo.id = -1,
191 .u.set_pminfo.type = XEN_PM_PDC,
192 };
193 uint32_t buf[3];
194 unsigned int ax, bx, cx, dx;
195 unsigned int mwait_mask;
196
197 /* We need to determine whether it is OK to expose the MWAIT
198 * capability to the kernel to harvest deeper than C3 states from ACPI
199 * _CST using the processor_harvest_xen.c module. For this to work, we
200 * need to gather the MWAIT_LEAF values (which the cstate.c code
201 * checks against). The hypervisor won't expose the MWAIT flag because
202 * it would break backwards compatibility; so we will find out directly
203 * from the hardware and hypercall.
204 */
205 if (!xen_initial_domain())
206 return false;
207
208 /*
209 * When running under platform earlier than Xen4.2, do not expose
210 * mwait, to avoid the risk of loading native acpi pad driver
211 */
212 if (!xen_running_on_version_or_later(4, 2))
213 return false;
214
215 ax = 1;
216 cx = 0;
217
218 native_cpuid(&ax, &bx, &cx, &dx);
219
220 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
221 (1 << (X86_FEATURE_MWAIT % 32));
222
223 if ((cx & mwait_mask) != mwait_mask)
224 return false;
225
226 /* We need to emulate the MWAIT_LEAF and for that we need both
227 * ecx and edx. The hypercall provides only partial information.
228 */
229
230 ax = CPUID_MWAIT_LEAF;
231 bx = 0;
232 cx = 0;
233 dx = 0;
234
235 native_cpuid(&ax, &bx, &cx, &dx);
236
237 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
238 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
239 */
240 buf[0] = ACPI_PDC_REVISION_ID;
241 buf[1] = 1;
242 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
243
244 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
245
246 if ((HYPERVISOR_platform_op(&op) == 0) &&
247 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
248 cpuid_leaf5_ecx_val = cx;
249 cpuid_leaf5_edx_val = dx;
250 }
251 return true;
252#else
253 return false;
254#endif
255}
256
257static bool __init xen_check_xsave(void)
258{
259 unsigned int cx, xsave_mask;
260
261 cx = cpuid_ecx(1);
262
263 xsave_mask = (1 << (X86_FEATURE_XSAVE % 32)) |
264 (1 << (X86_FEATURE_OSXSAVE % 32));
265
266 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
267 return (cx & xsave_mask) == xsave_mask;
268}
269
270static void __init xen_init_capabilities(void)
271{
272 setup_force_cpu_cap(X86_FEATURE_XENPV);
273 setup_clear_cpu_cap(X86_FEATURE_DCA);
274 setup_clear_cpu_cap(X86_FEATURE_APERFMPERF);
275 setup_clear_cpu_cap(X86_FEATURE_MTRR);
276 setup_clear_cpu_cap(X86_FEATURE_ACC);
277 setup_clear_cpu_cap(X86_FEATURE_X2APIC);
278 setup_clear_cpu_cap(X86_FEATURE_SME);
279
280 /*
281 * Xen PV would need some work to support PCID: CR3 handling as well
282 * as xen_flush_tlb_others() would need updating.
283 */
284 setup_clear_cpu_cap(X86_FEATURE_PCID);
285
286 if (!xen_initial_domain())
287 setup_clear_cpu_cap(X86_FEATURE_ACPI);
288
289 if (xen_check_mwait())
290 setup_force_cpu_cap(X86_FEATURE_MWAIT);
291 else
292 setup_clear_cpu_cap(X86_FEATURE_MWAIT);
293
294 if (!xen_check_xsave()) {
295 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
296 setup_clear_cpu_cap(X86_FEATURE_OSXSAVE);
297 }
298}
299
300static noinstr void xen_set_debugreg(int reg, unsigned long val)
301{
302 HYPERVISOR_set_debugreg(reg, val);
303}
304
305static noinstr unsigned long xen_get_debugreg(int reg)
306{
307 return HYPERVISOR_get_debugreg(reg);
308}
309
310static void xen_end_context_switch(struct task_struct *next)
311{
312 xen_mc_flush();
313 paravirt_end_context_switch(next);
314}
315
316static unsigned long xen_store_tr(void)
317{
318 return 0;
319}
320
321/*
322 * Set the page permissions for a particular virtual address. If the
323 * address is a vmalloc mapping (or other non-linear mapping), then
324 * find the linear mapping of the page and also set its protections to
325 * match.
326 */
327static void set_aliased_prot(void *v, pgprot_t prot)
328{
329 int level;
330 pte_t *ptep;
331 pte_t pte;
332 unsigned long pfn;
333 unsigned char dummy;
334 void *va;
335
336 ptep = lookup_address((unsigned long)v, &level);
337 BUG_ON(ptep == NULL);
338
339 pfn = pte_pfn(*ptep);
340 pte = pfn_pte(pfn, prot);
341
342 /*
343 * Careful: update_va_mapping() will fail if the virtual address
344 * we're poking isn't populated in the page tables. We don't
345 * need to worry about the direct map (that's always in the page
346 * tables), but we need to be careful about vmap space. In
347 * particular, the top level page table can lazily propagate
348 * entries between processes, so if we've switched mms since we
349 * vmapped the target in the first place, we might not have the
350 * top-level page table entry populated.
351 *
352 * We disable preemption because we want the same mm active when
353 * we probe the target and when we issue the hypercall. We'll
354 * have the same nominal mm, but if we're a kernel thread, lazy
355 * mm dropping could change our pgd.
356 *
357 * Out of an abundance of caution, this uses __get_user() to fault
358 * in the target address just in case there's some obscure case
359 * in which the target address isn't readable.
360 */
361
362 preempt_disable();
363
364 copy_from_kernel_nofault(&dummy, v, 1);
365
366 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
367 BUG();
368
369 va = __va(PFN_PHYS(pfn));
370
371 if (va != v && HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
372 BUG();
373
374 preempt_enable();
375}
376
377static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
378{
379 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
380 int i;
381
382 /*
383 * We need to mark the all aliases of the LDT pages RO. We
384 * don't need to call vm_flush_aliases(), though, since that's
385 * only responsible for flushing aliases out the TLBs, not the
386 * page tables, and Xen will flush the TLB for us if needed.
387 *
388 * To avoid confusing future readers: none of this is necessary
389 * to load the LDT. The hypervisor only checks this when the
390 * LDT is faulted in due to subsequent descriptor access.
391 */
392
393 for (i = 0; i < entries; i += entries_per_page)
394 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
395}
396
397static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
398{
399 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
400 int i;
401
402 for (i = 0; i < entries; i += entries_per_page)
403 set_aliased_prot(ldt + i, PAGE_KERNEL);
404}
405
406static void xen_set_ldt(const void *addr, unsigned entries)
407{
408 struct mmuext_op *op;
409 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
410
411 trace_xen_cpu_set_ldt(addr, entries);
412
413 op = mcs.args;
414 op->cmd = MMUEXT_SET_LDT;
415 op->arg1.linear_addr = (unsigned long)addr;
416 op->arg2.nr_ents = entries;
417
418 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
419
420 xen_mc_issue(PARAVIRT_LAZY_CPU);
421}
422
423static void xen_load_gdt(const struct desc_ptr *dtr)
424{
425 unsigned long va = dtr->address;
426 unsigned int size = dtr->size + 1;
427 unsigned long pfn, mfn;
428 int level;
429 pte_t *ptep;
430 void *virt;
431
432 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
433 BUG_ON(size > PAGE_SIZE);
434 BUG_ON(va & ~PAGE_MASK);
435
436 /*
437 * The GDT is per-cpu and is in the percpu data area.
438 * That can be virtually mapped, so we need to do a
439 * page-walk to get the underlying MFN for the
440 * hypercall. The page can also be in the kernel's
441 * linear range, so we need to RO that mapping too.
442 */
443 ptep = lookup_address(va, &level);
444 BUG_ON(ptep == NULL);
445
446 pfn = pte_pfn(*ptep);
447 mfn = pfn_to_mfn(pfn);
448 virt = __va(PFN_PHYS(pfn));
449
450 make_lowmem_page_readonly((void *)va);
451 make_lowmem_page_readonly(virt);
452
453 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
454 BUG();
455}
456
457/*
458 * load_gdt for early boot, when the gdt is only mapped once
459 */
460static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
461{
462 unsigned long va = dtr->address;
463 unsigned int size = dtr->size + 1;
464 unsigned long pfn, mfn;
465 pte_t pte;
466
467 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
468 BUG_ON(size > PAGE_SIZE);
469 BUG_ON(va & ~PAGE_MASK);
470
471 pfn = virt_to_pfn(va);
472 mfn = pfn_to_mfn(pfn);
473
474 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
475
476 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
477 BUG();
478
479 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
480 BUG();
481}
482
483static inline bool desc_equal(const struct desc_struct *d1,
484 const struct desc_struct *d2)
485{
486 return !memcmp(d1, d2, sizeof(*d1));
487}
488
489static void load_TLS_descriptor(struct thread_struct *t,
490 unsigned int cpu, unsigned int i)
491{
492 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
493 struct desc_struct *gdt;
494 xmaddr_t maddr;
495 struct multicall_space mc;
496
497 if (desc_equal(shadow, &t->tls_array[i]))
498 return;
499
500 *shadow = t->tls_array[i];
501
502 gdt = get_cpu_gdt_rw(cpu);
503 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
504 mc = __xen_mc_entry(0);
505
506 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
507}
508
509static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
510{
511 /*
512 * In lazy mode we need to zero %fs, otherwise we may get an
513 * exception between the new %fs descriptor being loaded and
514 * %fs being effectively cleared at __switch_to().
515 */
516 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
517 loadsegment(fs, 0);
518
519 xen_mc_batch();
520
521 load_TLS_descriptor(t, cpu, 0);
522 load_TLS_descriptor(t, cpu, 1);
523 load_TLS_descriptor(t, cpu, 2);
524
525 xen_mc_issue(PARAVIRT_LAZY_CPU);
526}
527
528static void xen_load_gs_index(unsigned int idx)
529{
530 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
531 BUG();
532}
533
534static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
535 const void *ptr)
536{
537 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
538 u64 entry = *(u64 *)ptr;
539
540 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
541
542 preempt_disable();
543
544 xen_mc_flush();
545 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
546 BUG();
547
548 preempt_enable();
549}
550
551void noist_exc_debug(struct pt_regs *regs);
552
553DEFINE_IDTENTRY_RAW(xenpv_exc_nmi)
554{
555 /* On Xen PV, NMI doesn't use IST. The C part is the same as native. */
556 exc_nmi(regs);
557}
558
559DEFINE_IDTENTRY_RAW_ERRORCODE(xenpv_exc_double_fault)
560{
561 /* On Xen PV, DF doesn't use IST. The C part is the same as native. */
562 exc_double_fault(regs, error_code);
563}
564
565DEFINE_IDTENTRY_RAW(xenpv_exc_debug)
566{
567 /*
568 * There's no IST on Xen PV, but we still need to dispatch
569 * to the correct handler.
570 */
571 if (user_mode(regs))
572 noist_exc_debug(regs);
573 else
574 exc_debug(regs);
575}
576
577DEFINE_IDTENTRY_RAW(exc_xen_unknown_trap)
578{
579 /* This should never happen and there is no way to handle it. */
580 instrumentation_begin();
581 pr_err("Unknown trap in Xen PV mode.");
582 BUG();
583 instrumentation_end();
584}
585
586#ifdef CONFIG_X86_MCE
587DEFINE_IDTENTRY_RAW(xenpv_exc_machine_check)
588{
589 /*
590 * There's no IST on Xen PV, but we still need to dispatch
591 * to the correct handler.
592 */
593 if (user_mode(regs))
594 noist_exc_machine_check(regs);
595 else
596 exc_machine_check(regs);
597}
598#endif
599
600struct trap_array_entry {
601 void (*orig)(void);
602 void (*xen)(void);
603 bool ist_okay;
604};
605
606#define TRAP_ENTRY(func, ist_ok) { \
607 .orig = asm_##func, \
608 .xen = xen_asm_##func, \
609 .ist_okay = ist_ok }
610
611#define TRAP_ENTRY_REDIR(func, ist_ok) { \
612 .orig = asm_##func, \
613 .xen = xen_asm_xenpv_##func, \
614 .ist_okay = ist_ok }
615
616static struct trap_array_entry trap_array[] = {
617 TRAP_ENTRY_REDIR(exc_debug, true ),
618 TRAP_ENTRY_REDIR(exc_double_fault, true ),
619#ifdef CONFIG_X86_MCE
620 TRAP_ENTRY_REDIR(exc_machine_check, true ),
621#endif
622 TRAP_ENTRY_REDIR(exc_nmi, true ),
623 TRAP_ENTRY(exc_int3, false ),
624 TRAP_ENTRY(exc_overflow, false ),
625#ifdef CONFIG_IA32_EMULATION
626 { entry_INT80_compat, xen_entry_INT80_compat, false },
627#endif
628 TRAP_ENTRY(exc_page_fault, false ),
629 TRAP_ENTRY(exc_divide_error, false ),
630 TRAP_ENTRY(exc_bounds, false ),
631 TRAP_ENTRY(exc_invalid_op, false ),
632 TRAP_ENTRY(exc_device_not_available, false ),
633 TRAP_ENTRY(exc_coproc_segment_overrun, false ),
634 TRAP_ENTRY(exc_invalid_tss, false ),
635 TRAP_ENTRY(exc_segment_not_present, false ),
636 TRAP_ENTRY(exc_stack_segment, false ),
637 TRAP_ENTRY(exc_general_protection, false ),
638 TRAP_ENTRY(exc_spurious_interrupt_bug, false ),
639 TRAP_ENTRY(exc_coprocessor_error, false ),
640 TRAP_ENTRY(exc_alignment_check, false ),
641 TRAP_ENTRY(exc_simd_coprocessor_error, false ),
642#ifdef CONFIG_X86_KERNEL_IBT
643 TRAP_ENTRY(exc_control_protection, false ),
644#endif
645};
646
647static bool __ref get_trap_addr(void **addr, unsigned int ist)
648{
649 unsigned int nr;
650 bool ist_okay = false;
651 bool found = false;
652
653 /*
654 * Replace trap handler addresses by Xen specific ones.
655 * Check for known traps using IST and whitelist them.
656 * The debugger ones are the only ones we care about.
657 * Xen will handle faults like double_fault, so we should never see
658 * them. Warn if there's an unexpected IST-using fault handler.
659 */
660 for (nr = 0; nr < ARRAY_SIZE(trap_array); nr++) {
661 struct trap_array_entry *entry = trap_array + nr;
662
663 if (*addr == entry->orig) {
664 *addr = entry->xen;
665 ist_okay = entry->ist_okay;
666 found = true;
667 break;
668 }
669 }
670
671 if (nr == ARRAY_SIZE(trap_array) &&
672 *addr >= (void *)early_idt_handler_array[0] &&
673 *addr < (void *)early_idt_handler_array[NUM_EXCEPTION_VECTORS]) {
674 nr = (*addr - (void *)early_idt_handler_array[0]) /
675 EARLY_IDT_HANDLER_SIZE;
676 *addr = (void *)xen_early_idt_handler_array[nr];
677 found = true;
678 }
679
680 if (!found)
681 *addr = (void *)xen_asm_exc_xen_unknown_trap;
682
683 if (WARN_ON(found && ist != 0 && !ist_okay))
684 return false;
685
686 return true;
687}
688
689static int cvt_gate_to_trap(int vector, const gate_desc *val,
690 struct trap_info *info)
691{
692 unsigned long addr;
693
694 if (val->bits.type != GATE_TRAP && val->bits.type != GATE_INTERRUPT)
695 return 0;
696
697 info->vector = vector;
698
699 addr = gate_offset(val);
700 if (!get_trap_addr((void **)&addr, val->bits.ist))
701 return 0;
702 info->address = addr;
703
704 info->cs = gate_segment(val);
705 info->flags = val->bits.dpl;
706 /* interrupt gates clear IF */
707 if (val->bits.type == GATE_INTERRUPT)
708 info->flags |= 1 << 2;
709
710 return 1;
711}
712
713/* Locations of each CPU's IDT */
714static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
715
716/* Set an IDT entry. If the entry is part of the current IDT, then
717 also update Xen. */
718static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
719{
720 unsigned long p = (unsigned long)&dt[entrynum];
721 unsigned long start, end;
722
723 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
724
725 preempt_disable();
726
727 start = __this_cpu_read(idt_desc.address);
728 end = start + __this_cpu_read(idt_desc.size) + 1;
729
730 xen_mc_flush();
731
732 native_write_idt_entry(dt, entrynum, g);
733
734 if (p >= start && (p + 8) <= end) {
735 struct trap_info info[2];
736
737 info[1].address = 0;
738
739 if (cvt_gate_to_trap(entrynum, g, &info[0]))
740 if (HYPERVISOR_set_trap_table(info))
741 BUG();
742 }
743
744 preempt_enable();
745}
746
747static unsigned xen_convert_trap_info(const struct desc_ptr *desc,
748 struct trap_info *traps, bool full)
749{
750 unsigned in, out, count;
751
752 count = (desc->size+1) / sizeof(gate_desc);
753 BUG_ON(count > 256);
754
755 for (in = out = 0; in < count; in++) {
756 gate_desc *entry = (gate_desc *)(desc->address) + in;
757
758 if (cvt_gate_to_trap(in, entry, &traps[out]) || full)
759 out++;
760 }
761
762 return out;
763}
764
765void xen_copy_trap_info(struct trap_info *traps)
766{
767 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
768
769 xen_convert_trap_info(desc, traps, true);
770}
771
772/* Load a new IDT into Xen. In principle this can be per-CPU, so we
773 hold a spinlock to protect the static traps[] array (static because
774 it avoids allocation, and saves stack space). */
775static void xen_load_idt(const struct desc_ptr *desc)
776{
777 static DEFINE_SPINLOCK(lock);
778 static struct trap_info traps[257];
779 static const struct trap_info zero = { };
780 unsigned out;
781
782 trace_xen_cpu_load_idt(desc);
783
784 spin_lock(&lock);
785
786 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
787
788 out = xen_convert_trap_info(desc, traps, false);
789 traps[out] = zero;
790
791 xen_mc_flush();
792 if (HYPERVISOR_set_trap_table(traps))
793 BUG();
794
795 spin_unlock(&lock);
796}
797
798/* Write a GDT descriptor entry. Ignore LDT descriptors, since
799 they're handled differently. */
800static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
801 const void *desc, int type)
802{
803 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
804
805 preempt_disable();
806
807 switch (type) {
808 case DESC_LDT:
809 case DESC_TSS:
810 /* ignore */
811 break;
812
813 default: {
814 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
815
816 xen_mc_flush();
817 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
818 BUG();
819 }
820
821 }
822
823 preempt_enable();
824}
825
826/*
827 * Version of write_gdt_entry for use at early boot-time needed to
828 * update an entry as simply as possible.
829 */
830static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
831 const void *desc, int type)
832{
833 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
834
835 switch (type) {
836 case DESC_LDT:
837 case DESC_TSS:
838 /* ignore */
839 break;
840
841 default: {
842 xmaddr_t maddr = virt_to_machine(&dt[entry]);
843
844 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
845 dt[entry] = *(struct desc_struct *)desc;
846 }
847
848 }
849}
850
851static void xen_load_sp0(unsigned long sp0)
852{
853 struct multicall_space mcs;
854
855 mcs = xen_mc_entry(0);
856 MULTI_stack_switch(mcs.mc, __KERNEL_DS, sp0);
857 xen_mc_issue(PARAVIRT_LAZY_CPU);
858 this_cpu_write(cpu_tss_rw.x86_tss.sp0, sp0);
859}
860
861#ifdef CONFIG_X86_IOPL_IOPERM
862static void xen_invalidate_io_bitmap(void)
863{
864 struct physdev_set_iobitmap iobitmap = {
865 .bitmap = NULL,
866 .nr_ports = 0,
867 };
868
869 native_tss_invalidate_io_bitmap();
870 HYPERVISOR_physdev_op(PHYSDEVOP_set_iobitmap, &iobitmap);
871}
872
873static void xen_update_io_bitmap(void)
874{
875 struct physdev_set_iobitmap iobitmap;
876 struct tss_struct *tss = this_cpu_ptr(&cpu_tss_rw);
877
878 native_tss_update_io_bitmap();
879
880 iobitmap.bitmap = (uint8_t *)(&tss->x86_tss) +
881 tss->x86_tss.io_bitmap_base;
882 if (tss->x86_tss.io_bitmap_base == IO_BITMAP_OFFSET_INVALID)
883 iobitmap.nr_ports = 0;
884 else
885 iobitmap.nr_ports = IO_BITMAP_BITS;
886
887 HYPERVISOR_physdev_op(PHYSDEVOP_set_iobitmap, &iobitmap);
888}
889#endif
890
891static void xen_io_delay(void)
892{
893}
894
895static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
896
897static unsigned long xen_read_cr0(void)
898{
899 unsigned long cr0 = this_cpu_read(xen_cr0_value);
900
901 if (unlikely(cr0 == 0)) {
902 cr0 = native_read_cr0();
903 this_cpu_write(xen_cr0_value, cr0);
904 }
905
906 return cr0;
907}
908
909static void xen_write_cr0(unsigned long cr0)
910{
911 struct multicall_space mcs;
912
913 this_cpu_write(xen_cr0_value, cr0);
914
915 /* Only pay attention to cr0.TS; everything else is
916 ignored. */
917 mcs = xen_mc_entry(0);
918
919 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
920
921 xen_mc_issue(PARAVIRT_LAZY_CPU);
922}
923
924static void xen_write_cr4(unsigned long cr4)
925{
926 cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
927
928 native_write_cr4(cr4);
929}
930
931static u64 xen_do_read_msr(unsigned int msr, int *err)
932{
933 u64 val = 0; /* Avoid uninitialized value for safe variant. */
934
935 if (pmu_msr_read(msr, &val, err))
936 return val;
937
938 if (err)
939 val = native_read_msr_safe(msr, err);
940 else
941 val = native_read_msr(msr);
942
943 switch (msr) {
944 case MSR_IA32_APICBASE:
945 val &= ~X2APIC_ENABLE;
946 break;
947 }
948 return val;
949}
950
951static void set_seg(unsigned int which, unsigned int low, unsigned int high,
952 int *err)
953{
954 u64 base = ((u64)high << 32) | low;
955
956 if (HYPERVISOR_set_segment_base(which, base) == 0)
957 return;
958
959 if (err)
960 *err = -EIO;
961 else
962 WARN(1, "Xen set_segment_base(%u, %llx) failed\n", which, base);
963}
964
965/*
966 * Support write_msr_safe() and write_msr() semantics.
967 * With err == NULL write_msr() semantics are selected.
968 * Supplying an err pointer requires err to be pre-initialized with 0.
969 */
970static void xen_do_write_msr(unsigned int msr, unsigned int low,
971 unsigned int high, int *err)
972{
973 switch (msr) {
974 case MSR_FS_BASE:
975 set_seg(SEGBASE_FS, low, high, err);
976 break;
977
978 case MSR_KERNEL_GS_BASE:
979 set_seg(SEGBASE_GS_USER, low, high, err);
980 break;
981
982 case MSR_GS_BASE:
983 set_seg(SEGBASE_GS_KERNEL, low, high, err);
984 break;
985
986 case MSR_STAR:
987 case MSR_CSTAR:
988 case MSR_LSTAR:
989 case MSR_SYSCALL_MASK:
990 case MSR_IA32_SYSENTER_CS:
991 case MSR_IA32_SYSENTER_ESP:
992 case MSR_IA32_SYSENTER_EIP:
993 /* Fast syscall setup is all done in hypercalls, so
994 these are all ignored. Stub them out here to stop
995 Xen console noise. */
996 break;
997
998 default:
999 if (!pmu_msr_write(msr, low, high, err)) {
1000 if (err)
1001 *err = native_write_msr_safe(msr, low, high);
1002 else
1003 native_write_msr(msr, low, high);
1004 }
1005 }
1006}
1007
1008static u64 xen_read_msr_safe(unsigned int msr, int *err)
1009{
1010 return xen_do_read_msr(msr, err);
1011}
1012
1013static int xen_write_msr_safe(unsigned int msr, unsigned int low,
1014 unsigned int high)
1015{
1016 int err = 0;
1017
1018 xen_do_write_msr(msr, low, high, &err);
1019
1020 return err;
1021}
1022
1023static u64 xen_read_msr(unsigned int msr)
1024{
1025 int err;
1026
1027 return xen_do_read_msr(msr, xen_msr_safe ? &err : NULL);
1028}
1029
1030static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
1031{
1032 int err;
1033
1034 xen_do_write_msr(msr, low, high, xen_msr_safe ? &err : NULL);
1035}
1036
1037/* This is called once we have the cpu_possible_mask */
1038void __init xen_setup_vcpu_info_placement(void)
1039{
1040 int cpu;
1041
1042 for_each_possible_cpu(cpu) {
1043 /* Set up direct vCPU id mapping for PV guests. */
1044 per_cpu(xen_vcpu_id, cpu) = cpu;
1045 xen_vcpu_setup(cpu);
1046 }
1047
1048 pv_ops.irq.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1049 pv_ops.irq.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1050 pv_ops.irq.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1051 pv_ops.mmu.read_cr2 = __PV_IS_CALLEE_SAVE(xen_read_cr2_direct);
1052}
1053
1054static const struct pv_info xen_info __initconst = {
1055 .extra_user_64bit_cs = FLAT_USER_CS64,
1056 .name = "Xen",
1057};
1058
1059static const typeof(pv_ops) xen_cpu_ops __initconst = {
1060 .cpu = {
1061 .cpuid = xen_cpuid,
1062
1063 .set_debugreg = xen_set_debugreg,
1064 .get_debugreg = xen_get_debugreg,
1065
1066 .read_cr0 = xen_read_cr0,
1067 .write_cr0 = xen_write_cr0,
1068
1069 .write_cr4 = xen_write_cr4,
1070
1071 .wbinvd = native_wbinvd,
1072
1073 .read_msr = xen_read_msr,
1074 .write_msr = xen_write_msr,
1075
1076 .read_msr_safe = xen_read_msr_safe,
1077 .write_msr_safe = xen_write_msr_safe,
1078
1079 .read_pmc = xen_read_pmc,
1080
1081 .load_tr_desc = paravirt_nop,
1082 .set_ldt = xen_set_ldt,
1083 .load_gdt = xen_load_gdt,
1084 .load_idt = xen_load_idt,
1085 .load_tls = xen_load_tls,
1086 .load_gs_index = xen_load_gs_index,
1087
1088 .alloc_ldt = xen_alloc_ldt,
1089 .free_ldt = xen_free_ldt,
1090
1091 .store_tr = xen_store_tr,
1092
1093 .write_ldt_entry = xen_write_ldt_entry,
1094 .write_gdt_entry = xen_write_gdt_entry,
1095 .write_idt_entry = xen_write_idt_entry,
1096 .load_sp0 = xen_load_sp0,
1097
1098#ifdef CONFIG_X86_IOPL_IOPERM
1099 .invalidate_io_bitmap = xen_invalidate_io_bitmap,
1100 .update_io_bitmap = xen_update_io_bitmap,
1101#endif
1102 .io_delay = xen_io_delay,
1103
1104 .start_context_switch = paravirt_start_context_switch,
1105 .end_context_switch = xen_end_context_switch,
1106 },
1107};
1108
1109static void xen_restart(char *msg)
1110{
1111 xen_reboot(SHUTDOWN_reboot);
1112}
1113
1114static void xen_machine_halt(void)
1115{
1116 xen_reboot(SHUTDOWN_poweroff);
1117}
1118
1119static void xen_machine_power_off(void)
1120{
1121 do_kernel_power_off();
1122 xen_reboot(SHUTDOWN_poweroff);
1123}
1124
1125static void xen_crash_shutdown(struct pt_regs *regs)
1126{
1127 xen_reboot(SHUTDOWN_crash);
1128}
1129
1130static const struct machine_ops xen_machine_ops __initconst = {
1131 .restart = xen_restart,
1132 .halt = xen_machine_halt,
1133 .power_off = xen_machine_power_off,
1134 .shutdown = xen_machine_halt,
1135 .crash_shutdown = xen_crash_shutdown,
1136 .emergency_restart = xen_emergency_restart,
1137};
1138
1139static unsigned char xen_get_nmi_reason(void)
1140{
1141 unsigned char reason = 0;
1142
1143 /* Construct a value which looks like it came from port 0x61. */
1144 if (test_bit(_XEN_NMIREASON_io_error,
1145 &HYPERVISOR_shared_info->arch.nmi_reason))
1146 reason |= NMI_REASON_IOCHK;
1147 if (test_bit(_XEN_NMIREASON_pci_serr,
1148 &HYPERVISOR_shared_info->arch.nmi_reason))
1149 reason |= NMI_REASON_SERR;
1150
1151 return reason;
1152}
1153
1154static void __init xen_boot_params_init_edd(void)
1155{
1156#if IS_ENABLED(CONFIG_EDD)
1157 struct xen_platform_op op;
1158 struct edd_info *edd_info;
1159 u32 *mbr_signature;
1160 unsigned nr;
1161 int ret;
1162
1163 edd_info = boot_params.eddbuf;
1164 mbr_signature = boot_params.edd_mbr_sig_buffer;
1165
1166 op.cmd = XENPF_firmware_info;
1167
1168 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1169 for (nr = 0; nr < EDDMAXNR; nr++) {
1170 struct edd_info *info = edd_info + nr;
1171
1172 op.u.firmware_info.index = nr;
1173 info->params.length = sizeof(info->params);
1174 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1175 &info->params);
1176 ret = HYPERVISOR_platform_op(&op);
1177 if (ret)
1178 break;
1179
1180#define C(x) info->x = op.u.firmware_info.u.disk_info.x
1181 C(device);
1182 C(version);
1183 C(interface_support);
1184 C(legacy_max_cylinder);
1185 C(legacy_max_head);
1186 C(legacy_sectors_per_track);
1187#undef C
1188 }
1189 boot_params.eddbuf_entries = nr;
1190
1191 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1192 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1193 op.u.firmware_info.index = nr;
1194 ret = HYPERVISOR_platform_op(&op);
1195 if (ret)
1196 break;
1197 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1198 }
1199 boot_params.edd_mbr_sig_buf_entries = nr;
1200#endif
1201}
1202
1203/*
1204 * Set up the GDT and segment registers for -fstack-protector. Until
1205 * we do this, we have to be careful not to call any stack-protected
1206 * function, which is most of the kernel.
1207 */
1208static void __init xen_setup_gdt(int cpu)
1209{
1210 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry_boot;
1211 pv_ops.cpu.load_gdt = xen_load_gdt_boot;
1212
1213 switch_gdt_and_percpu_base(cpu);
1214
1215 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry;
1216 pv_ops.cpu.load_gdt = xen_load_gdt;
1217}
1218
1219static void __init xen_dom0_set_legacy_features(void)
1220{
1221 x86_platform.legacy.rtc = 1;
1222}
1223
1224static void __init xen_domu_set_legacy_features(void)
1225{
1226 x86_platform.legacy.rtc = 0;
1227}
1228
1229extern void early_xen_iret_patch(void);
1230
1231/* First C function to be called on Xen boot */
1232asmlinkage __visible void __init xen_start_kernel(struct start_info *si)
1233{
1234 struct physdev_set_iopl set_iopl;
1235 unsigned long initrd_start = 0;
1236 int rc;
1237
1238 if (!si)
1239 return;
1240
1241 clear_bss();
1242
1243 xen_start_info = si;
1244
1245 __text_gen_insn(&early_xen_iret_patch,
1246 JMP32_INSN_OPCODE, &early_xen_iret_patch, &xen_iret,
1247 JMP32_INSN_SIZE);
1248
1249 xen_domain_type = XEN_PV_DOMAIN;
1250 xen_start_flags = xen_start_info->flags;
1251
1252 xen_setup_features();
1253
1254 /* Install Xen paravirt ops */
1255 pv_info = xen_info;
1256 pv_ops.cpu = xen_cpu_ops.cpu;
1257 xen_init_irq_ops();
1258
1259 /*
1260 * Setup xen_vcpu early because it is needed for
1261 * local_irq_disable(), irqs_disabled(), e.g. in printk().
1262 *
1263 * Don't do the full vcpu_info placement stuff until we have
1264 * the cpu_possible_mask and a non-dummy shared_info.
1265 */
1266 xen_vcpu_info_reset(0);
1267
1268 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1269 x86_platform.realmode_reserve = x86_init_noop;
1270 x86_platform.realmode_init = x86_init_noop;
1271
1272 x86_init.resources.memory_setup = xen_memory_setup;
1273 x86_init.irqs.intr_mode_select = x86_init_noop;
1274 x86_init.irqs.intr_mode_init = x86_init_noop;
1275 x86_init.oem.arch_setup = xen_arch_setup;
1276 x86_init.oem.banner = xen_banner;
1277 x86_init.hyper.init_platform = xen_pv_init_platform;
1278 x86_init.hyper.guest_late_init = xen_pv_guest_late_init;
1279
1280 /*
1281 * Set up some pagetable state before starting to set any ptes.
1282 */
1283
1284 xen_setup_machphys_mapping();
1285 xen_init_mmu_ops();
1286
1287 /* Prevent unwanted bits from being set in PTEs. */
1288 __supported_pte_mask &= ~_PAGE_GLOBAL;
1289 __default_kernel_pte_mask &= ~_PAGE_GLOBAL;
1290
1291 /* Get mfn list */
1292 xen_build_dynamic_phys_to_machine();
1293
1294 /* Work out if we support NX */
1295 get_cpu_cap(&boot_cpu_data);
1296 x86_configure_nx();
1297
1298 /*
1299 * Set up kernel GDT and segment registers, mainly so that
1300 * -fstack-protector code can be executed.
1301 */
1302 xen_setup_gdt(0);
1303
1304 /* Determine virtual and physical address sizes */
1305 get_cpu_address_sizes(&boot_cpu_data);
1306
1307 /* Let's presume PV guests always boot on vCPU with id 0. */
1308 per_cpu(xen_vcpu_id, 0) = 0;
1309
1310 idt_setup_early_handler();
1311
1312 xen_init_capabilities();
1313
1314#ifdef CONFIG_X86_LOCAL_APIC
1315 /*
1316 * set up the basic apic ops.
1317 */
1318 xen_init_apic();
1319#endif
1320
1321 machine_ops = xen_machine_ops;
1322
1323 /*
1324 * The only reliable way to retain the initial address of the
1325 * percpu gdt_page is to remember it here, so we can go and
1326 * mark it RW later, when the initial percpu area is freed.
1327 */
1328 xen_initial_gdt = &per_cpu(gdt_page, 0);
1329
1330 xen_smp_init();
1331
1332#ifdef CONFIG_ACPI_NUMA
1333 /*
1334 * The pages we from Xen are not related to machine pages, so
1335 * any NUMA information the kernel tries to get from ACPI will
1336 * be meaningless. Prevent it from trying.
1337 */
1338 disable_srat();
1339#endif
1340 WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
1341
1342 local_irq_disable();
1343 early_boot_irqs_disabled = true;
1344
1345 xen_raw_console_write("mapping kernel into physical memory\n");
1346 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1347 xen_start_info->nr_pages);
1348 xen_reserve_special_pages();
1349
1350 /*
1351 * We used to do this in xen_arch_setup, but that is too late
1352 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1353 * early_amd_init which pokes 0xcf8 port.
1354 */
1355 set_iopl.iopl = 1;
1356 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1357 if (rc != 0)
1358 xen_raw_printk("physdev_op failed %d\n", rc);
1359
1360
1361 if (xen_start_info->mod_start) {
1362 if (xen_start_info->flags & SIF_MOD_START_PFN)
1363 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1364 else
1365 initrd_start = __pa(xen_start_info->mod_start);
1366 }
1367
1368 /* Poke various useful things into boot_params */
1369 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1370 boot_params.hdr.ramdisk_image = initrd_start;
1371 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1372 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1373 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1374
1375 if (!xen_initial_domain()) {
1376 if (pci_xen)
1377 x86_init.pci.arch_init = pci_xen_init;
1378 x86_platform.set_legacy_features =
1379 xen_domu_set_legacy_features;
1380 } else {
1381 const struct dom0_vga_console_info *info =
1382 (void *)((char *)xen_start_info +
1383 xen_start_info->console.dom0.info_off);
1384 struct xen_platform_op op = {
1385 .cmd = XENPF_firmware_info,
1386 .interface_version = XENPF_INTERFACE_VERSION,
1387 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1388 };
1389
1390 x86_platform.set_legacy_features =
1391 xen_dom0_set_legacy_features;
1392 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1393 xen_start_info->console.domU.mfn = 0;
1394 xen_start_info->console.domU.evtchn = 0;
1395
1396 if (HYPERVISOR_platform_op(&op) == 0)
1397 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1398
1399 /* Make sure ACS will be enabled */
1400 pci_request_acs();
1401
1402 xen_acpi_sleep_register();
1403
1404 xen_boot_params_init_edd();
1405
1406#ifdef CONFIG_ACPI
1407 /*
1408 * Disable selecting "Firmware First mode" for correctable
1409 * memory errors, as this is the duty of the hypervisor to
1410 * decide.
1411 */
1412 acpi_disable_cmcff = 1;
1413#endif
1414 }
1415
1416 xen_add_preferred_consoles();
1417
1418#ifdef CONFIG_PCI
1419 /* PCI BIOS service won't work from a PV guest. */
1420 pci_probe &= ~PCI_PROBE_BIOS;
1421#endif
1422 xen_raw_console_write("about to get started...\n");
1423
1424 /* We need this for printk timestamps */
1425 xen_setup_runstate_info(0);
1426
1427 xen_efi_init(&boot_params);
1428
1429 /* Start the world */
1430 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1431 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1432}
1433
1434static int xen_cpu_up_prepare_pv(unsigned int cpu)
1435{
1436 int rc;
1437
1438 if (per_cpu(xen_vcpu, cpu) == NULL)
1439 return -ENODEV;
1440
1441 xen_setup_timer(cpu);
1442
1443 rc = xen_smp_intr_init(cpu);
1444 if (rc) {
1445 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1446 cpu, rc);
1447 return rc;
1448 }
1449
1450 rc = xen_smp_intr_init_pv(cpu);
1451 if (rc) {
1452 WARN(1, "xen_smp_intr_init_pv() for CPU %d failed: %d\n",
1453 cpu, rc);
1454 return rc;
1455 }
1456
1457 return 0;
1458}
1459
1460static int xen_cpu_dead_pv(unsigned int cpu)
1461{
1462 xen_smp_intr_free(cpu);
1463 xen_smp_intr_free_pv(cpu);
1464
1465 xen_teardown_timer(cpu);
1466
1467 return 0;
1468}
1469
1470static uint32_t __init xen_platform_pv(void)
1471{
1472 if (xen_pv_domain())
1473 return xen_cpuid_base();
1474
1475 return 0;
1476}
1477
1478const __initconst struct hypervisor_x86 x86_hyper_xen_pv = {
1479 .name = "Xen PV",
1480 .detect = xen_platform_pv,
1481 .type = X86_HYPER_XEN_PV,
1482 .runtime.pin_vcpu = xen_pin_vcpu,
1483 .ignore_nopv = true,
1484};
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Core of Xen paravirt_ops implementation.
4 *
5 * This file contains the xen_paravirt_ops structure itself, and the
6 * implementations for:
7 * - privileged instructions
8 * - interrupt flags
9 * - segment operations
10 * - booting and setup
11 *
12 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
13 */
14
15#include <linux/cpu.h>
16#include <linux/kernel.h>
17#include <linux/init.h>
18#include <linux/smp.h>
19#include <linux/preempt.h>
20#include <linux/hardirq.h>
21#include <linux/percpu.h>
22#include <linux/delay.h>
23#include <linux/start_kernel.h>
24#include <linux/sched.h>
25#include <linux/kprobes.h>
26#include <linux/memblock.h>
27#include <linux/export.h>
28#include <linux/mm.h>
29#include <linux/page-flags.h>
30#include <linux/highmem.h>
31#include <linux/console.h>
32#include <linux/pci.h>
33#include <linux/gfp.h>
34#include <linux/edd.h>
35#include <linux/frame.h>
36
37#include <xen/xen.h>
38#include <xen/events.h>
39#include <xen/interface/xen.h>
40#include <xen/interface/version.h>
41#include <xen/interface/physdev.h>
42#include <xen/interface/vcpu.h>
43#include <xen/interface/memory.h>
44#include <xen/interface/nmi.h>
45#include <xen/interface/xen-mca.h>
46#include <xen/features.h>
47#include <xen/page.h>
48#include <xen/hvc-console.h>
49#include <xen/acpi.h>
50
51#include <asm/paravirt.h>
52#include <asm/apic.h>
53#include <asm/page.h>
54#include <asm/xen/pci.h>
55#include <asm/xen/hypercall.h>
56#include <asm/xen/hypervisor.h>
57#include <asm/xen/cpuid.h>
58#include <asm/fixmap.h>
59#include <asm/processor.h>
60#include <asm/proto.h>
61#include <asm/msr-index.h>
62#include <asm/traps.h>
63#include <asm/setup.h>
64#include <asm/desc.h>
65#include <asm/pgalloc.h>
66#include <asm/pgtable.h>
67#include <asm/tlbflush.h>
68#include <asm/reboot.h>
69#include <asm/stackprotector.h>
70#include <asm/hypervisor.h>
71#include <asm/mach_traps.h>
72#include <asm/mwait.h>
73#include <asm/pci_x86.h>
74#include <asm/cpu.h>
75
76#ifdef CONFIG_ACPI
77#include <linux/acpi.h>
78#include <asm/acpi.h>
79#include <acpi/pdc_intel.h>
80#include <acpi/processor.h>
81#include <xen/interface/platform.h>
82#endif
83
84#include "xen-ops.h"
85#include "mmu.h"
86#include "smp.h"
87#include "multicalls.h"
88#include "pmu.h"
89
90#include "../kernel/cpu/cpu.h" /* get_cpu_cap() */
91
92void *xen_initial_gdt;
93
94static int xen_cpu_up_prepare_pv(unsigned int cpu);
95static int xen_cpu_dead_pv(unsigned int cpu);
96
97struct tls_descs {
98 struct desc_struct desc[3];
99};
100
101/*
102 * Updating the 3 TLS descriptors in the GDT on every task switch is
103 * surprisingly expensive so we avoid updating them if they haven't
104 * changed. Since Xen writes different descriptors than the one
105 * passed in the update_descriptor hypercall we keep shadow copies to
106 * compare against.
107 */
108static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
109
110static void __init xen_banner(void)
111{
112 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
113 struct xen_extraversion extra;
114 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
115
116 pr_info("Booting paravirtualized kernel on %s\n", pv_info.name);
117 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
118 version >> 16, version & 0xffff, extra.extraversion,
119 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
120
121#ifdef CONFIG_X86_32
122 pr_warn("WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! WARNING!\n"
123 "Support for running as 32-bit PV-guest under Xen will soon be removed\n"
124 "from the Linux kernel!\n"
125 "Please use either a 64-bit kernel or switch to HVM or PVH mode!\n"
126 "WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! WARNING!\n");
127#endif
128}
129
130static void __init xen_pv_init_platform(void)
131{
132 populate_extra_pte(fix_to_virt(FIX_PARAVIRT_BOOTMAP));
133
134 set_fixmap(FIX_PARAVIRT_BOOTMAP, xen_start_info->shared_info);
135 HYPERVISOR_shared_info = (void *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
136
137 /* xen clock uses per-cpu vcpu_info, need to init it for boot cpu */
138 xen_vcpu_info_reset(0);
139
140 /* pvclock is in shared info area */
141 xen_init_time_ops();
142}
143
144static void __init xen_pv_guest_late_init(void)
145{
146#ifndef CONFIG_SMP
147 /* Setup shared vcpu info for non-smp configurations */
148 xen_setup_vcpu_info_placement();
149#endif
150}
151
152/* Check if running on Xen version (major, minor) or later */
153bool
154xen_running_on_version_or_later(unsigned int major, unsigned int minor)
155{
156 unsigned int version;
157
158 if (!xen_domain())
159 return false;
160
161 version = HYPERVISOR_xen_version(XENVER_version, NULL);
162 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
163 ((version >> 16) > major))
164 return true;
165 return false;
166}
167
168static __read_mostly unsigned int cpuid_leaf5_ecx_val;
169static __read_mostly unsigned int cpuid_leaf5_edx_val;
170
171static void xen_cpuid(unsigned int *ax, unsigned int *bx,
172 unsigned int *cx, unsigned int *dx)
173{
174 unsigned maskebx = ~0;
175
176 /*
177 * Mask out inconvenient features, to try and disable as many
178 * unsupported kernel subsystems as possible.
179 */
180 switch (*ax) {
181 case CPUID_MWAIT_LEAF:
182 /* Synthesize the values.. */
183 *ax = 0;
184 *bx = 0;
185 *cx = cpuid_leaf5_ecx_val;
186 *dx = cpuid_leaf5_edx_val;
187 return;
188
189 case 0xb:
190 /* Suppress extended topology stuff */
191 maskebx = 0;
192 break;
193 }
194
195 asm(XEN_EMULATE_PREFIX "cpuid"
196 : "=a" (*ax),
197 "=b" (*bx),
198 "=c" (*cx),
199 "=d" (*dx)
200 : "0" (*ax), "2" (*cx));
201
202 *bx &= maskebx;
203}
204STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
205
206static bool __init xen_check_mwait(void)
207{
208#ifdef CONFIG_ACPI
209 struct xen_platform_op op = {
210 .cmd = XENPF_set_processor_pminfo,
211 .u.set_pminfo.id = -1,
212 .u.set_pminfo.type = XEN_PM_PDC,
213 };
214 uint32_t buf[3];
215 unsigned int ax, bx, cx, dx;
216 unsigned int mwait_mask;
217
218 /* We need to determine whether it is OK to expose the MWAIT
219 * capability to the kernel to harvest deeper than C3 states from ACPI
220 * _CST using the processor_harvest_xen.c module. For this to work, we
221 * need to gather the MWAIT_LEAF values (which the cstate.c code
222 * checks against). The hypervisor won't expose the MWAIT flag because
223 * it would break backwards compatibility; so we will find out directly
224 * from the hardware and hypercall.
225 */
226 if (!xen_initial_domain())
227 return false;
228
229 /*
230 * When running under platform earlier than Xen4.2, do not expose
231 * mwait, to avoid the risk of loading native acpi pad driver
232 */
233 if (!xen_running_on_version_or_later(4, 2))
234 return false;
235
236 ax = 1;
237 cx = 0;
238
239 native_cpuid(&ax, &bx, &cx, &dx);
240
241 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
242 (1 << (X86_FEATURE_MWAIT % 32));
243
244 if ((cx & mwait_mask) != mwait_mask)
245 return false;
246
247 /* We need to emulate the MWAIT_LEAF and for that we need both
248 * ecx and edx. The hypercall provides only partial information.
249 */
250
251 ax = CPUID_MWAIT_LEAF;
252 bx = 0;
253 cx = 0;
254 dx = 0;
255
256 native_cpuid(&ax, &bx, &cx, &dx);
257
258 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
259 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
260 */
261 buf[0] = ACPI_PDC_REVISION_ID;
262 buf[1] = 1;
263 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
264
265 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
266
267 if ((HYPERVISOR_platform_op(&op) == 0) &&
268 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
269 cpuid_leaf5_ecx_val = cx;
270 cpuid_leaf5_edx_val = dx;
271 }
272 return true;
273#else
274 return false;
275#endif
276}
277
278static bool __init xen_check_xsave(void)
279{
280 unsigned int cx, xsave_mask;
281
282 cx = cpuid_ecx(1);
283
284 xsave_mask = (1 << (X86_FEATURE_XSAVE % 32)) |
285 (1 << (X86_FEATURE_OSXSAVE % 32));
286
287 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
288 return (cx & xsave_mask) == xsave_mask;
289}
290
291static void __init xen_init_capabilities(void)
292{
293 setup_force_cpu_cap(X86_FEATURE_XENPV);
294 setup_clear_cpu_cap(X86_FEATURE_DCA);
295 setup_clear_cpu_cap(X86_FEATURE_APERFMPERF);
296 setup_clear_cpu_cap(X86_FEATURE_MTRR);
297 setup_clear_cpu_cap(X86_FEATURE_ACC);
298 setup_clear_cpu_cap(X86_FEATURE_X2APIC);
299 setup_clear_cpu_cap(X86_FEATURE_SME);
300
301 /*
302 * Xen PV would need some work to support PCID: CR3 handling as well
303 * as xen_flush_tlb_others() would need updating.
304 */
305 setup_clear_cpu_cap(X86_FEATURE_PCID);
306
307 if (!xen_initial_domain())
308 setup_clear_cpu_cap(X86_FEATURE_ACPI);
309
310 if (xen_check_mwait())
311 setup_force_cpu_cap(X86_FEATURE_MWAIT);
312 else
313 setup_clear_cpu_cap(X86_FEATURE_MWAIT);
314
315 if (!xen_check_xsave()) {
316 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
317 setup_clear_cpu_cap(X86_FEATURE_OSXSAVE);
318 }
319}
320
321static void xen_set_debugreg(int reg, unsigned long val)
322{
323 HYPERVISOR_set_debugreg(reg, val);
324}
325
326static unsigned long xen_get_debugreg(int reg)
327{
328 return HYPERVISOR_get_debugreg(reg);
329}
330
331static void xen_end_context_switch(struct task_struct *next)
332{
333 xen_mc_flush();
334 paravirt_end_context_switch(next);
335}
336
337static unsigned long xen_store_tr(void)
338{
339 return 0;
340}
341
342/*
343 * Set the page permissions for a particular virtual address. If the
344 * address is a vmalloc mapping (or other non-linear mapping), then
345 * find the linear mapping of the page and also set its protections to
346 * match.
347 */
348static void set_aliased_prot(void *v, pgprot_t prot)
349{
350 int level;
351 pte_t *ptep;
352 pte_t pte;
353 unsigned long pfn;
354 struct page *page;
355 unsigned char dummy;
356
357 ptep = lookup_address((unsigned long)v, &level);
358 BUG_ON(ptep == NULL);
359
360 pfn = pte_pfn(*ptep);
361 page = pfn_to_page(pfn);
362
363 pte = pfn_pte(pfn, prot);
364
365 /*
366 * Careful: update_va_mapping() will fail if the virtual address
367 * we're poking isn't populated in the page tables. We don't
368 * need to worry about the direct map (that's always in the page
369 * tables), but we need to be careful about vmap space. In
370 * particular, the top level page table can lazily propagate
371 * entries between processes, so if we've switched mms since we
372 * vmapped the target in the first place, we might not have the
373 * top-level page table entry populated.
374 *
375 * We disable preemption because we want the same mm active when
376 * we probe the target and when we issue the hypercall. We'll
377 * have the same nominal mm, but if we're a kernel thread, lazy
378 * mm dropping could change our pgd.
379 *
380 * Out of an abundance of caution, this uses __get_user() to fault
381 * in the target address just in case there's some obscure case
382 * in which the target address isn't readable.
383 */
384
385 preempt_disable();
386
387 probe_kernel_read(&dummy, v, 1);
388
389 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
390 BUG();
391
392 if (!PageHighMem(page)) {
393 void *av = __va(PFN_PHYS(pfn));
394
395 if (av != v)
396 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
397 BUG();
398 } else
399 kmap_flush_unused();
400
401 preempt_enable();
402}
403
404static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
405{
406 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
407 int i;
408
409 /*
410 * We need to mark the all aliases of the LDT pages RO. We
411 * don't need to call vm_flush_aliases(), though, since that's
412 * only responsible for flushing aliases out the TLBs, not the
413 * page tables, and Xen will flush the TLB for us if needed.
414 *
415 * To avoid confusing future readers: none of this is necessary
416 * to load the LDT. The hypervisor only checks this when the
417 * LDT is faulted in due to subsequent descriptor access.
418 */
419
420 for (i = 0; i < entries; i += entries_per_page)
421 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
422}
423
424static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
425{
426 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
427 int i;
428
429 for (i = 0; i < entries; i += entries_per_page)
430 set_aliased_prot(ldt + i, PAGE_KERNEL);
431}
432
433static void xen_set_ldt(const void *addr, unsigned entries)
434{
435 struct mmuext_op *op;
436 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
437
438 trace_xen_cpu_set_ldt(addr, entries);
439
440 op = mcs.args;
441 op->cmd = MMUEXT_SET_LDT;
442 op->arg1.linear_addr = (unsigned long)addr;
443 op->arg2.nr_ents = entries;
444
445 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
446
447 xen_mc_issue(PARAVIRT_LAZY_CPU);
448}
449
450static void xen_load_gdt(const struct desc_ptr *dtr)
451{
452 unsigned long va = dtr->address;
453 unsigned int size = dtr->size + 1;
454 unsigned long pfn, mfn;
455 int level;
456 pte_t *ptep;
457 void *virt;
458
459 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
460 BUG_ON(size > PAGE_SIZE);
461 BUG_ON(va & ~PAGE_MASK);
462
463 /*
464 * The GDT is per-cpu and is in the percpu data area.
465 * That can be virtually mapped, so we need to do a
466 * page-walk to get the underlying MFN for the
467 * hypercall. The page can also be in the kernel's
468 * linear range, so we need to RO that mapping too.
469 */
470 ptep = lookup_address(va, &level);
471 BUG_ON(ptep == NULL);
472
473 pfn = pte_pfn(*ptep);
474 mfn = pfn_to_mfn(pfn);
475 virt = __va(PFN_PHYS(pfn));
476
477 make_lowmem_page_readonly((void *)va);
478 make_lowmem_page_readonly(virt);
479
480 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
481 BUG();
482}
483
484/*
485 * load_gdt for early boot, when the gdt is only mapped once
486 */
487static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
488{
489 unsigned long va = dtr->address;
490 unsigned int size = dtr->size + 1;
491 unsigned long pfn, mfn;
492 pte_t pte;
493
494 /* @size should be at most GDT_SIZE which is smaller than PAGE_SIZE. */
495 BUG_ON(size > PAGE_SIZE);
496 BUG_ON(va & ~PAGE_MASK);
497
498 pfn = virt_to_pfn(va);
499 mfn = pfn_to_mfn(pfn);
500
501 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
502
503 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
504 BUG();
505
506 if (HYPERVISOR_set_gdt(&mfn, size / sizeof(struct desc_struct)))
507 BUG();
508}
509
510static inline bool desc_equal(const struct desc_struct *d1,
511 const struct desc_struct *d2)
512{
513 return !memcmp(d1, d2, sizeof(*d1));
514}
515
516static void load_TLS_descriptor(struct thread_struct *t,
517 unsigned int cpu, unsigned int i)
518{
519 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
520 struct desc_struct *gdt;
521 xmaddr_t maddr;
522 struct multicall_space mc;
523
524 if (desc_equal(shadow, &t->tls_array[i]))
525 return;
526
527 *shadow = t->tls_array[i];
528
529 gdt = get_cpu_gdt_rw(cpu);
530 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
531 mc = __xen_mc_entry(0);
532
533 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
534}
535
536static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
537{
538 /*
539 * XXX sleazy hack: If we're being called in a lazy-cpu zone
540 * and lazy gs handling is enabled, it means we're in a
541 * context switch, and %gs has just been saved. This means we
542 * can zero it out to prevent faults on exit from the
543 * hypervisor if the next process has no %gs. Either way, it
544 * has been saved, and the new value will get loaded properly.
545 * This will go away as soon as Xen has been modified to not
546 * save/restore %gs for normal hypercalls.
547 *
548 * On x86_64, this hack is not used for %gs, because gs points
549 * to KERNEL_GS_BASE (and uses it for PDA references), so we
550 * must not zero %gs on x86_64
551 *
552 * For x86_64, we need to zero %fs, otherwise we may get an
553 * exception between the new %fs descriptor being loaded and
554 * %fs being effectively cleared at __switch_to().
555 */
556 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
557#ifdef CONFIG_X86_32
558 lazy_load_gs(0);
559#else
560 loadsegment(fs, 0);
561#endif
562 }
563
564 xen_mc_batch();
565
566 load_TLS_descriptor(t, cpu, 0);
567 load_TLS_descriptor(t, cpu, 1);
568 load_TLS_descriptor(t, cpu, 2);
569
570 xen_mc_issue(PARAVIRT_LAZY_CPU);
571}
572
573#ifdef CONFIG_X86_64
574static void xen_load_gs_index(unsigned int idx)
575{
576 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
577 BUG();
578}
579#endif
580
581static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
582 const void *ptr)
583{
584 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
585 u64 entry = *(u64 *)ptr;
586
587 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
588
589 preempt_disable();
590
591 xen_mc_flush();
592 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
593 BUG();
594
595 preempt_enable();
596}
597
598#ifdef CONFIG_X86_64
599struct trap_array_entry {
600 void (*orig)(void);
601 void (*xen)(void);
602 bool ist_okay;
603};
604
605static struct trap_array_entry trap_array[] = {
606 { debug, xen_xendebug, true },
607 { double_fault, xen_double_fault, true },
608#ifdef CONFIG_X86_MCE
609 { machine_check, xen_machine_check, true },
610#endif
611 { nmi, xen_xennmi, true },
612 { int3, xen_int3, false },
613 { overflow, xen_overflow, false },
614#ifdef CONFIG_IA32_EMULATION
615 { entry_INT80_compat, xen_entry_INT80_compat, false },
616#endif
617 { page_fault, xen_page_fault, false },
618 { divide_error, xen_divide_error, false },
619 { bounds, xen_bounds, false },
620 { invalid_op, xen_invalid_op, false },
621 { device_not_available, xen_device_not_available, false },
622 { coprocessor_segment_overrun, xen_coprocessor_segment_overrun, false },
623 { invalid_TSS, xen_invalid_TSS, false },
624 { segment_not_present, xen_segment_not_present, false },
625 { stack_segment, xen_stack_segment, false },
626 { general_protection, xen_general_protection, false },
627 { spurious_interrupt_bug, xen_spurious_interrupt_bug, false },
628 { coprocessor_error, xen_coprocessor_error, false },
629 { alignment_check, xen_alignment_check, false },
630 { simd_coprocessor_error, xen_simd_coprocessor_error, false },
631};
632
633static bool __ref get_trap_addr(void **addr, unsigned int ist)
634{
635 unsigned int nr;
636 bool ist_okay = false;
637
638 /*
639 * Replace trap handler addresses by Xen specific ones.
640 * Check for known traps using IST and whitelist them.
641 * The debugger ones are the only ones we care about.
642 * Xen will handle faults like double_fault, * so we should never see
643 * them. Warn if there's an unexpected IST-using fault handler.
644 */
645 for (nr = 0; nr < ARRAY_SIZE(trap_array); nr++) {
646 struct trap_array_entry *entry = trap_array + nr;
647
648 if (*addr == entry->orig) {
649 *addr = entry->xen;
650 ist_okay = entry->ist_okay;
651 break;
652 }
653 }
654
655 if (nr == ARRAY_SIZE(trap_array) &&
656 *addr >= (void *)early_idt_handler_array[0] &&
657 *addr < (void *)early_idt_handler_array[NUM_EXCEPTION_VECTORS]) {
658 nr = (*addr - (void *)early_idt_handler_array[0]) /
659 EARLY_IDT_HANDLER_SIZE;
660 *addr = (void *)xen_early_idt_handler_array[nr];
661 }
662
663 if (WARN_ON(ist != 0 && !ist_okay))
664 return false;
665
666 return true;
667}
668#endif
669
670static int cvt_gate_to_trap(int vector, const gate_desc *val,
671 struct trap_info *info)
672{
673 unsigned long addr;
674
675 if (val->bits.type != GATE_TRAP && val->bits.type != GATE_INTERRUPT)
676 return 0;
677
678 info->vector = vector;
679
680 addr = gate_offset(val);
681#ifdef CONFIG_X86_64
682 if (!get_trap_addr((void **)&addr, val->bits.ist))
683 return 0;
684#endif /* CONFIG_X86_64 */
685 info->address = addr;
686
687 info->cs = gate_segment(val);
688 info->flags = val->bits.dpl;
689 /* interrupt gates clear IF */
690 if (val->bits.type == GATE_INTERRUPT)
691 info->flags |= 1 << 2;
692
693 return 1;
694}
695
696/* Locations of each CPU's IDT */
697static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
698
699/* Set an IDT entry. If the entry is part of the current IDT, then
700 also update Xen. */
701static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
702{
703 unsigned long p = (unsigned long)&dt[entrynum];
704 unsigned long start, end;
705
706 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
707
708 preempt_disable();
709
710 start = __this_cpu_read(idt_desc.address);
711 end = start + __this_cpu_read(idt_desc.size) + 1;
712
713 xen_mc_flush();
714
715 native_write_idt_entry(dt, entrynum, g);
716
717 if (p >= start && (p + 8) <= end) {
718 struct trap_info info[2];
719
720 info[1].address = 0;
721
722 if (cvt_gate_to_trap(entrynum, g, &info[0]))
723 if (HYPERVISOR_set_trap_table(info))
724 BUG();
725 }
726
727 preempt_enable();
728}
729
730static void xen_convert_trap_info(const struct desc_ptr *desc,
731 struct trap_info *traps)
732{
733 unsigned in, out, count;
734
735 count = (desc->size+1) / sizeof(gate_desc);
736 BUG_ON(count > 256);
737
738 for (in = out = 0; in < count; in++) {
739 gate_desc *entry = (gate_desc *)(desc->address) + in;
740
741 if (cvt_gate_to_trap(in, entry, &traps[out]))
742 out++;
743 }
744 traps[out].address = 0;
745}
746
747void xen_copy_trap_info(struct trap_info *traps)
748{
749 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
750
751 xen_convert_trap_info(desc, traps);
752}
753
754/* Load a new IDT into Xen. In principle this can be per-CPU, so we
755 hold a spinlock to protect the static traps[] array (static because
756 it avoids allocation, and saves stack space). */
757static void xen_load_idt(const struct desc_ptr *desc)
758{
759 static DEFINE_SPINLOCK(lock);
760 static struct trap_info traps[257];
761
762 trace_xen_cpu_load_idt(desc);
763
764 spin_lock(&lock);
765
766 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
767
768 xen_convert_trap_info(desc, traps);
769
770 xen_mc_flush();
771 if (HYPERVISOR_set_trap_table(traps))
772 BUG();
773
774 spin_unlock(&lock);
775}
776
777/* Write a GDT descriptor entry. Ignore LDT descriptors, since
778 they're handled differently. */
779static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
780 const void *desc, int type)
781{
782 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
783
784 preempt_disable();
785
786 switch (type) {
787 case DESC_LDT:
788 case DESC_TSS:
789 /* ignore */
790 break;
791
792 default: {
793 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
794
795 xen_mc_flush();
796 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
797 BUG();
798 }
799
800 }
801
802 preempt_enable();
803}
804
805/*
806 * Version of write_gdt_entry for use at early boot-time needed to
807 * update an entry as simply as possible.
808 */
809static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
810 const void *desc, int type)
811{
812 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
813
814 switch (type) {
815 case DESC_LDT:
816 case DESC_TSS:
817 /* ignore */
818 break;
819
820 default: {
821 xmaddr_t maddr = virt_to_machine(&dt[entry]);
822
823 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
824 dt[entry] = *(struct desc_struct *)desc;
825 }
826
827 }
828}
829
830static void xen_load_sp0(unsigned long sp0)
831{
832 struct multicall_space mcs;
833
834 mcs = xen_mc_entry(0);
835 MULTI_stack_switch(mcs.mc, __KERNEL_DS, sp0);
836 xen_mc_issue(PARAVIRT_LAZY_CPU);
837 this_cpu_write(cpu_tss_rw.x86_tss.sp0, sp0);
838}
839
840void xen_set_iopl_mask(unsigned mask)
841{
842 struct physdev_set_iopl set_iopl;
843
844 /* Force the change at ring 0. */
845 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
846 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
847}
848
849static void xen_io_delay(void)
850{
851}
852
853static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
854
855static unsigned long xen_read_cr0(void)
856{
857 unsigned long cr0 = this_cpu_read(xen_cr0_value);
858
859 if (unlikely(cr0 == 0)) {
860 cr0 = native_read_cr0();
861 this_cpu_write(xen_cr0_value, cr0);
862 }
863
864 return cr0;
865}
866
867static void xen_write_cr0(unsigned long cr0)
868{
869 struct multicall_space mcs;
870
871 this_cpu_write(xen_cr0_value, cr0);
872
873 /* Only pay attention to cr0.TS; everything else is
874 ignored. */
875 mcs = xen_mc_entry(0);
876
877 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
878
879 xen_mc_issue(PARAVIRT_LAZY_CPU);
880}
881
882static void xen_write_cr4(unsigned long cr4)
883{
884 cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
885
886 native_write_cr4(cr4);
887}
888
889static u64 xen_read_msr_safe(unsigned int msr, int *err)
890{
891 u64 val;
892
893 if (pmu_msr_read(msr, &val, err))
894 return val;
895
896 val = native_read_msr_safe(msr, err);
897 switch (msr) {
898 case MSR_IA32_APICBASE:
899 val &= ~X2APIC_ENABLE;
900 break;
901 }
902 return val;
903}
904
905static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
906{
907 int ret;
908
909 ret = 0;
910
911 switch (msr) {
912#ifdef CONFIG_X86_64
913 unsigned which;
914 u64 base;
915
916 case MSR_FS_BASE: which = SEGBASE_FS; goto set;
917 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set;
918 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set;
919
920 set:
921 base = ((u64)high << 32) | low;
922 if (HYPERVISOR_set_segment_base(which, base) != 0)
923 ret = -EIO;
924 break;
925#endif
926
927 case MSR_STAR:
928 case MSR_CSTAR:
929 case MSR_LSTAR:
930 case MSR_SYSCALL_MASK:
931 case MSR_IA32_SYSENTER_CS:
932 case MSR_IA32_SYSENTER_ESP:
933 case MSR_IA32_SYSENTER_EIP:
934 /* Fast syscall setup is all done in hypercalls, so
935 these are all ignored. Stub them out here to stop
936 Xen console noise. */
937 break;
938
939 default:
940 if (!pmu_msr_write(msr, low, high, &ret))
941 ret = native_write_msr_safe(msr, low, high);
942 }
943
944 return ret;
945}
946
947static u64 xen_read_msr(unsigned int msr)
948{
949 /*
950 * This will silently swallow a #GP from RDMSR. It may be worth
951 * changing that.
952 */
953 int err;
954
955 return xen_read_msr_safe(msr, &err);
956}
957
958static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
959{
960 /*
961 * This will silently swallow a #GP from WRMSR. It may be worth
962 * changing that.
963 */
964 xen_write_msr_safe(msr, low, high);
965}
966
967/* This is called once we have the cpu_possible_mask */
968void __init xen_setup_vcpu_info_placement(void)
969{
970 int cpu;
971
972 for_each_possible_cpu(cpu) {
973 /* Set up direct vCPU id mapping for PV guests. */
974 per_cpu(xen_vcpu_id, cpu) = cpu;
975
976 /*
977 * xen_vcpu_setup(cpu) can fail -- in which case it
978 * falls back to the shared_info version for cpus
979 * where xen_vcpu_nr(cpu) < MAX_VIRT_CPUS.
980 *
981 * xen_cpu_up_prepare_pv() handles the rest by failing
982 * them in hotplug.
983 */
984 (void) xen_vcpu_setup(cpu);
985 }
986
987 /*
988 * xen_vcpu_setup managed to place the vcpu_info within the
989 * percpu area for all cpus, so make use of it.
990 */
991 if (xen_have_vcpu_info_placement) {
992 pv_ops.irq.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
993 pv_ops.irq.restore_fl =
994 __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
995 pv_ops.irq.irq_disable =
996 __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
997 pv_ops.irq.irq_enable =
998 __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
999 pv_ops.mmu.read_cr2 =
1000 __PV_IS_CALLEE_SAVE(xen_read_cr2_direct);
1001 }
1002}
1003
1004static const struct pv_info xen_info __initconst = {
1005 .shared_kernel_pmd = 0,
1006
1007#ifdef CONFIG_X86_64
1008 .extra_user_64bit_cs = FLAT_USER_CS64,
1009#endif
1010 .name = "Xen",
1011};
1012
1013static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1014 .cpuid = xen_cpuid,
1015
1016 .set_debugreg = xen_set_debugreg,
1017 .get_debugreg = xen_get_debugreg,
1018
1019 .read_cr0 = xen_read_cr0,
1020 .write_cr0 = xen_write_cr0,
1021
1022 .write_cr4 = xen_write_cr4,
1023
1024 .wbinvd = native_wbinvd,
1025
1026 .read_msr = xen_read_msr,
1027 .write_msr = xen_write_msr,
1028
1029 .read_msr_safe = xen_read_msr_safe,
1030 .write_msr_safe = xen_write_msr_safe,
1031
1032 .read_pmc = xen_read_pmc,
1033
1034 .iret = xen_iret,
1035#ifdef CONFIG_X86_64
1036 .usergs_sysret64 = xen_sysret64,
1037#endif
1038
1039 .load_tr_desc = paravirt_nop,
1040 .set_ldt = xen_set_ldt,
1041 .load_gdt = xen_load_gdt,
1042 .load_idt = xen_load_idt,
1043 .load_tls = xen_load_tls,
1044#ifdef CONFIG_X86_64
1045 .load_gs_index = xen_load_gs_index,
1046#endif
1047
1048 .alloc_ldt = xen_alloc_ldt,
1049 .free_ldt = xen_free_ldt,
1050
1051 .store_tr = xen_store_tr,
1052
1053 .write_ldt_entry = xen_write_ldt_entry,
1054 .write_gdt_entry = xen_write_gdt_entry,
1055 .write_idt_entry = xen_write_idt_entry,
1056 .load_sp0 = xen_load_sp0,
1057
1058 .set_iopl_mask = xen_set_iopl_mask,
1059 .io_delay = xen_io_delay,
1060
1061 /* Xen takes care of %gs when switching to usermode for us */
1062 .swapgs = paravirt_nop,
1063
1064 .start_context_switch = paravirt_start_context_switch,
1065 .end_context_switch = xen_end_context_switch,
1066};
1067
1068static void xen_restart(char *msg)
1069{
1070 xen_reboot(SHUTDOWN_reboot);
1071}
1072
1073static void xen_machine_halt(void)
1074{
1075 xen_reboot(SHUTDOWN_poweroff);
1076}
1077
1078static void xen_machine_power_off(void)
1079{
1080 if (pm_power_off)
1081 pm_power_off();
1082 xen_reboot(SHUTDOWN_poweroff);
1083}
1084
1085static void xen_crash_shutdown(struct pt_regs *regs)
1086{
1087 xen_reboot(SHUTDOWN_crash);
1088}
1089
1090static const struct machine_ops xen_machine_ops __initconst = {
1091 .restart = xen_restart,
1092 .halt = xen_machine_halt,
1093 .power_off = xen_machine_power_off,
1094 .shutdown = xen_machine_halt,
1095 .crash_shutdown = xen_crash_shutdown,
1096 .emergency_restart = xen_emergency_restart,
1097};
1098
1099static unsigned char xen_get_nmi_reason(void)
1100{
1101 unsigned char reason = 0;
1102
1103 /* Construct a value which looks like it came from port 0x61. */
1104 if (test_bit(_XEN_NMIREASON_io_error,
1105 &HYPERVISOR_shared_info->arch.nmi_reason))
1106 reason |= NMI_REASON_IOCHK;
1107 if (test_bit(_XEN_NMIREASON_pci_serr,
1108 &HYPERVISOR_shared_info->arch.nmi_reason))
1109 reason |= NMI_REASON_SERR;
1110
1111 return reason;
1112}
1113
1114static void __init xen_boot_params_init_edd(void)
1115{
1116#if IS_ENABLED(CONFIG_EDD)
1117 struct xen_platform_op op;
1118 struct edd_info *edd_info;
1119 u32 *mbr_signature;
1120 unsigned nr;
1121 int ret;
1122
1123 edd_info = boot_params.eddbuf;
1124 mbr_signature = boot_params.edd_mbr_sig_buffer;
1125
1126 op.cmd = XENPF_firmware_info;
1127
1128 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1129 for (nr = 0; nr < EDDMAXNR; nr++) {
1130 struct edd_info *info = edd_info + nr;
1131
1132 op.u.firmware_info.index = nr;
1133 info->params.length = sizeof(info->params);
1134 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1135 &info->params);
1136 ret = HYPERVISOR_platform_op(&op);
1137 if (ret)
1138 break;
1139
1140#define C(x) info->x = op.u.firmware_info.u.disk_info.x
1141 C(device);
1142 C(version);
1143 C(interface_support);
1144 C(legacy_max_cylinder);
1145 C(legacy_max_head);
1146 C(legacy_sectors_per_track);
1147#undef C
1148 }
1149 boot_params.eddbuf_entries = nr;
1150
1151 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1152 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1153 op.u.firmware_info.index = nr;
1154 ret = HYPERVISOR_platform_op(&op);
1155 if (ret)
1156 break;
1157 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1158 }
1159 boot_params.edd_mbr_sig_buf_entries = nr;
1160#endif
1161}
1162
1163/*
1164 * Set up the GDT and segment registers for -fstack-protector. Until
1165 * we do this, we have to be careful not to call any stack-protected
1166 * function, which is most of the kernel.
1167 */
1168static void __init xen_setup_gdt(int cpu)
1169{
1170 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry_boot;
1171 pv_ops.cpu.load_gdt = xen_load_gdt_boot;
1172
1173 setup_stack_canary_segment(cpu);
1174 switch_to_new_gdt(cpu);
1175
1176 pv_ops.cpu.write_gdt_entry = xen_write_gdt_entry;
1177 pv_ops.cpu.load_gdt = xen_load_gdt;
1178}
1179
1180static void __init xen_dom0_set_legacy_features(void)
1181{
1182 x86_platform.legacy.rtc = 1;
1183}
1184
1185/* First C function to be called on Xen boot */
1186asmlinkage __visible void __init xen_start_kernel(void)
1187{
1188 struct physdev_set_iopl set_iopl;
1189 unsigned long initrd_start = 0;
1190 int rc;
1191
1192 if (!xen_start_info)
1193 return;
1194
1195 xen_domain_type = XEN_PV_DOMAIN;
1196 xen_start_flags = xen_start_info->flags;
1197
1198 xen_setup_features();
1199
1200 /* Install Xen paravirt ops */
1201 pv_info = xen_info;
1202 pv_ops.init.patch = paravirt_patch_default;
1203 pv_ops.cpu = xen_cpu_ops;
1204 xen_init_irq_ops();
1205
1206 /*
1207 * Setup xen_vcpu early because it is needed for
1208 * local_irq_disable(), irqs_disabled(), e.g. in printk().
1209 *
1210 * Don't do the full vcpu_info placement stuff until we have
1211 * the cpu_possible_mask and a non-dummy shared_info.
1212 */
1213 xen_vcpu_info_reset(0);
1214
1215 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1216
1217 x86_init.resources.memory_setup = xen_memory_setup;
1218 x86_init.irqs.intr_mode_init = x86_init_noop;
1219 x86_init.oem.arch_setup = xen_arch_setup;
1220 x86_init.oem.banner = xen_banner;
1221 x86_init.hyper.init_platform = xen_pv_init_platform;
1222 x86_init.hyper.guest_late_init = xen_pv_guest_late_init;
1223
1224 /*
1225 * Set up some pagetable state before starting to set any ptes.
1226 */
1227
1228 xen_setup_machphys_mapping();
1229 xen_init_mmu_ops();
1230
1231 /* Prevent unwanted bits from being set in PTEs. */
1232 __supported_pte_mask &= ~_PAGE_GLOBAL;
1233 __default_kernel_pte_mask &= ~_PAGE_GLOBAL;
1234
1235 /*
1236 * Prevent page tables from being allocated in highmem, even
1237 * if CONFIG_HIGHPTE is enabled.
1238 */
1239 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1240
1241 /* Get mfn list */
1242 xen_build_dynamic_phys_to_machine();
1243
1244 /*
1245 * Set up kernel GDT and segment registers, mainly so that
1246 * -fstack-protector code can be executed.
1247 */
1248 xen_setup_gdt(0);
1249
1250 /* Work out if we support NX */
1251 get_cpu_cap(&boot_cpu_data);
1252 x86_configure_nx();
1253
1254 /* Determine virtual and physical address sizes */
1255 get_cpu_address_sizes(&boot_cpu_data);
1256
1257 /* Let's presume PV guests always boot on vCPU with id 0. */
1258 per_cpu(xen_vcpu_id, 0) = 0;
1259
1260 idt_setup_early_handler();
1261
1262 xen_init_capabilities();
1263
1264#ifdef CONFIG_X86_LOCAL_APIC
1265 /*
1266 * set up the basic apic ops.
1267 */
1268 xen_init_apic();
1269#endif
1270
1271 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1272 pv_ops.mmu.ptep_modify_prot_start =
1273 xen_ptep_modify_prot_start;
1274 pv_ops.mmu.ptep_modify_prot_commit =
1275 xen_ptep_modify_prot_commit;
1276 }
1277
1278 machine_ops = xen_machine_ops;
1279
1280 /*
1281 * The only reliable way to retain the initial address of the
1282 * percpu gdt_page is to remember it here, so we can go and
1283 * mark it RW later, when the initial percpu area is freed.
1284 */
1285 xen_initial_gdt = &per_cpu(gdt_page, 0);
1286
1287 xen_smp_init();
1288
1289#ifdef CONFIG_ACPI_NUMA
1290 /*
1291 * The pages we from Xen are not related to machine pages, so
1292 * any NUMA information the kernel tries to get from ACPI will
1293 * be meaningless. Prevent it from trying.
1294 */
1295 acpi_numa = -1;
1296#endif
1297 WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_pv, xen_cpu_dead_pv));
1298
1299 local_irq_disable();
1300 early_boot_irqs_disabled = true;
1301
1302 xen_raw_console_write("mapping kernel into physical memory\n");
1303 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1304 xen_start_info->nr_pages);
1305 xen_reserve_special_pages();
1306
1307 /* keep using Xen gdt for now; no urgent need to change it */
1308
1309#ifdef CONFIG_X86_32
1310 pv_info.kernel_rpl = 1;
1311 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1312 pv_info.kernel_rpl = 0;
1313#else
1314 pv_info.kernel_rpl = 0;
1315#endif
1316 /* set the limit of our address space */
1317 xen_reserve_top();
1318
1319 /*
1320 * We used to do this in xen_arch_setup, but that is too late
1321 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1322 * early_amd_init which pokes 0xcf8 port.
1323 */
1324 set_iopl.iopl = 1;
1325 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1326 if (rc != 0)
1327 xen_raw_printk("physdev_op failed %d\n", rc);
1328
1329#ifdef CONFIG_X86_32
1330 /* set up basic CPUID stuff */
1331 cpu_detect(&new_cpu_data);
1332 set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1333 new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
1334#endif
1335
1336 if (xen_start_info->mod_start) {
1337 if (xen_start_info->flags & SIF_MOD_START_PFN)
1338 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1339 else
1340 initrd_start = __pa(xen_start_info->mod_start);
1341 }
1342
1343 /* Poke various useful things into boot_params */
1344 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1345 boot_params.hdr.ramdisk_image = initrd_start;
1346 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1347 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1348 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1349
1350 if (!xen_initial_domain()) {
1351 add_preferred_console("xenboot", 0, NULL);
1352 if (pci_xen)
1353 x86_init.pci.arch_init = pci_xen_init;
1354 } else {
1355 const struct dom0_vga_console_info *info =
1356 (void *)((char *)xen_start_info +
1357 xen_start_info->console.dom0.info_off);
1358 struct xen_platform_op op = {
1359 .cmd = XENPF_firmware_info,
1360 .interface_version = XENPF_INTERFACE_VERSION,
1361 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1362 };
1363
1364 x86_platform.set_legacy_features =
1365 xen_dom0_set_legacy_features;
1366 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1367 xen_start_info->console.domU.mfn = 0;
1368 xen_start_info->console.domU.evtchn = 0;
1369
1370 if (HYPERVISOR_platform_op(&op) == 0)
1371 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1372
1373 /* Make sure ACS will be enabled */
1374 pci_request_acs();
1375
1376 xen_acpi_sleep_register();
1377
1378 /* Avoid searching for BIOS MP tables */
1379 x86_init.mpparse.find_smp_config = x86_init_noop;
1380 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1381
1382 xen_boot_params_init_edd();
1383 }
1384
1385 if (!boot_params.screen_info.orig_video_isVGA)
1386 add_preferred_console("tty", 0, NULL);
1387 add_preferred_console("hvc", 0, NULL);
1388 if (boot_params.screen_info.orig_video_isVGA)
1389 add_preferred_console("tty", 0, NULL);
1390
1391#ifdef CONFIG_PCI
1392 /* PCI BIOS service won't work from a PV guest. */
1393 pci_probe &= ~PCI_PROBE_BIOS;
1394#endif
1395 xen_raw_console_write("about to get started...\n");
1396
1397 /* We need this for printk timestamps */
1398 xen_setup_runstate_info(0);
1399
1400 xen_efi_init(&boot_params);
1401
1402 /* Start the world */
1403#ifdef CONFIG_X86_32
1404 i386_start_kernel();
1405#else
1406 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1407 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1408#endif
1409}
1410
1411static int xen_cpu_up_prepare_pv(unsigned int cpu)
1412{
1413 int rc;
1414
1415 if (per_cpu(xen_vcpu, cpu) == NULL)
1416 return -ENODEV;
1417
1418 xen_setup_timer(cpu);
1419
1420 rc = xen_smp_intr_init(cpu);
1421 if (rc) {
1422 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1423 cpu, rc);
1424 return rc;
1425 }
1426
1427 rc = xen_smp_intr_init_pv(cpu);
1428 if (rc) {
1429 WARN(1, "xen_smp_intr_init_pv() for CPU %d failed: %d\n",
1430 cpu, rc);
1431 return rc;
1432 }
1433
1434 return 0;
1435}
1436
1437static int xen_cpu_dead_pv(unsigned int cpu)
1438{
1439 xen_smp_intr_free(cpu);
1440 xen_smp_intr_free_pv(cpu);
1441
1442 xen_teardown_timer(cpu);
1443
1444 return 0;
1445}
1446
1447static uint32_t __init xen_platform_pv(void)
1448{
1449 if (xen_pv_domain())
1450 return xen_cpuid_base();
1451
1452 return 0;
1453}
1454
1455const __initconst struct hypervisor_x86 x86_hyper_xen_pv = {
1456 .name = "Xen PV",
1457 .detect = xen_platform_pv,
1458 .type = X86_HYPER_XEN_PV,
1459 .runtime.pin_vcpu = xen_pin_vcpu,
1460 .ignore_nopv = true,
1461};