75f33b2a59336a865d7f8abf1dada19763696b70
[linux-3.10.git] / arch / x86 / xen / enlighten.c
1 /*
2  * Core of Xen paravirt_ops implementation.
3  *
4  * This file contains the xen_paravirt_ops structure itself, and the
5  * implementations for:
6  * - privileged instructions
7  * - interrupt flags
8  * - segment operations
9  * - booting and setup
10  *
11  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
12  */
13
14 #include <linux/cpu.h>
15 #include <linux/kernel.h>
16 #include <linux/init.h>
17 #include <linux/smp.h>
18 #include <linux/preempt.h>
19 #include <linux/hardirq.h>
20 #include <linux/percpu.h>
21 #include <linux/delay.h>
22 #include <linux/start_kernel.h>
23 #include <linux/sched.h>
24 #include <linux/kprobes.h>
25 #include <linux/bootmem.h>
26 #include <linux/module.h>
27 #include <linux/mm.h>
28 #include <linux/page-flags.h>
29 #include <linux/highmem.h>
30 #include <linux/console.h>
31 #include <linux/pci.h>
32 #include <linux/gfp.h>
33 #include <linux/memblock.h>
34
35 #include <xen/xen.h>
36 #include <xen/interface/xen.h>
37 #include <xen/interface/version.h>
38 #include <xen/interface/physdev.h>
39 #include <xen/interface/vcpu.h>
40 #include <xen/interface/memory.h>
41 #include <xen/features.h>
42 #include <xen/page.h>
43 #include <xen/hvm.h>
44 #include <xen/hvc-console.h>
45 #include <xen/acpi.h>
46
47 #include <asm/paravirt.h>
48 #include <asm/apic.h>
49 #include <asm/page.h>
50 #include <asm/xen/pci.h>
51 #include <asm/xen/hypercall.h>
52 #include <asm/xen/hypervisor.h>
53 #include <asm/fixmap.h>
54 #include <asm/processor.h>
55 #include <asm/proto.h>
56 #include <asm/msr-index.h>
57 #include <asm/traps.h>
58 #include <asm/setup.h>
59 #include <asm/desc.h>
60 #include <asm/pgalloc.h>
61 #include <asm/pgtable.h>
62 #include <asm/tlbflush.h>
63 #include <asm/reboot.h>
64 #include <asm/stackprotector.h>
65 #include <asm/hypervisor.h>
66 #include <asm/mwait.h>
67 #include <asm/pci_x86.h>
68
69 #ifdef CONFIG_ACPI
70 #include <linux/acpi.h>
71 #include <asm/acpi.h>
72 #include <acpi/pdc_intel.h>
73 #include <acpi/processor.h>
74 #include <xen/interface/platform.h>
75 #endif
76
77 #include "xen-ops.h"
78 #include "mmu.h"
79 #include "smp.h"
80 #include "multicalls.h"
81
82 EXPORT_SYMBOL_GPL(hypercall_page);
83
84 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
85 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
86
87 enum xen_domain_type xen_domain_type = XEN_NATIVE;
88 EXPORT_SYMBOL_GPL(xen_domain_type);
89
90 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
91 EXPORT_SYMBOL(machine_to_phys_mapping);
92 unsigned long  machine_to_phys_nr;
93 EXPORT_SYMBOL(machine_to_phys_nr);
94
95 struct start_info *xen_start_info;
96 EXPORT_SYMBOL_GPL(xen_start_info);
97
98 struct shared_info xen_dummy_shared_info;
99
100 void *xen_initial_gdt;
101
102 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
103 __read_mostly int xen_have_vector_callback;
104 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
105
106 /*
107  * Point at some empty memory to start with. We map the real shared_info
108  * page as soon as fixmap is up and running.
109  */
110 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
111
112 /*
113  * Flag to determine whether vcpu info placement is available on all
114  * VCPUs.  We assume it is to start with, and then set it to zero on
115  * the first failure.  This is because it can succeed on some VCPUs
116  * and not others, since it can involve hypervisor memory allocation,
117  * or because the guest failed to guarantee all the appropriate
118  * constraints on all VCPUs (ie buffer can't cross a page boundary).
119  *
120  * Note that any particular CPU may be using a placed vcpu structure,
121  * but we can only optimise if the all are.
122  *
123  * 0: not available, 1: available
124  */
125 static int have_vcpu_info_placement = 1;
126
127 static void clamp_max_cpus(void)
128 {
129 #ifdef CONFIG_SMP
130         if (setup_max_cpus > MAX_VIRT_CPUS)
131                 setup_max_cpus = MAX_VIRT_CPUS;
132 #endif
133 }
134
135 static void xen_vcpu_setup(int cpu)
136 {
137         struct vcpu_register_vcpu_info info;
138         int err;
139         struct vcpu_info *vcpup;
140
141         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
142
143         if (cpu < MAX_VIRT_CPUS)
144                 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
145
146         if (!have_vcpu_info_placement) {
147                 if (cpu >= MAX_VIRT_CPUS)
148                         clamp_max_cpus();
149                 return;
150         }
151
152         vcpup = &per_cpu(xen_vcpu_info, cpu);
153         info.mfn = arbitrary_virt_to_mfn(vcpup);
154         info.offset = offset_in_page(vcpup);
155
156         /* Check to see if the hypervisor will put the vcpu_info
157            structure where we want it, which allows direct access via
158            a percpu-variable. */
159         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
160
161         if (err) {
162                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
163                 have_vcpu_info_placement = 0;
164                 clamp_max_cpus();
165         } else {
166                 /* This cpu is using the registered vcpu info, even if
167                    later ones fail to. */
168                 per_cpu(xen_vcpu, cpu) = vcpup;
169         }
170 }
171
172 /*
173  * On restore, set the vcpu placement up again.
174  * If it fails, then we're in a bad state, since
175  * we can't back out from using it...
176  */
177 void xen_vcpu_restore(void)
178 {
179         int cpu;
180
181         for_each_online_cpu(cpu) {
182                 bool other_cpu = (cpu != smp_processor_id());
183
184                 if (other_cpu &&
185                     HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
186                         BUG();
187
188                 xen_setup_runstate_info(cpu);
189
190                 if (have_vcpu_info_placement)
191                         xen_vcpu_setup(cpu);
192
193                 if (other_cpu &&
194                     HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
195                         BUG();
196         }
197 }
198
199 static void __init xen_banner(void)
200 {
201         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
202         struct xen_extraversion extra;
203         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
204
205         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
206                pv_info.name);
207         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
208                version >> 16, version & 0xffff, extra.extraversion,
209                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
210 }
211
212 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
213 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
214
215 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
216 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
217 static __read_mostly unsigned int cpuid_leaf5_edx_val;
218
219 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
220                       unsigned int *cx, unsigned int *dx)
221 {
222         unsigned maskebx = ~0;
223         unsigned maskecx = ~0;
224         unsigned maskedx = ~0;
225         unsigned setecx = 0;
226         /*
227          * Mask out inconvenient features, to try and disable as many
228          * unsupported kernel subsystems as possible.
229          */
230         switch (*ax) {
231         case 1:
232                 maskecx = cpuid_leaf1_ecx_mask;
233                 setecx = cpuid_leaf1_ecx_set_mask;
234                 maskedx = cpuid_leaf1_edx_mask;
235                 break;
236
237         case CPUID_MWAIT_LEAF:
238                 /* Synthesize the values.. */
239                 *ax = 0;
240                 *bx = 0;
241                 *cx = cpuid_leaf5_ecx_val;
242                 *dx = cpuid_leaf5_edx_val;
243                 return;
244
245         case 0xb:
246                 /* Suppress extended topology stuff */
247                 maskebx = 0;
248                 break;
249         }
250
251         asm(XEN_EMULATE_PREFIX "cpuid"
252                 : "=a" (*ax),
253                   "=b" (*bx),
254                   "=c" (*cx),
255                   "=d" (*dx)
256                 : "0" (*ax), "2" (*cx));
257
258         *bx &= maskebx;
259         *cx &= maskecx;
260         *cx |= setecx;
261         *dx &= maskedx;
262
263 }
264
265 static bool __init xen_check_mwait(void)
266 {
267 #if defined(CONFIG_ACPI) && !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR) && \
268         !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR_MODULE)
269         struct xen_platform_op op = {
270                 .cmd                    = XENPF_set_processor_pminfo,
271                 .u.set_pminfo.id        = -1,
272                 .u.set_pminfo.type      = XEN_PM_PDC,
273         };
274         uint32_t buf[3];
275         unsigned int ax, bx, cx, dx;
276         unsigned int mwait_mask;
277
278         /* We need to determine whether it is OK to expose the MWAIT
279          * capability to the kernel to harvest deeper than C3 states from ACPI
280          * _CST using the processor_harvest_xen.c module. For this to work, we
281          * need to gather the MWAIT_LEAF values (which the cstate.c code
282          * checks against). The hypervisor won't expose the MWAIT flag because
283          * it would break backwards compatibility; so we will find out directly
284          * from the hardware and hypercall.
285          */
286         if (!xen_initial_domain())
287                 return false;
288
289         ax = 1;
290         cx = 0;
291
292         native_cpuid(&ax, &bx, &cx, &dx);
293
294         mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
295                      (1 << (X86_FEATURE_MWAIT % 32));
296
297         if ((cx & mwait_mask) != mwait_mask)
298                 return false;
299
300         /* We need to emulate the MWAIT_LEAF and for that we need both
301          * ecx and edx. The hypercall provides only partial information.
302          */
303
304         ax = CPUID_MWAIT_LEAF;
305         bx = 0;
306         cx = 0;
307         dx = 0;
308
309         native_cpuid(&ax, &bx, &cx, &dx);
310
311         /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
312          * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
313          */
314         buf[0] = ACPI_PDC_REVISION_ID;
315         buf[1] = 1;
316         buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
317
318         set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
319
320         if ((HYPERVISOR_dom0_op(&op) == 0) &&
321             (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
322                 cpuid_leaf5_ecx_val = cx;
323                 cpuid_leaf5_edx_val = dx;
324         }
325         return true;
326 #else
327         return false;
328 #endif
329 }
330 static void __init xen_init_cpuid_mask(void)
331 {
332         unsigned int ax, bx, cx, dx;
333         unsigned int xsave_mask;
334
335         cpuid_leaf1_edx_mask =
336                 ~((1 << X86_FEATURE_MCE)  |  /* disable MCE */
337                   (1 << X86_FEATURE_MCA)  |  /* disable MCA */
338                   (1 << X86_FEATURE_MTRR) |  /* disable MTRR */
339                   (1 << X86_FEATURE_ACC));   /* thermal monitoring */
340
341         if (!xen_initial_domain())
342                 cpuid_leaf1_edx_mask &=
343                         ~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
344                           (1 << X86_FEATURE_ACPI));  /* disable ACPI */
345         ax = 1;
346         cx = 0;
347         xen_cpuid(&ax, &bx, &cx, &dx);
348
349         xsave_mask =
350                 (1 << (X86_FEATURE_XSAVE % 32)) |
351                 (1 << (X86_FEATURE_OSXSAVE % 32));
352
353         /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
354         if ((cx & xsave_mask) != xsave_mask)
355                 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
356         if (xen_check_mwait())
357                 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
358 }
359
360 static void xen_set_debugreg(int reg, unsigned long val)
361 {
362         HYPERVISOR_set_debugreg(reg, val);
363 }
364
365 static unsigned long xen_get_debugreg(int reg)
366 {
367         return HYPERVISOR_get_debugreg(reg);
368 }
369
370 static void xen_end_context_switch(struct task_struct *next)
371 {
372         xen_mc_flush();
373         paravirt_end_context_switch(next);
374 }
375
376 static unsigned long xen_store_tr(void)
377 {
378         return 0;
379 }
380
381 /*
382  * Set the page permissions for a particular virtual address.  If the
383  * address is a vmalloc mapping (or other non-linear mapping), then
384  * find the linear mapping of the page and also set its protections to
385  * match.
386  */
387 static void set_aliased_prot(void *v, pgprot_t prot)
388 {
389         int level;
390         pte_t *ptep;
391         pte_t pte;
392         unsigned long pfn;
393         struct page *page;
394
395         ptep = lookup_address((unsigned long)v, &level);
396         BUG_ON(ptep == NULL);
397
398         pfn = pte_pfn(*ptep);
399         page = pfn_to_page(pfn);
400
401         pte = pfn_pte(pfn, prot);
402
403         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
404                 BUG();
405
406         if (!PageHighMem(page)) {
407                 void *av = __va(PFN_PHYS(pfn));
408
409                 if (av != v)
410                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
411                                 BUG();
412         } else
413                 kmap_flush_unused();
414 }
415
416 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
417 {
418         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
419         int i;
420
421         for(i = 0; i < entries; i += entries_per_page)
422                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
423 }
424
425 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
426 {
427         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
428         int i;
429
430         for(i = 0; i < entries; i += entries_per_page)
431                 set_aliased_prot(ldt + i, PAGE_KERNEL);
432 }
433
434 static void xen_set_ldt(const void *addr, unsigned entries)
435 {
436         struct mmuext_op *op;
437         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
438
439         trace_xen_cpu_set_ldt(addr, entries);
440
441         op = mcs.args;
442         op->cmd = MMUEXT_SET_LDT;
443         op->arg1.linear_addr = (unsigned long)addr;
444         op->arg2.nr_ents = entries;
445
446         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
447
448         xen_mc_issue(PARAVIRT_LAZY_CPU);
449 }
450
451 static void xen_load_gdt(const struct desc_ptr *dtr)
452 {
453         unsigned long va = dtr->address;
454         unsigned int size = dtr->size + 1;
455         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
456         unsigned long frames[pages];
457         int f;
458
459         /*
460          * A GDT can be up to 64k in size, which corresponds to 8192
461          * 8-byte entries, or 16 4k pages..
462          */
463
464         BUG_ON(size > 65536);
465         BUG_ON(va & ~PAGE_MASK);
466
467         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
468                 int level;
469                 pte_t *ptep;
470                 unsigned long pfn, mfn;
471                 void *virt;
472
473                 /*
474                  * The GDT is per-cpu and is in the percpu data area.
475                  * That can be virtually mapped, so we need to do a
476                  * page-walk to get the underlying MFN for the
477                  * hypercall.  The page can also be in the kernel's
478                  * linear range, so we need to RO that mapping too.
479                  */
480                 ptep = lookup_address(va, &level);
481                 BUG_ON(ptep == NULL);
482
483                 pfn = pte_pfn(*ptep);
484                 mfn = pfn_to_mfn(pfn);
485                 virt = __va(PFN_PHYS(pfn));
486
487                 frames[f] = mfn;
488
489                 make_lowmem_page_readonly((void *)va);
490                 make_lowmem_page_readonly(virt);
491         }
492
493         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
494                 BUG();
495 }
496
497 /*
498  * load_gdt for early boot, when the gdt is only mapped once
499  */
500 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
501 {
502         unsigned long va = dtr->address;
503         unsigned int size = dtr->size + 1;
504         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
505         unsigned long frames[pages];
506         int f;
507
508         /*
509          * A GDT can be up to 64k in size, which corresponds to 8192
510          * 8-byte entries, or 16 4k pages..
511          */
512
513         BUG_ON(size > 65536);
514         BUG_ON(va & ~PAGE_MASK);
515
516         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
517                 pte_t pte;
518                 unsigned long pfn, mfn;
519
520                 pfn = virt_to_pfn(va);
521                 mfn = pfn_to_mfn(pfn);
522
523                 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
524
525                 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
526                         BUG();
527
528                 frames[f] = mfn;
529         }
530
531         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
532                 BUG();
533 }
534
535 static void load_TLS_descriptor(struct thread_struct *t,
536                                 unsigned int cpu, unsigned int i)
537 {
538         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
539         xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
540         struct multicall_space mc = __xen_mc_entry(0);
541
542         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
543 }
544
545 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
546 {
547         /*
548          * XXX sleazy hack: If we're being called in a lazy-cpu zone
549          * and lazy gs handling is enabled, it means we're in a
550          * context switch, and %gs has just been saved.  This means we
551          * can zero it out to prevent faults on exit from the
552          * hypervisor if the next process has no %gs.  Either way, it
553          * has been saved, and the new value will get loaded properly.
554          * This will go away as soon as Xen has been modified to not
555          * save/restore %gs for normal hypercalls.
556          *
557          * On x86_64, this hack is not used for %gs, because gs points
558          * to KERNEL_GS_BASE (and uses it for PDA references), so we
559          * must not zero %gs on x86_64
560          *
561          * For x86_64, we need to zero %fs, otherwise we may get an
562          * exception between the new %fs descriptor being loaded and
563          * %fs being effectively cleared at __switch_to().
564          */
565         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
566 #ifdef CONFIG_X86_32
567                 lazy_load_gs(0);
568 #else
569                 loadsegment(fs, 0);
570 #endif
571         }
572
573         xen_mc_batch();
574
575         load_TLS_descriptor(t, cpu, 0);
576         load_TLS_descriptor(t, cpu, 1);
577         load_TLS_descriptor(t, cpu, 2);
578
579         xen_mc_issue(PARAVIRT_LAZY_CPU);
580 }
581
582 #ifdef CONFIG_X86_64
583 static void xen_load_gs_index(unsigned int idx)
584 {
585         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
586                 BUG();
587 }
588 #endif
589
590 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
591                                 const void *ptr)
592 {
593         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
594         u64 entry = *(u64 *)ptr;
595
596         trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
597
598         preempt_disable();
599
600         xen_mc_flush();
601         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
602                 BUG();
603
604         preempt_enable();
605 }
606
607 static int cvt_gate_to_trap(int vector, const gate_desc *val,
608                             struct trap_info *info)
609 {
610         unsigned long addr;
611
612         if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
613                 return 0;
614
615         info->vector = vector;
616
617         addr = gate_offset(*val);
618 #ifdef CONFIG_X86_64
619         /*
620          * Look for known traps using IST, and substitute them
621          * appropriately.  The debugger ones are the only ones we care
622          * about.  Xen will handle faults like double_fault and
623          * machine_check, so we should never see them.  Warn if
624          * there's an unexpected IST-using fault handler.
625          */
626         if (addr == (unsigned long)debug)
627                 addr = (unsigned long)xen_debug;
628         else if (addr == (unsigned long)int3)
629                 addr = (unsigned long)xen_int3;
630         else if (addr == (unsigned long)stack_segment)
631                 addr = (unsigned long)xen_stack_segment;
632         else if (addr == (unsigned long)double_fault ||
633                  addr == (unsigned long)nmi) {
634                 /* Don't need to handle these */
635                 return 0;
636 #ifdef CONFIG_X86_MCE
637         } else if (addr == (unsigned long)machine_check) {
638                 return 0;
639 #endif
640         } else {
641                 /* Some other trap using IST? */
642                 if (WARN_ON(val->ist != 0))
643                         return 0;
644         }
645 #endif  /* CONFIG_X86_64 */
646         info->address = addr;
647
648         info->cs = gate_segment(*val);
649         info->flags = val->dpl;
650         /* interrupt gates clear IF */
651         if (val->type == GATE_INTERRUPT)
652                 info->flags |= 1 << 2;
653
654         return 1;
655 }
656
657 /* Locations of each CPU's IDT */
658 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
659
660 /* Set an IDT entry.  If the entry is part of the current IDT, then
661    also update Xen. */
662 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
663 {
664         unsigned long p = (unsigned long)&dt[entrynum];
665         unsigned long start, end;
666
667         trace_xen_cpu_write_idt_entry(dt, entrynum, g);
668
669         preempt_disable();
670
671         start = __this_cpu_read(idt_desc.address);
672         end = start + __this_cpu_read(idt_desc.size) + 1;
673
674         xen_mc_flush();
675
676         native_write_idt_entry(dt, entrynum, g);
677
678         if (p >= start && (p + 8) <= end) {
679                 struct trap_info info[2];
680
681                 info[1].address = 0;
682
683                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
684                         if (HYPERVISOR_set_trap_table(info))
685                                 BUG();
686         }
687
688         preempt_enable();
689 }
690
691 static void xen_convert_trap_info(const struct desc_ptr *desc,
692                                   struct trap_info *traps)
693 {
694         unsigned in, out, count;
695
696         count = (desc->size+1) / sizeof(gate_desc);
697         BUG_ON(count > 256);
698
699         for (in = out = 0; in < count; in++) {
700                 gate_desc *entry = (gate_desc*)(desc->address) + in;
701
702                 if (cvt_gate_to_trap(in, entry, &traps[out]))
703                         out++;
704         }
705         traps[out].address = 0;
706 }
707
708 void xen_copy_trap_info(struct trap_info *traps)
709 {
710         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
711
712         xen_convert_trap_info(desc, traps);
713 }
714
715 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
716    hold a spinlock to protect the static traps[] array (static because
717    it avoids allocation, and saves stack space). */
718 static void xen_load_idt(const struct desc_ptr *desc)
719 {
720         static DEFINE_SPINLOCK(lock);
721         static struct trap_info traps[257];
722
723         trace_xen_cpu_load_idt(desc);
724
725         spin_lock(&lock);
726
727         __get_cpu_var(idt_desc) = *desc;
728
729         xen_convert_trap_info(desc, traps);
730
731         xen_mc_flush();
732         if (HYPERVISOR_set_trap_table(traps))
733                 BUG();
734
735         spin_unlock(&lock);
736 }
737
738 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
739    they're handled differently. */
740 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
741                                 const void *desc, int type)
742 {
743         trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
744
745         preempt_disable();
746
747         switch (type) {
748         case DESC_LDT:
749         case DESC_TSS:
750                 /* ignore */
751                 break;
752
753         default: {
754                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
755
756                 xen_mc_flush();
757                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
758                         BUG();
759         }
760
761         }
762
763         preempt_enable();
764 }
765
766 /*
767  * Version of write_gdt_entry for use at early boot-time needed to
768  * update an entry as simply as possible.
769  */
770 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
771                                             const void *desc, int type)
772 {
773         trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
774
775         switch (type) {
776         case DESC_LDT:
777         case DESC_TSS:
778                 /* ignore */
779                 break;
780
781         default: {
782                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
783
784                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
785                         dt[entry] = *(struct desc_struct *)desc;
786         }
787
788         }
789 }
790
791 static void xen_load_sp0(struct tss_struct *tss,
792                          struct thread_struct *thread)
793 {
794         struct multicall_space mcs;
795
796         mcs = xen_mc_entry(0);
797         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
798         xen_mc_issue(PARAVIRT_LAZY_CPU);
799 }
800
801 static void xen_set_iopl_mask(unsigned mask)
802 {
803         struct physdev_set_iopl set_iopl;
804
805         /* Force the change at ring 0. */
806         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
807         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
808 }
809
810 static void xen_io_delay(void)
811 {
812 }
813
814 #ifdef CONFIG_X86_LOCAL_APIC
815 static unsigned long xen_set_apic_id(unsigned int x)
816 {
817         WARN_ON(1);
818         return x;
819 }
820 static unsigned int xen_get_apic_id(unsigned long x)
821 {
822         return ((x)>>24) & 0xFFu;
823 }
824 static u32 xen_apic_read(u32 reg)
825 {
826         struct xen_platform_op op = {
827                 .cmd = XENPF_get_cpuinfo,
828                 .interface_version = XENPF_INTERFACE_VERSION,
829                 .u.pcpu_info.xen_cpuid = 0,
830         };
831         int ret = 0;
832
833         /* Shouldn't need this as APIC is turned off for PV, and we only
834          * get called on the bootup processor. But just in case. */
835         if (!xen_initial_domain() || smp_processor_id())
836                 return 0;
837
838         if (reg == APIC_LVR)
839                 return 0x10;
840
841         if (reg != APIC_ID)
842                 return 0;
843
844         ret = HYPERVISOR_dom0_op(&op);
845         if (ret)
846                 return 0;
847
848         return op.u.pcpu_info.apic_id << 24;
849 }
850
851 static void xen_apic_write(u32 reg, u32 val)
852 {
853         /* Warn to see if there's any stray references */
854         WARN_ON(1);
855 }
856
857 static u64 xen_apic_icr_read(void)
858 {
859         return 0;
860 }
861
862 static void xen_apic_icr_write(u32 low, u32 id)
863 {
864         /* Warn to see if there's any stray references */
865         WARN_ON(1);
866 }
867
868 static void xen_apic_wait_icr_idle(void)
869 {
870         return;
871 }
872
873 static u32 xen_safe_apic_wait_icr_idle(void)
874 {
875         return 0;
876 }
877
878 static void set_xen_basic_apic_ops(void)
879 {
880         apic->read = xen_apic_read;
881         apic->write = xen_apic_write;
882         apic->icr_read = xen_apic_icr_read;
883         apic->icr_write = xen_apic_icr_write;
884         apic->wait_icr_idle = xen_apic_wait_icr_idle;
885         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
886         apic->set_apic_id = xen_set_apic_id;
887         apic->get_apic_id = xen_get_apic_id;
888
889 #ifdef CONFIG_SMP
890         apic->send_IPI_allbutself = xen_send_IPI_allbutself;
891         apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
892         apic->send_IPI_mask = xen_send_IPI_mask;
893         apic->send_IPI_all = xen_send_IPI_all;
894         apic->send_IPI_self = xen_send_IPI_self;
895 #endif
896 }
897
898 #endif
899
900 static void xen_clts(void)
901 {
902         struct multicall_space mcs;
903
904         mcs = xen_mc_entry(0);
905
906         MULTI_fpu_taskswitch(mcs.mc, 0);
907
908         xen_mc_issue(PARAVIRT_LAZY_CPU);
909 }
910
911 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
912
913 static unsigned long xen_read_cr0(void)
914 {
915         unsigned long cr0 = this_cpu_read(xen_cr0_value);
916
917         if (unlikely(cr0 == 0)) {
918                 cr0 = native_read_cr0();
919                 this_cpu_write(xen_cr0_value, cr0);
920         }
921
922         return cr0;
923 }
924
925 static void xen_write_cr0(unsigned long cr0)
926 {
927         struct multicall_space mcs;
928
929         this_cpu_write(xen_cr0_value, cr0);
930
931         /* Only pay attention to cr0.TS; everything else is
932            ignored. */
933         mcs = xen_mc_entry(0);
934
935         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
936
937         xen_mc_issue(PARAVIRT_LAZY_CPU);
938 }
939
940 static void xen_write_cr4(unsigned long cr4)
941 {
942         cr4 &= ~X86_CR4_PGE;
943         cr4 &= ~X86_CR4_PSE;
944
945         native_write_cr4(cr4);
946 }
947
948 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
949 {
950         int ret;
951
952         ret = 0;
953
954         switch (msr) {
955 #ifdef CONFIG_X86_64
956                 unsigned which;
957                 u64 base;
958
959         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
960         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
961         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
962
963         set:
964                 base = ((u64)high << 32) | low;
965                 if (HYPERVISOR_set_segment_base(which, base) != 0)
966                         ret = -EIO;
967                 break;
968 #endif
969
970         case MSR_STAR:
971         case MSR_CSTAR:
972         case MSR_LSTAR:
973         case MSR_SYSCALL_MASK:
974         case MSR_IA32_SYSENTER_CS:
975         case MSR_IA32_SYSENTER_ESP:
976         case MSR_IA32_SYSENTER_EIP:
977                 /* Fast syscall setup is all done in hypercalls, so
978                    these are all ignored.  Stub them out here to stop
979                    Xen console noise. */
980                 break;
981
982         case MSR_IA32_CR_PAT:
983                 if (smp_processor_id() == 0)
984                         xen_set_pat(((u64)high << 32) | low);
985                 break;
986
987         default:
988                 ret = native_write_msr_safe(msr, low, high);
989         }
990
991         return ret;
992 }
993
994 void xen_setup_shared_info(void)
995 {
996         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
997                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
998                            xen_start_info->shared_info);
999
1000                 HYPERVISOR_shared_info =
1001                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1002         } else
1003                 HYPERVISOR_shared_info =
1004                         (struct shared_info *)__va(xen_start_info->shared_info);
1005
1006 #ifndef CONFIG_SMP
1007         /* In UP this is as good a place as any to set up shared info */
1008         xen_setup_vcpu_info_placement();
1009 #endif
1010
1011         xen_setup_mfn_list_list();
1012 }
1013
1014 /* This is called once we have the cpu_possible_mask */
1015 void xen_setup_vcpu_info_placement(void)
1016 {
1017         int cpu;
1018
1019         for_each_possible_cpu(cpu)
1020                 xen_vcpu_setup(cpu);
1021
1022         /* xen_vcpu_setup managed to place the vcpu_info within the
1023            percpu area for all cpus, so make use of it */
1024         if (have_vcpu_info_placement) {
1025                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1026                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1027                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1028                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1029                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1030         }
1031 }
1032
1033 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1034                           unsigned long addr, unsigned len)
1035 {
1036         char *start, *end, *reloc;
1037         unsigned ret;
1038
1039         start = end = reloc = NULL;
1040
1041 #define SITE(op, x)                                                     \
1042         case PARAVIRT_PATCH(op.x):                                      \
1043         if (have_vcpu_info_placement) {                                 \
1044                 start = (char *)xen_##x##_direct;                       \
1045                 end = xen_##x##_direct_end;                             \
1046                 reloc = xen_##x##_direct_reloc;                         \
1047         }                                                               \
1048         goto patch_site
1049
1050         switch (type) {
1051                 SITE(pv_irq_ops, irq_enable);
1052                 SITE(pv_irq_ops, irq_disable);
1053                 SITE(pv_irq_ops, save_fl);
1054                 SITE(pv_irq_ops, restore_fl);
1055 #undef SITE
1056
1057         patch_site:
1058                 if (start == NULL || (end-start) > len)
1059                         goto default_patch;
1060
1061                 ret = paravirt_patch_insns(insnbuf, len, start, end);
1062
1063                 /* Note: because reloc is assigned from something that
1064                    appears to be an array, gcc assumes it's non-null,
1065                    but doesn't know its relationship with start and
1066                    end. */
1067                 if (reloc > start && reloc < end) {
1068                         int reloc_off = reloc - start;
1069                         long *relocp = (long *)(insnbuf + reloc_off);
1070                         long delta = start - (char *)addr;
1071
1072                         *relocp += delta;
1073                 }
1074                 break;
1075
1076         default_patch:
1077         default:
1078                 ret = paravirt_patch_default(type, clobbers, insnbuf,
1079                                              addr, len);
1080                 break;
1081         }
1082
1083         return ret;
1084 }
1085
1086 static const struct pv_info xen_info __initconst = {
1087         .paravirt_enabled = 1,
1088         .shared_kernel_pmd = 0,
1089
1090 #ifdef CONFIG_X86_64
1091         .extra_user_64bit_cs = FLAT_USER_CS64,
1092 #endif
1093
1094         .name = "Xen",
1095 };
1096
1097 static const struct pv_init_ops xen_init_ops __initconst = {
1098         .patch = xen_patch,
1099 };
1100
1101 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1102         .cpuid = xen_cpuid,
1103
1104         .set_debugreg = xen_set_debugreg,
1105         .get_debugreg = xen_get_debugreg,
1106
1107         .clts = xen_clts,
1108
1109         .read_cr0 = xen_read_cr0,
1110         .write_cr0 = xen_write_cr0,
1111
1112         .read_cr4 = native_read_cr4,
1113         .read_cr4_safe = native_read_cr4_safe,
1114         .write_cr4 = xen_write_cr4,
1115
1116         .wbinvd = native_wbinvd,
1117
1118         .read_msr = native_read_msr_safe,
1119         .write_msr = xen_write_msr_safe,
1120         .read_tsc = native_read_tsc,
1121         .read_pmc = native_read_pmc,
1122
1123         .iret = xen_iret,
1124         .irq_enable_sysexit = xen_sysexit,
1125 #ifdef CONFIG_X86_64
1126         .usergs_sysret32 = xen_sysret32,
1127         .usergs_sysret64 = xen_sysret64,
1128 #endif
1129
1130         .load_tr_desc = paravirt_nop,
1131         .set_ldt = xen_set_ldt,
1132         .load_gdt = xen_load_gdt,
1133         .load_idt = xen_load_idt,
1134         .load_tls = xen_load_tls,
1135 #ifdef CONFIG_X86_64
1136         .load_gs_index = xen_load_gs_index,
1137 #endif
1138
1139         .alloc_ldt = xen_alloc_ldt,
1140         .free_ldt = xen_free_ldt,
1141
1142         .store_gdt = native_store_gdt,
1143         .store_idt = native_store_idt,
1144         .store_tr = xen_store_tr,
1145
1146         .write_ldt_entry = xen_write_ldt_entry,
1147         .write_gdt_entry = xen_write_gdt_entry,
1148         .write_idt_entry = xen_write_idt_entry,
1149         .load_sp0 = xen_load_sp0,
1150
1151         .set_iopl_mask = xen_set_iopl_mask,
1152         .io_delay = xen_io_delay,
1153
1154         /* Xen takes care of %gs when switching to usermode for us */
1155         .swapgs = paravirt_nop,
1156
1157         .start_context_switch = paravirt_start_context_switch,
1158         .end_context_switch = xen_end_context_switch,
1159 };
1160
1161 static const struct pv_apic_ops xen_apic_ops __initconst = {
1162 #ifdef CONFIG_X86_LOCAL_APIC
1163         .startup_ipi_hook = paravirt_nop,
1164 #endif
1165 };
1166
1167 static void xen_reboot(int reason)
1168 {
1169         struct sched_shutdown r = { .reason = reason };
1170
1171         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1172                 BUG();
1173 }
1174
1175 static void xen_restart(char *msg)
1176 {
1177         xen_reboot(SHUTDOWN_reboot);
1178 }
1179
1180 static void xen_emergency_restart(void)
1181 {
1182         xen_reboot(SHUTDOWN_reboot);
1183 }
1184
1185 static void xen_machine_halt(void)
1186 {
1187         xen_reboot(SHUTDOWN_poweroff);
1188 }
1189
1190 static void xen_machine_power_off(void)
1191 {
1192         if (pm_power_off)
1193                 pm_power_off();
1194         xen_reboot(SHUTDOWN_poweroff);
1195 }
1196
1197 static void xen_crash_shutdown(struct pt_regs *regs)
1198 {
1199         xen_reboot(SHUTDOWN_crash);
1200 }
1201
1202 static int
1203 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1204 {
1205         xen_reboot(SHUTDOWN_crash);
1206         return NOTIFY_DONE;
1207 }
1208
1209 static struct notifier_block xen_panic_block = {
1210         .notifier_call= xen_panic_event,
1211 };
1212
1213 int xen_panic_handler_init(void)
1214 {
1215         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1216         return 0;
1217 }
1218
1219 static const struct machine_ops xen_machine_ops __initconst = {
1220         .restart = xen_restart,
1221         .halt = xen_machine_halt,
1222         .power_off = xen_machine_power_off,
1223         .shutdown = xen_machine_halt,
1224         .crash_shutdown = xen_crash_shutdown,
1225         .emergency_restart = xen_emergency_restart,
1226 };
1227
1228 /*
1229  * Set up the GDT and segment registers for -fstack-protector.  Until
1230  * we do this, we have to be careful not to call any stack-protected
1231  * function, which is most of the kernel.
1232  */
1233 static void __init xen_setup_stackprotector(void)
1234 {
1235         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1236         pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1237
1238         setup_stack_canary_segment(0);
1239         switch_to_new_gdt(0);
1240
1241         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1242         pv_cpu_ops.load_gdt = xen_load_gdt;
1243 }
1244
1245 /* First C function to be called on Xen boot */
1246 asmlinkage void __init xen_start_kernel(void)
1247 {
1248         struct physdev_set_iopl set_iopl;
1249         int rc;
1250         pgd_t *pgd;
1251
1252         if (!xen_start_info)
1253                 return;
1254
1255         xen_domain_type = XEN_PV_DOMAIN;
1256
1257         xen_setup_machphys_mapping();
1258
1259         /* Install Xen paravirt ops */
1260         pv_info = xen_info;
1261         pv_init_ops = xen_init_ops;
1262         pv_cpu_ops = xen_cpu_ops;
1263         pv_apic_ops = xen_apic_ops;
1264
1265         x86_init.resources.memory_setup = xen_memory_setup;
1266         x86_init.oem.arch_setup = xen_arch_setup;
1267         x86_init.oem.banner = xen_banner;
1268
1269         xen_init_time_ops();
1270
1271         /*
1272          * Set up some pagetable state before starting to set any ptes.
1273          */
1274
1275         xen_init_mmu_ops();
1276
1277         /* Prevent unwanted bits from being set in PTEs. */
1278         __supported_pte_mask &= ~_PAGE_GLOBAL;
1279 #if 0
1280         if (!xen_initial_domain())
1281 #endif
1282                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1283
1284         __supported_pte_mask |= _PAGE_IOMAP;
1285
1286         /*
1287          * Prevent page tables from being allocated in highmem, even
1288          * if CONFIG_HIGHPTE is enabled.
1289          */
1290         __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1291
1292         /* Work out if we support NX */
1293         x86_configure_nx();
1294
1295         xen_setup_features();
1296
1297         /* Get mfn list */
1298         if (!xen_feature(XENFEAT_auto_translated_physmap))
1299                 xen_build_dynamic_phys_to_machine();
1300
1301         /*
1302          * Set up kernel GDT and segment registers, mainly so that
1303          * -fstack-protector code can be executed.
1304          */
1305         xen_setup_stackprotector();
1306
1307         xen_init_irq_ops();
1308         xen_init_cpuid_mask();
1309
1310 #ifdef CONFIG_X86_LOCAL_APIC
1311         /*
1312          * set up the basic apic ops.
1313          */
1314         set_xen_basic_apic_ops();
1315 #endif
1316
1317         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1318                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1319                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1320         }
1321
1322         machine_ops = xen_machine_ops;
1323
1324         /*
1325          * The only reliable way to retain the initial address of the
1326          * percpu gdt_page is to remember it here, so we can go and
1327          * mark it RW later, when the initial percpu area is freed.
1328          */
1329         xen_initial_gdt = &per_cpu(gdt_page, 0);
1330
1331         xen_smp_init();
1332
1333 #ifdef CONFIG_ACPI_NUMA
1334         /*
1335          * The pages we from Xen are not related to machine pages, so
1336          * any NUMA information the kernel tries to get from ACPI will
1337          * be meaningless.  Prevent it from trying.
1338          */
1339         acpi_numa = -1;
1340 #endif
1341
1342         pgd = (pgd_t *)xen_start_info->pt_base;
1343
1344         /* Don't do the full vcpu_info placement stuff until we have a
1345            possible map and a non-dummy shared_info. */
1346         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1347
1348         local_irq_disable();
1349         early_boot_irqs_disabled = true;
1350
1351         xen_raw_console_write("mapping kernel into physical memory\n");
1352         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1353
1354         /* Allocate and initialize top and mid mfn levels for p2m structure */
1355         xen_build_mfn_list_list();
1356
1357         /* keep using Xen gdt for now; no urgent need to change it */
1358
1359 #ifdef CONFIG_X86_32
1360         pv_info.kernel_rpl = 1;
1361         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1362                 pv_info.kernel_rpl = 0;
1363 #else
1364         pv_info.kernel_rpl = 0;
1365 #endif
1366         /* set the limit of our address space */
1367         xen_reserve_top();
1368
1369         /* We used to do this in xen_arch_setup, but that is too late on AMD
1370          * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1371          * which pokes 0xcf8 port.
1372          */
1373         set_iopl.iopl = 1;
1374         rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1375         if (rc != 0)
1376                 xen_raw_printk("physdev_op failed %d\n", rc);
1377
1378 #ifdef CONFIG_X86_32
1379         /* set up basic CPUID stuff */
1380         cpu_detect(&new_cpu_data);
1381         new_cpu_data.hard_math = 1;
1382         new_cpu_data.wp_works_ok = 1;
1383         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1384 #endif
1385
1386         /* Poke various useful things into boot_params */
1387         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1388         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1389                 ? __pa(xen_start_info->mod_start) : 0;
1390         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1391         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1392
1393         if (!xen_initial_domain()) {
1394                 add_preferred_console("xenboot", 0, NULL);
1395                 add_preferred_console("tty", 0, NULL);
1396                 add_preferred_console("hvc", 0, NULL);
1397                 if (pci_xen)
1398                         x86_init.pci.arch_init = pci_xen_init;
1399         } else {
1400                 const struct dom0_vga_console_info *info =
1401                         (void *)((char *)xen_start_info +
1402                                  xen_start_info->console.dom0.info_off);
1403
1404                 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1405                 xen_start_info->console.domU.mfn = 0;
1406                 xen_start_info->console.domU.evtchn = 0;
1407
1408                 xen_init_apic();
1409
1410                 /* Make sure ACS will be enabled */
1411                 pci_request_acs();
1412
1413                 xen_acpi_sleep_register();
1414         }
1415 #ifdef CONFIG_PCI
1416         /* PCI BIOS service won't work from a PV guest. */
1417         pci_probe &= ~PCI_PROBE_BIOS;
1418 #endif
1419         xen_raw_console_write("about to get started...\n");
1420
1421         xen_setup_runstate_info(0);
1422
1423         /* Start the world */
1424 #ifdef CONFIG_X86_32
1425         i386_start_kernel();
1426 #else
1427         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1428 #endif
1429 }
1430
1431 static int init_hvm_pv_info(int *major, int *minor)
1432 {
1433         uint32_t eax, ebx, ecx, edx, pages, msr, base;
1434         u64 pfn;
1435
1436         base = xen_cpuid_base();
1437         cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1438
1439         *major = eax >> 16;
1440         *minor = eax & 0xffff;
1441         printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1442
1443         cpuid(base + 2, &pages, &msr, &ecx, &edx);
1444
1445         pfn = __pa(hypercall_page);
1446         wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1447
1448         xen_setup_features();
1449
1450         pv_info.name = "Xen HVM";
1451
1452         xen_domain_type = XEN_HVM_DOMAIN;
1453
1454         return 0;
1455 }
1456
1457 void __ref xen_hvm_init_shared_info(void)
1458 {
1459         int cpu;
1460         struct xen_add_to_physmap xatp;
1461         static struct shared_info *shared_info_page = 0;
1462
1463         if (!shared_info_page)
1464                 shared_info_page = (struct shared_info *)
1465                         extend_brk(PAGE_SIZE, PAGE_SIZE);
1466         xatp.domid = DOMID_SELF;
1467         xatp.idx = 0;
1468         xatp.space = XENMAPSPACE_shared_info;
1469         xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1470         if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1471                 BUG();
1472
1473         HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1474
1475         /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1476          * page, we use it in the event channel upcall and in some pvclock
1477          * related functions. We don't need the vcpu_info placement
1478          * optimizations because we don't use any pv_mmu or pv_irq op on
1479          * HVM.
1480          * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1481          * online but xen_hvm_init_shared_info is run at resume time too and
1482          * in that case multiple vcpus might be online. */
1483         for_each_online_cpu(cpu) {
1484                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1485         }
1486 }
1487
1488 #ifdef CONFIG_XEN_PVHVM
1489 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1490                                     unsigned long action, void *hcpu)
1491 {
1492         int cpu = (long)hcpu;
1493         switch (action) {
1494         case CPU_UP_PREPARE:
1495                 xen_vcpu_setup(cpu);
1496                 if (xen_have_vector_callback)
1497                         xen_init_lock_cpu(cpu);
1498                 break;
1499         default:
1500                 break;
1501         }
1502         return NOTIFY_OK;
1503 }
1504
1505 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1506         .notifier_call  = xen_hvm_cpu_notify,
1507 };
1508
1509 static void __init xen_hvm_guest_init(void)
1510 {
1511         int r;
1512         int major, minor;
1513
1514         r = init_hvm_pv_info(&major, &minor);
1515         if (r < 0)
1516                 return;
1517
1518         xen_hvm_init_shared_info();
1519
1520         if (xen_feature(XENFEAT_hvm_callback_vector))
1521                 xen_have_vector_callback = 1;
1522         xen_hvm_smp_init();
1523         register_cpu_notifier(&xen_hvm_cpu_notifier);
1524         xen_unplug_emulated_devices();
1525         x86_init.irqs.intr_init = xen_init_IRQ;
1526         xen_hvm_init_time_ops();
1527         xen_hvm_init_mmu_ops();
1528 }
1529
1530 static bool __init xen_hvm_platform(void)
1531 {
1532         if (xen_pv_domain())
1533                 return false;
1534
1535         if (!xen_cpuid_base())
1536                 return false;
1537
1538         return true;
1539 }
1540
1541 bool xen_hvm_need_lapic(void)
1542 {
1543         if (xen_pv_domain())
1544                 return false;
1545         if (!xen_hvm_domain())
1546                 return false;
1547         if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1548                 return false;
1549         return true;
1550 }
1551 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1552
1553 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1554         .name                   = "Xen HVM",
1555         .detect                 = xen_hvm_platform,
1556         .init_platform          = xen_hvm_guest_init,
1557 };
1558 EXPORT_SYMBOL(x86_hyper_xen_hvm);
1559 #endif