Merge branch 'x86-apic-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[~shefty/rdma-dev.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/events.h>
37 #include <xen/interface/xen.h>
38 #include <xen/interface/version.h>
39 #include <xen/interface/physdev.h>
40 #include <xen/interface/vcpu.h>
41 #include <xen/interface/memory.h>
42 #include <xen/interface/xen-mca.h>
43 #include <xen/features.h>
44 #include <xen/page.h>
45 #include <xen/hvm.h>
46 #include <xen/hvc-console.h>
47 #include <xen/acpi.h>
48
49 #include <asm/paravirt.h>
50 #include <asm/apic.h>
51 #include <asm/page.h>
52 #include <asm/xen/pci.h>
53 #include <asm/xen/hypercall.h>
54 #include <asm/xen/hypervisor.h>
55 #include <asm/fixmap.h>
56 #include <asm/processor.h>
57 #include <asm/proto.h>
58 #include <asm/msr-index.h>
59 #include <asm/traps.h>
60 #include <asm/setup.h>
61 #include <asm/desc.h>
62 #include <asm/pgalloc.h>
63 #include <asm/pgtable.h>
64 #include <asm/tlbflush.h>
65 #include <asm/reboot.h>
66 #include <asm/stackprotector.h>
67 #include <asm/hypervisor.h>
68 #include <asm/mwait.h>
69 #include <asm/pci_x86.h>
70
71 #ifdef CONFIG_ACPI
72 #include <linux/acpi.h>
73 #include <asm/acpi.h>
74 #include <acpi/pdc_intel.h>
75 #include <acpi/processor.h>
76 #include <xen/interface/platform.h>
77 #endif
78
79 #include "xen-ops.h"
80 #include "mmu.h"
81 #include "smp.h"
82 #include "multicalls.h"
83
84 EXPORT_SYMBOL_GPL(hypercall_page);
85
86 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
87 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
88
89 enum xen_domain_type xen_domain_type = XEN_NATIVE;
90 EXPORT_SYMBOL_GPL(xen_domain_type);
91
92 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
93 EXPORT_SYMBOL(machine_to_phys_mapping);
94 unsigned long  machine_to_phys_nr;
95 EXPORT_SYMBOL(machine_to_phys_nr);
96
97 struct start_info *xen_start_info;
98 EXPORT_SYMBOL_GPL(xen_start_info);
99
100 struct shared_info xen_dummy_shared_info;
101
102 void *xen_initial_gdt;
103
104 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
105 __read_mostly int xen_have_vector_callback;
106 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
107
108 /*
109  * Point at some empty memory to start with. We map the real shared_info
110  * page as soon as fixmap is up and running.
111  */
112 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
113
114 /*
115  * Flag to determine whether vcpu info placement is available on all
116  * VCPUs.  We assume it is to start with, and then set it to zero on
117  * the first failure.  This is because it can succeed on some VCPUs
118  * and not others, since it can involve hypervisor memory allocation,
119  * or because the guest failed to guarantee all the appropriate
120  * constraints on all VCPUs (ie buffer can't cross a page boundary).
121  *
122  * Note that any particular CPU may be using a placed vcpu structure,
123  * but we can only optimise if the all are.
124  *
125  * 0: not available, 1: available
126  */
127 static int have_vcpu_info_placement = 1;
128
129 struct tls_descs {
130         struct desc_struct desc[3];
131 };
132
133 /*
134  * Updating the 3 TLS descriptors in the GDT on every task switch is
135  * surprisingly expensive so we avoid updating them if they haven't
136  * changed.  Since Xen writes different descriptors than the one
137  * passed in the update_descriptor hypercall we keep shadow copies to
138  * compare against.
139  */
140 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
141
142 static void clamp_max_cpus(void)
143 {
144 #ifdef CONFIG_SMP
145         if (setup_max_cpus > MAX_VIRT_CPUS)
146                 setup_max_cpus = MAX_VIRT_CPUS;
147 #endif
148 }
149
150 static void xen_vcpu_setup(int cpu)
151 {
152         struct vcpu_register_vcpu_info info;
153         int err;
154         struct vcpu_info *vcpup;
155
156         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
157
158         if (cpu < MAX_VIRT_CPUS)
159                 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
160
161         if (!have_vcpu_info_placement) {
162                 if (cpu >= MAX_VIRT_CPUS)
163                         clamp_max_cpus();
164                 return;
165         }
166
167         vcpup = &per_cpu(xen_vcpu_info, cpu);
168         info.mfn = arbitrary_virt_to_mfn(vcpup);
169         info.offset = offset_in_page(vcpup);
170
171         /* Check to see if the hypervisor will put the vcpu_info
172            structure where we want it, which allows direct access via
173            a percpu-variable. */
174         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
175
176         if (err) {
177                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
178                 have_vcpu_info_placement = 0;
179                 clamp_max_cpus();
180         } else {
181                 /* This cpu is using the registered vcpu info, even if
182                    later ones fail to. */
183                 per_cpu(xen_vcpu, cpu) = vcpup;
184         }
185 }
186
187 /*
188  * On restore, set the vcpu placement up again.
189  * If it fails, then we're in a bad state, since
190  * we can't back out from using it...
191  */
192 void xen_vcpu_restore(void)
193 {
194         int cpu;
195
196         for_each_possible_cpu(cpu) {
197                 bool other_cpu = (cpu != smp_processor_id());
198                 bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL);
199
200                 if (other_cpu && is_up &&
201                     HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
202                         BUG();
203
204                 xen_setup_runstate_info(cpu);
205
206                 if (have_vcpu_info_placement)
207                         xen_vcpu_setup(cpu);
208
209                 if (other_cpu && is_up &&
210                     HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
211                         BUG();
212         }
213 }
214
215 static void __init xen_banner(void)
216 {
217         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
218         struct xen_extraversion extra;
219         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
220
221         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
222                pv_info.name);
223         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
224                version >> 16, version & 0xffff, extra.extraversion,
225                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
226 }
227 /* Check if running on Xen version (major, minor) or later */
228 bool
229 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
230 {
231         unsigned int version;
232
233         if (!xen_domain())
234                 return false;
235
236         version = HYPERVISOR_xen_version(XENVER_version, NULL);
237         if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
238                 ((version >> 16) > major))
239                 return true;
240         return false;
241 }
242
243 #define CPUID_THERM_POWER_LEAF 6
244 #define APERFMPERF_PRESENT 0
245
246 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
247 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
248
249 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
250 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
251 static __read_mostly unsigned int cpuid_leaf5_edx_val;
252
253 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
254                       unsigned int *cx, unsigned int *dx)
255 {
256         unsigned maskebx = ~0;
257         unsigned maskecx = ~0;
258         unsigned maskedx = ~0;
259         unsigned setecx = 0;
260         /*
261          * Mask out inconvenient features, to try and disable as many
262          * unsupported kernel subsystems as possible.
263          */
264         switch (*ax) {
265         case 1:
266                 maskecx = cpuid_leaf1_ecx_mask;
267                 setecx = cpuid_leaf1_ecx_set_mask;
268                 maskedx = cpuid_leaf1_edx_mask;
269                 break;
270
271         case CPUID_MWAIT_LEAF:
272                 /* Synthesize the values.. */
273                 *ax = 0;
274                 *bx = 0;
275                 *cx = cpuid_leaf5_ecx_val;
276                 *dx = cpuid_leaf5_edx_val;
277                 return;
278
279         case CPUID_THERM_POWER_LEAF:
280                 /* Disabling APERFMPERF for kernel usage */
281                 maskecx = ~(1 << APERFMPERF_PRESENT);
282                 break;
283
284         case 0xb:
285                 /* Suppress extended topology stuff */
286                 maskebx = 0;
287                 break;
288         }
289
290         asm(XEN_EMULATE_PREFIX "cpuid"
291                 : "=a" (*ax),
292                   "=b" (*bx),
293                   "=c" (*cx),
294                   "=d" (*dx)
295                 : "0" (*ax), "2" (*cx));
296
297         *bx &= maskebx;
298         *cx &= maskecx;
299         *cx |= setecx;
300         *dx &= maskedx;
301
302 }
303
304 static bool __init xen_check_mwait(void)
305 {
306 #ifdef CONFIG_ACPI
307         struct xen_platform_op op = {
308                 .cmd                    = XENPF_set_processor_pminfo,
309                 .u.set_pminfo.id        = -1,
310                 .u.set_pminfo.type      = XEN_PM_PDC,
311         };
312         uint32_t buf[3];
313         unsigned int ax, bx, cx, dx;
314         unsigned int mwait_mask;
315
316         /* We need to determine whether it is OK to expose the MWAIT
317          * capability to the kernel to harvest deeper than C3 states from ACPI
318          * _CST using the processor_harvest_xen.c module. For this to work, we
319          * need to gather the MWAIT_LEAF values (which the cstate.c code
320          * checks against). The hypervisor won't expose the MWAIT flag because
321          * it would break backwards compatibility; so we will find out directly
322          * from the hardware and hypercall.
323          */
324         if (!xen_initial_domain())
325                 return false;
326
327         /*
328          * When running under platform earlier than Xen4.2, do not expose
329          * mwait, to avoid the risk of loading native acpi pad driver
330          */
331         if (!xen_running_on_version_or_later(4, 2))
332                 return false;
333
334         ax = 1;
335         cx = 0;
336
337         native_cpuid(&ax, &bx, &cx, &dx);
338
339         mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
340                      (1 << (X86_FEATURE_MWAIT % 32));
341
342         if ((cx & mwait_mask) != mwait_mask)
343                 return false;
344
345         /* We need to emulate the MWAIT_LEAF and for that we need both
346          * ecx and edx. The hypercall provides only partial information.
347          */
348
349         ax = CPUID_MWAIT_LEAF;
350         bx = 0;
351         cx = 0;
352         dx = 0;
353
354         native_cpuid(&ax, &bx, &cx, &dx);
355
356         /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
357          * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
358          */
359         buf[0] = ACPI_PDC_REVISION_ID;
360         buf[1] = 1;
361         buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
362
363         set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
364
365         if ((HYPERVISOR_dom0_op(&op) == 0) &&
366             (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
367                 cpuid_leaf5_ecx_val = cx;
368                 cpuid_leaf5_edx_val = dx;
369         }
370         return true;
371 #else
372         return false;
373 #endif
374 }
375 static void __init xen_init_cpuid_mask(void)
376 {
377         unsigned int ax, bx, cx, dx;
378         unsigned int xsave_mask;
379
380         cpuid_leaf1_edx_mask =
381                 ~((1 << X86_FEATURE_MTRR) |  /* disable MTRR */
382                   (1 << X86_FEATURE_ACC));   /* thermal monitoring */
383
384         if (!xen_initial_domain())
385                 cpuid_leaf1_edx_mask &=
386                         ~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
387                           (1 << X86_FEATURE_ACPI));  /* disable ACPI */
388         ax = 1;
389         cx = 0;
390         xen_cpuid(&ax, &bx, &cx, &dx);
391
392         xsave_mask =
393                 (1 << (X86_FEATURE_XSAVE % 32)) |
394                 (1 << (X86_FEATURE_OSXSAVE % 32));
395
396         /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
397         if ((cx & xsave_mask) != xsave_mask)
398                 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
399         if (xen_check_mwait())
400                 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
401 }
402
403 static void xen_set_debugreg(int reg, unsigned long val)
404 {
405         HYPERVISOR_set_debugreg(reg, val);
406 }
407
408 static unsigned long xen_get_debugreg(int reg)
409 {
410         return HYPERVISOR_get_debugreg(reg);
411 }
412
413 static void xen_end_context_switch(struct task_struct *next)
414 {
415         xen_mc_flush();
416         paravirt_end_context_switch(next);
417 }
418
419 static unsigned long xen_store_tr(void)
420 {
421         return 0;
422 }
423
424 /*
425  * Set the page permissions for a particular virtual address.  If the
426  * address is a vmalloc mapping (or other non-linear mapping), then
427  * find the linear mapping of the page and also set its protections to
428  * match.
429  */
430 static void set_aliased_prot(void *v, pgprot_t prot)
431 {
432         int level;
433         pte_t *ptep;
434         pte_t pte;
435         unsigned long pfn;
436         struct page *page;
437
438         ptep = lookup_address((unsigned long)v, &level);
439         BUG_ON(ptep == NULL);
440
441         pfn = pte_pfn(*ptep);
442         page = pfn_to_page(pfn);
443
444         pte = pfn_pte(pfn, prot);
445
446         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
447                 BUG();
448
449         if (!PageHighMem(page)) {
450                 void *av = __va(PFN_PHYS(pfn));
451
452                 if (av != v)
453                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
454                                 BUG();
455         } else
456                 kmap_flush_unused();
457 }
458
459 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
460 {
461         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
462         int i;
463
464         for(i = 0; i < entries; i += entries_per_page)
465                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
466 }
467
468 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
469 {
470         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
471         int i;
472
473         for(i = 0; i < entries; i += entries_per_page)
474                 set_aliased_prot(ldt + i, PAGE_KERNEL);
475 }
476
477 static void xen_set_ldt(const void *addr, unsigned entries)
478 {
479         struct mmuext_op *op;
480         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
481
482         trace_xen_cpu_set_ldt(addr, entries);
483
484         op = mcs.args;
485         op->cmd = MMUEXT_SET_LDT;
486         op->arg1.linear_addr = (unsigned long)addr;
487         op->arg2.nr_ents = entries;
488
489         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
490
491         xen_mc_issue(PARAVIRT_LAZY_CPU);
492 }
493
494 static void xen_load_gdt(const struct desc_ptr *dtr)
495 {
496         unsigned long va = dtr->address;
497         unsigned int size = dtr->size + 1;
498         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
499         unsigned long frames[pages];
500         int f;
501
502         /*
503          * A GDT can be up to 64k in size, which corresponds to 8192
504          * 8-byte entries, or 16 4k pages..
505          */
506
507         BUG_ON(size > 65536);
508         BUG_ON(va & ~PAGE_MASK);
509
510         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
511                 int level;
512                 pte_t *ptep;
513                 unsigned long pfn, mfn;
514                 void *virt;
515
516                 /*
517                  * The GDT is per-cpu and is in the percpu data area.
518                  * That can be virtually mapped, so we need to do a
519                  * page-walk to get the underlying MFN for the
520                  * hypercall.  The page can also be in the kernel's
521                  * linear range, so we need to RO that mapping too.
522                  */
523                 ptep = lookup_address(va, &level);
524                 BUG_ON(ptep == NULL);
525
526                 pfn = pte_pfn(*ptep);
527                 mfn = pfn_to_mfn(pfn);
528                 virt = __va(PFN_PHYS(pfn));
529
530                 frames[f] = mfn;
531
532                 make_lowmem_page_readonly((void *)va);
533                 make_lowmem_page_readonly(virt);
534         }
535
536         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
537                 BUG();
538 }
539
540 /*
541  * load_gdt for early boot, when the gdt is only mapped once
542  */
543 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
544 {
545         unsigned long va = dtr->address;
546         unsigned int size = dtr->size + 1;
547         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
548         unsigned long frames[pages];
549         int f;
550
551         /*
552          * A GDT can be up to 64k in size, which corresponds to 8192
553          * 8-byte entries, or 16 4k pages..
554          */
555
556         BUG_ON(size > 65536);
557         BUG_ON(va & ~PAGE_MASK);
558
559         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
560                 pte_t pte;
561                 unsigned long pfn, mfn;
562
563                 pfn = virt_to_pfn(va);
564                 mfn = pfn_to_mfn(pfn);
565
566                 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
567
568                 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
569                         BUG();
570
571                 frames[f] = mfn;
572         }
573
574         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
575                 BUG();
576 }
577
578 static inline bool desc_equal(const struct desc_struct *d1,
579                               const struct desc_struct *d2)
580 {
581         return d1->a == d2->a && d1->b == d2->b;
582 }
583
584 static void load_TLS_descriptor(struct thread_struct *t,
585                                 unsigned int cpu, unsigned int i)
586 {
587         struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
588         struct desc_struct *gdt;
589         xmaddr_t maddr;
590         struct multicall_space mc;
591
592         if (desc_equal(shadow, &t->tls_array[i]))
593                 return;
594
595         *shadow = t->tls_array[i];
596
597         gdt = get_cpu_gdt_table(cpu);
598         maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
599         mc = __xen_mc_entry(0);
600
601         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
602 }
603
604 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
605 {
606         /*
607          * XXX sleazy hack: If we're being called in a lazy-cpu zone
608          * and lazy gs handling is enabled, it means we're in a
609          * context switch, and %gs has just been saved.  This means we
610          * can zero it out to prevent faults on exit from the
611          * hypervisor if the next process has no %gs.  Either way, it
612          * has been saved, and the new value will get loaded properly.
613          * This will go away as soon as Xen has been modified to not
614          * save/restore %gs for normal hypercalls.
615          *
616          * On x86_64, this hack is not used for %gs, because gs points
617          * to KERNEL_GS_BASE (and uses it for PDA references), so we
618          * must not zero %gs on x86_64
619          *
620          * For x86_64, we need to zero %fs, otherwise we may get an
621          * exception between the new %fs descriptor being loaded and
622          * %fs being effectively cleared at __switch_to().
623          */
624         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
625 #ifdef CONFIG_X86_32
626                 lazy_load_gs(0);
627 #else
628                 loadsegment(fs, 0);
629 #endif
630         }
631
632         xen_mc_batch();
633
634         load_TLS_descriptor(t, cpu, 0);
635         load_TLS_descriptor(t, cpu, 1);
636         load_TLS_descriptor(t, cpu, 2);
637
638         xen_mc_issue(PARAVIRT_LAZY_CPU);
639 }
640
641 #ifdef CONFIG_X86_64
642 static void xen_load_gs_index(unsigned int idx)
643 {
644         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
645                 BUG();
646 }
647 #endif
648
649 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
650                                 const void *ptr)
651 {
652         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
653         u64 entry = *(u64 *)ptr;
654
655         trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
656
657         preempt_disable();
658
659         xen_mc_flush();
660         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
661                 BUG();
662
663         preempt_enable();
664 }
665
666 static int cvt_gate_to_trap(int vector, const gate_desc *val,
667                             struct trap_info *info)
668 {
669         unsigned long addr;
670
671         if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
672                 return 0;
673
674         info->vector = vector;
675
676         addr = gate_offset(*val);
677 #ifdef CONFIG_X86_64
678         /*
679          * Look for known traps using IST, and substitute them
680          * appropriately.  The debugger ones are the only ones we care
681          * about.  Xen will handle faults like double_fault,
682          * so we should never see them.  Warn if
683          * there's an unexpected IST-using fault handler.
684          */
685         if (addr == (unsigned long)debug)
686                 addr = (unsigned long)xen_debug;
687         else if (addr == (unsigned long)int3)
688                 addr = (unsigned long)xen_int3;
689         else if (addr == (unsigned long)stack_segment)
690                 addr = (unsigned long)xen_stack_segment;
691         else if (addr == (unsigned long)double_fault ||
692                  addr == (unsigned long)nmi) {
693                 /* Don't need to handle these */
694                 return 0;
695 #ifdef CONFIG_X86_MCE
696         } else if (addr == (unsigned long)machine_check) {
697                 /*
698                  * when xen hypervisor inject vMCE to guest,
699                  * use native mce handler to handle it
700                  */
701                 ;
702 #endif
703         } else {
704                 /* Some other trap using IST? */
705                 if (WARN_ON(val->ist != 0))
706                         return 0;
707         }
708 #endif  /* CONFIG_X86_64 */
709         info->address = addr;
710
711         info->cs = gate_segment(*val);
712         info->flags = val->dpl;
713         /* interrupt gates clear IF */
714         if (val->type == GATE_INTERRUPT)
715                 info->flags |= 1 << 2;
716
717         return 1;
718 }
719
720 /* Locations of each CPU's IDT */
721 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
722
723 /* Set an IDT entry.  If the entry is part of the current IDT, then
724    also update Xen. */
725 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
726 {
727         unsigned long p = (unsigned long)&dt[entrynum];
728         unsigned long start, end;
729
730         trace_xen_cpu_write_idt_entry(dt, entrynum, g);
731
732         preempt_disable();
733
734         start = __this_cpu_read(idt_desc.address);
735         end = start + __this_cpu_read(idt_desc.size) + 1;
736
737         xen_mc_flush();
738
739         native_write_idt_entry(dt, entrynum, g);
740
741         if (p >= start && (p + 8) <= end) {
742                 struct trap_info info[2];
743
744                 info[1].address = 0;
745
746                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
747                         if (HYPERVISOR_set_trap_table(info))
748                                 BUG();
749         }
750
751         preempt_enable();
752 }
753
754 static void xen_convert_trap_info(const struct desc_ptr *desc,
755                                   struct trap_info *traps)
756 {
757         unsigned in, out, count;
758
759         count = (desc->size+1) / sizeof(gate_desc);
760         BUG_ON(count > 256);
761
762         for (in = out = 0; in < count; in++) {
763                 gate_desc *entry = (gate_desc*)(desc->address) + in;
764
765                 if (cvt_gate_to_trap(in, entry, &traps[out]))
766                         out++;
767         }
768         traps[out].address = 0;
769 }
770
771 void xen_copy_trap_info(struct trap_info *traps)
772 {
773         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
774
775         xen_convert_trap_info(desc, traps);
776 }
777
778 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
779    hold a spinlock to protect the static traps[] array (static because
780    it avoids allocation, and saves stack space). */
781 static void xen_load_idt(const struct desc_ptr *desc)
782 {
783         static DEFINE_SPINLOCK(lock);
784         static struct trap_info traps[257];
785
786         trace_xen_cpu_load_idt(desc);
787
788         spin_lock(&lock);
789
790         __get_cpu_var(idt_desc) = *desc;
791
792         xen_convert_trap_info(desc, traps);
793
794         xen_mc_flush();
795         if (HYPERVISOR_set_trap_table(traps))
796                 BUG();
797
798         spin_unlock(&lock);
799 }
800
801 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
802    they're handled differently. */
803 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
804                                 const void *desc, int type)
805 {
806         trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
807
808         preempt_disable();
809
810         switch (type) {
811         case DESC_LDT:
812         case DESC_TSS:
813                 /* ignore */
814                 break;
815
816         default: {
817                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
818
819                 xen_mc_flush();
820                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
821                         BUG();
822         }
823
824         }
825
826         preempt_enable();
827 }
828
829 /*
830  * Version of write_gdt_entry for use at early boot-time needed to
831  * update an entry as simply as possible.
832  */
833 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
834                                             const void *desc, int type)
835 {
836         trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
837
838         switch (type) {
839         case DESC_LDT:
840         case DESC_TSS:
841                 /* ignore */
842                 break;
843
844         default: {
845                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
846
847                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
848                         dt[entry] = *(struct desc_struct *)desc;
849         }
850
851         }
852 }
853
854 static void xen_load_sp0(struct tss_struct *tss,
855                          struct thread_struct *thread)
856 {
857         struct multicall_space mcs;
858
859         mcs = xen_mc_entry(0);
860         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
861         xen_mc_issue(PARAVIRT_LAZY_CPU);
862 }
863
864 static void xen_set_iopl_mask(unsigned mask)
865 {
866         struct physdev_set_iopl set_iopl;
867
868         /* Force the change at ring 0. */
869         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
870         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
871 }
872
873 static void xen_io_delay(void)
874 {
875 }
876
877 #ifdef CONFIG_X86_LOCAL_APIC
878 static unsigned long xen_set_apic_id(unsigned int x)
879 {
880         WARN_ON(1);
881         return x;
882 }
883 static unsigned int xen_get_apic_id(unsigned long x)
884 {
885         return ((x)>>24) & 0xFFu;
886 }
887 static u32 xen_apic_read(u32 reg)
888 {
889         struct xen_platform_op op = {
890                 .cmd = XENPF_get_cpuinfo,
891                 .interface_version = XENPF_INTERFACE_VERSION,
892                 .u.pcpu_info.xen_cpuid = 0,
893         };
894         int ret = 0;
895
896         /* Shouldn't need this as APIC is turned off for PV, and we only
897          * get called on the bootup processor. But just in case. */
898         if (!xen_initial_domain() || smp_processor_id())
899                 return 0;
900
901         if (reg == APIC_LVR)
902                 return 0x10;
903
904         if (reg != APIC_ID)
905                 return 0;
906
907         ret = HYPERVISOR_dom0_op(&op);
908         if (ret)
909                 return 0;
910
911         return op.u.pcpu_info.apic_id << 24;
912 }
913
914 static void xen_apic_write(u32 reg, u32 val)
915 {
916         /* Warn to see if there's any stray references */
917         WARN_ON(1);
918 }
919
920 static u64 xen_apic_icr_read(void)
921 {
922         return 0;
923 }
924
925 static void xen_apic_icr_write(u32 low, u32 id)
926 {
927         /* Warn to see if there's any stray references */
928         WARN_ON(1);
929 }
930
931 static void xen_apic_wait_icr_idle(void)
932 {
933         return;
934 }
935
936 static u32 xen_safe_apic_wait_icr_idle(void)
937 {
938         return 0;
939 }
940
941 static void set_xen_basic_apic_ops(void)
942 {
943         apic->read = xen_apic_read;
944         apic->write = xen_apic_write;
945         apic->icr_read = xen_apic_icr_read;
946         apic->icr_write = xen_apic_icr_write;
947         apic->wait_icr_idle = xen_apic_wait_icr_idle;
948         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
949         apic->set_apic_id = xen_set_apic_id;
950         apic->get_apic_id = xen_get_apic_id;
951
952 #ifdef CONFIG_SMP
953         apic->send_IPI_allbutself = xen_send_IPI_allbutself;
954         apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
955         apic->send_IPI_mask = xen_send_IPI_mask;
956         apic->send_IPI_all = xen_send_IPI_all;
957         apic->send_IPI_self = xen_send_IPI_self;
958 #endif
959 }
960
961 #endif
962
963 static void xen_clts(void)
964 {
965         struct multicall_space mcs;
966
967         mcs = xen_mc_entry(0);
968
969         MULTI_fpu_taskswitch(mcs.mc, 0);
970
971         xen_mc_issue(PARAVIRT_LAZY_CPU);
972 }
973
974 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
975
976 static unsigned long xen_read_cr0(void)
977 {
978         unsigned long cr0 = this_cpu_read(xen_cr0_value);
979
980         if (unlikely(cr0 == 0)) {
981                 cr0 = native_read_cr0();
982                 this_cpu_write(xen_cr0_value, cr0);
983         }
984
985         return cr0;
986 }
987
988 static void xen_write_cr0(unsigned long cr0)
989 {
990         struct multicall_space mcs;
991
992         this_cpu_write(xen_cr0_value, cr0);
993
994         /* Only pay attention to cr0.TS; everything else is
995            ignored. */
996         mcs = xen_mc_entry(0);
997
998         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
999
1000         xen_mc_issue(PARAVIRT_LAZY_CPU);
1001 }
1002
1003 static void xen_write_cr4(unsigned long cr4)
1004 {
1005         cr4 &= ~X86_CR4_PGE;
1006         cr4 &= ~X86_CR4_PSE;
1007
1008         native_write_cr4(cr4);
1009 }
1010 #ifdef CONFIG_X86_64
1011 static inline unsigned long xen_read_cr8(void)
1012 {
1013         return 0;
1014 }
1015 static inline void xen_write_cr8(unsigned long val)
1016 {
1017         BUG_ON(val);
1018 }
1019 #endif
1020 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
1021 {
1022         int ret;
1023
1024         ret = 0;
1025
1026         switch (msr) {
1027 #ifdef CONFIG_X86_64
1028                 unsigned which;
1029                 u64 base;
1030
1031         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
1032         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
1033         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
1034
1035         set:
1036                 base = ((u64)high << 32) | low;
1037                 if (HYPERVISOR_set_segment_base(which, base) != 0)
1038                         ret = -EIO;
1039                 break;
1040 #endif
1041
1042         case MSR_STAR:
1043         case MSR_CSTAR:
1044         case MSR_LSTAR:
1045         case MSR_SYSCALL_MASK:
1046         case MSR_IA32_SYSENTER_CS:
1047         case MSR_IA32_SYSENTER_ESP:
1048         case MSR_IA32_SYSENTER_EIP:
1049                 /* Fast syscall setup is all done in hypercalls, so
1050                    these are all ignored.  Stub them out here to stop
1051                    Xen console noise. */
1052                 break;
1053
1054         case MSR_IA32_CR_PAT:
1055                 if (smp_processor_id() == 0)
1056                         xen_set_pat(((u64)high << 32) | low);
1057                 break;
1058
1059         default:
1060                 ret = native_write_msr_safe(msr, low, high);
1061         }
1062
1063         return ret;
1064 }
1065
1066 void xen_setup_shared_info(void)
1067 {
1068         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1069                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1070                            xen_start_info->shared_info);
1071
1072                 HYPERVISOR_shared_info =
1073                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1074         } else
1075                 HYPERVISOR_shared_info =
1076                         (struct shared_info *)__va(xen_start_info->shared_info);
1077
1078 #ifndef CONFIG_SMP
1079         /* In UP this is as good a place as any to set up shared info */
1080         xen_setup_vcpu_info_placement();
1081 #endif
1082
1083         xen_setup_mfn_list_list();
1084 }
1085
1086 /* This is called once we have the cpu_possible_mask */
1087 void xen_setup_vcpu_info_placement(void)
1088 {
1089         int cpu;
1090
1091         for_each_possible_cpu(cpu)
1092                 xen_vcpu_setup(cpu);
1093
1094         /* xen_vcpu_setup managed to place the vcpu_info within the
1095            percpu area for all cpus, so make use of it */
1096         if (have_vcpu_info_placement) {
1097                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1098                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1099                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1100                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1101                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1102         }
1103 }
1104
1105 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1106                           unsigned long addr, unsigned len)
1107 {
1108         char *start, *end, *reloc;
1109         unsigned ret;
1110
1111         start = end = reloc = NULL;
1112
1113 #define SITE(op, x)                                                     \
1114         case PARAVIRT_PATCH(op.x):                                      \
1115         if (have_vcpu_info_placement) {                                 \
1116                 start = (char *)xen_##x##_direct;                       \
1117                 end = xen_##x##_direct_end;                             \
1118                 reloc = xen_##x##_direct_reloc;                         \
1119         }                                                               \
1120         goto patch_site
1121
1122         switch (type) {
1123                 SITE(pv_irq_ops, irq_enable);
1124                 SITE(pv_irq_ops, irq_disable);
1125                 SITE(pv_irq_ops, save_fl);
1126                 SITE(pv_irq_ops, restore_fl);
1127 #undef SITE
1128
1129         patch_site:
1130                 if (start == NULL || (end-start) > len)
1131                         goto default_patch;
1132
1133                 ret = paravirt_patch_insns(insnbuf, len, start, end);
1134
1135                 /* Note: because reloc is assigned from something that
1136                    appears to be an array, gcc assumes it's non-null,
1137                    but doesn't know its relationship with start and
1138                    end. */
1139                 if (reloc > start && reloc < end) {
1140                         int reloc_off = reloc - start;
1141                         long *relocp = (long *)(insnbuf + reloc_off);
1142                         long delta = start - (char *)addr;
1143
1144                         *relocp += delta;
1145                 }
1146                 break;
1147
1148         default_patch:
1149         default:
1150                 ret = paravirt_patch_default(type, clobbers, insnbuf,
1151                                              addr, len);
1152                 break;
1153         }
1154
1155         return ret;
1156 }
1157
1158 static const struct pv_info xen_info __initconst = {
1159         .paravirt_enabled = 1,
1160         .shared_kernel_pmd = 0,
1161
1162 #ifdef CONFIG_X86_64
1163         .extra_user_64bit_cs = FLAT_USER_CS64,
1164 #endif
1165
1166         .name = "Xen",
1167 };
1168
1169 static const struct pv_init_ops xen_init_ops __initconst = {
1170         .patch = xen_patch,
1171 };
1172
1173 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1174         .cpuid = xen_cpuid,
1175
1176         .set_debugreg = xen_set_debugreg,
1177         .get_debugreg = xen_get_debugreg,
1178
1179         .clts = xen_clts,
1180
1181         .read_cr0 = xen_read_cr0,
1182         .write_cr0 = xen_write_cr0,
1183
1184         .read_cr4 = native_read_cr4,
1185         .read_cr4_safe = native_read_cr4_safe,
1186         .write_cr4 = xen_write_cr4,
1187
1188 #ifdef CONFIG_X86_64
1189         .read_cr8 = xen_read_cr8,
1190         .write_cr8 = xen_write_cr8,
1191 #endif
1192
1193         .wbinvd = native_wbinvd,
1194
1195         .read_msr = native_read_msr_safe,
1196         .write_msr = xen_write_msr_safe,
1197
1198         .read_tsc = native_read_tsc,
1199         .read_pmc = native_read_pmc,
1200
1201         .read_tscp = native_read_tscp,
1202
1203         .iret = xen_iret,
1204         .irq_enable_sysexit = xen_sysexit,
1205 #ifdef CONFIG_X86_64
1206         .usergs_sysret32 = xen_sysret32,
1207         .usergs_sysret64 = xen_sysret64,
1208 #endif
1209
1210         .load_tr_desc = paravirt_nop,
1211         .set_ldt = xen_set_ldt,
1212         .load_gdt = xen_load_gdt,
1213         .load_idt = xen_load_idt,
1214         .load_tls = xen_load_tls,
1215 #ifdef CONFIG_X86_64
1216         .load_gs_index = xen_load_gs_index,
1217 #endif
1218
1219         .alloc_ldt = xen_alloc_ldt,
1220         .free_ldt = xen_free_ldt,
1221
1222         .store_gdt = native_store_gdt,
1223         .store_idt = native_store_idt,
1224         .store_tr = xen_store_tr,
1225
1226         .write_ldt_entry = xen_write_ldt_entry,
1227         .write_gdt_entry = xen_write_gdt_entry,
1228         .write_idt_entry = xen_write_idt_entry,
1229         .load_sp0 = xen_load_sp0,
1230
1231         .set_iopl_mask = xen_set_iopl_mask,
1232         .io_delay = xen_io_delay,
1233
1234         /* Xen takes care of %gs when switching to usermode for us */
1235         .swapgs = paravirt_nop,
1236
1237         .start_context_switch = paravirt_start_context_switch,
1238         .end_context_switch = xen_end_context_switch,
1239 };
1240
1241 static const struct pv_apic_ops xen_apic_ops __initconst = {
1242 #ifdef CONFIG_X86_LOCAL_APIC
1243         .startup_ipi_hook = paravirt_nop,
1244 #endif
1245 };
1246
1247 static void xen_reboot(int reason)
1248 {
1249         struct sched_shutdown r = { .reason = reason };
1250
1251         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1252                 BUG();
1253 }
1254
1255 static void xen_restart(char *msg)
1256 {
1257         xen_reboot(SHUTDOWN_reboot);
1258 }
1259
1260 static void xen_emergency_restart(void)
1261 {
1262         xen_reboot(SHUTDOWN_reboot);
1263 }
1264
1265 static void xen_machine_halt(void)
1266 {
1267         xen_reboot(SHUTDOWN_poweroff);
1268 }
1269
1270 static void xen_machine_power_off(void)
1271 {
1272         if (pm_power_off)
1273                 pm_power_off();
1274         xen_reboot(SHUTDOWN_poweroff);
1275 }
1276
1277 static void xen_crash_shutdown(struct pt_regs *regs)
1278 {
1279         xen_reboot(SHUTDOWN_crash);
1280 }
1281
1282 static int
1283 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1284 {
1285         xen_reboot(SHUTDOWN_crash);
1286         return NOTIFY_DONE;
1287 }
1288
1289 static struct notifier_block xen_panic_block = {
1290         .notifier_call= xen_panic_event,
1291 };
1292
1293 int xen_panic_handler_init(void)
1294 {
1295         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1296         return 0;
1297 }
1298
1299 static const struct machine_ops xen_machine_ops __initconst = {
1300         .restart = xen_restart,
1301         .halt = xen_machine_halt,
1302         .power_off = xen_machine_power_off,
1303         .shutdown = xen_machine_halt,
1304         .crash_shutdown = xen_crash_shutdown,
1305         .emergency_restart = xen_emergency_restart,
1306 };
1307
1308 /*
1309  * Set up the GDT and segment registers for -fstack-protector.  Until
1310  * we do this, we have to be careful not to call any stack-protected
1311  * function, which is most of the kernel.
1312  */
1313 static void __init xen_setup_stackprotector(void)
1314 {
1315         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1316         pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1317
1318         setup_stack_canary_segment(0);
1319         switch_to_new_gdt(0);
1320
1321         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1322         pv_cpu_ops.load_gdt = xen_load_gdt;
1323 }
1324
1325 /* First C function to be called on Xen boot */
1326 asmlinkage void __init xen_start_kernel(void)
1327 {
1328         struct physdev_set_iopl set_iopl;
1329         int rc;
1330
1331         if (!xen_start_info)
1332                 return;
1333
1334         xen_domain_type = XEN_PV_DOMAIN;
1335
1336         xen_setup_machphys_mapping();
1337
1338         /* Install Xen paravirt ops */
1339         pv_info = xen_info;
1340         pv_init_ops = xen_init_ops;
1341         pv_cpu_ops = xen_cpu_ops;
1342         pv_apic_ops = xen_apic_ops;
1343
1344         x86_init.resources.memory_setup = xen_memory_setup;
1345         x86_init.oem.arch_setup = xen_arch_setup;
1346         x86_init.oem.banner = xen_banner;
1347
1348         xen_init_time_ops();
1349
1350         /*
1351          * Set up some pagetable state before starting to set any ptes.
1352          */
1353
1354         xen_init_mmu_ops();
1355
1356         /* Prevent unwanted bits from being set in PTEs. */
1357         __supported_pte_mask &= ~_PAGE_GLOBAL;
1358 #if 0
1359         if (!xen_initial_domain())
1360 #endif
1361                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1362
1363         __supported_pte_mask |= _PAGE_IOMAP;
1364
1365         /*
1366          * Prevent page tables from being allocated in highmem, even
1367          * if CONFIG_HIGHPTE is enabled.
1368          */
1369         __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1370
1371         /* Work out if we support NX */
1372         x86_configure_nx();
1373
1374         xen_setup_features();
1375
1376         /* Get mfn list */
1377         if (!xen_feature(XENFEAT_auto_translated_physmap))
1378                 xen_build_dynamic_phys_to_machine();
1379
1380         /*
1381          * Set up kernel GDT and segment registers, mainly so that
1382          * -fstack-protector code can be executed.
1383          */
1384         xen_setup_stackprotector();
1385
1386         xen_init_irq_ops();
1387         xen_init_cpuid_mask();
1388
1389 #ifdef CONFIG_X86_LOCAL_APIC
1390         /*
1391          * set up the basic apic ops.
1392          */
1393         set_xen_basic_apic_ops();
1394 #endif
1395
1396         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1397                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1398                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1399         }
1400
1401         machine_ops = xen_machine_ops;
1402
1403         /*
1404          * The only reliable way to retain the initial address of the
1405          * percpu gdt_page is to remember it here, so we can go and
1406          * mark it RW later, when the initial percpu area is freed.
1407          */
1408         xen_initial_gdt = &per_cpu(gdt_page, 0);
1409
1410         xen_smp_init();
1411
1412 #ifdef CONFIG_ACPI_NUMA
1413         /*
1414          * The pages we from Xen are not related to machine pages, so
1415          * any NUMA information the kernel tries to get from ACPI will
1416          * be meaningless.  Prevent it from trying.
1417          */
1418         acpi_numa = -1;
1419 #endif
1420
1421         /* Don't do the full vcpu_info placement stuff until we have a
1422            possible map and a non-dummy shared_info. */
1423         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1424
1425         local_irq_disable();
1426         early_boot_irqs_disabled = true;
1427
1428         xen_raw_console_write("mapping kernel into physical memory\n");
1429         xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base, xen_start_info->nr_pages);
1430
1431         /* Allocate and initialize top and mid mfn levels for p2m structure */
1432         xen_build_mfn_list_list();
1433
1434         /* keep using Xen gdt for now; no urgent need to change it */
1435
1436 #ifdef CONFIG_X86_32
1437         pv_info.kernel_rpl = 1;
1438         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1439                 pv_info.kernel_rpl = 0;
1440 #else
1441         pv_info.kernel_rpl = 0;
1442 #endif
1443         /* set the limit of our address space */
1444         xen_reserve_top();
1445
1446         /* We used to do this in xen_arch_setup, but that is too late on AMD
1447          * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1448          * which pokes 0xcf8 port.
1449          */
1450         set_iopl.iopl = 1;
1451         rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1452         if (rc != 0)
1453                 xen_raw_printk("physdev_op failed %d\n", rc);
1454
1455 #ifdef CONFIG_X86_32
1456         /* set up basic CPUID stuff */
1457         cpu_detect(&new_cpu_data);
1458         new_cpu_data.hard_math = 1;
1459         new_cpu_data.wp_works_ok = 1;
1460         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1461 #endif
1462
1463         /* Poke various useful things into boot_params */
1464         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1465         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1466                 ? __pa(xen_start_info->mod_start) : 0;
1467         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1468         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1469
1470         if (!xen_initial_domain()) {
1471                 add_preferred_console("xenboot", 0, NULL);
1472                 add_preferred_console("tty", 0, NULL);
1473                 add_preferred_console("hvc", 0, NULL);
1474                 if (pci_xen)
1475                         x86_init.pci.arch_init = pci_xen_init;
1476         } else {
1477                 const struct dom0_vga_console_info *info =
1478                         (void *)((char *)xen_start_info +
1479                                  xen_start_info->console.dom0.info_off);
1480                 struct xen_platform_op op = {
1481                         .cmd = XENPF_firmware_info,
1482                         .interface_version = XENPF_INTERFACE_VERSION,
1483                         .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1484                 };
1485
1486                 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1487                 xen_start_info->console.domU.mfn = 0;
1488                 xen_start_info->console.domU.evtchn = 0;
1489
1490                 if (HYPERVISOR_dom0_op(&op) == 0)
1491                         boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1492
1493                 xen_init_apic();
1494
1495                 /* Make sure ACS will be enabled */
1496                 pci_request_acs();
1497
1498                 xen_acpi_sleep_register();
1499
1500                 /* Avoid searching for BIOS MP tables */
1501                 x86_init.mpparse.find_smp_config = x86_init_noop;
1502                 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1503         }
1504 #ifdef CONFIG_PCI
1505         /* PCI BIOS service won't work from a PV guest. */
1506         pci_probe &= ~PCI_PROBE_BIOS;
1507 #endif
1508         xen_raw_console_write("about to get started...\n");
1509
1510         xen_setup_runstate_info(0);
1511
1512         /* Start the world */
1513 #ifdef CONFIG_X86_32
1514         i386_start_kernel();
1515 #else
1516         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1517 #endif
1518 }
1519
1520 void __ref xen_hvm_init_shared_info(void)
1521 {
1522         int cpu;
1523         struct xen_add_to_physmap xatp;
1524         static struct shared_info *shared_info_page = 0;
1525
1526         if (!shared_info_page)
1527                 shared_info_page = (struct shared_info *)
1528                         extend_brk(PAGE_SIZE, PAGE_SIZE);
1529         xatp.domid = DOMID_SELF;
1530         xatp.idx = 0;
1531         xatp.space = XENMAPSPACE_shared_info;
1532         xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1533         if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1534                 BUG();
1535
1536         HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1537
1538         /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1539          * page, we use it in the event channel upcall and in some pvclock
1540          * related functions. We don't need the vcpu_info placement
1541          * optimizations because we don't use any pv_mmu or pv_irq op on
1542          * HVM.
1543          * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1544          * online but xen_hvm_init_shared_info is run at resume time too and
1545          * in that case multiple vcpus might be online. */
1546         for_each_online_cpu(cpu) {
1547                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1548         }
1549 }
1550
1551 #ifdef CONFIG_XEN_PVHVM
1552 static void __init init_hvm_pv_info(void)
1553 {
1554         int major, minor;
1555         uint32_t eax, ebx, ecx, edx, pages, msr, base;
1556         u64 pfn;
1557
1558         base = xen_cpuid_base();
1559         cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1560
1561         major = eax >> 16;
1562         minor = eax & 0xffff;
1563         printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1564
1565         cpuid(base + 2, &pages, &msr, &ecx, &edx);
1566
1567         pfn = __pa(hypercall_page);
1568         wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1569
1570         xen_setup_features();
1571
1572         pv_info.name = "Xen HVM";
1573
1574         xen_domain_type = XEN_HVM_DOMAIN;
1575 }
1576
1577 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1578                                     unsigned long action, void *hcpu)
1579 {
1580         int cpu = (long)hcpu;
1581         switch (action) {
1582         case CPU_UP_PREPARE:
1583                 xen_vcpu_setup(cpu);
1584                 if (xen_have_vector_callback)
1585                         xen_init_lock_cpu(cpu);
1586                 break;
1587         default:
1588                 break;
1589         }
1590         return NOTIFY_OK;
1591 }
1592
1593 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1594         .notifier_call  = xen_hvm_cpu_notify,
1595 };
1596
1597 static void __init xen_hvm_guest_init(void)
1598 {
1599         init_hvm_pv_info();
1600
1601         xen_hvm_init_shared_info();
1602
1603         if (xen_feature(XENFEAT_hvm_callback_vector))
1604                 xen_have_vector_callback = 1;
1605         xen_hvm_smp_init();
1606         register_cpu_notifier(&xen_hvm_cpu_notifier);
1607         xen_unplug_emulated_devices();
1608         x86_init.irqs.intr_init = xen_init_IRQ;
1609         xen_hvm_init_time_ops();
1610         xen_hvm_init_mmu_ops();
1611 }
1612
1613 static bool __init xen_hvm_platform(void)
1614 {
1615         if (xen_pv_domain())
1616                 return false;
1617
1618         if (!xen_cpuid_base())
1619                 return false;
1620
1621         return true;
1622 }
1623
1624 bool xen_hvm_need_lapic(void)
1625 {
1626         if (xen_pv_domain())
1627                 return false;
1628         if (!xen_hvm_domain())
1629                 return false;
1630         if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1631                 return false;
1632         return true;
1633 }
1634 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1635
1636 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1637         .name                   = "Xen HVM",
1638         .detect                 = xen_hvm_platform,
1639         .init_platform          = xen_hvm_guest_init,
1640         .x2apic_available       = xen_x2apic_para_available,
1641 };
1642 EXPORT_SYMBOL(x86_hyper_xen_hvm);
1643 #endif