Merge branch 'stable/bug.fixes' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6.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
46 #include <asm/paravirt.h>
47 #include <asm/apic.h>
48 #include <asm/page.h>
49 #include <asm/xen/pci.h>
50 #include <asm/xen/hypercall.h>
51 #include <asm/xen/hypervisor.h>
52 #include <asm/fixmap.h>
53 #include <asm/processor.h>
54 #include <asm/proto.h>
55 #include <asm/msr-index.h>
56 #include <asm/traps.h>
57 #include <asm/setup.h>
58 #include <asm/desc.h>
59 #include <asm/pgalloc.h>
60 #include <asm/pgtable.h>
61 #include <asm/tlbflush.h>
62 #include <asm/reboot.h>
63 #include <asm/stackprotector.h>
64 #include <asm/hypervisor.h>
65
66 #include "xen-ops.h"
67 #include "mmu.h"
68 #include "multicalls.h"
69
70 EXPORT_SYMBOL_GPL(hypercall_page);
71
72 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
73 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
74
75 enum xen_domain_type xen_domain_type = XEN_NATIVE;
76 EXPORT_SYMBOL_GPL(xen_domain_type);
77
78 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
79 EXPORT_SYMBOL(machine_to_phys_mapping);
80 unsigned int   machine_to_phys_order;
81 EXPORT_SYMBOL(machine_to_phys_order);
82
83 struct start_info *xen_start_info;
84 EXPORT_SYMBOL_GPL(xen_start_info);
85
86 struct shared_info xen_dummy_shared_info;
87
88 void *xen_initial_gdt;
89
90 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
91 __read_mostly int xen_have_vector_callback;
92 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
93
94 /*
95  * Point at some empty memory to start with. We map the real shared_info
96  * page as soon as fixmap is up and running.
97  */
98 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
99
100 /*
101  * Flag to determine whether vcpu info placement is available on all
102  * VCPUs.  We assume it is to start with, and then set it to zero on
103  * the first failure.  This is because it can succeed on some VCPUs
104  * and not others, since it can involve hypervisor memory allocation,
105  * or because the guest failed to guarantee all the appropriate
106  * constraints on all VCPUs (ie buffer can't cross a page boundary).
107  *
108  * Note that any particular CPU may be using a placed vcpu structure,
109  * but we can only optimise if the all are.
110  *
111  * 0: not available, 1: available
112  */
113 static int have_vcpu_info_placement = 1;
114
115 static void clamp_max_cpus(void)
116 {
117 #ifdef CONFIG_SMP
118         if (setup_max_cpus > MAX_VIRT_CPUS)
119                 setup_max_cpus = MAX_VIRT_CPUS;
120 #endif
121 }
122
123 static void xen_vcpu_setup(int cpu)
124 {
125         struct vcpu_register_vcpu_info info;
126         int err;
127         struct vcpu_info *vcpup;
128
129         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
130
131         if (cpu < MAX_VIRT_CPUS)
132                 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
133
134         if (!have_vcpu_info_placement) {
135                 if (cpu >= MAX_VIRT_CPUS)
136                         clamp_max_cpus();
137                 return;
138         }
139
140         vcpup = &per_cpu(xen_vcpu_info, cpu);
141         info.mfn = arbitrary_virt_to_mfn(vcpup);
142         info.offset = offset_in_page(vcpup);
143
144         /* Check to see if the hypervisor will put the vcpu_info
145            structure where we want it, which allows direct access via
146            a percpu-variable. */
147         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
148
149         if (err) {
150                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
151                 have_vcpu_info_placement = 0;
152                 clamp_max_cpus();
153         } else {
154                 /* This cpu is using the registered vcpu info, even if
155                    later ones fail to. */
156                 per_cpu(xen_vcpu, cpu) = vcpup;
157         }
158 }
159
160 /*
161  * On restore, set the vcpu placement up again.
162  * If it fails, then we're in a bad state, since
163  * we can't back out from using it...
164  */
165 void xen_vcpu_restore(void)
166 {
167         int cpu;
168
169         for_each_online_cpu(cpu) {
170                 bool other_cpu = (cpu != smp_processor_id());
171
172                 if (other_cpu &&
173                     HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
174                         BUG();
175
176                 xen_setup_runstate_info(cpu);
177
178                 if (have_vcpu_info_placement)
179                         xen_vcpu_setup(cpu);
180
181                 if (other_cpu &&
182                     HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
183                         BUG();
184         }
185 }
186
187 static void __init xen_banner(void)
188 {
189         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
190         struct xen_extraversion extra;
191         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
192
193         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
194                pv_info.name);
195         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
196                version >> 16, version & 0xffff, extra.extraversion,
197                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
198 }
199
200 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
201 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
202
203 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
204                       unsigned int *cx, unsigned int *dx)
205 {
206         unsigned maskebx = ~0;
207         unsigned maskecx = ~0;
208         unsigned maskedx = ~0;
209
210         /*
211          * Mask out inconvenient features, to try and disable as many
212          * unsupported kernel subsystems as possible.
213          */
214         switch (*ax) {
215         case 1:
216                 maskecx = cpuid_leaf1_ecx_mask;
217                 maskedx = cpuid_leaf1_edx_mask;
218                 break;
219
220         case 0xb:
221                 /* Suppress extended topology stuff */
222                 maskebx = 0;
223                 break;
224         }
225
226         asm(XEN_EMULATE_PREFIX "cpuid"
227                 : "=a" (*ax),
228                   "=b" (*bx),
229                   "=c" (*cx),
230                   "=d" (*dx)
231                 : "0" (*ax), "2" (*cx));
232
233         *bx &= maskebx;
234         *cx &= maskecx;
235         *dx &= maskedx;
236 }
237
238 static void __init xen_init_cpuid_mask(void)
239 {
240         unsigned int ax, bx, cx, dx;
241         unsigned int xsave_mask;
242
243         cpuid_leaf1_edx_mask =
244                 ~((1 << X86_FEATURE_MCE)  |  /* disable MCE */
245                   (1 << X86_FEATURE_MCA)  |  /* disable MCA */
246                   (1 << X86_FEATURE_MTRR) |  /* disable MTRR */
247                   (1 << X86_FEATURE_ACC));   /* thermal monitoring */
248
249         if (!xen_initial_domain())
250                 cpuid_leaf1_edx_mask &=
251                         ~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
252                           (1 << X86_FEATURE_ACPI));  /* disable ACPI */
253         ax = 1;
254         xen_cpuid(&ax, &bx, &cx, &dx);
255
256         xsave_mask =
257                 (1 << (X86_FEATURE_XSAVE % 32)) |
258                 (1 << (X86_FEATURE_OSXSAVE % 32));
259
260         /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
261         if ((cx & xsave_mask) != xsave_mask)
262                 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
263 }
264
265 static void xen_set_debugreg(int reg, unsigned long val)
266 {
267         HYPERVISOR_set_debugreg(reg, val);
268 }
269
270 static unsigned long xen_get_debugreg(int reg)
271 {
272         return HYPERVISOR_get_debugreg(reg);
273 }
274
275 static void xen_end_context_switch(struct task_struct *next)
276 {
277         xen_mc_flush();
278         paravirt_end_context_switch(next);
279 }
280
281 static unsigned long xen_store_tr(void)
282 {
283         return 0;
284 }
285
286 /*
287  * Set the page permissions for a particular virtual address.  If the
288  * address is a vmalloc mapping (or other non-linear mapping), then
289  * find the linear mapping of the page and also set its protections to
290  * match.
291  */
292 static void set_aliased_prot(void *v, pgprot_t prot)
293 {
294         int level;
295         pte_t *ptep;
296         pte_t pte;
297         unsigned long pfn;
298         struct page *page;
299
300         ptep = lookup_address((unsigned long)v, &level);
301         BUG_ON(ptep == NULL);
302
303         pfn = pte_pfn(*ptep);
304         page = pfn_to_page(pfn);
305
306         pte = pfn_pte(pfn, prot);
307
308         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
309                 BUG();
310
311         if (!PageHighMem(page)) {
312                 void *av = __va(PFN_PHYS(pfn));
313
314                 if (av != v)
315                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
316                                 BUG();
317         } else
318                 kmap_flush_unused();
319 }
320
321 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
322 {
323         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
324         int i;
325
326         for(i = 0; i < entries; i += entries_per_page)
327                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
328 }
329
330 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
331 {
332         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
333         int i;
334
335         for(i = 0; i < entries; i += entries_per_page)
336                 set_aliased_prot(ldt + i, PAGE_KERNEL);
337 }
338
339 static void xen_set_ldt(const void *addr, unsigned entries)
340 {
341         struct mmuext_op *op;
342         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
343
344         op = mcs.args;
345         op->cmd = MMUEXT_SET_LDT;
346         op->arg1.linear_addr = (unsigned long)addr;
347         op->arg2.nr_ents = entries;
348
349         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
350
351         xen_mc_issue(PARAVIRT_LAZY_CPU);
352 }
353
354 static void xen_load_gdt(const struct desc_ptr *dtr)
355 {
356         unsigned long va = dtr->address;
357         unsigned int size = dtr->size + 1;
358         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
359         unsigned long frames[pages];
360         int f;
361
362         /*
363          * A GDT can be up to 64k in size, which corresponds to 8192
364          * 8-byte entries, or 16 4k pages..
365          */
366
367         BUG_ON(size > 65536);
368         BUG_ON(va & ~PAGE_MASK);
369
370         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
371                 int level;
372                 pte_t *ptep;
373                 unsigned long pfn, mfn;
374                 void *virt;
375
376                 /*
377                  * The GDT is per-cpu and is in the percpu data area.
378                  * That can be virtually mapped, so we need to do a
379                  * page-walk to get the underlying MFN for the
380                  * hypercall.  The page can also be in the kernel's
381                  * linear range, so we need to RO that mapping too.
382                  */
383                 ptep = lookup_address(va, &level);
384                 BUG_ON(ptep == NULL);
385
386                 pfn = pte_pfn(*ptep);
387                 mfn = pfn_to_mfn(pfn);
388                 virt = __va(PFN_PHYS(pfn));
389
390                 frames[f] = mfn;
391
392                 make_lowmem_page_readonly((void *)va);
393                 make_lowmem_page_readonly(virt);
394         }
395
396         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
397                 BUG();
398 }
399
400 /*
401  * load_gdt for early boot, when the gdt is only mapped once
402  */
403 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
404 {
405         unsigned long va = dtr->address;
406         unsigned int size = dtr->size + 1;
407         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
408         unsigned long frames[pages];
409         int f;
410
411         /*
412          * A GDT can be up to 64k in size, which corresponds to 8192
413          * 8-byte entries, or 16 4k pages..
414          */
415
416         BUG_ON(size > 65536);
417         BUG_ON(va & ~PAGE_MASK);
418
419         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
420                 pte_t pte;
421                 unsigned long pfn, mfn;
422
423                 pfn = virt_to_pfn(va);
424                 mfn = pfn_to_mfn(pfn);
425
426                 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
427
428                 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
429                         BUG();
430
431                 frames[f] = mfn;
432         }
433
434         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
435                 BUG();
436 }
437
438 static void load_TLS_descriptor(struct thread_struct *t,
439                                 unsigned int cpu, unsigned int i)
440 {
441         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
442         xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
443         struct multicall_space mc = __xen_mc_entry(0);
444
445         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
446 }
447
448 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
449 {
450         /*
451          * XXX sleazy hack: If we're being called in a lazy-cpu zone
452          * and lazy gs handling is enabled, it means we're in a
453          * context switch, and %gs has just been saved.  This means we
454          * can zero it out to prevent faults on exit from the
455          * hypervisor if the next process has no %gs.  Either way, it
456          * has been saved, and the new value will get loaded properly.
457          * This will go away as soon as Xen has been modified to not
458          * save/restore %gs for normal hypercalls.
459          *
460          * On x86_64, this hack is not used for %gs, because gs points
461          * to KERNEL_GS_BASE (and uses it for PDA references), so we
462          * must not zero %gs on x86_64
463          *
464          * For x86_64, we need to zero %fs, otherwise we may get an
465          * exception between the new %fs descriptor being loaded and
466          * %fs being effectively cleared at __switch_to().
467          */
468         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
469 #ifdef CONFIG_X86_32
470                 lazy_load_gs(0);
471 #else
472                 loadsegment(fs, 0);
473 #endif
474         }
475
476         xen_mc_batch();
477
478         load_TLS_descriptor(t, cpu, 0);
479         load_TLS_descriptor(t, cpu, 1);
480         load_TLS_descriptor(t, cpu, 2);
481
482         xen_mc_issue(PARAVIRT_LAZY_CPU);
483 }
484
485 #ifdef CONFIG_X86_64
486 static void xen_load_gs_index(unsigned int idx)
487 {
488         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
489                 BUG();
490 }
491 #endif
492
493 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
494                                 const void *ptr)
495 {
496         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
497         u64 entry = *(u64 *)ptr;
498
499         preempt_disable();
500
501         xen_mc_flush();
502         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
503                 BUG();
504
505         preempt_enable();
506 }
507
508 static int cvt_gate_to_trap(int vector, const gate_desc *val,
509                             struct trap_info *info)
510 {
511         unsigned long addr;
512
513         if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
514                 return 0;
515
516         info->vector = vector;
517
518         addr = gate_offset(*val);
519 #ifdef CONFIG_X86_64
520         /*
521          * Look for known traps using IST, and substitute them
522          * appropriately.  The debugger ones are the only ones we care
523          * about.  Xen will handle faults like double_fault and
524          * machine_check, so we should never see them.  Warn if
525          * there's an unexpected IST-using fault handler.
526          */
527         if (addr == (unsigned long)debug)
528                 addr = (unsigned long)xen_debug;
529         else if (addr == (unsigned long)int3)
530                 addr = (unsigned long)xen_int3;
531         else if (addr == (unsigned long)stack_segment)
532                 addr = (unsigned long)xen_stack_segment;
533         else if (addr == (unsigned long)double_fault ||
534                  addr == (unsigned long)nmi) {
535                 /* Don't need to handle these */
536                 return 0;
537 #ifdef CONFIG_X86_MCE
538         } else if (addr == (unsigned long)machine_check) {
539                 return 0;
540 #endif
541         } else {
542                 /* Some other trap using IST? */
543                 if (WARN_ON(val->ist != 0))
544                         return 0;
545         }
546 #endif  /* CONFIG_X86_64 */
547         info->address = addr;
548
549         info->cs = gate_segment(*val);
550         info->flags = val->dpl;
551         /* interrupt gates clear IF */
552         if (val->type == GATE_INTERRUPT)
553                 info->flags |= 1 << 2;
554
555         return 1;
556 }
557
558 /* Locations of each CPU's IDT */
559 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
560
561 /* Set an IDT entry.  If the entry is part of the current IDT, then
562    also update Xen. */
563 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
564 {
565         unsigned long p = (unsigned long)&dt[entrynum];
566         unsigned long start, end;
567
568         preempt_disable();
569
570         start = __this_cpu_read(idt_desc.address);
571         end = start + __this_cpu_read(idt_desc.size) + 1;
572
573         xen_mc_flush();
574
575         native_write_idt_entry(dt, entrynum, g);
576
577         if (p >= start && (p + 8) <= end) {
578                 struct trap_info info[2];
579
580                 info[1].address = 0;
581
582                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
583                         if (HYPERVISOR_set_trap_table(info))
584                                 BUG();
585         }
586
587         preempt_enable();
588 }
589
590 static void xen_convert_trap_info(const struct desc_ptr *desc,
591                                   struct trap_info *traps)
592 {
593         unsigned in, out, count;
594
595         count = (desc->size+1) / sizeof(gate_desc);
596         BUG_ON(count > 256);
597
598         for (in = out = 0; in < count; in++) {
599                 gate_desc *entry = (gate_desc*)(desc->address) + in;
600
601                 if (cvt_gate_to_trap(in, entry, &traps[out]))
602                         out++;
603         }
604         traps[out].address = 0;
605 }
606
607 void xen_copy_trap_info(struct trap_info *traps)
608 {
609         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
610
611         xen_convert_trap_info(desc, traps);
612 }
613
614 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
615    hold a spinlock to protect the static traps[] array (static because
616    it avoids allocation, and saves stack space). */
617 static void xen_load_idt(const struct desc_ptr *desc)
618 {
619         static DEFINE_SPINLOCK(lock);
620         static struct trap_info traps[257];
621
622         spin_lock(&lock);
623
624         __get_cpu_var(idt_desc) = *desc;
625
626         xen_convert_trap_info(desc, traps);
627
628         xen_mc_flush();
629         if (HYPERVISOR_set_trap_table(traps))
630                 BUG();
631
632         spin_unlock(&lock);
633 }
634
635 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
636    they're handled differently. */
637 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
638                                 const void *desc, int type)
639 {
640         preempt_disable();
641
642         switch (type) {
643         case DESC_LDT:
644         case DESC_TSS:
645                 /* ignore */
646                 break;
647
648         default: {
649                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
650
651                 xen_mc_flush();
652                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
653                         BUG();
654         }
655
656         }
657
658         preempt_enable();
659 }
660
661 /*
662  * Version of write_gdt_entry for use at early boot-time needed to
663  * update an entry as simply as possible.
664  */
665 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
666                                             const void *desc, int type)
667 {
668         switch (type) {
669         case DESC_LDT:
670         case DESC_TSS:
671                 /* ignore */
672                 break;
673
674         default: {
675                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
676
677                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
678                         dt[entry] = *(struct desc_struct *)desc;
679         }
680
681         }
682 }
683
684 static void xen_load_sp0(struct tss_struct *tss,
685                          struct thread_struct *thread)
686 {
687         struct multicall_space mcs = xen_mc_entry(0);
688         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
689         xen_mc_issue(PARAVIRT_LAZY_CPU);
690 }
691
692 static void xen_set_iopl_mask(unsigned mask)
693 {
694         struct physdev_set_iopl set_iopl;
695
696         /* Force the change at ring 0. */
697         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
698         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
699 }
700
701 static void xen_io_delay(void)
702 {
703 }
704
705 #ifdef CONFIG_X86_LOCAL_APIC
706 static u32 xen_apic_read(u32 reg)
707 {
708         return 0;
709 }
710
711 static void xen_apic_write(u32 reg, u32 val)
712 {
713         /* Warn to see if there's any stray references */
714         WARN_ON(1);
715 }
716
717 static u64 xen_apic_icr_read(void)
718 {
719         return 0;
720 }
721
722 static void xen_apic_icr_write(u32 low, u32 id)
723 {
724         /* Warn to see if there's any stray references */
725         WARN_ON(1);
726 }
727
728 static void xen_apic_wait_icr_idle(void)
729 {
730         return;
731 }
732
733 static u32 xen_safe_apic_wait_icr_idle(void)
734 {
735         return 0;
736 }
737
738 static void set_xen_basic_apic_ops(void)
739 {
740         apic->read = xen_apic_read;
741         apic->write = xen_apic_write;
742         apic->icr_read = xen_apic_icr_read;
743         apic->icr_write = xen_apic_icr_write;
744         apic->wait_icr_idle = xen_apic_wait_icr_idle;
745         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
746 }
747
748 #endif
749
750 static void xen_clts(void)
751 {
752         struct multicall_space mcs;
753
754         mcs = xen_mc_entry(0);
755
756         MULTI_fpu_taskswitch(mcs.mc, 0);
757
758         xen_mc_issue(PARAVIRT_LAZY_CPU);
759 }
760
761 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
762
763 static unsigned long xen_read_cr0(void)
764 {
765         unsigned long cr0 = percpu_read(xen_cr0_value);
766
767         if (unlikely(cr0 == 0)) {
768                 cr0 = native_read_cr0();
769                 percpu_write(xen_cr0_value, cr0);
770         }
771
772         return cr0;
773 }
774
775 static void xen_write_cr0(unsigned long cr0)
776 {
777         struct multicall_space mcs;
778
779         percpu_write(xen_cr0_value, cr0);
780
781         /* Only pay attention to cr0.TS; everything else is
782            ignored. */
783         mcs = xen_mc_entry(0);
784
785         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
786
787         xen_mc_issue(PARAVIRT_LAZY_CPU);
788 }
789
790 static void xen_write_cr4(unsigned long cr4)
791 {
792         cr4 &= ~X86_CR4_PGE;
793         cr4 &= ~X86_CR4_PSE;
794
795         native_write_cr4(cr4);
796 }
797
798 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
799 {
800         int ret;
801
802         ret = 0;
803
804         switch (msr) {
805 #ifdef CONFIG_X86_64
806                 unsigned which;
807                 u64 base;
808
809         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
810         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
811         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
812
813         set:
814                 base = ((u64)high << 32) | low;
815                 if (HYPERVISOR_set_segment_base(which, base) != 0)
816                         ret = -EIO;
817                 break;
818 #endif
819
820         case MSR_STAR:
821         case MSR_CSTAR:
822         case MSR_LSTAR:
823         case MSR_SYSCALL_MASK:
824         case MSR_IA32_SYSENTER_CS:
825         case MSR_IA32_SYSENTER_ESP:
826         case MSR_IA32_SYSENTER_EIP:
827                 /* Fast syscall setup is all done in hypercalls, so
828                    these are all ignored.  Stub them out here to stop
829                    Xen console noise. */
830                 break;
831
832         case MSR_IA32_CR_PAT:
833                 if (smp_processor_id() == 0)
834                         xen_set_pat(((u64)high << 32) | low);
835                 break;
836
837         default:
838                 ret = native_write_msr_safe(msr, low, high);
839         }
840
841         return ret;
842 }
843
844 void xen_setup_shared_info(void)
845 {
846         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
847                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
848                            xen_start_info->shared_info);
849
850                 HYPERVISOR_shared_info =
851                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
852         } else
853                 HYPERVISOR_shared_info =
854                         (struct shared_info *)__va(xen_start_info->shared_info);
855
856 #ifndef CONFIG_SMP
857         /* In UP this is as good a place as any to set up shared info */
858         xen_setup_vcpu_info_placement();
859 #endif
860
861         xen_setup_mfn_list_list();
862 }
863
864 /* This is called once we have the cpu_possible_map */
865 void xen_setup_vcpu_info_placement(void)
866 {
867         int cpu;
868
869         for_each_possible_cpu(cpu)
870                 xen_vcpu_setup(cpu);
871
872         /* xen_vcpu_setup managed to place the vcpu_info within the
873            percpu area for all cpus, so make use of it */
874         if (have_vcpu_info_placement) {
875                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
876                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
877                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
878                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
879                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
880         }
881 }
882
883 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
884                           unsigned long addr, unsigned len)
885 {
886         char *start, *end, *reloc;
887         unsigned ret;
888
889         start = end = reloc = NULL;
890
891 #define SITE(op, x)                                                     \
892         case PARAVIRT_PATCH(op.x):                                      \
893         if (have_vcpu_info_placement) {                                 \
894                 start = (char *)xen_##x##_direct;                       \
895                 end = xen_##x##_direct_end;                             \
896                 reloc = xen_##x##_direct_reloc;                         \
897         }                                                               \
898         goto patch_site
899
900         switch (type) {
901                 SITE(pv_irq_ops, irq_enable);
902                 SITE(pv_irq_ops, irq_disable);
903                 SITE(pv_irq_ops, save_fl);
904                 SITE(pv_irq_ops, restore_fl);
905 #undef SITE
906
907         patch_site:
908                 if (start == NULL || (end-start) > len)
909                         goto default_patch;
910
911                 ret = paravirt_patch_insns(insnbuf, len, start, end);
912
913                 /* Note: because reloc is assigned from something that
914                    appears to be an array, gcc assumes it's non-null,
915                    but doesn't know its relationship with start and
916                    end. */
917                 if (reloc > start && reloc < end) {
918                         int reloc_off = reloc - start;
919                         long *relocp = (long *)(insnbuf + reloc_off);
920                         long delta = start - (char *)addr;
921
922                         *relocp += delta;
923                 }
924                 break;
925
926         default_patch:
927         default:
928                 ret = paravirt_patch_default(type, clobbers, insnbuf,
929                                              addr, len);
930                 break;
931         }
932
933         return ret;
934 }
935
936 static const struct pv_info xen_info __initconst = {
937         .paravirt_enabled = 1,
938         .shared_kernel_pmd = 0,
939
940         .name = "Xen",
941 };
942
943 static const struct pv_init_ops xen_init_ops __initconst = {
944         .patch = xen_patch,
945 };
946
947 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
948         .cpuid = xen_cpuid,
949
950         .set_debugreg = xen_set_debugreg,
951         .get_debugreg = xen_get_debugreg,
952
953         .clts = xen_clts,
954
955         .read_cr0 = xen_read_cr0,
956         .write_cr0 = xen_write_cr0,
957
958         .read_cr4 = native_read_cr4,
959         .read_cr4_safe = native_read_cr4_safe,
960         .write_cr4 = xen_write_cr4,
961
962         .wbinvd = native_wbinvd,
963
964         .read_msr = native_read_msr_safe,
965         .write_msr = xen_write_msr_safe,
966         .read_tsc = native_read_tsc,
967         .read_pmc = native_read_pmc,
968
969         .iret = xen_iret,
970         .irq_enable_sysexit = xen_sysexit,
971 #ifdef CONFIG_X86_64
972         .usergs_sysret32 = xen_sysret32,
973         .usergs_sysret64 = xen_sysret64,
974 #endif
975
976         .load_tr_desc = paravirt_nop,
977         .set_ldt = xen_set_ldt,
978         .load_gdt = xen_load_gdt,
979         .load_idt = xen_load_idt,
980         .load_tls = xen_load_tls,
981 #ifdef CONFIG_X86_64
982         .load_gs_index = xen_load_gs_index,
983 #endif
984
985         .alloc_ldt = xen_alloc_ldt,
986         .free_ldt = xen_free_ldt,
987
988         .store_gdt = native_store_gdt,
989         .store_idt = native_store_idt,
990         .store_tr = xen_store_tr,
991
992         .write_ldt_entry = xen_write_ldt_entry,
993         .write_gdt_entry = xen_write_gdt_entry,
994         .write_idt_entry = xen_write_idt_entry,
995         .load_sp0 = xen_load_sp0,
996
997         .set_iopl_mask = xen_set_iopl_mask,
998         .io_delay = xen_io_delay,
999
1000         /* Xen takes care of %gs when switching to usermode for us */
1001         .swapgs = paravirt_nop,
1002
1003         .start_context_switch = paravirt_start_context_switch,
1004         .end_context_switch = xen_end_context_switch,
1005 };
1006
1007 static const struct pv_apic_ops xen_apic_ops __initconst = {
1008 #ifdef CONFIG_X86_LOCAL_APIC
1009         .startup_ipi_hook = paravirt_nop,
1010 #endif
1011 };
1012
1013 static void xen_reboot(int reason)
1014 {
1015         struct sched_shutdown r = { .reason = reason };
1016
1017         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1018                 BUG();
1019 }
1020
1021 static void xen_restart(char *msg)
1022 {
1023         xen_reboot(SHUTDOWN_reboot);
1024 }
1025
1026 static void xen_emergency_restart(void)
1027 {
1028         xen_reboot(SHUTDOWN_reboot);
1029 }
1030
1031 static void xen_machine_halt(void)
1032 {
1033         xen_reboot(SHUTDOWN_poweroff);
1034 }
1035
1036 static void xen_machine_power_off(void)
1037 {
1038         if (pm_power_off)
1039                 pm_power_off();
1040         xen_reboot(SHUTDOWN_poweroff);
1041 }
1042
1043 static void xen_crash_shutdown(struct pt_regs *regs)
1044 {
1045         xen_reboot(SHUTDOWN_crash);
1046 }
1047
1048 static int
1049 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1050 {
1051         xen_reboot(SHUTDOWN_crash);
1052         return NOTIFY_DONE;
1053 }
1054
1055 static struct notifier_block xen_panic_block = {
1056         .notifier_call= xen_panic_event,
1057 };
1058
1059 int xen_panic_handler_init(void)
1060 {
1061         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1062         return 0;
1063 }
1064
1065 static const struct machine_ops xen_machine_ops __initconst = {
1066         .restart = xen_restart,
1067         .halt = xen_machine_halt,
1068         .power_off = xen_machine_power_off,
1069         .shutdown = xen_machine_halt,
1070         .crash_shutdown = xen_crash_shutdown,
1071         .emergency_restart = xen_emergency_restart,
1072 };
1073
1074 /*
1075  * Set up the GDT and segment registers for -fstack-protector.  Until
1076  * we do this, we have to be careful not to call any stack-protected
1077  * function, which is most of the kernel.
1078  */
1079 static void __init xen_setup_stackprotector(void)
1080 {
1081         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1082         pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1083
1084         setup_stack_canary_segment(0);
1085         switch_to_new_gdt(0);
1086
1087         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1088         pv_cpu_ops.load_gdt = xen_load_gdt;
1089 }
1090
1091 /* First C function to be called on Xen boot */
1092 asmlinkage void __init xen_start_kernel(void)
1093 {
1094         struct physdev_set_iopl set_iopl;
1095         int rc;
1096         pgd_t *pgd;
1097
1098         if (!xen_start_info)
1099                 return;
1100
1101         xen_domain_type = XEN_PV_DOMAIN;
1102
1103         xen_setup_machphys_mapping();
1104
1105         /* Install Xen paravirt ops */
1106         pv_info = xen_info;
1107         pv_init_ops = xen_init_ops;
1108         pv_cpu_ops = xen_cpu_ops;
1109         pv_apic_ops = xen_apic_ops;
1110
1111         x86_init.resources.memory_setup = xen_memory_setup;
1112         x86_init.oem.arch_setup = xen_arch_setup;
1113         x86_init.oem.banner = xen_banner;
1114
1115         xen_init_time_ops();
1116
1117         /*
1118          * Set up some pagetable state before starting to set any ptes.
1119          */
1120
1121         xen_init_mmu_ops();
1122
1123         /* Prevent unwanted bits from being set in PTEs. */
1124         __supported_pte_mask &= ~_PAGE_GLOBAL;
1125         if (!xen_initial_domain())
1126                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1127
1128         __supported_pte_mask |= _PAGE_IOMAP;
1129
1130         /*
1131          * Prevent page tables from being allocated in highmem, even
1132          * if CONFIG_HIGHPTE is enabled.
1133          */
1134         __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1135
1136         /* Work out if we support NX */
1137         x86_configure_nx();
1138
1139         xen_setup_features();
1140
1141         /* Get mfn list */
1142         if (!xen_feature(XENFEAT_auto_translated_physmap))
1143                 xen_build_dynamic_phys_to_machine();
1144
1145         /*
1146          * Set up kernel GDT and segment registers, mainly so that
1147          * -fstack-protector code can be executed.
1148          */
1149         xen_setup_stackprotector();
1150
1151         xen_init_irq_ops();
1152         xen_init_cpuid_mask();
1153
1154 #ifdef CONFIG_X86_LOCAL_APIC
1155         /*
1156          * set up the basic apic ops.
1157          */
1158         set_xen_basic_apic_ops();
1159 #endif
1160
1161         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1162                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1163                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1164         }
1165
1166         machine_ops = xen_machine_ops;
1167
1168         /*
1169          * The only reliable way to retain the initial address of the
1170          * percpu gdt_page is to remember it here, so we can go and
1171          * mark it RW later, when the initial percpu area is freed.
1172          */
1173         xen_initial_gdt = &per_cpu(gdt_page, 0);
1174
1175         xen_smp_init();
1176
1177 #ifdef CONFIG_ACPI_NUMA
1178         /*
1179          * The pages we from Xen are not related to machine pages, so
1180          * any NUMA information the kernel tries to get from ACPI will
1181          * be meaningless.  Prevent it from trying.
1182          */
1183         acpi_numa = -1;
1184 #endif
1185
1186         pgd = (pgd_t *)xen_start_info->pt_base;
1187
1188         if (!xen_initial_domain())
1189                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1190
1191         __supported_pte_mask |= _PAGE_IOMAP;
1192         /* Don't do the full vcpu_info placement stuff until we have a
1193            possible map and a non-dummy shared_info. */
1194         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1195
1196         local_irq_disable();
1197         early_boot_irqs_disabled = true;
1198
1199         memblock_init();
1200
1201         xen_raw_console_write("mapping kernel into physical memory\n");
1202         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1203         xen_ident_map_ISA();
1204
1205         /* Allocate and initialize top and mid mfn levels for p2m structure */
1206         xen_build_mfn_list_list();
1207
1208         /* keep using Xen gdt for now; no urgent need to change it */
1209
1210 #ifdef CONFIG_X86_32
1211         pv_info.kernel_rpl = 1;
1212         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1213                 pv_info.kernel_rpl = 0;
1214 #else
1215         pv_info.kernel_rpl = 0;
1216 #endif
1217         /* set the limit of our address space */
1218         xen_reserve_top();
1219
1220         /* We used to do this in xen_arch_setup, but that is too late on AMD
1221          * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1222          * which pokes 0xcf8 port.
1223          */
1224         set_iopl.iopl = 1;
1225         rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1226         if (rc != 0)
1227                 xen_raw_printk("physdev_op failed %d\n", rc);
1228
1229 #ifdef CONFIG_X86_32
1230         /* set up basic CPUID stuff */
1231         cpu_detect(&new_cpu_data);
1232         new_cpu_data.hard_math = 1;
1233         new_cpu_data.wp_works_ok = 1;
1234         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1235 #endif
1236
1237         /* Poke various useful things into boot_params */
1238         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1239         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1240                 ? __pa(xen_start_info->mod_start) : 0;
1241         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1242         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1243
1244         if (!xen_initial_domain()) {
1245                 add_preferred_console("xenboot", 0, NULL);
1246                 add_preferred_console("tty", 0, NULL);
1247                 add_preferred_console("hvc", 0, NULL);
1248                 if (pci_xen)
1249                         x86_init.pci.arch_init = pci_xen_init;
1250         } else {
1251                 /* Make sure ACS will be enabled */
1252                 pci_request_acs();
1253         }
1254                 
1255
1256         xen_raw_console_write("about to get started...\n");
1257
1258         xen_setup_runstate_info(0);
1259
1260         /* Start the world */
1261 #ifdef CONFIG_X86_32
1262         i386_start_kernel();
1263 #else
1264         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1265 #endif
1266 }
1267
1268 static int init_hvm_pv_info(int *major, int *minor)
1269 {
1270         uint32_t eax, ebx, ecx, edx, pages, msr, base;
1271         u64 pfn;
1272
1273         base = xen_cpuid_base();
1274         cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1275
1276         *major = eax >> 16;
1277         *minor = eax & 0xffff;
1278         printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1279
1280         cpuid(base + 2, &pages, &msr, &ecx, &edx);
1281
1282         pfn = __pa(hypercall_page);
1283         wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1284
1285         xen_setup_features();
1286
1287         pv_info.name = "Xen HVM";
1288
1289         xen_domain_type = XEN_HVM_DOMAIN;
1290
1291         return 0;
1292 }
1293
1294 void __ref xen_hvm_init_shared_info(void)
1295 {
1296         int cpu;
1297         struct xen_add_to_physmap xatp;
1298         static struct shared_info *shared_info_page = 0;
1299
1300         if (!shared_info_page)
1301                 shared_info_page = (struct shared_info *)
1302                         extend_brk(PAGE_SIZE, PAGE_SIZE);
1303         xatp.domid = DOMID_SELF;
1304         xatp.idx = 0;
1305         xatp.space = XENMAPSPACE_shared_info;
1306         xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1307         if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1308                 BUG();
1309
1310         HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1311
1312         /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1313          * page, we use it in the event channel upcall and in some pvclock
1314          * related functions. We don't need the vcpu_info placement
1315          * optimizations because we don't use any pv_mmu or pv_irq op on
1316          * HVM.
1317          * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1318          * online but xen_hvm_init_shared_info is run at resume time too and
1319          * in that case multiple vcpus might be online. */
1320         for_each_online_cpu(cpu) {
1321                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1322         }
1323 }
1324
1325 #ifdef CONFIG_XEN_PVHVM
1326 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1327                                     unsigned long action, void *hcpu)
1328 {
1329         int cpu = (long)hcpu;
1330         switch (action) {
1331         case CPU_UP_PREPARE:
1332                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1333                 if (xen_have_vector_callback)
1334                         xen_init_lock_cpu(cpu);
1335                 break;
1336         default:
1337                 break;
1338         }
1339         return NOTIFY_OK;
1340 }
1341
1342 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1343         .notifier_call  = xen_hvm_cpu_notify,
1344 };
1345
1346 static void __init xen_hvm_guest_init(void)
1347 {
1348         int r;
1349         int major, minor;
1350
1351         r = init_hvm_pv_info(&major, &minor);
1352         if (r < 0)
1353                 return;
1354
1355         xen_hvm_init_shared_info();
1356
1357         if (xen_feature(XENFEAT_hvm_callback_vector))
1358                 xen_have_vector_callback = 1;
1359         xen_hvm_smp_init();
1360         register_cpu_notifier(&xen_hvm_cpu_notifier);
1361         xen_unplug_emulated_devices();
1362         have_vcpu_info_placement = 0;
1363         x86_init.irqs.intr_init = xen_init_IRQ;
1364         xen_hvm_init_time_ops();
1365         xen_hvm_init_mmu_ops();
1366 }
1367
1368 static bool __init xen_hvm_platform(void)
1369 {
1370         if (xen_pv_domain())
1371                 return false;
1372
1373         if (!xen_cpuid_base())
1374                 return false;
1375
1376         return true;
1377 }
1378
1379 bool xen_hvm_need_lapic(void)
1380 {
1381         if (xen_pv_domain())
1382                 return false;
1383         if (!xen_hvm_domain())
1384                 return false;
1385         if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1386                 return false;
1387         return true;
1388 }
1389 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1390
1391 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1392         .name                   = "Xen HVM",
1393         .detect                 = xen_hvm_platform,
1394         .init_platform          = xen_hvm_guest_init,
1395 };
1396 EXPORT_SYMBOL(x86_hyper_xen_hvm);
1397 #endif