Merge branches 'upstream/xenfs' and 'upstream/core' of git://git.kernel.org/pub/scm...
[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/hypercall.h>
50 #include <asm/xen/hypervisor.h>
51 #include <asm/fixmap.h>
52 #include <asm/processor.h>
53 #include <asm/proto.h>
54 #include <asm/msr-index.h>
55 #include <asm/traps.h>
56 #include <asm/setup.h>
57 #include <asm/desc.h>
58 #include <asm/pgalloc.h>
59 #include <asm/pgtable.h>
60 #include <asm/tlbflush.h>
61 #include <asm/reboot.h>
62 #include <asm/setup.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 struct start_info *xen_start_info;
79 EXPORT_SYMBOL_GPL(xen_start_info);
80
81 struct shared_info xen_dummy_shared_info;
82
83 void *xen_initial_gdt;
84
85 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
86 __read_mostly int xen_have_vector_callback;
87 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
88
89 /*
90  * Point at some empty memory to start with. We map the real shared_info
91  * page as soon as fixmap is up and running.
92  */
93 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
94
95 /*
96  * Flag to determine whether vcpu info placement is available on all
97  * VCPUs.  We assume it is to start with, and then set it to zero on
98  * the first failure.  This is because it can succeed on some VCPUs
99  * and not others, since it can involve hypervisor memory allocation,
100  * or because the guest failed to guarantee all the appropriate
101  * constraints on all VCPUs (ie buffer can't cross a page boundary).
102  *
103  * Note that any particular CPU may be using a placed vcpu structure,
104  * but we can only optimise if the all are.
105  *
106  * 0: not available, 1: available
107  */
108 static int have_vcpu_info_placement = 1;
109
110 static void clamp_max_cpus(void)
111 {
112 #ifdef CONFIG_SMP
113         if (setup_max_cpus > MAX_VIRT_CPUS)
114                 setup_max_cpus = MAX_VIRT_CPUS;
115 #endif
116 }
117
118 static void xen_vcpu_setup(int cpu)
119 {
120         struct vcpu_register_vcpu_info info;
121         int err;
122         struct vcpu_info *vcpup;
123
124         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
125
126         if (cpu < MAX_VIRT_CPUS)
127                 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
128
129         if (!have_vcpu_info_placement) {
130                 if (cpu >= MAX_VIRT_CPUS)
131                         clamp_max_cpus();
132                 return;
133         }
134
135         vcpup = &per_cpu(xen_vcpu_info, cpu);
136         info.mfn = arbitrary_virt_to_mfn(vcpup);
137         info.offset = offset_in_page(vcpup);
138
139         /* Check to see if the hypervisor will put the vcpu_info
140            structure where we want it, which allows direct access via
141            a percpu-variable. */
142         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
143
144         if (err) {
145                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
146                 have_vcpu_info_placement = 0;
147                 clamp_max_cpus();
148         } else {
149                 /* This cpu is using the registered vcpu info, even if
150                    later ones fail to. */
151                 per_cpu(xen_vcpu, cpu) = vcpup;
152         }
153 }
154
155 /*
156  * On restore, set the vcpu placement up again.
157  * If it fails, then we're in a bad state, since
158  * we can't back out from using it...
159  */
160 void xen_vcpu_restore(void)
161 {
162         int cpu;
163
164         for_each_online_cpu(cpu) {
165                 bool other_cpu = (cpu != smp_processor_id());
166
167                 if (other_cpu &&
168                     HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
169                         BUG();
170
171                 xen_setup_runstate_info(cpu);
172
173                 if (have_vcpu_info_placement)
174                         xen_vcpu_setup(cpu);
175
176                 if (other_cpu &&
177                     HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
178                         BUG();
179         }
180 }
181
182 static void __init xen_banner(void)
183 {
184         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
185         struct xen_extraversion extra;
186         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
187
188         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
189                pv_info.name);
190         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
191                version >> 16, version & 0xffff, extra.extraversion,
192                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
193 }
194
195 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
196 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
197
198 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
199                       unsigned int *cx, unsigned int *dx)
200 {
201         unsigned maskebx = ~0;
202         unsigned maskecx = ~0;
203         unsigned maskedx = ~0;
204
205         /*
206          * Mask out inconvenient features, to try and disable as many
207          * unsupported kernel subsystems as possible.
208          */
209         switch (*ax) {
210         case 1:
211                 maskecx = cpuid_leaf1_ecx_mask;
212                 maskedx = cpuid_leaf1_edx_mask;
213                 break;
214
215         case 0xb:
216                 /* Suppress extended topology stuff */
217                 maskebx = 0;
218                 break;
219         }
220
221         asm(XEN_EMULATE_PREFIX "cpuid"
222                 : "=a" (*ax),
223                   "=b" (*bx),
224                   "=c" (*cx),
225                   "=d" (*dx)
226                 : "0" (*ax), "2" (*cx));
227
228         *bx &= maskebx;
229         *cx &= maskecx;
230         *dx &= maskedx;
231 }
232
233 static __init void xen_init_cpuid_mask(void)
234 {
235         unsigned int ax, bx, cx, dx;
236
237         cpuid_leaf1_edx_mask =
238                 ~((1 << X86_FEATURE_MCE)  |  /* disable MCE */
239                   (1 << X86_FEATURE_MCA)  |  /* disable MCA */
240                   (1 << X86_FEATURE_ACC));   /* thermal monitoring */
241
242         if (!xen_initial_domain())
243                 cpuid_leaf1_edx_mask &=
244                         ~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
245                           (1 << X86_FEATURE_ACPI));  /* disable ACPI */
246
247         ax = 1;
248         cx = 0;
249         xen_cpuid(&ax, &bx, &cx, &dx);
250
251         /* cpuid claims we support xsave; try enabling it to see what happens */
252         if (cx & (1 << (X86_FEATURE_XSAVE % 32))) {
253                 unsigned long cr4;
254
255                 set_in_cr4(X86_CR4_OSXSAVE);
256                 
257                 cr4 = read_cr4();
258
259                 if ((cr4 & X86_CR4_OSXSAVE) == 0)
260                         cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_XSAVE % 32));
261
262                 clear_in_cr4(X86_CR4_OSXSAVE);
263         }
264 }
265
266 static void xen_set_debugreg(int reg, unsigned long val)
267 {
268         HYPERVISOR_set_debugreg(reg, val);
269 }
270
271 static unsigned long xen_get_debugreg(int reg)
272 {
273         return HYPERVISOR_get_debugreg(reg);
274 }
275
276 static void xen_end_context_switch(struct task_struct *next)
277 {
278         xen_mc_flush();
279         paravirt_end_context_switch(next);
280 }
281
282 static unsigned long xen_store_tr(void)
283 {
284         return 0;
285 }
286
287 /*
288  * Set the page permissions for a particular virtual address.  If the
289  * address is a vmalloc mapping (or other non-linear mapping), then
290  * find the linear mapping of the page and also set its protections to
291  * match.
292  */
293 static void set_aliased_prot(void *v, pgprot_t prot)
294 {
295         int level;
296         pte_t *ptep;
297         pte_t pte;
298         unsigned long pfn;
299         struct page *page;
300
301         ptep = lookup_address((unsigned long)v, &level);
302         BUG_ON(ptep == NULL);
303
304         pfn = pte_pfn(*ptep);
305         page = pfn_to_page(pfn);
306
307         pte = pfn_pte(pfn, prot);
308
309         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
310                 BUG();
311
312         if (!PageHighMem(page)) {
313                 void *av = __va(PFN_PHYS(pfn));
314
315                 if (av != v)
316                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
317                                 BUG();
318         } else
319                 kmap_flush_unused();
320 }
321
322 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
323 {
324         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
325         int i;
326
327         for(i = 0; i < entries; i += entries_per_page)
328                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
329 }
330
331 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
332 {
333         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
334         int i;
335
336         for(i = 0; i < entries; i += entries_per_page)
337                 set_aliased_prot(ldt + i, PAGE_KERNEL);
338 }
339
340 static void xen_set_ldt(const void *addr, unsigned entries)
341 {
342         struct mmuext_op *op;
343         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
344
345         op = mcs.args;
346         op->cmd = MMUEXT_SET_LDT;
347         op->arg1.linear_addr = (unsigned long)addr;
348         op->arg2.nr_ents = entries;
349
350         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
351
352         xen_mc_issue(PARAVIRT_LAZY_CPU);
353 }
354
355 static void xen_load_gdt(const struct desc_ptr *dtr)
356 {
357         unsigned long va = dtr->address;
358         unsigned int size = dtr->size + 1;
359         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
360         unsigned long frames[pages];
361         int f;
362
363         /*
364          * A GDT can be up to 64k in size, which corresponds to 8192
365          * 8-byte entries, or 16 4k pages..
366          */
367
368         BUG_ON(size > 65536);
369         BUG_ON(va & ~PAGE_MASK);
370
371         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
372                 int level;
373                 pte_t *ptep;
374                 unsigned long pfn, mfn;
375                 void *virt;
376
377                 /*
378                  * The GDT is per-cpu and is in the percpu data area.
379                  * That can be virtually mapped, so we need to do a
380                  * page-walk to get the underlying MFN for the
381                  * hypercall.  The page can also be in the kernel's
382                  * linear range, so we need to RO that mapping too.
383                  */
384                 ptep = lookup_address(va, &level);
385                 BUG_ON(ptep == NULL);
386
387                 pfn = pte_pfn(*ptep);
388                 mfn = pfn_to_mfn(pfn);
389                 virt = __va(PFN_PHYS(pfn));
390
391                 frames[f] = mfn;
392
393                 make_lowmem_page_readonly((void *)va);
394                 make_lowmem_page_readonly(virt);
395         }
396
397         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
398                 BUG();
399 }
400
401 /*
402  * load_gdt for early boot, when the gdt is only mapped once
403  */
404 static __init void xen_load_gdt_boot(const struct desc_ptr *dtr)
405 {
406         unsigned long va = dtr->address;
407         unsigned int size = dtr->size + 1;
408         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
409         unsigned long frames[pages];
410         int f;
411
412         /*
413          * A GDT can be up to 64k in size, which corresponds to 8192
414          * 8-byte entries, or 16 4k pages..
415          */
416
417         BUG_ON(size > 65536);
418         BUG_ON(va & ~PAGE_MASK);
419
420         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
421                 pte_t pte;
422                 unsigned long pfn, mfn;
423
424                 pfn = virt_to_pfn(va);
425                 mfn = pfn_to_mfn(pfn);
426
427                 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
428
429                 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
430                         BUG();
431
432                 frames[f] = mfn;
433         }
434
435         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
436                 BUG();
437 }
438
439 static void load_TLS_descriptor(struct thread_struct *t,
440                                 unsigned int cpu, unsigned int i)
441 {
442         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
443         xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
444         struct multicall_space mc = __xen_mc_entry(0);
445
446         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
447 }
448
449 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
450 {
451         /*
452          * XXX sleazy hack: If we're being called in a lazy-cpu zone
453          * and lazy gs handling is enabled, it means we're in a
454          * context switch, and %gs has just been saved.  This means we
455          * can zero it out to prevent faults on exit from the
456          * hypervisor if the next process has no %gs.  Either way, it
457          * has been saved, and the new value will get loaded properly.
458          * This will go away as soon as Xen has been modified to not
459          * save/restore %gs for normal hypercalls.
460          *
461          * On x86_64, this hack is not used for %gs, because gs points
462          * to KERNEL_GS_BASE (and uses it for PDA references), so we
463          * must not zero %gs on x86_64
464          *
465          * For x86_64, we need to zero %fs, otherwise we may get an
466          * exception between the new %fs descriptor being loaded and
467          * %fs being effectively cleared at __switch_to().
468          */
469         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
470 #ifdef CONFIG_X86_32
471                 lazy_load_gs(0);
472 #else
473                 loadsegment(fs, 0);
474 #endif
475         }
476
477         xen_mc_batch();
478
479         load_TLS_descriptor(t, cpu, 0);
480         load_TLS_descriptor(t, cpu, 1);
481         load_TLS_descriptor(t, cpu, 2);
482
483         xen_mc_issue(PARAVIRT_LAZY_CPU);
484 }
485
486 #ifdef CONFIG_X86_64
487 static void xen_load_gs_index(unsigned int idx)
488 {
489         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
490                 BUG();
491 }
492 #endif
493
494 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
495                                 const void *ptr)
496 {
497         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
498         u64 entry = *(u64 *)ptr;
499
500         preempt_disable();
501
502         xen_mc_flush();
503         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
504                 BUG();
505
506         preempt_enable();
507 }
508
509 static int cvt_gate_to_trap(int vector, const gate_desc *val,
510                             struct trap_info *info)
511 {
512         unsigned long addr;
513
514         if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
515                 return 0;
516
517         info->vector = vector;
518
519         addr = gate_offset(*val);
520 #ifdef CONFIG_X86_64
521         /*
522          * Look for known traps using IST, and substitute them
523          * appropriately.  The debugger ones are the only ones we care
524          * about.  Xen will handle faults like double_fault and
525          * machine_check, so we should never see them.  Warn if
526          * there's an unexpected IST-using fault handler.
527          */
528         if (addr == (unsigned long)debug)
529                 addr = (unsigned long)xen_debug;
530         else if (addr == (unsigned long)int3)
531                 addr = (unsigned long)xen_int3;
532         else if (addr == (unsigned long)stack_segment)
533                 addr = (unsigned long)xen_stack_segment;
534         else if (addr == (unsigned long)double_fault ||
535                  addr == (unsigned long)nmi) {
536                 /* Don't need to handle these */
537                 return 0;
538 #ifdef CONFIG_X86_MCE
539         } else if (addr == (unsigned long)machine_check) {
540                 return 0;
541 #endif
542         } else {
543                 /* Some other trap using IST? */
544                 if (WARN_ON(val->ist != 0))
545                         return 0;
546         }
547 #endif  /* CONFIG_X86_64 */
548         info->address = addr;
549
550         info->cs = gate_segment(*val);
551         info->flags = val->dpl;
552         /* interrupt gates clear IF */
553         if (val->type == GATE_INTERRUPT)
554                 info->flags |= 1 << 2;
555
556         return 1;
557 }
558
559 /* Locations of each CPU's IDT */
560 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
561
562 /* Set an IDT entry.  If the entry is part of the current IDT, then
563    also update Xen. */
564 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
565 {
566         unsigned long p = (unsigned long)&dt[entrynum];
567         unsigned long start, end;
568
569         preempt_disable();
570
571         start = __get_cpu_var(idt_desc).address;
572         end = start + __get_cpu_var(idt_desc).size + 1;
573
574         xen_mc_flush();
575
576         native_write_idt_entry(dt, entrynum, g);
577
578         if (p >= start && (p + 8) <= end) {
579                 struct trap_info info[2];
580
581                 info[1].address = 0;
582
583                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
584                         if (HYPERVISOR_set_trap_table(info))
585                                 BUG();
586         }
587
588         preempt_enable();
589 }
590
591 static void xen_convert_trap_info(const struct desc_ptr *desc,
592                                   struct trap_info *traps)
593 {
594         unsigned in, out, count;
595
596         count = (desc->size+1) / sizeof(gate_desc);
597         BUG_ON(count > 256);
598
599         for (in = out = 0; in < count; in++) {
600                 gate_desc *entry = (gate_desc*)(desc->address) + in;
601
602                 if (cvt_gate_to_trap(in, entry, &traps[out]))
603                         out++;
604         }
605         traps[out].address = 0;
606 }
607
608 void xen_copy_trap_info(struct trap_info *traps)
609 {
610         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
611
612         xen_convert_trap_info(desc, traps);
613 }
614
615 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
616    hold a spinlock to protect the static traps[] array (static because
617    it avoids allocation, and saves stack space). */
618 static void xen_load_idt(const struct desc_ptr *desc)
619 {
620         static DEFINE_SPINLOCK(lock);
621         static struct trap_info traps[257];
622
623         spin_lock(&lock);
624
625         __get_cpu_var(idt_desc) = *desc;
626
627         xen_convert_trap_info(desc, traps);
628
629         xen_mc_flush();
630         if (HYPERVISOR_set_trap_table(traps))
631                 BUG();
632
633         spin_unlock(&lock);
634 }
635
636 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
637    they're handled differently. */
638 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
639                                 const void *desc, int type)
640 {
641         preempt_disable();
642
643         switch (type) {
644         case DESC_LDT:
645         case DESC_TSS:
646                 /* ignore */
647                 break;
648
649         default: {
650                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
651
652                 xen_mc_flush();
653                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
654                         BUG();
655         }
656
657         }
658
659         preempt_enable();
660 }
661
662 /*
663  * Version of write_gdt_entry for use at early boot-time needed to
664  * update an entry as simply as possible.
665  */
666 static __init void xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
667                                             const void *desc, int type)
668 {
669         switch (type) {
670         case DESC_LDT:
671         case DESC_TSS:
672                 /* ignore */
673                 break;
674
675         default: {
676                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
677
678                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
679                         dt[entry] = *(struct desc_struct *)desc;
680         }
681
682         }
683 }
684
685 static void xen_load_sp0(struct tss_struct *tss,
686                          struct thread_struct *thread)
687 {
688         struct multicall_space mcs = xen_mc_entry(0);
689         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
690         xen_mc_issue(PARAVIRT_LAZY_CPU);
691 }
692
693 static void xen_set_iopl_mask(unsigned mask)
694 {
695         struct physdev_set_iopl set_iopl;
696
697         /* Force the change at ring 0. */
698         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
699         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
700 }
701
702 static void xen_io_delay(void)
703 {
704 }
705
706 #ifdef CONFIG_X86_LOCAL_APIC
707 static u32 xen_apic_read(u32 reg)
708 {
709         return 0;
710 }
711
712 static void xen_apic_write(u32 reg, u32 val)
713 {
714         /* Warn to see if there's any stray references */
715         WARN_ON(1);
716 }
717
718 static u64 xen_apic_icr_read(void)
719 {
720         return 0;
721 }
722
723 static void xen_apic_icr_write(u32 low, u32 id)
724 {
725         /* Warn to see if there's any stray references */
726         WARN_ON(1);
727 }
728
729 static void xen_apic_wait_icr_idle(void)
730 {
731         return;
732 }
733
734 static u32 xen_safe_apic_wait_icr_idle(void)
735 {
736         return 0;
737 }
738
739 static void set_xen_basic_apic_ops(void)
740 {
741         apic->read = xen_apic_read;
742         apic->write = xen_apic_write;
743         apic->icr_read = xen_apic_icr_read;
744         apic->icr_write = xen_apic_icr_write;
745         apic->wait_icr_idle = xen_apic_wait_icr_idle;
746         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
747 }
748
749 #endif
750
751 static void xen_clts(void)
752 {
753         struct multicall_space mcs;
754
755         mcs = xen_mc_entry(0);
756
757         MULTI_fpu_taskswitch(mcs.mc, 0);
758
759         xen_mc_issue(PARAVIRT_LAZY_CPU);
760 }
761
762 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
763
764 static unsigned long xen_read_cr0(void)
765 {
766         unsigned long cr0 = percpu_read(xen_cr0_value);
767
768         if (unlikely(cr0 == 0)) {
769                 cr0 = native_read_cr0();
770                 percpu_write(xen_cr0_value, cr0);
771         }
772
773         return cr0;
774 }
775
776 static void xen_write_cr0(unsigned long cr0)
777 {
778         struct multicall_space mcs;
779
780         percpu_write(xen_cr0_value, cr0);
781
782         /* Only pay attention to cr0.TS; everything else is
783            ignored. */
784         mcs = xen_mc_entry(0);
785
786         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
787
788         xen_mc_issue(PARAVIRT_LAZY_CPU);
789 }
790
791 static void xen_write_cr4(unsigned long cr4)
792 {
793         cr4 &= ~X86_CR4_PGE;
794         cr4 &= ~X86_CR4_PSE;
795
796         native_write_cr4(cr4);
797 }
798
799 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
800 {
801         int ret;
802
803         ret = 0;
804
805         switch (msr) {
806 #ifdef CONFIG_X86_64
807                 unsigned which;
808                 u64 base;
809
810         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
811         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
812         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
813
814         set:
815                 base = ((u64)high << 32) | low;
816                 if (HYPERVISOR_set_segment_base(which, base) != 0)
817                         ret = -EIO;
818                 break;
819 #endif
820
821         case MSR_STAR:
822         case MSR_CSTAR:
823         case MSR_LSTAR:
824         case MSR_SYSCALL_MASK:
825         case MSR_IA32_SYSENTER_CS:
826         case MSR_IA32_SYSENTER_ESP:
827         case MSR_IA32_SYSENTER_EIP:
828                 /* Fast syscall setup is all done in hypercalls, so
829                    these are all ignored.  Stub them out here to stop
830                    Xen console noise. */
831                 break;
832
833         case MSR_IA32_CR_PAT:
834                 if (smp_processor_id() == 0)
835                         xen_set_pat(((u64)high << 32) | low);
836                 break;
837
838         default:
839                 ret = native_write_msr_safe(msr, low, high);
840         }
841
842         return ret;
843 }
844
845 void xen_setup_shared_info(void)
846 {
847         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
848                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
849                            xen_start_info->shared_info);
850
851                 HYPERVISOR_shared_info =
852                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
853         } else
854                 HYPERVISOR_shared_info =
855                         (struct shared_info *)__va(xen_start_info->shared_info);
856
857 #ifndef CONFIG_SMP
858         /* In UP this is as good a place as any to set up shared info */
859         xen_setup_vcpu_info_placement();
860 #endif
861
862         xen_setup_mfn_list_list();
863 }
864
865 /* This is called once we have the cpu_possible_map */
866 void xen_setup_vcpu_info_placement(void)
867 {
868         int cpu;
869
870         for_each_possible_cpu(cpu)
871                 xen_vcpu_setup(cpu);
872
873         /* xen_vcpu_setup managed to place the vcpu_info within the
874            percpu area for all cpus, so make use of it */
875         if (have_vcpu_info_placement) {
876                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
877                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
878                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
879                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
880                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
881         }
882 }
883
884 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
885                           unsigned long addr, unsigned len)
886 {
887         char *start, *end, *reloc;
888         unsigned ret;
889
890         start = end = reloc = NULL;
891
892 #define SITE(op, x)                                                     \
893         case PARAVIRT_PATCH(op.x):                                      \
894         if (have_vcpu_info_placement) {                                 \
895                 start = (char *)xen_##x##_direct;                       \
896                 end = xen_##x##_direct_end;                             \
897                 reloc = xen_##x##_direct_reloc;                         \
898         }                                                               \
899         goto patch_site
900
901         switch (type) {
902                 SITE(pv_irq_ops, irq_enable);
903                 SITE(pv_irq_ops, irq_disable);
904                 SITE(pv_irq_ops, save_fl);
905                 SITE(pv_irq_ops, restore_fl);
906 #undef SITE
907
908         patch_site:
909                 if (start == NULL || (end-start) > len)
910                         goto default_patch;
911
912                 ret = paravirt_patch_insns(insnbuf, len, start, end);
913
914                 /* Note: because reloc is assigned from something that
915                    appears to be an array, gcc assumes it's non-null,
916                    but doesn't know its relationship with start and
917                    end. */
918                 if (reloc > start && reloc < end) {
919                         int reloc_off = reloc - start;
920                         long *relocp = (long *)(insnbuf + reloc_off);
921                         long delta = start - (char *)addr;
922
923                         *relocp += delta;
924                 }
925                 break;
926
927         default_patch:
928         default:
929                 ret = paravirt_patch_default(type, clobbers, insnbuf,
930                                              addr, len);
931                 break;
932         }
933
934         return ret;
935 }
936
937 static const struct pv_info xen_info __initdata = {
938         .paravirt_enabled = 1,
939         .shared_kernel_pmd = 0,
940
941         .name = "Xen",
942 };
943
944 static const struct pv_init_ops xen_init_ops __initdata = {
945         .patch = xen_patch,
946 };
947
948 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
949         .cpuid = xen_cpuid,
950
951         .set_debugreg = xen_set_debugreg,
952         .get_debugreg = xen_get_debugreg,
953
954         .clts = xen_clts,
955
956         .read_cr0 = xen_read_cr0,
957         .write_cr0 = xen_write_cr0,
958
959         .read_cr4 = native_read_cr4,
960         .read_cr4_safe = native_read_cr4_safe,
961         .write_cr4 = xen_write_cr4,
962
963         .wbinvd = native_wbinvd,
964
965         .read_msr = native_read_msr_safe,
966         .write_msr = xen_write_msr_safe,
967         .read_tsc = native_read_tsc,
968         .read_pmc = native_read_pmc,
969
970         .iret = xen_iret,
971         .irq_enable_sysexit = xen_sysexit,
972 #ifdef CONFIG_X86_64
973         .usergs_sysret32 = xen_sysret32,
974         .usergs_sysret64 = xen_sysret64,
975 #endif
976
977         .load_tr_desc = paravirt_nop,
978         .set_ldt = xen_set_ldt,
979         .load_gdt = xen_load_gdt,
980         .load_idt = xen_load_idt,
981         .load_tls = xen_load_tls,
982 #ifdef CONFIG_X86_64
983         .load_gs_index = xen_load_gs_index,
984 #endif
985
986         .alloc_ldt = xen_alloc_ldt,
987         .free_ldt = xen_free_ldt,
988
989         .store_gdt = native_store_gdt,
990         .store_idt = native_store_idt,
991         .store_tr = xen_store_tr,
992
993         .write_ldt_entry = xen_write_ldt_entry,
994         .write_gdt_entry = xen_write_gdt_entry,
995         .write_idt_entry = xen_write_idt_entry,
996         .load_sp0 = xen_load_sp0,
997
998         .set_iopl_mask = xen_set_iopl_mask,
999         .io_delay = xen_io_delay,
1000
1001         /* Xen takes care of %gs when switching to usermode for us */
1002         .swapgs = paravirt_nop,
1003
1004         .start_context_switch = paravirt_start_context_switch,
1005         .end_context_switch = xen_end_context_switch,
1006 };
1007
1008 static const struct pv_apic_ops xen_apic_ops __initdata = {
1009 #ifdef CONFIG_X86_LOCAL_APIC
1010         .startup_ipi_hook = paravirt_nop,
1011 #endif
1012 };
1013
1014 static void xen_reboot(int reason)
1015 {
1016         struct sched_shutdown r = { .reason = reason };
1017
1018 #ifdef CONFIG_SMP
1019         smp_send_stop();
1020 #endif
1021
1022         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1023                 BUG();
1024 }
1025
1026 static void xen_restart(char *msg)
1027 {
1028         xen_reboot(SHUTDOWN_reboot);
1029 }
1030
1031 static void xen_emergency_restart(void)
1032 {
1033         xen_reboot(SHUTDOWN_reboot);
1034 }
1035
1036 static void xen_machine_halt(void)
1037 {
1038         xen_reboot(SHUTDOWN_poweroff);
1039 }
1040
1041 static void xen_crash_shutdown(struct pt_regs *regs)
1042 {
1043         xen_reboot(SHUTDOWN_crash);
1044 }
1045
1046 static int
1047 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1048 {
1049         xen_reboot(SHUTDOWN_crash);
1050         return NOTIFY_DONE;
1051 }
1052
1053 static struct notifier_block xen_panic_block = {
1054         .notifier_call= xen_panic_event,
1055 };
1056
1057 int xen_panic_handler_init(void)
1058 {
1059         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1060         return 0;
1061 }
1062
1063 static const struct machine_ops __initdata xen_machine_ops = {
1064         .restart = xen_restart,
1065         .halt = xen_machine_halt,
1066         .power_off = xen_machine_halt,
1067         .shutdown = xen_machine_halt,
1068         .crash_shutdown = xen_crash_shutdown,
1069         .emergency_restart = xen_emergency_restart,
1070 };
1071
1072 /*
1073  * Set up the GDT and segment registers for -fstack-protector.  Until
1074  * we do this, we have to be careful not to call any stack-protected
1075  * function, which is most of the kernel.
1076  */
1077 static void __init xen_setup_stackprotector(void)
1078 {
1079         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1080         pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1081
1082         setup_stack_canary_segment(0);
1083         switch_to_new_gdt(0);
1084
1085         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1086         pv_cpu_ops.load_gdt = xen_load_gdt;
1087 }
1088
1089 /* First C function to be called on Xen boot */
1090 asmlinkage void __init xen_start_kernel(void)
1091 {
1092         pgd_t *pgd;
1093
1094         if (!xen_start_info)
1095                 return;
1096
1097         xen_domain_type = XEN_PV_DOMAIN;
1098
1099         /* Install Xen paravirt ops */
1100         pv_info = xen_info;
1101         pv_init_ops = xen_init_ops;
1102         pv_cpu_ops = xen_cpu_ops;
1103         pv_apic_ops = xen_apic_ops;
1104
1105         x86_init.resources.memory_setup = xen_memory_setup;
1106         x86_init.oem.arch_setup = xen_arch_setup;
1107         x86_init.oem.banner = xen_banner;
1108
1109         xen_init_time_ops();
1110
1111         /*
1112          * Set up some pagetable state before starting to set any ptes.
1113          */
1114
1115         xen_init_mmu_ops();
1116
1117         /* Prevent unwanted bits from being set in PTEs. */
1118         __supported_pte_mask &= ~_PAGE_GLOBAL;
1119         if (!xen_initial_domain())
1120                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1121
1122         __supported_pte_mask |= _PAGE_IOMAP;
1123
1124         /*
1125          * Prevent page tables from being allocated in highmem, even
1126          * if CONFIG_HIGHPTE is enabled.
1127          */
1128         __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1129
1130         /* Work out if we support NX */
1131         x86_configure_nx();
1132
1133         xen_setup_features();
1134
1135         /* Get mfn list */
1136         if (!xen_feature(XENFEAT_auto_translated_physmap))
1137                 xen_build_dynamic_phys_to_machine();
1138
1139         /*
1140          * Set up kernel GDT and segment registers, mainly so that
1141          * -fstack-protector code can be executed.
1142          */
1143         xen_setup_stackprotector();
1144
1145         xen_init_irq_ops();
1146         xen_init_cpuid_mask();
1147
1148 #ifdef CONFIG_X86_LOCAL_APIC
1149         /*
1150          * set up the basic apic ops.
1151          */
1152         set_xen_basic_apic_ops();
1153 #endif
1154
1155         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1156                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1157                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1158         }
1159
1160         machine_ops = xen_machine_ops;
1161
1162         /*
1163          * The only reliable way to retain the initial address of the
1164          * percpu gdt_page is to remember it here, so we can go and
1165          * mark it RW later, when the initial percpu area is freed.
1166          */
1167         xen_initial_gdt = &per_cpu(gdt_page, 0);
1168
1169         xen_smp_init();
1170
1171         pgd = (pgd_t *)xen_start_info->pt_base;
1172
1173         if (!xen_initial_domain())
1174                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1175
1176         __supported_pte_mask |= _PAGE_IOMAP;
1177         /* Don't do the full vcpu_info placement stuff until we have a
1178            possible map and a non-dummy shared_info. */
1179         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1180
1181         local_irq_disable();
1182         early_boot_irqs_off();
1183
1184         memblock_init();
1185
1186         xen_raw_console_write("mapping kernel into physical memory\n");
1187         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1188
1189         /* Allocate and initialize top and mid mfn levels for p2m structure */
1190         xen_build_mfn_list_list();
1191
1192         init_mm.pgd = pgd;
1193
1194         /* keep using Xen gdt for now; no urgent need to change it */
1195
1196 #ifdef CONFIG_X86_32
1197         pv_info.kernel_rpl = 1;
1198         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1199                 pv_info.kernel_rpl = 0;
1200 #else
1201         pv_info.kernel_rpl = 0;
1202 #endif
1203
1204         /* set the limit of our address space */
1205         xen_reserve_top();
1206
1207 #ifdef CONFIG_X86_32
1208         /* set up basic CPUID stuff */
1209         cpu_detect(&new_cpu_data);
1210         new_cpu_data.hard_math = 1;
1211         new_cpu_data.wp_works_ok = 1;
1212         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1213 #endif
1214
1215         /* Poke various useful things into boot_params */
1216         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1217         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1218                 ? __pa(xen_start_info->mod_start) : 0;
1219         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1220         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1221
1222         if (!xen_initial_domain()) {
1223                 add_preferred_console("xenboot", 0, NULL);
1224                 add_preferred_console("tty", 0, NULL);
1225                 add_preferred_console("hvc", 0, NULL);
1226         } else {
1227                 /* Make sure ACS will be enabled */
1228                 pci_request_acs();
1229         }
1230                 
1231
1232         xen_raw_console_write("about to get started...\n");
1233
1234         xen_setup_runstate_info(0);
1235
1236         /* Start the world */
1237 #ifdef CONFIG_X86_32
1238         i386_start_kernel();
1239 #else
1240         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1241 #endif
1242 }
1243
1244 static uint32_t xen_cpuid_base(void)
1245 {
1246         uint32_t base, eax, ebx, ecx, edx;
1247         char signature[13];
1248
1249         for (base = 0x40000000; base < 0x40010000; base += 0x100) {
1250                 cpuid(base, &eax, &ebx, &ecx, &edx);
1251                 *(uint32_t *)(signature + 0) = ebx;
1252                 *(uint32_t *)(signature + 4) = ecx;
1253                 *(uint32_t *)(signature + 8) = edx;
1254                 signature[12] = 0;
1255
1256                 if (!strcmp("XenVMMXenVMM", signature) && ((eax - base) >= 2))
1257                         return base;
1258         }
1259
1260         return 0;
1261 }
1262
1263 static int init_hvm_pv_info(int *major, int *minor)
1264 {
1265         uint32_t eax, ebx, ecx, edx, pages, msr, base;
1266         u64 pfn;
1267
1268         base = xen_cpuid_base();
1269         cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1270
1271         *major = eax >> 16;
1272         *minor = eax & 0xffff;
1273         printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor);
1274
1275         cpuid(base + 2, &pages, &msr, &ecx, &edx);
1276
1277         pfn = __pa(hypercall_page);
1278         wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1279
1280         xen_setup_features();
1281
1282         pv_info = xen_info;
1283         pv_info.kernel_rpl = 0;
1284
1285         xen_domain_type = XEN_HVM_DOMAIN;
1286
1287         return 0;
1288 }
1289
1290 void xen_hvm_init_shared_info(void)
1291 {
1292         int cpu;
1293         struct xen_add_to_physmap xatp;
1294         static struct shared_info *shared_info_page = 0;
1295
1296         if (!shared_info_page)
1297                 shared_info_page = (struct shared_info *)
1298                         extend_brk(PAGE_SIZE, PAGE_SIZE);
1299         xatp.domid = DOMID_SELF;
1300         xatp.idx = 0;
1301         xatp.space = XENMAPSPACE_shared_info;
1302         xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1303         if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1304                 BUG();
1305
1306         HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1307
1308         /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1309          * page, we use it in the event channel upcall and in some pvclock
1310          * related functions. We don't need the vcpu_info placement
1311          * optimizations because we don't use any pv_mmu or pv_irq op on
1312          * HVM.
1313          * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1314          * online but xen_hvm_init_shared_info is run at resume time too and
1315          * in that case multiple vcpus might be online. */
1316         for_each_online_cpu(cpu) {
1317                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1318         }
1319 }
1320
1321 #ifdef CONFIG_XEN_PVHVM
1322 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1323                                     unsigned long action, void *hcpu)
1324 {
1325         int cpu = (long)hcpu;
1326         switch (action) {
1327         case CPU_UP_PREPARE:
1328                 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1329                 break;
1330         default:
1331                 break;
1332         }
1333         return NOTIFY_OK;
1334 }
1335
1336 static struct notifier_block __cpuinitdata xen_hvm_cpu_notifier = {
1337         .notifier_call  = xen_hvm_cpu_notify,
1338 };
1339
1340 static void __init xen_hvm_guest_init(void)
1341 {
1342         int r;
1343         int major, minor;
1344
1345         r = init_hvm_pv_info(&major, &minor);
1346         if (r < 0)
1347                 return;
1348
1349         xen_hvm_init_shared_info();
1350
1351         if (xen_feature(XENFEAT_hvm_callback_vector))
1352                 xen_have_vector_callback = 1;
1353         register_cpu_notifier(&xen_hvm_cpu_notifier);
1354         xen_unplug_emulated_devices();
1355         have_vcpu_info_placement = 0;
1356         x86_init.irqs.intr_init = xen_init_IRQ;
1357         xen_hvm_init_time_ops();
1358         xen_hvm_init_mmu_ops();
1359 }
1360
1361 static bool __init xen_hvm_platform(void)
1362 {
1363         if (xen_pv_domain())
1364                 return false;
1365
1366         if (!xen_cpuid_base())
1367                 return false;
1368
1369         return true;
1370 }
1371
1372 const __refconst struct hypervisor_x86 x86_hyper_xen_hvm = {
1373         .name                   = "Xen HVM",
1374         .detect                 = xen_hvm_platform,
1375         .init_platform          = xen_hvm_guest_init,
1376 };
1377 EXPORT_SYMBOL(x86_hyper_xen_hvm);
1378 #endif