xen/x86/PCI: Add support for the Xen PCI subsystem
[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
34 #include <xen/xen.h>
35 #include <xen/interface/xen.h>
36 #include <xen/interface/version.h>
37 #include <xen/interface/physdev.h>
38 #include <xen/interface/vcpu.h>
39 #include <xen/interface/memory.h>
40 #include <xen/features.h>
41 #include <xen/page.h>
42 #include <xen/hvm.h>
43 #include <xen/hvc-console.h>
44
45 #include <asm/paravirt.h>
46 #include <asm/apic.h>
47 #include <asm/page.h>
48 #include <asm/xen/pci.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         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
140                cpu, vcpup, info.mfn, info.offset);
141
142         /* Check to see if the hypervisor will put the vcpu_info
143            structure where we want it, which allows direct access via
144            a percpu-variable. */
145         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
146
147         if (err) {
148                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
149                 have_vcpu_info_placement = 0;
150                 clamp_max_cpus();
151         } else {
152                 /* This cpu is using the registered vcpu info, even if
153                    later ones fail to. */
154                 per_cpu(xen_vcpu, cpu) = vcpup;
155
156                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
157                        cpu, vcpup);
158         }
159 }
160
161 /*
162  * On restore, set the vcpu placement up again.
163  * If it fails, then we're in a bad state, since
164  * we can't back out from using it...
165  */
166 void xen_vcpu_restore(void)
167 {
168         int cpu;
169
170         for_each_online_cpu(cpu) {
171                 bool other_cpu = (cpu != smp_processor_id());
172
173                 if (other_cpu &&
174                     HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
175                         BUG();
176
177                 xen_setup_runstate_info(cpu);
178
179                 if (have_vcpu_info_placement)
180                         xen_vcpu_setup(cpu);
181
182                 if (other_cpu &&
183                     HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
184                         BUG();
185         }
186 }
187
188 static void __init xen_banner(void)
189 {
190         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
191         struct xen_extraversion extra;
192         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
193
194         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
195                pv_info.name);
196         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
197                version >> 16, version & 0xffff, extra.extraversion,
198                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
199 }
200
201 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
202 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
203
204 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
205                       unsigned int *cx, unsigned int *dx)
206 {
207         unsigned maskebx = ~0;
208         unsigned maskecx = ~0;
209         unsigned maskedx = ~0;
210
211         /*
212          * Mask out inconvenient features, to try and disable as many
213          * unsupported kernel subsystems as possible.
214          */
215         switch (*ax) {
216         case 1:
217                 maskecx = cpuid_leaf1_ecx_mask;
218                 maskedx = cpuid_leaf1_edx_mask;
219                 break;
220
221         case 0xb:
222                 /* Suppress extended topology stuff */
223                 maskebx = 0;
224                 break;
225         }
226
227         asm(XEN_EMULATE_PREFIX "cpuid"
228                 : "=a" (*ax),
229                   "=b" (*bx),
230                   "=c" (*cx),
231                   "=d" (*dx)
232                 : "0" (*ax), "2" (*cx));
233
234         *bx &= maskebx;
235         *cx &= maskecx;
236         *dx &= maskedx;
237 }
238
239 static __init void xen_init_cpuid_mask(void)
240 {
241         unsigned int ax, bx, cx, dx;
242
243         cpuid_leaf1_edx_mask =
244                 ~((1 << X86_FEATURE_MCE)  |  /* disable MCE */
245                   (1 << X86_FEATURE_MCA)  |  /* disable MCA */
246                   (1 << X86_FEATURE_ACC));   /* thermal monitoring */
247
248         if (!xen_initial_domain())
249                 cpuid_leaf1_edx_mask &=
250                         ~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
251                           (1 << X86_FEATURE_ACPI));  /* disable ACPI */
252
253         ax = 1;
254         cx = 0;
255         xen_cpuid(&ax, &bx, &cx, &dx);
256
257         /* cpuid claims we support xsave; try enabling it to see what happens */
258         if (cx & (1 << (X86_FEATURE_XSAVE % 32))) {
259                 unsigned long cr4;
260
261                 set_in_cr4(X86_CR4_OSXSAVE);
262                 
263                 cr4 = read_cr4();
264
265                 if ((cr4 & X86_CR4_OSXSAVE) == 0)
266                         cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_XSAVE % 32));
267
268                 clear_in_cr4(X86_CR4_OSXSAVE);
269         }
270 }
271
272 static void xen_set_debugreg(int reg, unsigned long val)
273 {
274         HYPERVISOR_set_debugreg(reg, val);
275 }
276
277 static unsigned long xen_get_debugreg(int reg)
278 {
279         return HYPERVISOR_get_debugreg(reg);
280 }
281
282 static void xen_end_context_switch(struct task_struct *next)
283 {
284         xen_mc_flush();
285         paravirt_end_context_switch(next);
286 }
287
288 static unsigned long xen_store_tr(void)
289 {
290         return 0;
291 }
292
293 /*
294  * Set the page permissions for a particular virtual address.  If the
295  * address is a vmalloc mapping (or other non-linear mapping), then
296  * find the linear mapping of the page and also set its protections to
297  * match.
298  */
299 static void set_aliased_prot(void *v, pgprot_t prot)
300 {
301         int level;
302         pte_t *ptep;
303         pte_t pte;
304         unsigned long pfn;
305         struct page *page;
306
307         ptep = lookup_address((unsigned long)v, &level);
308         BUG_ON(ptep == NULL);
309
310         pfn = pte_pfn(*ptep);
311         page = pfn_to_page(pfn);
312
313         pte = pfn_pte(pfn, prot);
314
315         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
316                 BUG();
317
318         if (!PageHighMem(page)) {
319                 void *av = __va(PFN_PHYS(pfn));
320
321                 if (av != v)
322                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
323                                 BUG();
324         } else
325                 kmap_flush_unused();
326 }
327
328 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
329 {
330         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
331         int i;
332
333         for(i = 0; i < entries; i += entries_per_page)
334                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
335 }
336
337 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
338 {
339         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
340         int i;
341
342         for(i = 0; i < entries; i += entries_per_page)
343                 set_aliased_prot(ldt + i, PAGE_KERNEL);
344 }
345
346 static void xen_set_ldt(const void *addr, unsigned entries)
347 {
348         struct mmuext_op *op;
349         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
350
351         op = mcs.args;
352         op->cmd = MMUEXT_SET_LDT;
353         op->arg1.linear_addr = (unsigned long)addr;
354         op->arg2.nr_ents = entries;
355
356         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
357
358         xen_mc_issue(PARAVIRT_LAZY_CPU);
359 }
360
361 static void xen_load_gdt(const struct desc_ptr *dtr)
362 {
363         unsigned long va = dtr->address;
364         unsigned int size = dtr->size + 1;
365         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
366         unsigned long frames[pages];
367         int f;
368
369         /*
370          * A GDT can be up to 64k in size, which corresponds to 8192
371          * 8-byte entries, or 16 4k pages..
372          */
373
374         BUG_ON(size > 65536);
375         BUG_ON(va & ~PAGE_MASK);
376
377         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
378                 int level;
379                 pte_t *ptep;
380                 unsigned long pfn, mfn;
381                 void *virt;
382
383                 /*
384                  * The GDT is per-cpu and is in the percpu data area.
385                  * That can be virtually mapped, so we need to do a
386                  * page-walk to get the underlying MFN for the
387                  * hypercall.  The page can also be in the kernel's
388                  * linear range, so we need to RO that mapping too.
389                  */
390                 ptep = lookup_address(va, &level);
391                 BUG_ON(ptep == NULL);
392
393                 pfn = pte_pfn(*ptep);
394                 mfn = pfn_to_mfn(pfn);
395                 virt = __va(PFN_PHYS(pfn));
396
397                 frames[f] = mfn;
398
399                 make_lowmem_page_readonly((void *)va);
400                 make_lowmem_page_readonly(virt);
401         }
402
403         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
404                 BUG();
405 }
406
407 /*
408  * load_gdt for early boot, when the gdt is only mapped once
409  */
410 static __init void xen_load_gdt_boot(const struct desc_ptr *dtr)
411 {
412         unsigned long va = dtr->address;
413         unsigned int size = dtr->size + 1;
414         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
415         unsigned long frames[pages];
416         int f;
417
418         /*
419          * A GDT can be up to 64k in size, which corresponds to 8192
420          * 8-byte entries, or 16 4k pages..
421          */
422
423         BUG_ON(size > 65536);
424         BUG_ON(va & ~PAGE_MASK);
425
426         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
427                 pte_t pte;
428                 unsigned long pfn, mfn;
429
430                 pfn = virt_to_pfn(va);
431                 mfn = pfn_to_mfn(pfn);
432
433                 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
434
435                 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
436                         BUG();
437
438                 frames[f] = mfn;
439         }
440
441         if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
442                 BUG();
443 }
444
445 static void load_TLS_descriptor(struct thread_struct *t,
446                                 unsigned int cpu, unsigned int i)
447 {
448         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
449         xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
450         struct multicall_space mc = __xen_mc_entry(0);
451
452         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
453 }
454
455 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
456 {
457         /*
458          * XXX sleazy hack: If we're being called in a lazy-cpu zone
459          * and lazy gs handling is enabled, it means we're in a
460          * context switch, and %gs has just been saved.  This means we
461          * can zero it out to prevent faults on exit from the
462          * hypervisor if the next process has no %gs.  Either way, it
463          * has been saved, and the new value will get loaded properly.
464          * This will go away as soon as Xen has been modified to not
465          * save/restore %gs for normal hypercalls.
466          *
467          * On x86_64, this hack is not used for %gs, because gs points
468          * to KERNEL_GS_BASE (and uses it for PDA references), so we
469          * must not zero %gs on x86_64
470          *
471          * For x86_64, we need to zero %fs, otherwise we may get an
472          * exception between the new %fs descriptor being loaded and
473          * %fs being effectively cleared at __switch_to().
474          */
475         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
476 #ifdef CONFIG_X86_32
477                 lazy_load_gs(0);
478 #else
479                 loadsegment(fs, 0);
480 #endif
481         }
482
483         xen_mc_batch();
484
485         load_TLS_descriptor(t, cpu, 0);
486         load_TLS_descriptor(t, cpu, 1);
487         load_TLS_descriptor(t, cpu, 2);
488
489         xen_mc_issue(PARAVIRT_LAZY_CPU);
490 }
491
492 #ifdef CONFIG_X86_64
493 static void xen_load_gs_index(unsigned int idx)
494 {
495         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
496                 BUG();
497 }
498 #endif
499
500 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
501                                 const void *ptr)
502 {
503         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
504         u64 entry = *(u64 *)ptr;
505
506         preempt_disable();
507
508         xen_mc_flush();
509         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
510                 BUG();
511
512         preempt_enable();
513 }
514
515 static int cvt_gate_to_trap(int vector, const gate_desc *val,
516                             struct trap_info *info)
517 {
518         unsigned long addr;
519
520         if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
521                 return 0;
522
523         info->vector = vector;
524
525         addr = gate_offset(*val);
526 #ifdef CONFIG_X86_64
527         /*
528          * Look for known traps using IST, and substitute them
529          * appropriately.  The debugger ones are the only ones we care
530          * about.  Xen will handle faults like double_fault and
531          * machine_check, so we should never see them.  Warn if
532          * there's an unexpected IST-using fault handler.
533          */
534         if (addr == (unsigned long)debug)
535                 addr = (unsigned long)xen_debug;
536         else if (addr == (unsigned long)int3)
537                 addr = (unsigned long)xen_int3;
538         else if (addr == (unsigned long)stack_segment)
539                 addr = (unsigned long)xen_stack_segment;
540         else if (addr == (unsigned long)double_fault ||
541                  addr == (unsigned long)nmi) {
542                 /* Don't need to handle these */
543                 return 0;
544 #ifdef CONFIG_X86_MCE
545         } else if (addr == (unsigned long)machine_check) {
546                 return 0;
547 #endif
548         } else {
549                 /* Some other trap using IST? */
550                 if (WARN_ON(val->ist != 0))
551                         return 0;
552         }
553 #endif  /* CONFIG_X86_64 */
554         info->address = addr;
555
556         info->cs = gate_segment(*val);
557         info->flags = val->dpl;
558         /* interrupt gates clear IF */
559         if (val->type == GATE_INTERRUPT)
560                 info->flags |= 1 << 2;
561
562         return 1;
563 }
564
565 /* Locations of each CPU's IDT */
566 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
567
568 /* Set an IDT entry.  If the entry is part of the current IDT, then
569    also update Xen. */
570 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
571 {
572         unsigned long p = (unsigned long)&dt[entrynum];
573         unsigned long start, end;
574
575         preempt_disable();
576
577         start = __get_cpu_var(idt_desc).address;
578         end = start + __get_cpu_var(idt_desc).size + 1;
579
580         xen_mc_flush();
581
582         native_write_idt_entry(dt, entrynum, g);
583
584         if (p >= start && (p + 8) <= end) {
585                 struct trap_info info[2];
586
587                 info[1].address = 0;
588
589                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
590                         if (HYPERVISOR_set_trap_table(info))
591                                 BUG();
592         }
593
594         preempt_enable();
595 }
596
597 static void xen_convert_trap_info(const struct desc_ptr *desc,
598                                   struct trap_info *traps)
599 {
600         unsigned in, out, count;
601
602         count = (desc->size+1) / sizeof(gate_desc);
603         BUG_ON(count > 256);
604
605         for (in = out = 0; in < count; in++) {
606                 gate_desc *entry = (gate_desc*)(desc->address) + in;
607
608                 if (cvt_gate_to_trap(in, entry, &traps[out]))
609                         out++;
610         }
611         traps[out].address = 0;
612 }
613
614 void xen_copy_trap_info(struct trap_info *traps)
615 {
616         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
617
618         xen_convert_trap_info(desc, traps);
619 }
620
621 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
622    hold a spinlock to protect the static traps[] array (static because
623    it avoids allocation, and saves stack space). */
624 static void xen_load_idt(const struct desc_ptr *desc)
625 {
626         static DEFINE_SPINLOCK(lock);
627         static struct trap_info traps[257];
628
629         spin_lock(&lock);
630
631         __get_cpu_var(idt_desc) = *desc;
632
633         xen_convert_trap_info(desc, traps);
634
635         xen_mc_flush();
636         if (HYPERVISOR_set_trap_table(traps))
637                 BUG();
638
639         spin_unlock(&lock);
640 }
641
642 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
643    they're handled differently. */
644 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
645                                 const void *desc, int type)
646 {
647         preempt_disable();
648
649         switch (type) {
650         case DESC_LDT:
651         case DESC_TSS:
652                 /* ignore */
653                 break;
654
655         default: {
656                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
657
658                 xen_mc_flush();
659                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
660                         BUG();
661         }
662
663         }
664
665         preempt_enable();
666 }
667
668 /*
669  * Version of write_gdt_entry for use at early boot-time needed to
670  * update an entry as simply as possible.
671  */
672 static __init void xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
673                                             const void *desc, int type)
674 {
675         switch (type) {
676         case DESC_LDT:
677         case DESC_TSS:
678                 /* ignore */
679                 break;
680
681         default: {
682                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
683
684                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
685                         dt[entry] = *(struct desc_struct *)desc;
686         }
687
688         }
689 }
690
691 static void xen_load_sp0(struct tss_struct *tss,
692                          struct thread_struct *thread)
693 {
694         struct multicall_space mcs = xen_mc_entry(0);
695         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
696         xen_mc_issue(PARAVIRT_LAZY_CPU);
697 }
698
699 static void xen_set_iopl_mask(unsigned mask)
700 {
701         struct physdev_set_iopl set_iopl;
702
703         /* Force the change at ring 0. */
704         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
705         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
706 }
707
708 static void xen_io_delay(void)
709 {
710 }
711
712 #ifdef CONFIG_X86_LOCAL_APIC
713 static u32 xen_apic_read(u32 reg)
714 {
715         return 0;
716 }
717
718 static void xen_apic_write(u32 reg, u32 val)
719 {
720         /* Warn to see if there's any stray references */
721         WARN_ON(1);
722 }
723
724 static u64 xen_apic_icr_read(void)
725 {
726         return 0;
727 }
728
729 static void xen_apic_icr_write(u32 low, u32 id)
730 {
731         /* Warn to see if there's any stray references */
732         WARN_ON(1);
733 }
734
735 static void xen_apic_wait_icr_idle(void)
736 {
737         return;
738 }
739
740 static u32 xen_safe_apic_wait_icr_idle(void)
741 {
742         return 0;
743 }
744
745 static void set_xen_basic_apic_ops(void)
746 {
747         apic->read = xen_apic_read;
748         apic->write = xen_apic_write;
749         apic->icr_read = xen_apic_icr_read;
750         apic->icr_write = xen_apic_icr_write;
751         apic->wait_icr_idle = xen_apic_wait_icr_idle;
752         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
753 }
754
755 #endif
756
757 static void xen_clts(void)
758 {
759         struct multicall_space mcs;
760
761         mcs = xen_mc_entry(0);
762
763         MULTI_fpu_taskswitch(mcs.mc, 0);
764
765         xen_mc_issue(PARAVIRT_LAZY_CPU);
766 }
767
768 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
769
770 static unsigned long xen_read_cr0(void)
771 {
772         unsigned long cr0 = percpu_read(xen_cr0_value);
773
774         if (unlikely(cr0 == 0)) {
775                 cr0 = native_read_cr0();
776                 percpu_write(xen_cr0_value, cr0);
777         }
778
779         return cr0;
780 }
781
782 static void xen_write_cr0(unsigned long cr0)
783 {
784         struct multicall_space mcs;
785
786         percpu_write(xen_cr0_value, cr0);
787
788         /* Only pay attention to cr0.TS; everything else is
789            ignored. */
790         mcs = xen_mc_entry(0);
791
792         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
793
794         xen_mc_issue(PARAVIRT_LAZY_CPU);
795 }
796
797 static void xen_write_cr4(unsigned long cr4)
798 {
799         cr4 &= ~X86_CR4_PGE;
800         cr4 &= ~X86_CR4_PSE;
801
802         native_write_cr4(cr4);
803 }
804
805 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
806 {
807         int ret;
808
809         ret = 0;
810
811         switch (msr) {
812 #ifdef CONFIG_X86_64
813                 unsigned which;
814                 u64 base;
815
816         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
817         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
818         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
819
820         set:
821                 base = ((u64)high << 32) | low;
822                 if (HYPERVISOR_set_segment_base(which, base) != 0)
823                         ret = -EIO;
824                 break;
825 #endif
826
827         case MSR_STAR:
828         case MSR_CSTAR:
829         case MSR_LSTAR:
830         case MSR_SYSCALL_MASK:
831         case MSR_IA32_SYSENTER_CS:
832         case MSR_IA32_SYSENTER_ESP:
833         case MSR_IA32_SYSENTER_EIP:
834                 /* Fast syscall setup is all done in hypercalls, so
835                    these are all ignored.  Stub them out here to stop
836                    Xen console noise. */
837                 break;
838
839         default:
840                 ret = native_write_msr_safe(msr, low, high);
841         }
842
843         return ret;
844 }
845
846 void xen_setup_shared_info(void)
847 {
848         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
849                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
850                            xen_start_info->shared_info);
851
852                 HYPERVISOR_shared_info =
853                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
854         } else
855                 HYPERVISOR_shared_info =
856                         (struct shared_info *)__va(xen_start_info->shared_info);
857
858 #ifndef CONFIG_SMP
859         /* In UP this is as good a place as any to set up shared info */
860         xen_setup_vcpu_info_placement();
861 #endif
862
863         xen_setup_mfn_list_list();
864 }
865
866 /* This is called once we have the cpu_possible_map */
867 void xen_setup_vcpu_info_placement(void)
868 {
869         int cpu;
870
871         for_each_possible_cpu(cpu)
872                 xen_vcpu_setup(cpu);
873
874         /* xen_vcpu_setup managed to place the vcpu_info within the
875            percpu area for all cpus, so make use of it */
876         if (have_vcpu_info_placement) {
877                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
878
879                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
880                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
881                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
882                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
883                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
884         }
885 }
886
887 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
888                           unsigned long addr, unsigned len)
889 {
890         char *start, *end, *reloc;
891         unsigned ret;
892
893         start = end = reloc = NULL;
894
895 #define SITE(op, x)                                                     \
896         case PARAVIRT_PATCH(op.x):                                      \
897         if (have_vcpu_info_placement) {                                 \
898                 start = (char *)xen_##x##_direct;                       \
899                 end = xen_##x##_direct_end;                             \
900                 reloc = xen_##x##_direct_reloc;                         \
901         }                                                               \
902         goto patch_site
903
904         switch (type) {
905                 SITE(pv_irq_ops, irq_enable);
906                 SITE(pv_irq_ops, irq_disable);
907                 SITE(pv_irq_ops, save_fl);
908                 SITE(pv_irq_ops, restore_fl);
909 #undef SITE
910
911         patch_site:
912                 if (start == NULL || (end-start) > len)
913                         goto default_patch;
914
915                 ret = paravirt_patch_insns(insnbuf, len, start, end);
916
917                 /* Note: because reloc is assigned from something that
918                    appears to be an array, gcc assumes it's non-null,
919                    but doesn't know its relationship with start and
920                    end. */
921                 if (reloc > start && reloc < end) {
922                         int reloc_off = reloc - start;
923                         long *relocp = (long *)(insnbuf + reloc_off);
924                         long delta = start - (char *)addr;
925
926                         *relocp += delta;
927                 }
928                 break;
929
930         default_patch:
931         default:
932                 ret = paravirt_patch_default(type, clobbers, insnbuf,
933                                              addr, len);
934                 break;
935         }
936
937         return ret;
938 }
939
940 static const struct pv_info xen_info __initdata = {
941         .paravirt_enabled = 1,
942         .shared_kernel_pmd = 0,
943
944         .name = "Xen",
945 };
946
947 static const struct pv_init_ops xen_init_ops __initdata = {
948         .patch = xen_patch,
949 };
950
951 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
952         .cpuid = xen_cpuid,
953
954         .set_debugreg = xen_set_debugreg,
955         .get_debugreg = xen_get_debugreg,
956
957         .clts = xen_clts,
958
959         .read_cr0 = xen_read_cr0,
960         .write_cr0 = xen_write_cr0,
961
962         .read_cr4 = native_read_cr4,
963         .read_cr4_safe = native_read_cr4_safe,
964         .write_cr4 = xen_write_cr4,
965
966         .wbinvd = native_wbinvd,
967
968         .read_msr = native_read_msr_safe,
969         .write_msr = xen_write_msr_safe,
970         .read_tsc = native_read_tsc,
971         .read_pmc = native_read_pmc,
972
973         .iret = xen_iret,
974         .irq_enable_sysexit = xen_sysexit,
975 #ifdef CONFIG_X86_64
976         .usergs_sysret32 = xen_sysret32,
977         .usergs_sysret64 = xen_sysret64,
978 #endif
979
980         .load_tr_desc = paravirt_nop,
981         .set_ldt = xen_set_ldt,
982         .load_gdt = xen_load_gdt,
983         .load_idt = xen_load_idt,
984         .load_tls = xen_load_tls,
985 #ifdef CONFIG_X86_64
986         .load_gs_index = xen_load_gs_index,
987 #endif
988
989         .alloc_ldt = xen_alloc_ldt,
990         .free_ldt = xen_free_ldt,
991
992         .store_gdt = native_store_gdt,
993         .store_idt = native_store_idt,
994         .store_tr = xen_store_tr,
995
996         .write_ldt_entry = xen_write_ldt_entry,
997         .write_gdt_entry = xen_write_gdt_entry,
998         .write_idt_entry = xen_write_idt_entry,
999         .load_sp0 = xen_load_sp0,
1000
1001         .set_iopl_mask = xen_set_iopl_mask,
1002         .io_delay = xen_io_delay,
1003
1004         /* Xen takes care of %gs when switching to usermode for us */
1005         .swapgs = paravirt_nop,
1006
1007         .start_context_switch = paravirt_start_context_switch,
1008         .end_context_switch = xen_end_context_switch,
1009 };
1010
1011 static const struct pv_apic_ops xen_apic_ops __initdata = {
1012 #ifdef CONFIG_X86_LOCAL_APIC
1013         .startup_ipi_hook = paravirt_nop,
1014 #endif
1015 };
1016
1017 static void xen_reboot(int reason)
1018 {
1019         struct sched_shutdown r = { .reason = reason };
1020
1021 #ifdef CONFIG_SMP
1022         smp_send_stop();
1023 #endif
1024
1025         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1026                 BUG();
1027 }
1028
1029 static void xen_restart(char *msg)
1030 {
1031         xen_reboot(SHUTDOWN_reboot);
1032 }
1033
1034 static void xen_emergency_restart(void)
1035 {
1036         xen_reboot(SHUTDOWN_reboot);
1037 }
1038
1039 static void xen_machine_halt(void)
1040 {
1041         xen_reboot(SHUTDOWN_poweroff);
1042 }
1043
1044 static void xen_crash_shutdown(struct pt_regs *regs)
1045 {
1046         xen_reboot(SHUTDOWN_crash);
1047 }
1048
1049 static int
1050 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1051 {
1052         xen_reboot(SHUTDOWN_crash);
1053         return NOTIFY_DONE;
1054 }
1055
1056 static struct notifier_block xen_panic_block = {
1057         .notifier_call= xen_panic_event,
1058 };
1059
1060 int xen_panic_handler_init(void)
1061 {
1062         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1063         return 0;
1064 }
1065
1066 static const struct machine_ops __initdata xen_machine_ops = {
1067         .restart = xen_restart,
1068         .halt = xen_machine_halt,
1069         .power_off = xen_machine_halt,
1070         .shutdown = xen_machine_halt,
1071         .crash_shutdown = xen_crash_shutdown,
1072         .emergency_restart = xen_emergency_restart,
1073 };
1074
1075 /*
1076  * Set up the GDT and segment registers for -fstack-protector.  Until
1077  * we do this, we have to be careful not to call any stack-protected
1078  * function, which is most of the kernel.
1079  */
1080 static void __init xen_setup_stackprotector(void)
1081 {
1082         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1083         pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1084
1085         setup_stack_canary_segment(0);
1086         switch_to_new_gdt(0);
1087
1088         pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1089         pv_cpu_ops.load_gdt = xen_load_gdt;
1090 }
1091
1092 /* First C function to be called on Xen boot */
1093 asmlinkage void __init xen_start_kernel(void)
1094 {
1095         pgd_t *pgd;
1096
1097         if (!xen_start_info)
1098                 return;
1099
1100         xen_domain_type = XEN_PV_DOMAIN;
1101
1102         /* Install Xen paravirt ops */
1103         pv_info = xen_info;
1104         pv_init_ops = xen_init_ops;
1105         pv_cpu_ops = xen_cpu_ops;
1106         pv_apic_ops = xen_apic_ops;
1107
1108         x86_init.resources.memory_setup = xen_memory_setup;
1109         x86_init.oem.arch_setup = xen_arch_setup;
1110         x86_init.oem.banner = xen_banner;
1111
1112         xen_init_time_ops();
1113
1114         /*
1115          * Set up some pagetable state before starting to set any ptes.
1116          */
1117
1118         xen_init_mmu_ops();
1119
1120         /* Prevent unwanted bits from being set in PTEs. */
1121         __supported_pte_mask &= ~_PAGE_GLOBAL;
1122         if (!xen_initial_domain())
1123                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1124
1125         __supported_pte_mask |= _PAGE_IOMAP;
1126
1127         /*
1128          * Prevent page tables from being allocated in highmem, even
1129          * if CONFIG_HIGHPTE is enabled.
1130          */
1131         __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1132
1133         /* Work out if we support NX */
1134         x86_configure_nx();
1135
1136         xen_setup_features();
1137
1138         /* Get mfn list */
1139         if (!xen_feature(XENFEAT_auto_translated_physmap))
1140                 xen_build_dynamic_phys_to_machine();
1141
1142         /*
1143          * Set up kernel GDT and segment registers, mainly so that
1144          * -fstack-protector code can be executed.
1145          */
1146         xen_setup_stackprotector();
1147
1148         xen_init_irq_ops();
1149         xen_init_cpuid_mask();
1150
1151 #ifdef CONFIG_X86_LOCAL_APIC
1152         /*
1153          * set up the basic apic ops.
1154          */
1155         set_xen_basic_apic_ops();
1156 #endif
1157
1158         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1159                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1160                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1161         }
1162
1163         machine_ops = xen_machine_ops;
1164
1165         /*
1166          * The only reliable way to retain the initial address of the
1167          * percpu gdt_page is to remember it here, so we can go and
1168          * mark it RW later, when the initial percpu area is freed.
1169          */
1170         xen_initial_gdt = &per_cpu(gdt_page, 0);
1171
1172         xen_smp_init();
1173
1174         pgd = (pgd_t *)xen_start_info->pt_base;
1175
1176         if (!xen_initial_domain())
1177                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1178
1179         __supported_pte_mask |= _PAGE_IOMAP;
1180         /* Don't do the full vcpu_info placement stuff until we have a
1181            possible map and a non-dummy shared_info. */
1182         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1183
1184         local_irq_disable();
1185         early_boot_irqs_off();
1186
1187         xen_raw_console_write("mapping kernel into physical memory\n");
1188         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1189
1190         init_mm.pgd = pgd;
1191
1192         /* keep using Xen gdt for now; no urgent need to change it */
1193
1194 #ifdef CONFIG_X86_32
1195         pv_info.kernel_rpl = 1;
1196         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1197                 pv_info.kernel_rpl = 0;
1198 #else
1199         pv_info.kernel_rpl = 0;
1200 #endif
1201
1202         /* set the limit of our address space */
1203         xen_reserve_top();
1204
1205 #ifdef CONFIG_X86_32
1206         /* set up basic CPUID stuff */
1207         cpu_detect(&new_cpu_data);
1208         new_cpu_data.hard_math = 1;
1209         new_cpu_data.wp_works_ok = 1;
1210         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1211 #endif
1212
1213         /* Poke various useful things into boot_params */
1214         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1215         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1216                 ? __pa(xen_start_info->mod_start) : 0;
1217         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1218         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1219
1220         if (!xen_initial_domain()) {
1221                 add_preferred_console("xenboot", 0, NULL);
1222                 add_preferred_console("tty", 0, NULL);
1223                 add_preferred_console("hvc", 0, NULL);
1224                 if (pci_xen)
1225                         x86_init.pci.arch_init = pci_xen_init;
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