Merge commit 'v2.6.28-rc9' into next
[linux-2.6] / 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/kernel.h>
15 #include <linux/init.h>
16 #include <linux/smp.h>
17 #include <linux/preempt.h>
18 #include <linux/hardirq.h>
19 #include <linux/percpu.h>
20 #include <linux/delay.h>
21 #include <linux/start_kernel.h>
22 #include <linux/sched.h>
23 #include <linux/bootmem.h>
24 #include <linux/module.h>
25 #include <linux/mm.h>
26 #include <linux/page-flags.h>
27 #include <linux/highmem.h>
28 #include <linux/console.h>
29
30 #include <xen/interface/xen.h>
31 #include <xen/interface/physdev.h>
32 #include <xen/interface/vcpu.h>
33 #include <xen/features.h>
34 #include <xen/page.h>
35 #include <xen/hvc-console.h>
36
37 #include <asm/paravirt.h>
38 #include <asm/apic.h>
39 #include <asm/page.h>
40 #include <asm/xen/hypercall.h>
41 #include <asm/xen/hypervisor.h>
42 #include <asm/fixmap.h>
43 #include <asm/processor.h>
44 #include <asm/msr-index.h>
45 #include <asm/setup.h>
46 #include <asm/desc.h>
47 #include <asm/pgtable.h>
48 #include <asm/tlbflush.h>
49 #include <asm/reboot.h>
50
51 #include "xen-ops.h"
52 #include "mmu.h"
53 #include "multicalls.h"
54
55 EXPORT_SYMBOL_GPL(hypercall_page);
56
57 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
58 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
59
60 enum xen_domain_type xen_domain_type = XEN_NATIVE;
61 EXPORT_SYMBOL_GPL(xen_domain_type);
62
63 /*
64  * Identity map, in addition to plain kernel map.  This needs to be
65  * large enough to allocate page table pages to allocate the rest.
66  * Each page can map 2MB.
67  */
68 static pte_t level1_ident_pgt[PTRS_PER_PTE * 4] __page_aligned_bss;
69
70 #ifdef CONFIG_X86_64
71 /* l3 pud for userspace vsyscall mapping */
72 static pud_t level3_user_vsyscall[PTRS_PER_PUD] __page_aligned_bss;
73 #endif /* CONFIG_X86_64 */
74
75 /*
76  * Note about cr3 (pagetable base) values:
77  *
78  * xen_cr3 contains the current logical cr3 value; it contains the
79  * last set cr3.  This may not be the current effective cr3, because
80  * its update may be being lazily deferred.  However, a vcpu looking
81  * at its own cr3 can use this value knowing that it everything will
82  * be self-consistent.
83  *
84  * xen_current_cr3 contains the actual vcpu cr3; it is set once the
85  * hypercall to set the vcpu cr3 is complete (so it may be a little
86  * out of date, but it will never be set early).  If one vcpu is
87  * looking at another vcpu's cr3 value, it should use this variable.
88  */
89 DEFINE_PER_CPU(unsigned long, xen_cr3);  /* cr3 stored as physaddr */
90 DEFINE_PER_CPU(unsigned long, xen_current_cr3);  /* actual vcpu cr3 */
91
92 struct start_info *xen_start_info;
93 EXPORT_SYMBOL_GPL(xen_start_info);
94
95 struct shared_info xen_dummy_shared_info;
96
97 /*
98  * Point at some empty memory to start with. We map the real shared_info
99  * page as soon as fixmap is up and running.
100  */
101 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
102
103 /*
104  * Flag to determine whether vcpu info placement is available on all
105  * VCPUs.  We assume it is to start with, and then set it to zero on
106  * the first failure.  This is because it can succeed on some VCPUs
107  * and not others, since it can involve hypervisor memory allocation,
108  * or because the guest failed to guarantee all the appropriate
109  * constraints on all VCPUs (ie buffer can't cross a page boundary).
110  *
111  * Note that any particular CPU may be using a placed vcpu structure,
112  * but we can only optimise if the all are.
113  *
114  * 0: not available, 1: available
115  */
116 static int have_vcpu_info_placement =
117 #ifdef CONFIG_X86_32
118         1
119 #else
120         0
121 #endif
122         ;
123
124
125 static void xen_vcpu_setup(int cpu)
126 {
127         struct vcpu_register_vcpu_info info;
128         int err;
129         struct vcpu_info *vcpup;
130
131         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
132         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
133
134         if (!have_vcpu_info_placement)
135                 return;         /* already tested, not available */
136
137         vcpup = &per_cpu(xen_vcpu_info, cpu);
138
139         info.mfn = virt_to_mfn(vcpup);
140         info.offset = offset_in_page(vcpup);
141
142         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
143                cpu, vcpup, info.mfn, info.offset);
144
145         /* Check to see if the hypervisor will put the vcpu_info
146            structure where we want it, which allows direct access via
147            a percpu-variable. */
148         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
149
150         if (err) {
151                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
152                 have_vcpu_info_placement = 0;
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                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
159                        cpu, vcpup);
160         }
161 }
162
163 /*
164  * On restore, set the vcpu placement up again.
165  * If it fails, then we're in a bad state, since
166  * we can't back out from using it...
167  */
168 void xen_vcpu_restore(void)
169 {
170         if (have_vcpu_info_placement) {
171                 int cpu;
172
173                 for_each_online_cpu(cpu) {
174                         bool other_cpu = (cpu != smp_processor_id());
175
176                         if (other_cpu &&
177                             HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
178                                 BUG();
179
180                         xen_vcpu_setup(cpu);
181
182                         if (other_cpu &&
183                             HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
184                                 BUG();
185                 }
186
187                 BUG_ON(!have_vcpu_info_placement);
188         }
189 }
190
191 static void __init xen_banner(void)
192 {
193         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
194         struct xen_extraversion extra;
195         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
196
197         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
198                pv_info.name);
199         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
200                version >> 16, version & 0xffff, extra.extraversion,
201                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
202 }
203
204 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
205                       unsigned int *cx, unsigned int *dx)
206 {
207         unsigned maskedx = ~0;
208
209         /*
210          * Mask out inconvenient features, to try and disable as many
211          * unsupported kernel subsystems as possible.
212          */
213         if (*ax == 1)
214                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
215                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
216                             (1 << X86_FEATURE_MCE)  |  /* disable MCE */
217                             (1 << X86_FEATURE_MCA)  |  /* disable MCA */
218                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
219
220         asm(XEN_EMULATE_PREFIX "cpuid"
221                 : "=a" (*ax),
222                   "=b" (*bx),
223                   "=c" (*cx),
224                   "=d" (*dx)
225                 : "0" (*ax), "2" (*cx));
226         *dx &= maskedx;
227 }
228
229 static void xen_set_debugreg(int reg, unsigned long val)
230 {
231         HYPERVISOR_set_debugreg(reg, val);
232 }
233
234 static unsigned long xen_get_debugreg(int reg)
235 {
236         return HYPERVISOR_get_debugreg(reg);
237 }
238
239 static void xen_leave_lazy(void)
240 {
241         paravirt_leave_lazy(paravirt_get_lazy_mode());
242         xen_mc_flush();
243 }
244
245 static unsigned long xen_store_tr(void)
246 {
247         return 0;
248 }
249
250 /*
251  * Set the page permissions for a particular virtual address.  If the
252  * address is a vmalloc mapping (or other non-linear mapping), then
253  * find the linear mapping of the page and also set its protections to
254  * match.
255  */
256 static void set_aliased_prot(void *v, pgprot_t prot)
257 {
258         int level;
259         pte_t *ptep;
260         pte_t pte;
261         unsigned long pfn;
262         struct page *page;
263
264         ptep = lookup_address((unsigned long)v, &level);
265         BUG_ON(ptep == NULL);
266
267         pfn = pte_pfn(*ptep);
268         page = pfn_to_page(pfn);
269
270         pte = pfn_pte(pfn, prot);
271
272         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
273                 BUG();
274
275         if (!PageHighMem(page)) {
276                 void *av = __va(PFN_PHYS(pfn));
277
278                 if (av != v)
279                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
280                                 BUG();
281         } else
282                 kmap_flush_unused();
283 }
284
285 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
286 {
287         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
288         int i;
289
290         for(i = 0; i < entries; i += entries_per_page)
291                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
292 }
293
294 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
295 {
296         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
297         int i;
298
299         for(i = 0; i < entries; i += entries_per_page)
300                 set_aliased_prot(ldt + i, PAGE_KERNEL);
301 }
302
303 static void xen_set_ldt(const void *addr, unsigned entries)
304 {
305         struct mmuext_op *op;
306         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
307
308         op = mcs.args;
309         op->cmd = MMUEXT_SET_LDT;
310         op->arg1.linear_addr = (unsigned long)addr;
311         op->arg2.nr_ents = entries;
312
313         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
314
315         xen_mc_issue(PARAVIRT_LAZY_CPU);
316 }
317
318 static void xen_load_gdt(const struct desc_ptr *dtr)
319 {
320         unsigned long *frames;
321         unsigned long va = dtr->address;
322         unsigned int size = dtr->size + 1;
323         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
324         int f;
325         struct multicall_space mcs;
326
327         /* A GDT can be up to 64k in size, which corresponds to 8192
328            8-byte entries, or 16 4k pages.. */
329
330         BUG_ON(size > 65536);
331         BUG_ON(va & ~PAGE_MASK);
332
333         mcs = xen_mc_entry(sizeof(*frames) * pages);
334         frames = mcs.args;
335
336         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
337                 frames[f] = virt_to_mfn(va);
338                 make_lowmem_page_readonly((void *)va);
339         }
340
341         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
342
343         xen_mc_issue(PARAVIRT_LAZY_CPU);
344 }
345
346 static void load_TLS_descriptor(struct thread_struct *t,
347                                 unsigned int cpu, unsigned int i)
348 {
349         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
350         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
351         struct multicall_space mc = __xen_mc_entry(0);
352
353         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
354 }
355
356 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
357 {
358         /*
359          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
360          * it means we're in a context switch, and %gs has just been
361          * saved.  This means we can zero it out to prevent faults on
362          * exit from the hypervisor if the next process has no %gs.
363          * Either way, it has been saved, and the new value will get
364          * loaded properly.  This will go away as soon as Xen has been
365          * modified to not save/restore %gs for normal hypercalls.
366          *
367          * On x86_64, this hack is not used for %gs, because gs points
368          * to KERNEL_GS_BASE (and uses it for PDA references), so we
369          * must not zero %gs on x86_64
370          *
371          * For x86_64, we need to zero %fs, otherwise we may get an
372          * exception between the new %fs descriptor being loaded and
373          * %fs being effectively cleared at __switch_to().
374          */
375         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
376 #ifdef CONFIG_X86_32
377                 loadsegment(gs, 0);
378 #else
379                 loadsegment(fs, 0);
380 #endif
381         }
382
383         xen_mc_batch();
384
385         load_TLS_descriptor(t, cpu, 0);
386         load_TLS_descriptor(t, cpu, 1);
387         load_TLS_descriptor(t, cpu, 2);
388
389         xen_mc_issue(PARAVIRT_LAZY_CPU);
390 }
391
392 #ifdef CONFIG_X86_64
393 static void xen_load_gs_index(unsigned int idx)
394 {
395         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
396                 BUG();
397 }
398 #endif
399
400 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
401                                 const void *ptr)
402 {
403         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
404         u64 entry = *(u64 *)ptr;
405
406         preempt_disable();
407
408         xen_mc_flush();
409         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
410                 BUG();
411
412         preempt_enable();
413 }
414
415 static int cvt_gate_to_trap(int vector, const gate_desc *val,
416                             struct trap_info *info)
417 {
418         if (val->type != 0xf && val->type != 0xe)
419                 return 0;
420
421         info->vector = vector;
422         info->address = gate_offset(*val);
423         info->cs = gate_segment(*val);
424         info->flags = val->dpl;
425         /* interrupt gates clear IF */
426         if (val->type == 0xe)
427                 info->flags |= 4;
428
429         return 1;
430 }
431
432 /* Locations of each CPU's IDT */
433 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
434
435 /* Set an IDT entry.  If the entry is part of the current IDT, then
436    also update Xen. */
437 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
438 {
439         unsigned long p = (unsigned long)&dt[entrynum];
440         unsigned long start, end;
441
442         preempt_disable();
443
444         start = __get_cpu_var(idt_desc).address;
445         end = start + __get_cpu_var(idt_desc).size + 1;
446
447         xen_mc_flush();
448
449         native_write_idt_entry(dt, entrynum, g);
450
451         if (p >= start && (p + 8) <= end) {
452                 struct trap_info info[2];
453
454                 info[1].address = 0;
455
456                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
457                         if (HYPERVISOR_set_trap_table(info))
458                                 BUG();
459         }
460
461         preempt_enable();
462 }
463
464 static void xen_convert_trap_info(const struct desc_ptr *desc,
465                                   struct trap_info *traps)
466 {
467         unsigned in, out, count;
468
469         count = (desc->size+1) / sizeof(gate_desc);
470         BUG_ON(count > 256);
471
472         for (in = out = 0; in < count; in++) {
473                 gate_desc *entry = (gate_desc*)(desc->address) + in;
474
475                 if (cvt_gate_to_trap(in, entry, &traps[out]))
476                         out++;
477         }
478         traps[out].address = 0;
479 }
480
481 void xen_copy_trap_info(struct trap_info *traps)
482 {
483         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
484
485         xen_convert_trap_info(desc, traps);
486 }
487
488 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
489    hold a spinlock to protect the static traps[] array (static because
490    it avoids allocation, and saves stack space). */
491 static void xen_load_idt(const struct desc_ptr *desc)
492 {
493         static DEFINE_SPINLOCK(lock);
494         static struct trap_info traps[257];
495
496         spin_lock(&lock);
497
498         __get_cpu_var(idt_desc) = *desc;
499
500         xen_convert_trap_info(desc, traps);
501
502         xen_mc_flush();
503         if (HYPERVISOR_set_trap_table(traps))
504                 BUG();
505
506         spin_unlock(&lock);
507 }
508
509 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
510    they're handled differently. */
511 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
512                                 const void *desc, int type)
513 {
514         preempt_disable();
515
516         switch (type) {
517         case DESC_LDT:
518         case DESC_TSS:
519                 /* ignore */
520                 break;
521
522         default: {
523                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
524
525                 xen_mc_flush();
526                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
527                         BUG();
528         }
529
530         }
531
532         preempt_enable();
533 }
534
535 static void xen_load_sp0(struct tss_struct *tss,
536                          struct thread_struct *thread)
537 {
538         struct multicall_space mcs = xen_mc_entry(0);
539         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
540         xen_mc_issue(PARAVIRT_LAZY_CPU);
541 }
542
543 static void xen_set_iopl_mask(unsigned mask)
544 {
545         struct physdev_set_iopl set_iopl;
546
547         /* Force the change at ring 0. */
548         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
549         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
550 }
551
552 static void xen_io_delay(void)
553 {
554 }
555
556 #ifdef CONFIG_X86_LOCAL_APIC
557 static u32 xen_apic_read(u32 reg)
558 {
559         return 0;
560 }
561
562 static void xen_apic_write(u32 reg, u32 val)
563 {
564         /* Warn to see if there's any stray references */
565         WARN_ON(1);
566 }
567
568 static u64 xen_apic_icr_read(void)
569 {
570         return 0;
571 }
572
573 static void xen_apic_icr_write(u32 low, u32 id)
574 {
575         /* Warn to see if there's any stray references */
576         WARN_ON(1);
577 }
578
579 static void xen_apic_wait_icr_idle(void)
580 {
581         return;
582 }
583
584 static u32 xen_safe_apic_wait_icr_idle(void)
585 {
586         return 0;
587 }
588
589 static struct apic_ops xen_basic_apic_ops = {
590         .read = xen_apic_read,
591         .write = xen_apic_write,
592         .icr_read = xen_apic_icr_read,
593         .icr_write = xen_apic_icr_write,
594         .wait_icr_idle = xen_apic_wait_icr_idle,
595         .safe_wait_icr_idle = xen_safe_apic_wait_icr_idle,
596 };
597
598 #endif
599
600 static void xen_flush_tlb(void)
601 {
602         struct mmuext_op *op;
603         struct multicall_space mcs;
604
605         preempt_disable();
606
607         mcs = xen_mc_entry(sizeof(*op));
608
609         op = mcs.args;
610         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
611         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
612
613         xen_mc_issue(PARAVIRT_LAZY_MMU);
614
615         preempt_enable();
616 }
617
618 static void xen_flush_tlb_single(unsigned long addr)
619 {
620         struct mmuext_op *op;
621         struct multicall_space mcs;
622
623         preempt_disable();
624
625         mcs = xen_mc_entry(sizeof(*op));
626         op = mcs.args;
627         op->cmd = MMUEXT_INVLPG_LOCAL;
628         op->arg1.linear_addr = addr & PAGE_MASK;
629         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
630
631         xen_mc_issue(PARAVIRT_LAZY_MMU);
632
633         preempt_enable();
634 }
635
636 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
637                                  unsigned long va)
638 {
639         struct {
640                 struct mmuext_op op;
641                 cpumask_t mask;
642         } *args;
643         cpumask_t cpumask = *cpus;
644         struct multicall_space mcs;
645
646         /*
647          * A couple of (to be removed) sanity checks:
648          *
649          * - current CPU must not be in mask
650          * - mask must exist :)
651          */
652         BUG_ON(cpus_empty(cpumask));
653         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
654         BUG_ON(!mm);
655
656         /* If a CPU which we ran on has gone down, OK. */
657         cpus_and(cpumask, cpumask, cpu_online_map);
658         if (cpus_empty(cpumask))
659                 return;
660
661         mcs = xen_mc_entry(sizeof(*args));
662         args = mcs.args;
663         args->mask = cpumask;
664         args->op.arg2.vcpumask = &args->mask;
665
666         if (va == TLB_FLUSH_ALL) {
667                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
668         } else {
669                 args->op.cmd = MMUEXT_INVLPG_MULTI;
670                 args->op.arg1.linear_addr = va;
671         }
672
673         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
674
675         xen_mc_issue(PARAVIRT_LAZY_MMU);
676 }
677
678 static void xen_clts(void)
679 {
680         struct multicall_space mcs;
681
682         mcs = xen_mc_entry(0);
683
684         MULTI_fpu_taskswitch(mcs.mc, 0);
685
686         xen_mc_issue(PARAVIRT_LAZY_CPU);
687 }
688
689 static void xen_write_cr0(unsigned long cr0)
690 {
691         struct multicall_space mcs;
692
693         /* Only pay attention to cr0.TS; everything else is
694            ignored. */
695         mcs = xen_mc_entry(0);
696
697         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
698
699         xen_mc_issue(PARAVIRT_LAZY_CPU);
700 }
701
702 static void xen_write_cr2(unsigned long cr2)
703 {
704         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
705 }
706
707 static unsigned long xen_read_cr2(void)
708 {
709         return x86_read_percpu(xen_vcpu)->arch.cr2;
710 }
711
712 static unsigned long xen_read_cr2_direct(void)
713 {
714         return x86_read_percpu(xen_vcpu_info.arch.cr2);
715 }
716
717 static void xen_write_cr4(unsigned long cr4)
718 {
719         cr4 &= ~X86_CR4_PGE;
720         cr4 &= ~X86_CR4_PSE;
721
722         native_write_cr4(cr4);
723 }
724
725 static unsigned long xen_read_cr3(void)
726 {
727         return x86_read_percpu(xen_cr3);
728 }
729
730 static void set_current_cr3(void *v)
731 {
732         x86_write_percpu(xen_current_cr3, (unsigned long)v);
733 }
734
735 static void __xen_write_cr3(bool kernel, unsigned long cr3)
736 {
737         struct mmuext_op *op;
738         struct multicall_space mcs;
739         unsigned long mfn;
740
741         if (cr3)
742                 mfn = pfn_to_mfn(PFN_DOWN(cr3));
743         else
744                 mfn = 0;
745
746         WARN_ON(mfn == 0 && kernel);
747
748         mcs = __xen_mc_entry(sizeof(*op));
749
750         op = mcs.args;
751         op->cmd = kernel ? MMUEXT_NEW_BASEPTR : MMUEXT_NEW_USER_BASEPTR;
752         op->arg1.mfn = mfn;
753
754         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
755
756         if (kernel) {
757                 x86_write_percpu(xen_cr3, cr3);
758
759                 /* Update xen_current_cr3 once the batch has actually
760                    been submitted. */
761                 xen_mc_callback(set_current_cr3, (void *)cr3);
762         }
763 }
764
765 static void xen_write_cr3(unsigned long cr3)
766 {
767         BUG_ON(preemptible());
768
769         xen_mc_batch();  /* disables interrupts */
770
771         /* Update while interrupts are disabled, so its atomic with
772            respect to ipis */
773         x86_write_percpu(xen_cr3, cr3);
774
775         __xen_write_cr3(true, cr3);
776
777 #ifdef CONFIG_X86_64
778         {
779                 pgd_t *user_pgd = xen_get_user_pgd(__va(cr3));
780                 if (user_pgd)
781                         __xen_write_cr3(false, __pa(user_pgd));
782                 else
783                         __xen_write_cr3(false, 0);
784         }
785 #endif
786
787         xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
788 }
789
790 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
791 {
792         int ret;
793
794         ret = 0;
795
796         switch(msr) {
797 #ifdef CONFIG_X86_64
798                 unsigned which;
799                 u64 base;
800
801         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
802         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
803         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
804
805         set:
806                 base = ((u64)high << 32) | low;
807                 if (HYPERVISOR_set_segment_base(which, base) != 0)
808                         ret = -EFAULT;
809                 break;
810 #endif
811
812         case MSR_STAR:
813         case MSR_CSTAR:
814         case MSR_LSTAR:
815         case MSR_SYSCALL_MASK:
816         case MSR_IA32_SYSENTER_CS:
817         case MSR_IA32_SYSENTER_ESP:
818         case MSR_IA32_SYSENTER_EIP:
819                 /* Fast syscall setup is all done in hypercalls, so
820                    these are all ignored.  Stub them out here to stop
821                    Xen console noise. */
822                 break;
823
824         default:
825                 ret = native_write_msr_safe(msr, low, high);
826         }
827
828         return ret;
829 }
830
831 /* Early in boot, while setting up the initial pagetable, assume
832    everything is pinned. */
833 static __init void xen_alloc_pte_init(struct mm_struct *mm, unsigned long pfn)
834 {
835 #ifdef CONFIG_FLATMEM
836         BUG_ON(mem_map);        /* should only be used early */
837 #endif
838         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
839 }
840
841 /* Early release_pte assumes that all pts are pinned, since there's
842    only init_mm and anything attached to that is pinned. */
843 static void xen_release_pte_init(unsigned long pfn)
844 {
845         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
846 }
847
848 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
849 {
850         struct mmuext_op op;
851         op.cmd = cmd;
852         op.arg1.mfn = pfn_to_mfn(pfn);
853         if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
854                 BUG();
855 }
856
857 /* This needs to make sure the new pte page is pinned iff its being
858    attached to a pinned pagetable. */
859 static void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn, unsigned level)
860 {
861         struct page *page = pfn_to_page(pfn);
862
863         if (PagePinned(virt_to_page(mm->pgd))) {
864                 SetPagePinned(page);
865
866                 vm_unmap_aliases();
867                 if (!PageHighMem(page)) {
868                         make_lowmem_page_readonly(__va(PFN_PHYS((unsigned long)pfn)));
869                         if (level == PT_PTE && USE_SPLIT_PTLOCKS)
870                                 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
871                 } else {
872                         /* make sure there are no stray mappings of
873                            this page */
874                         kmap_flush_unused();
875                 }
876         }
877 }
878
879 static void xen_alloc_pte(struct mm_struct *mm, unsigned long pfn)
880 {
881         xen_alloc_ptpage(mm, pfn, PT_PTE);
882 }
883
884 static void xen_alloc_pmd(struct mm_struct *mm, unsigned long pfn)
885 {
886         xen_alloc_ptpage(mm, pfn, PT_PMD);
887 }
888
889 static int xen_pgd_alloc(struct mm_struct *mm)
890 {
891         pgd_t *pgd = mm->pgd;
892         int ret = 0;
893
894         BUG_ON(PagePinned(virt_to_page(pgd)));
895
896 #ifdef CONFIG_X86_64
897         {
898                 struct page *page = virt_to_page(pgd);
899                 pgd_t *user_pgd;
900
901                 BUG_ON(page->private != 0);
902
903                 ret = -ENOMEM;
904
905                 user_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
906                 page->private = (unsigned long)user_pgd;
907
908                 if (user_pgd != NULL) {
909                         user_pgd[pgd_index(VSYSCALL_START)] =
910                                 __pgd(__pa(level3_user_vsyscall) | _PAGE_TABLE);
911                         ret = 0;
912                 }
913
914                 BUG_ON(PagePinned(virt_to_page(xen_get_user_pgd(pgd))));
915         }
916 #endif
917
918         return ret;
919 }
920
921 static void xen_pgd_free(struct mm_struct *mm, pgd_t *pgd)
922 {
923 #ifdef CONFIG_X86_64
924         pgd_t *user_pgd = xen_get_user_pgd(pgd);
925
926         if (user_pgd)
927                 free_page((unsigned long)user_pgd);
928 #endif
929 }
930
931 /* This should never happen until we're OK to use struct page */
932 static void xen_release_ptpage(unsigned long pfn, unsigned level)
933 {
934         struct page *page = pfn_to_page(pfn);
935
936         if (PagePinned(page)) {
937                 if (!PageHighMem(page)) {
938                         if (level == PT_PTE && USE_SPLIT_PTLOCKS)
939                                 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
940                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
941                 }
942                 ClearPagePinned(page);
943         }
944 }
945
946 static void xen_release_pte(unsigned long pfn)
947 {
948         xen_release_ptpage(pfn, PT_PTE);
949 }
950
951 static void xen_release_pmd(unsigned long pfn)
952 {
953         xen_release_ptpage(pfn, PT_PMD);
954 }
955
956 #if PAGETABLE_LEVELS == 4
957 static void xen_alloc_pud(struct mm_struct *mm, unsigned long pfn)
958 {
959         xen_alloc_ptpage(mm, pfn, PT_PUD);
960 }
961
962 static void xen_release_pud(unsigned long pfn)
963 {
964         xen_release_ptpage(pfn, PT_PUD);
965 }
966 #endif
967
968 #ifdef CONFIG_HIGHPTE
969 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
970 {
971         pgprot_t prot = PAGE_KERNEL;
972
973         if (PagePinned(page))
974                 prot = PAGE_KERNEL_RO;
975
976         if (0 && PageHighMem(page))
977                 printk("mapping highpte %lx type %d prot %s\n",
978                        page_to_pfn(page), type,
979                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
980
981         return kmap_atomic_prot(page, type, prot);
982 }
983 #endif
984
985 #ifdef CONFIG_X86_32
986 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
987 {
988         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
989         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
990                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
991                                pte_val_ma(pte));
992
993         return pte;
994 }
995
996 /* Init-time set_pte while constructing initial pagetables, which
997    doesn't allow RO pagetable pages to be remapped RW */
998 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
999 {
1000         pte = mask_rw_pte(ptep, pte);
1001
1002         xen_set_pte(ptep, pte);
1003 }
1004 #endif
1005
1006 static __init void xen_pagetable_setup_start(pgd_t *base)
1007 {
1008 }
1009
1010 void xen_setup_shared_info(void)
1011 {
1012         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1013                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1014                            xen_start_info->shared_info);
1015
1016                 HYPERVISOR_shared_info =
1017                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1018         } else
1019                 HYPERVISOR_shared_info =
1020                         (struct shared_info *)__va(xen_start_info->shared_info);
1021
1022 #ifndef CONFIG_SMP
1023         /* In UP this is as good a place as any to set up shared info */
1024         xen_setup_vcpu_info_placement();
1025 #endif
1026
1027         xen_setup_mfn_list_list();
1028 }
1029
1030 static __init void xen_pagetable_setup_done(pgd_t *base)
1031 {
1032         xen_setup_shared_info();
1033 }
1034
1035 static __init void xen_post_allocator_init(void)
1036 {
1037         pv_mmu_ops.set_pte = xen_set_pte;
1038         pv_mmu_ops.set_pmd = xen_set_pmd;
1039         pv_mmu_ops.set_pud = xen_set_pud;
1040 #if PAGETABLE_LEVELS == 4
1041         pv_mmu_ops.set_pgd = xen_set_pgd;
1042 #endif
1043
1044         /* This will work as long as patching hasn't happened yet
1045            (which it hasn't) */
1046         pv_mmu_ops.alloc_pte = xen_alloc_pte;
1047         pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
1048         pv_mmu_ops.release_pte = xen_release_pte;
1049         pv_mmu_ops.release_pmd = xen_release_pmd;
1050 #if PAGETABLE_LEVELS == 4
1051         pv_mmu_ops.alloc_pud = xen_alloc_pud;
1052         pv_mmu_ops.release_pud = xen_release_pud;
1053 #endif
1054
1055 #ifdef CONFIG_X86_64
1056         SetPagePinned(virt_to_page(level3_user_vsyscall));
1057 #endif
1058         xen_mark_init_mm_pinned();
1059 }
1060
1061 /* This is called once we have the cpu_possible_map */
1062 void xen_setup_vcpu_info_placement(void)
1063 {
1064         int cpu;
1065
1066         for_each_possible_cpu(cpu)
1067                 xen_vcpu_setup(cpu);
1068
1069         /* xen_vcpu_setup managed to place the vcpu_info within the
1070            percpu area for all cpus, so make use of it */
1071         if (have_vcpu_info_placement) {
1072                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
1073
1074                 pv_irq_ops.save_fl = xen_save_fl_direct;
1075                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
1076                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
1077                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
1078                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1079         }
1080 }
1081
1082 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1083                           unsigned long addr, unsigned len)
1084 {
1085         char *start, *end, *reloc;
1086         unsigned ret;
1087
1088         start = end = reloc = NULL;
1089
1090 #define SITE(op, x)                                                     \
1091         case PARAVIRT_PATCH(op.x):                                      \
1092         if (have_vcpu_info_placement) {                                 \
1093                 start = (char *)xen_##x##_direct;                       \
1094                 end = xen_##x##_direct_end;                             \
1095                 reloc = xen_##x##_direct_reloc;                         \
1096         }                                                               \
1097         goto patch_site
1098
1099         switch (type) {
1100                 SITE(pv_irq_ops, irq_enable);
1101                 SITE(pv_irq_ops, irq_disable);
1102                 SITE(pv_irq_ops, save_fl);
1103                 SITE(pv_irq_ops, restore_fl);
1104 #undef SITE
1105
1106         patch_site:
1107                 if (start == NULL || (end-start) > len)
1108                         goto default_patch;
1109
1110                 ret = paravirt_patch_insns(insnbuf, len, start, end);
1111
1112                 /* Note: because reloc is assigned from something that
1113                    appears to be an array, gcc assumes it's non-null,
1114                    but doesn't know its relationship with start and
1115                    end. */
1116                 if (reloc > start && reloc < end) {
1117                         int reloc_off = reloc - start;
1118                         long *relocp = (long *)(insnbuf + reloc_off);
1119                         long delta = start - (char *)addr;
1120
1121                         *relocp += delta;
1122                 }
1123                 break;
1124
1125         default_patch:
1126         default:
1127                 ret = paravirt_patch_default(type, clobbers, insnbuf,
1128                                              addr, len);
1129                 break;
1130         }
1131
1132         return ret;
1133 }
1134
1135 static void xen_set_fixmap(unsigned idx, unsigned long phys, pgprot_t prot)
1136 {
1137         pte_t pte;
1138
1139         phys >>= PAGE_SHIFT;
1140
1141         switch (idx) {
1142         case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
1143 #ifdef CONFIG_X86_F00F_BUG
1144         case FIX_F00F_IDT:
1145 #endif
1146 #ifdef CONFIG_X86_32
1147         case FIX_WP_TEST:
1148         case FIX_VDSO:
1149 # ifdef CONFIG_HIGHMEM
1150         case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1151 # endif
1152 #else
1153         case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1154 #endif
1155 #ifdef CONFIG_X86_LOCAL_APIC
1156         case FIX_APIC_BASE:     /* maps dummy local APIC */
1157 #endif
1158                 pte = pfn_pte(phys, prot);
1159                 break;
1160
1161         default:
1162                 pte = mfn_pte(phys, prot);
1163                 break;
1164         }
1165
1166         __native_set_fixmap(idx, pte);
1167
1168 #ifdef CONFIG_X86_64
1169         /* Replicate changes to map the vsyscall page into the user
1170            pagetable vsyscall mapping. */
1171         if (idx >= VSYSCALL_LAST_PAGE && idx <= VSYSCALL_FIRST_PAGE) {
1172                 unsigned long vaddr = __fix_to_virt(idx);
1173                 set_pte_vaddr_pud(level3_user_vsyscall, vaddr, pte);
1174         }
1175 #endif
1176 }
1177
1178 static const struct pv_info xen_info __initdata = {
1179         .paravirt_enabled = 1,
1180         .shared_kernel_pmd = 0,
1181
1182         .name = "Xen",
1183 };
1184
1185 static const struct pv_init_ops xen_init_ops __initdata = {
1186         .patch = xen_patch,
1187
1188         .banner = xen_banner,
1189         .memory_setup = xen_memory_setup,
1190         .arch_setup = xen_arch_setup,
1191         .post_allocator_init = xen_post_allocator_init,
1192 };
1193
1194 static const struct pv_time_ops xen_time_ops __initdata = {
1195         .time_init = xen_time_init,
1196
1197         .set_wallclock = xen_set_wallclock,
1198         .get_wallclock = xen_get_wallclock,
1199         .get_tsc_khz = xen_tsc_khz,
1200         .sched_clock = xen_sched_clock,
1201 };
1202
1203 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
1204         .cpuid = xen_cpuid,
1205
1206         .set_debugreg = xen_set_debugreg,
1207         .get_debugreg = xen_get_debugreg,
1208
1209         .clts = xen_clts,
1210
1211         .read_cr0 = native_read_cr0,
1212         .write_cr0 = xen_write_cr0,
1213
1214         .read_cr4 = native_read_cr4,
1215         .read_cr4_safe = native_read_cr4_safe,
1216         .write_cr4 = xen_write_cr4,
1217
1218         .wbinvd = native_wbinvd,
1219
1220         .read_msr = native_read_msr_safe,
1221         .write_msr = xen_write_msr_safe,
1222         .read_tsc = native_read_tsc,
1223         .read_pmc = native_read_pmc,
1224
1225         .iret = xen_iret,
1226         .irq_enable_sysexit = xen_sysexit,
1227 #ifdef CONFIG_X86_64
1228         .usergs_sysret32 = xen_sysret32,
1229         .usergs_sysret64 = xen_sysret64,
1230 #endif
1231
1232         .load_tr_desc = paravirt_nop,
1233         .set_ldt = xen_set_ldt,
1234         .load_gdt = xen_load_gdt,
1235         .load_idt = xen_load_idt,
1236         .load_tls = xen_load_tls,
1237 #ifdef CONFIG_X86_64
1238         .load_gs_index = xen_load_gs_index,
1239 #endif
1240
1241         .alloc_ldt = xen_alloc_ldt,
1242         .free_ldt = xen_free_ldt,
1243
1244         .store_gdt = native_store_gdt,
1245         .store_idt = native_store_idt,
1246         .store_tr = xen_store_tr,
1247
1248         .write_ldt_entry = xen_write_ldt_entry,
1249         .write_gdt_entry = xen_write_gdt_entry,
1250         .write_idt_entry = xen_write_idt_entry,
1251         .load_sp0 = xen_load_sp0,
1252
1253         .set_iopl_mask = xen_set_iopl_mask,
1254         .io_delay = xen_io_delay,
1255
1256         /* Xen takes care of %gs when switching to usermode for us */
1257         .swapgs = paravirt_nop,
1258
1259         .lazy_mode = {
1260                 .enter = paravirt_enter_lazy_cpu,
1261                 .leave = xen_leave_lazy,
1262         },
1263 };
1264
1265 static const struct pv_apic_ops xen_apic_ops __initdata = {
1266 #ifdef CONFIG_X86_LOCAL_APIC
1267         .setup_boot_clock = paravirt_nop,
1268         .setup_secondary_clock = paravirt_nop,
1269         .startup_ipi_hook = paravirt_nop,
1270 #endif
1271 };
1272
1273 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1274         .pagetable_setup_start = xen_pagetable_setup_start,
1275         .pagetable_setup_done = xen_pagetable_setup_done,
1276
1277         .read_cr2 = xen_read_cr2,
1278         .write_cr2 = xen_write_cr2,
1279
1280         .read_cr3 = xen_read_cr3,
1281         .write_cr3 = xen_write_cr3,
1282
1283         .flush_tlb_user = xen_flush_tlb,
1284         .flush_tlb_kernel = xen_flush_tlb,
1285         .flush_tlb_single = xen_flush_tlb_single,
1286         .flush_tlb_others = xen_flush_tlb_others,
1287
1288         .pte_update = paravirt_nop,
1289         .pte_update_defer = paravirt_nop,
1290
1291         .pgd_alloc = xen_pgd_alloc,
1292         .pgd_free = xen_pgd_free,
1293
1294         .alloc_pte = xen_alloc_pte_init,
1295         .release_pte = xen_release_pte_init,
1296         .alloc_pmd = xen_alloc_pte_init,
1297         .alloc_pmd_clone = paravirt_nop,
1298         .release_pmd = xen_release_pte_init,
1299
1300 #ifdef CONFIG_HIGHPTE
1301         .kmap_atomic_pte = xen_kmap_atomic_pte,
1302 #endif
1303
1304 #ifdef CONFIG_X86_64
1305         .set_pte = xen_set_pte,
1306 #else
1307         .set_pte = xen_set_pte_init,
1308 #endif
1309         .set_pte_at = xen_set_pte_at,
1310         .set_pmd = xen_set_pmd_hyper,
1311
1312         .ptep_modify_prot_start = __ptep_modify_prot_start,
1313         .ptep_modify_prot_commit = __ptep_modify_prot_commit,
1314
1315         .pte_val = xen_pte_val,
1316         .pte_flags = native_pte_flags,
1317         .pgd_val = xen_pgd_val,
1318
1319         .make_pte = xen_make_pte,
1320         .make_pgd = xen_make_pgd,
1321
1322 #ifdef CONFIG_X86_PAE
1323         .set_pte_atomic = xen_set_pte_atomic,
1324         .set_pte_present = xen_set_pte_at,
1325         .pte_clear = xen_pte_clear,
1326         .pmd_clear = xen_pmd_clear,
1327 #endif  /* CONFIG_X86_PAE */
1328         .set_pud = xen_set_pud_hyper,
1329
1330         .make_pmd = xen_make_pmd,
1331         .pmd_val = xen_pmd_val,
1332
1333 #if PAGETABLE_LEVELS == 4
1334         .pud_val = xen_pud_val,
1335         .make_pud = xen_make_pud,
1336         .set_pgd = xen_set_pgd_hyper,
1337
1338         .alloc_pud = xen_alloc_pte_init,
1339         .release_pud = xen_release_pte_init,
1340 #endif  /* PAGETABLE_LEVELS == 4 */
1341
1342         .activate_mm = xen_activate_mm,
1343         .dup_mmap = xen_dup_mmap,
1344         .exit_mmap = xen_exit_mmap,
1345
1346         .lazy_mode = {
1347                 .enter = paravirt_enter_lazy_mmu,
1348                 .leave = xen_leave_lazy,
1349         },
1350
1351         .set_fixmap = xen_set_fixmap,
1352 };
1353
1354 static void xen_reboot(int reason)
1355 {
1356         struct sched_shutdown r = { .reason = reason };
1357
1358 #ifdef CONFIG_SMP
1359         smp_send_stop();
1360 #endif
1361
1362         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1363                 BUG();
1364 }
1365
1366 static void xen_restart(char *msg)
1367 {
1368         xen_reboot(SHUTDOWN_reboot);
1369 }
1370
1371 static void xen_emergency_restart(void)
1372 {
1373         xen_reboot(SHUTDOWN_reboot);
1374 }
1375
1376 static void xen_machine_halt(void)
1377 {
1378         xen_reboot(SHUTDOWN_poweroff);
1379 }
1380
1381 static void xen_crash_shutdown(struct pt_regs *regs)
1382 {
1383         xen_reboot(SHUTDOWN_crash);
1384 }
1385
1386 static const struct machine_ops __initdata xen_machine_ops = {
1387         .restart = xen_restart,
1388         .halt = xen_machine_halt,
1389         .power_off = xen_machine_halt,
1390         .shutdown = xen_machine_halt,
1391         .crash_shutdown = xen_crash_shutdown,
1392         .emergency_restart = xen_emergency_restart,
1393 };
1394
1395
1396 static void __init xen_reserve_top(void)
1397 {
1398 #ifdef CONFIG_X86_32
1399         unsigned long top = HYPERVISOR_VIRT_START;
1400         struct xen_platform_parameters pp;
1401
1402         if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1403                 top = pp.virt_start;
1404
1405         reserve_top_address(-top);
1406 #endif  /* CONFIG_X86_32 */
1407 }
1408
1409 /*
1410  * Like __va(), but returns address in the kernel mapping (which is
1411  * all we have until the physical memory mapping has been set up.
1412  */
1413 static void *__ka(phys_addr_t paddr)
1414 {
1415 #ifdef CONFIG_X86_64
1416         return (void *)(paddr + __START_KERNEL_map);
1417 #else
1418         return __va(paddr);
1419 #endif
1420 }
1421
1422 /* Convert a machine address to physical address */
1423 static unsigned long m2p(phys_addr_t maddr)
1424 {
1425         phys_addr_t paddr;
1426
1427         maddr &= PTE_PFN_MASK;
1428         paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1429
1430         return paddr;
1431 }
1432
1433 /* Convert a machine address to kernel virtual */
1434 static void *m2v(phys_addr_t maddr)
1435 {
1436         return __ka(m2p(maddr));
1437 }
1438
1439 static void set_page_prot(void *addr, pgprot_t prot)
1440 {
1441         unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1442         pte_t pte = pfn_pte(pfn, prot);
1443
1444         if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1445                 BUG();
1446 }
1447
1448 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1449 {
1450         unsigned pmdidx, pteidx;
1451         unsigned ident_pte;
1452         unsigned long pfn;
1453
1454         ident_pte = 0;
1455         pfn = 0;
1456         for(pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1457                 pte_t *pte_page;
1458
1459                 /* Reuse or allocate a page of ptes */
1460                 if (pmd_present(pmd[pmdidx]))
1461                         pte_page = m2v(pmd[pmdidx].pmd);
1462                 else {
1463                         /* Check for free pte pages */
1464                         if (ident_pte == ARRAY_SIZE(level1_ident_pgt))
1465                                 break;
1466
1467                         pte_page = &level1_ident_pgt[ident_pte];
1468                         ident_pte += PTRS_PER_PTE;
1469
1470                         pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1471                 }
1472
1473                 /* Install mappings */
1474                 for(pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1475                         pte_t pte;
1476
1477                         if (pfn > max_pfn_mapped)
1478                                 max_pfn_mapped = pfn;
1479
1480                         if (!pte_none(pte_page[pteidx]))
1481                                 continue;
1482
1483                         pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1484                         pte_page[pteidx] = pte;
1485                 }
1486         }
1487
1488         for(pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1489                 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1490
1491         set_page_prot(pmd, PAGE_KERNEL_RO);
1492 }
1493
1494 #ifdef CONFIG_X86_64
1495 static void convert_pfn_mfn(void *v)
1496 {
1497         pte_t *pte = v;
1498         int i;
1499
1500         /* All levels are converted the same way, so just treat them
1501            as ptes. */
1502         for(i = 0; i < PTRS_PER_PTE; i++)
1503                 pte[i] = xen_make_pte(pte[i].pte);
1504 }
1505
1506 /*
1507  * Set up the inital kernel pagetable.
1508  *
1509  * We can construct this by grafting the Xen provided pagetable into
1510  * head_64.S's preconstructed pagetables.  We copy the Xen L2's into
1511  * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt.  This
1512  * means that only the kernel has a physical mapping to start with -
1513  * but that's enough to get __va working.  We need to fill in the rest
1514  * of the physical mapping once some sort of allocator has been set
1515  * up.
1516  */
1517 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1518 {
1519         pud_t *l3;
1520         pmd_t *l2;
1521
1522         /* Zap identity mapping */
1523         init_level4_pgt[0] = __pgd(0);
1524
1525         /* Pre-constructed entries are in pfn, so convert to mfn */
1526         convert_pfn_mfn(init_level4_pgt);
1527         convert_pfn_mfn(level3_ident_pgt);
1528         convert_pfn_mfn(level3_kernel_pgt);
1529
1530         l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1531         l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1532
1533         memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1534         memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1535
1536         l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1537         l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1538         memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1539
1540         /* Set up identity map */
1541         xen_map_identity_early(level2_ident_pgt, max_pfn);
1542
1543         /* Make pagetable pieces RO */
1544         set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1545         set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1546         set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1547         set_page_prot(level3_user_vsyscall, PAGE_KERNEL_RO);
1548         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1549         set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1550
1551         /* Pin down new L4 */
1552         pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1553                           PFN_DOWN(__pa_symbol(init_level4_pgt)));
1554
1555         /* Unpin Xen-provided one */
1556         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1557
1558         /* Switch over */
1559         pgd = init_level4_pgt;
1560
1561         /*
1562          * At this stage there can be no user pgd, and no page
1563          * structure to attach it to, so make sure we just set kernel
1564          * pgd.
1565          */
1566         xen_mc_batch();
1567         __xen_write_cr3(true, __pa(pgd));
1568         xen_mc_issue(PARAVIRT_LAZY_CPU);
1569
1570         reserve_early(__pa(xen_start_info->pt_base),
1571                       __pa(xen_start_info->pt_base +
1572                            xen_start_info->nr_pt_frames * PAGE_SIZE),
1573                       "XEN PAGETABLES");
1574
1575         return pgd;
1576 }
1577 #else   /* !CONFIG_X86_64 */
1578 static pmd_t level2_kernel_pgt[PTRS_PER_PMD] __page_aligned_bss;
1579
1580 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1581 {
1582         pmd_t *kernel_pmd;
1583
1584         init_pg_tables_start = __pa(pgd);
1585         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1586         max_pfn_mapped = PFN_DOWN(init_pg_tables_end + 512*1024);
1587
1588         kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1589         memcpy(level2_kernel_pgt, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1590
1591         xen_map_identity_early(level2_kernel_pgt, max_pfn);
1592
1593         memcpy(swapper_pg_dir, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1594         set_pgd(&swapper_pg_dir[KERNEL_PGD_BOUNDARY],
1595                         __pgd(__pa(level2_kernel_pgt) | _PAGE_PRESENT));
1596
1597         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1598         set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1599         set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1600
1601         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1602
1603         xen_write_cr3(__pa(swapper_pg_dir));
1604
1605         pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(swapper_pg_dir)));
1606
1607         return swapper_pg_dir;
1608 }
1609 #endif  /* CONFIG_X86_64 */
1610
1611 /* First C function to be called on Xen boot */
1612 asmlinkage void __init xen_start_kernel(void)
1613 {
1614         pgd_t *pgd;
1615
1616         if (!xen_start_info)
1617                 return;
1618
1619         xen_domain_type = XEN_PV_DOMAIN;
1620
1621         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
1622
1623         xen_setup_features();
1624
1625         /* Install Xen paravirt ops */
1626         pv_info = xen_info;
1627         pv_init_ops = xen_init_ops;
1628         pv_time_ops = xen_time_ops;
1629         pv_cpu_ops = xen_cpu_ops;
1630         pv_apic_ops = xen_apic_ops;
1631         pv_mmu_ops = xen_mmu_ops;
1632
1633         xen_init_irq_ops();
1634
1635 #ifdef CONFIG_X86_LOCAL_APIC
1636         /*
1637          * set up the basic apic ops.
1638          */
1639         apic_ops = &xen_basic_apic_ops;
1640 #endif
1641
1642         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1643                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1644                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1645         }
1646
1647         machine_ops = xen_machine_ops;
1648
1649 #ifdef CONFIG_X86_64
1650         /* Disable until direct per-cpu data access. */
1651         have_vcpu_info_placement = 0;
1652         x86_64_init_pda();
1653 #endif
1654
1655         xen_smp_init();
1656
1657         /* Get mfn list */
1658         if (!xen_feature(XENFEAT_auto_translated_physmap))
1659                 xen_build_dynamic_phys_to_machine();
1660
1661         pgd = (pgd_t *)xen_start_info->pt_base;
1662
1663         /* Prevent unwanted bits from being set in PTEs. */
1664         __supported_pte_mask &= ~_PAGE_GLOBAL;
1665         if (!xen_initial_domain())
1666                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1667
1668         /* Don't do the full vcpu_info placement stuff until we have a
1669            possible map and a non-dummy shared_info. */
1670         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1671
1672         xen_raw_console_write("mapping kernel into physical memory\n");
1673         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1674
1675         init_mm.pgd = pgd;
1676
1677         /* keep using Xen gdt for now; no urgent need to change it */
1678
1679         pv_info.kernel_rpl = 1;
1680         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1681                 pv_info.kernel_rpl = 0;
1682
1683         /* set the limit of our address space */
1684         xen_reserve_top();
1685
1686 #ifdef CONFIG_X86_32
1687         /* set up basic CPUID stuff */
1688         cpu_detect(&new_cpu_data);
1689         new_cpu_data.hard_math = 1;
1690         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1691 #endif
1692
1693         /* Poke various useful things into boot_params */
1694         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1695         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1696                 ? __pa(xen_start_info->mod_start) : 0;
1697         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1698         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1699
1700         if (!xen_initial_domain()) {
1701                 add_preferred_console("xenboot", 0, NULL);
1702                 add_preferred_console("tty", 0, NULL);
1703                 add_preferred_console("hvc", 0, NULL);
1704         }
1705
1706         xen_raw_console_write("about to get started...\n");
1707
1708         /* Start the world */
1709 #ifdef CONFIG_X86_32
1710         i386_start_kernel();
1711 #else
1712         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1713 #endif
1714 }