x86, agpgart: scary messages are fortunately obsolete
[sfrench/cifs-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/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
29 #include <xen/interface/xen.h>
30 #include <xen/interface/physdev.h>
31 #include <xen/interface/vcpu.h>
32 #include <xen/interface/sched.h>
33 #include <xen/features.h>
34 #include <xen/page.h>
35
36 #include <asm/paravirt.h>
37 #include <asm/page.h>
38 #include <asm/xen/hypercall.h>
39 #include <asm/xen/hypervisor.h>
40 #include <asm/fixmap.h>
41 #include <asm/processor.h>
42 #include <asm/setup.h>
43 #include <asm/desc.h>
44 #include <asm/pgtable.h>
45 #include <asm/tlbflush.h>
46 #include <asm/reboot.h>
47
48 #include "xen-ops.h"
49 #include "mmu.h"
50 #include "multicalls.h"
51
52 EXPORT_SYMBOL_GPL(hypercall_page);
53
54 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
55 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
56
57 /*
58  * Note about cr3 (pagetable base) values:
59  *
60  * xen_cr3 contains the current logical cr3 value; it contains the
61  * last set cr3.  This may not be the current effective cr3, because
62  * its update may be being lazily deferred.  However, a vcpu looking
63  * at its own cr3 can use this value knowing that it everything will
64  * be self-consistent.
65  *
66  * xen_current_cr3 contains the actual vcpu cr3; it is set once the
67  * hypercall to set the vcpu cr3 is complete (so it may be a little
68  * out of date, but it will never be set early).  If one vcpu is
69  * looking at another vcpu's cr3 value, it should use this variable.
70  */
71 DEFINE_PER_CPU(unsigned long, xen_cr3);  /* cr3 stored as physaddr */
72 DEFINE_PER_CPU(unsigned long, xen_current_cr3);  /* actual vcpu cr3 */
73
74 struct start_info *xen_start_info;
75 EXPORT_SYMBOL_GPL(xen_start_info);
76
77 static /* __initdata */ struct shared_info dummy_shared_info;
78
79 /*
80  * Point at some empty memory to start with. We map the real shared_info
81  * page as soon as fixmap is up and running.
82  */
83 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
84
85 /*
86  * Flag to determine whether vcpu info placement is available on all
87  * VCPUs.  We assume it is to start with, and then set it to zero on
88  * the first failure.  This is because it can succeed on some VCPUs
89  * and not others, since it can involve hypervisor memory allocation,
90  * or because the guest failed to guarantee all the appropriate
91  * constraints on all VCPUs (ie buffer can't cross a page boundary).
92  *
93  * Note that any particular CPU may be using a placed vcpu structure,
94  * but we can only optimise if the all are.
95  *
96  * 0: not available, 1: available
97  */
98 static int have_vcpu_info_placement = 1;
99
100 static void __init xen_vcpu_setup(int cpu)
101 {
102         struct vcpu_register_vcpu_info info;
103         int err;
104         struct vcpu_info *vcpup;
105
106         BUG_ON(HYPERVISOR_shared_info == &dummy_shared_info);
107         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
108
109         if (!have_vcpu_info_placement)
110                 return;         /* already tested, not available */
111
112         vcpup = &per_cpu(xen_vcpu_info, cpu);
113
114         info.mfn = virt_to_mfn(vcpup);
115         info.offset = offset_in_page(vcpup);
116
117         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
118                cpu, vcpup, info.mfn, info.offset);
119
120         /* Check to see if the hypervisor will put the vcpu_info
121            structure where we want it, which allows direct access via
122            a percpu-variable. */
123         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
124
125         if (err) {
126                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
127                 have_vcpu_info_placement = 0;
128         } else {
129                 /* This cpu is using the registered vcpu info, even if
130                    later ones fail to. */
131                 per_cpu(xen_vcpu, cpu) = vcpup;
132
133                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
134                        cpu, vcpup);
135         }
136 }
137
138 static void __init xen_banner(void)
139 {
140         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
141                pv_info.name);
142         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
143 }
144
145 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
146                       unsigned int *cx, unsigned int *dx)
147 {
148         unsigned maskedx = ~0;
149
150         /*
151          * Mask out inconvenient features, to try and disable as many
152          * unsupported kernel subsystems as possible.
153          */
154         if (*ax == 1)
155                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
156                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
157                             (1 << X86_FEATURE_SEP)  |  /* disable SEP */
158                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
159
160         asm(XEN_EMULATE_PREFIX "cpuid"
161                 : "=a" (*ax),
162                   "=b" (*bx),
163                   "=c" (*cx),
164                   "=d" (*dx)
165                 : "0" (*ax), "2" (*cx));
166         *dx &= maskedx;
167 }
168
169 static void xen_set_debugreg(int reg, unsigned long val)
170 {
171         HYPERVISOR_set_debugreg(reg, val);
172 }
173
174 static unsigned long xen_get_debugreg(int reg)
175 {
176         return HYPERVISOR_get_debugreg(reg);
177 }
178
179 static unsigned long xen_save_fl(void)
180 {
181         struct vcpu_info *vcpu;
182         unsigned long flags;
183
184         vcpu = x86_read_percpu(xen_vcpu);
185
186         /* flag has opposite sense of mask */
187         flags = !vcpu->evtchn_upcall_mask;
188
189         /* convert to IF type flag
190            -0 -> 0x00000000
191            -1 -> 0xffffffff
192         */
193         return (-flags) & X86_EFLAGS_IF;
194 }
195
196 static void xen_restore_fl(unsigned long flags)
197 {
198         struct vcpu_info *vcpu;
199
200         /* convert from IF type flag */
201         flags = !(flags & X86_EFLAGS_IF);
202
203         /* There's a one instruction preempt window here.  We need to
204            make sure we're don't switch CPUs between getting the vcpu
205            pointer and updating the mask. */
206         preempt_disable();
207         vcpu = x86_read_percpu(xen_vcpu);
208         vcpu->evtchn_upcall_mask = flags;
209         preempt_enable_no_resched();
210
211         /* Doesn't matter if we get preempted here, because any
212            pending event will get dealt with anyway. */
213
214         if (flags == 0) {
215                 preempt_check_resched();
216                 barrier(); /* unmask then check (avoid races) */
217                 if (unlikely(vcpu->evtchn_upcall_pending))
218                         force_evtchn_callback();
219         }
220 }
221
222 static void xen_irq_disable(void)
223 {
224         /* There's a one instruction preempt window here.  We need to
225            make sure we're don't switch CPUs between getting the vcpu
226            pointer and updating the mask. */
227         preempt_disable();
228         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
229         preempt_enable_no_resched();
230 }
231
232 static void xen_irq_enable(void)
233 {
234         struct vcpu_info *vcpu;
235
236         /* There's a one instruction preempt window here.  We need to
237            make sure we're don't switch CPUs between getting the vcpu
238            pointer and updating the mask. */
239         preempt_disable();
240         vcpu = x86_read_percpu(xen_vcpu);
241         vcpu->evtchn_upcall_mask = 0;
242         preempt_enable_no_resched();
243
244         /* Doesn't matter if we get preempted here, because any
245            pending event will get dealt with anyway. */
246
247         barrier(); /* unmask then check (avoid races) */
248         if (unlikely(vcpu->evtchn_upcall_pending))
249                 force_evtchn_callback();
250 }
251
252 static void xen_safe_halt(void)
253 {
254         /* Blocking includes an implicit local_irq_enable(). */
255         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
256                 BUG();
257 }
258
259 static void xen_halt(void)
260 {
261         if (irqs_disabled())
262                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
263         else
264                 xen_safe_halt();
265 }
266
267 static void xen_leave_lazy(void)
268 {
269         paravirt_leave_lazy(paravirt_get_lazy_mode());
270         xen_mc_flush();
271 }
272
273 static unsigned long xen_store_tr(void)
274 {
275         return 0;
276 }
277
278 static void xen_set_ldt(const void *addr, unsigned entries)
279 {
280         struct mmuext_op *op;
281         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
282
283         op = mcs.args;
284         op->cmd = MMUEXT_SET_LDT;
285         op->arg1.linear_addr = (unsigned long)addr;
286         op->arg2.nr_ents = entries;
287
288         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
289
290         xen_mc_issue(PARAVIRT_LAZY_CPU);
291 }
292
293 static void xen_load_gdt(const struct desc_ptr *dtr)
294 {
295         unsigned long *frames;
296         unsigned long va = dtr->address;
297         unsigned int size = dtr->size + 1;
298         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
299         int f;
300         struct multicall_space mcs;
301
302         /* A GDT can be up to 64k in size, which corresponds to 8192
303            8-byte entries, or 16 4k pages.. */
304
305         BUG_ON(size > 65536);
306         BUG_ON(va & ~PAGE_MASK);
307
308         mcs = xen_mc_entry(sizeof(*frames) * pages);
309         frames = mcs.args;
310
311         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
312                 frames[f] = virt_to_mfn(va);
313                 make_lowmem_page_readonly((void *)va);
314         }
315
316         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
317
318         xen_mc_issue(PARAVIRT_LAZY_CPU);
319 }
320
321 static void load_TLS_descriptor(struct thread_struct *t,
322                                 unsigned int cpu, unsigned int i)
323 {
324         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
325         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
326         struct multicall_space mc = __xen_mc_entry(0);
327
328         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
329 }
330
331 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
332 {
333         xen_mc_batch();
334
335         load_TLS_descriptor(t, cpu, 0);
336         load_TLS_descriptor(t, cpu, 1);
337         load_TLS_descriptor(t, cpu, 2);
338
339         xen_mc_issue(PARAVIRT_LAZY_CPU);
340
341         /*
342          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
343          * it means we're in a context switch, and %gs has just been
344          * saved.  This means we can zero it out to prevent faults on
345          * exit from the hypervisor if the next process has no %gs.
346          * Either way, it has been saved, and the new value will get
347          * loaded properly.  This will go away as soon as Xen has been
348          * modified to not save/restore %gs for normal hypercalls.
349          */
350         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
351                 loadsegment(gs, 0);
352 }
353
354 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
355                                 const void *ptr)
356 {
357         unsigned long lp = (unsigned long)&dt[entrynum];
358         xmaddr_t mach_lp = virt_to_machine(lp);
359         u64 entry = *(u64 *)ptr;
360
361         preempt_disable();
362
363         xen_mc_flush();
364         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
365                 BUG();
366
367         preempt_enable();
368 }
369
370 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
371                             struct trap_info *info)
372 {
373         u8 type, dpl;
374
375         type = (high >> 8) & 0x1f;
376         dpl = (high >> 13) & 3;
377
378         if (type != 0xf && type != 0xe)
379                 return 0;
380
381         info->vector = vector;
382         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
383         info->cs = low >> 16;
384         info->flags = dpl;
385         /* interrupt gates clear IF */
386         if (type == 0xe)
387                 info->flags |= 4;
388
389         return 1;
390 }
391
392 /* Locations of each CPU's IDT */
393 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
394
395 /* Set an IDT entry.  If the entry is part of the current IDT, then
396    also update Xen. */
397 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
398 {
399         unsigned long p = (unsigned long)&dt[entrynum];
400         unsigned long start, end;
401
402         preempt_disable();
403
404         start = __get_cpu_var(idt_desc).address;
405         end = start + __get_cpu_var(idt_desc).size + 1;
406
407         xen_mc_flush();
408
409         native_write_idt_entry(dt, entrynum, g);
410
411         if (p >= start && (p + 8) <= end) {
412                 struct trap_info info[2];
413                 u32 *desc = (u32 *)g;
414
415                 info[1].address = 0;
416
417                 if (cvt_gate_to_trap(entrynum, desc[0], desc[1], &info[0]))
418                         if (HYPERVISOR_set_trap_table(info))
419                                 BUG();
420         }
421
422         preempt_enable();
423 }
424
425 static void xen_convert_trap_info(const struct desc_ptr *desc,
426                                   struct trap_info *traps)
427 {
428         unsigned in, out, count;
429
430         count = (desc->size+1) / 8;
431         BUG_ON(count > 256);
432
433         for (in = out = 0; in < count; in++) {
434                 const u32 *entry = (u32 *)(desc->address + in * 8);
435
436                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
437                         out++;
438         }
439         traps[out].address = 0;
440 }
441
442 void xen_copy_trap_info(struct trap_info *traps)
443 {
444         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
445
446         xen_convert_trap_info(desc, traps);
447 }
448
449 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
450    hold a spinlock to protect the static traps[] array (static because
451    it avoids allocation, and saves stack space). */
452 static void xen_load_idt(const struct desc_ptr *desc)
453 {
454         static DEFINE_SPINLOCK(lock);
455         static struct trap_info traps[257];
456
457         spin_lock(&lock);
458
459         __get_cpu_var(idt_desc) = *desc;
460
461         xen_convert_trap_info(desc, traps);
462
463         xen_mc_flush();
464         if (HYPERVISOR_set_trap_table(traps))
465                 BUG();
466
467         spin_unlock(&lock);
468 }
469
470 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
471    they're handled differently. */
472 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
473                                 const void *desc, int type)
474 {
475         preempt_disable();
476
477         switch (type) {
478         case DESC_LDT:
479         case DESC_TSS:
480                 /* ignore */
481                 break;
482
483         default: {
484                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
485
486                 xen_mc_flush();
487                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
488                         BUG();
489         }
490
491         }
492
493         preempt_enable();
494 }
495
496 static void xen_load_sp0(struct tss_struct *tss,
497                           struct thread_struct *thread)
498 {
499         struct multicall_space mcs = xen_mc_entry(0);
500         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
501         xen_mc_issue(PARAVIRT_LAZY_CPU);
502 }
503
504 static void xen_set_iopl_mask(unsigned mask)
505 {
506         struct physdev_set_iopl set_iopl;
507
508         /* Force the change at ring 0. */
509         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
510         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
511 }
512
513 static void xen_io_delay(void)
514 {
515 }
516
517 #ifdef CONFIG_X86_LOCAL_APIC
518 static u32 xen_apic_read(unsigned long reg)
519 {
520         return 0;
521 }
522
523 static void xen_apic_write(unsigned long reg, u32 val)
524 {
525         /* Warn to see if there's any stray references */
526         WARN_ON(1);
527 }
528 #endif
529
530 static void xen_flush_tlb(void)
531 {
532         struct mmuext_op *op;
533         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
534
535         op = mcs.args;
536         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
537         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
538
539         xen_mc_issue(PARAVIRT_LAZY_MMU);
540 }
541
542 static void xen_flush_tlb_single(unsigned long addr)
543 {
544         struct mmuext_op *op;
545         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
546
547         op = mcs.args;
548         op->cmd = MMUEXT_INVLPG_LOCAL;
549         op->arg1.linear_addr = addr & PAGE_MASK;
550         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
551
552         xen_mc_issue(PARAVIRT_LAZY_MMU);
553 }
554
555 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
556                                  unsigned long va)
557 {
558         struct {
559                 struct mmuext_op op;
560                 cpumask_t mask;
561         } *args;
562         cpumask_t cpumask = *cpus;
563         struct multicall_space mcs;
564
565         /*
566          * A couple of (to be removed) sanity checks:
567          *
568          * - current CPU must not be in mask
569          * - mask must exist :)
570          */
571         BUG_ON(cpus_empty(cpumask));
572         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
573         BUG_ON(!mm);
574
575         /* If a CPU which we ran on has gone down, OK. */
576         cpus_and(cpumask, cpumask, cpu_online_map);
577         if (cpus_empty(cpumask))
578                 return;
579
580         mcs = xen_mc_entry(sizeof(*args));
581         args = mcs.args;
582         args->mask = cpumask;
583         args->op.arg2.vcpumask = &args->mask;
584
585         if (va == TLB_FLUSH_ALL) {
586                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
587         } else {
588                 args->op.cmd = MMUEXT_INVLPG_MULTI;
589                 args->op.arg1.linear_addr = va;
590         }
591
592         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
593
594         xen_mc_issue(PARAVIRT_LAZY_MMU);
595 }
596
597 static void xen_write_cr2(unsigned long cr2)
598 {
599         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
600 }
601
602 static unsigned long xen_read_cr2(void)
603 {
604         return x86_read_percpu(xen_vcpu)->arch.cr2;
605 }
606
607 static unsigned long xen_read_cr2_direct(void)
608 {
609         return x86_read_percpu(xen_vcpu_info.arch.cr2);
610 }
611
612 static void xen_write_cr4(unsigned long cr4)
613 {
614         /* Just ignore cr4 changes; Xen doesn't allow us to do
615            anything anyway. */
616 }
617
618 static unsigned long xen_read_cr3(void)
619 {
620         return x86_read_percpu(xen_cr3);
621 }
622
623 static void set_current_cr3(void *v)
624 {
625         x86_write_percpu(xen_current_cr3, (unsigned long)v);
626 }
627
628 static void xen_write_cr3(unsigned long cr3)
629 {
630         struct mmuext_op *op;
631         struct multicall_space mcs;
632         unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
633
634         BUG_ON(preemptible());
635
636         mcs = xen_mc_entry(sizeof(*op));  /* disables interrupts */
637
638         /* Update while interrupts are disabled, so its atomic with
639            respect to ipis */
640         x86_write_percpu(xen_cr3, cr3);
641
642         op = mcs.args;
643         op->cmd = MMUEXT_NEW_BASEPTR;
644         op->arg1.mfn = mfn;
645
646         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
647
648         /* Update xen_update_cr3 once the batch has actually
649            been submitted. */
650         xen_mc_callback(set_current_cr3, (void *)cr3);
651
652         xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
653 }
654
655 /* Early in boot, while setting up the initial pagetable, assume
656    everything is pinned. */
657 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
658 {
659         BUG_ON(mem_map);        /* should only be used early */
660         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
661 }
662
663 /* Early release_pt assumes that all pts are pinned, since there's
664    only init_mm and anything attached to that is pinned. */
665 static void xen_release_pt_init(u32 pfn)
666 {
667         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
668 }
669
670 static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
671 {
672         struct mmuext_op op;
673         op.cmd = level;
674         op.arg1.mfn = pfn_to_mfn(pfn);
675         if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
676                 BUG();
677 }
678
679 /* This needs to make sure the new pte page is pinned iff its being
680    attached to a pinned pagetable. */
681 static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
682 {
683         struct page *page = pfn_to_page(pfn);
684
685         if (PagePinned(virt_to_page(mm->pgd))) {
686                 SetPagePinned(page);
687
688                 if (!PageHighMem(page)) {
689                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
690                         pin_pagetable_pfn(level, pfn);
691                 } else
692                         /* make sure there are no stray mappings of
693                            this page */
694                         kmap_flush_unused();
695         }
696 }
697
698 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
699 {
700         xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L1_TABLE);
701 }
702
703 static void xen_alloc_pd(struct mm_struct *mm, u32 pfn)
704 {
705         xen_alloc_ptpage(mm, pfn, MMUEXT_PIN_L2_TABLE);
706 }
707
708 /* This should never happen until we're OK to use struct page */
709 static void xen_release_pt(u32 pfn)
710 {
711         struct page *page = pfn_to_page(pfn);
712
713         if (PagePinned(page)) {
714                 if (!PageHighMem(page)) {
715                         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
716                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
717                 }
718         }
719 }
720
721 #ifdef CONFIG_HIGHPTE
722 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
723 {
724         pgprot_t prot = PAGE_KERNEL;
725
726         if (PagePinned(page))
727                 prot = PAGE_KERNEL_RO;
728
729         if (0 && PageHighMem(page))
730                 printk("mapping highpte %lx type %d prot %s\n",
731                        page_to_pfn(page), type,
732                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
733
734         return kmap_atomic_prot(page, type, prot);
735 }
736 #endif
737
738 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
739 {
740         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
741         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
742                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
743                                pte_val_ma(pte));
744
745         return pte;
746 }
747
748 /* Init-time set_pte while constructing initial pagetables, which
749    doesn't allow RO pagetable pages to be remapped RW */
750 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
751 {
752         pte = mask_rw_pte(ptep, pte);
753
754         xen_set_pte(ptep, pte);
755 }
756
757 static __init void xen_pagetable_setup_start(pgd_t *base)
758 {
759         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
760
761         /* special set_pte for pagetable initialization */
762         pv_mmu_ops.set_pte = xen_set_pte_init;
763
764         init_mm.pgd = base;
765         /*
766          * copy top-level of Xen-supplied pagetable into place.  For
767          * !PAE we can use this as-is, but for PAE it is a stand-in
768          * while we copy the pmd pages.
769          */
770         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
771
772         if (PTRS_PER_PMD > 1) {
773                 int i;
774                 /*
775                  * For PAE, need to allocate new pmds, rather than
776                  * share Xen's, since Xen doesn't like pmd's being
777                  * shared between address spaces.
778                  */
779                 for (i = 0; i < PTRS_PER_PGD; i++) {
780                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
781                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
782
783                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
784                                        PAGE_SIZE);
785
786                                 make_lowmem_page_readonly(pmd);
787
788                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
789                         } else
790                                 pgd_clear(&base[i]);
791                 }
792         }
793
794         /* make sure zero_page is mapped RO so we can use it in pagetables */
795         make_lowmem_page_readonly(empty_zero_page);
796         make_lowmem_page_readonly(base);
797         /*
798          * Switch to new pagetable.  This is done before
799          * pagetable_init has done anything so that the new pages
800          * added to the table can be prepared properly for Xen.
801          */
802         xen_write_cr3(__pa(base));
803
804         /* Unpin initial Xen pagetable */
805         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE,
806                           PFN_DOWN(__pa(xen_start_info->pt_base)));
807 }
808
809 static __init void setup_shared_info(void)
810 {
811         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
812                 unsigned long addr = fix_to_virt(FIX_PARAVIRT_BOOTMAP);
813
814                 /*
815                  * Create a mapping for the shared info page.
816                  * Should be set_fixmap(), but shared_info is a machine
817                  * address with no corresponding pseudo-phys address.
818                  */
819                 set_pte_mfn(addr,
820                             PFN_DOWN(xen_start_info->shared_info),
821                             PAGE_KERNEL);
822
823                 HYPERVISOR_shared_info = (struct shared_info *)addr;
824         } else
825                 HYPERVISOR_shared_info =
826                         (struct shared_info *)__va(xen_start_info->shared_info);
827
828 #ifndef CONFIG_SMP
829         /* In UP this is as good a place as any to set up shared info */
830         xen_setup_vcpu_info_placement();
831 #endif
832 }
833
834 static __init void xen_pagetable_setup_done(pgd_t *base)
835 {
836         /* This will work as long as patching hasn't happened yet
837            (which it hasn't) */
838         pv_mmu_ops.alloc_pt = xen_alloc_pt;
839         pv_mmu_ops.alloc_pd = xen_alloc_pd;
840         pv_mmu_ops.release_pt = xen_release_pt;
841         pv_mmu_ops.release_pd = xen_release_pt;
842         pv_mmu_ops.set_pte = xen_set_pte;
843
844         setup_shared_info();
845
846         /* Actually pin the pagetable down, but we can't set PG_pinned
847            yet because the page structures don't exist yet. */
848         {
849                 unsigned level;
850
851 #ifdef CONFIG_X86_PAE
852                 level = MMUEXT_PIN_L3_TABLE;
853 #else
854                 level = MMUEXT_PIN_L2_TABLE;
855 #endif
856
857                 pin_pagetable_pfn(level, PFN_DOWN(__pa(base)));
858         }
859 }
860
861 /* This is called once we have the cpu_possible_map */
862 void __init xen_setup_vcpu_info_placement(void)
863 {
864         int cpu;
865
866         for_each_possible_cpu(cpu)
867                 xen_vcpu_setup(cpu);
868
869         /* xen_vcpu_setup managed to place the vcpu_info within the
870            percpu area for all cpus, so make use of it */
871         if (have_vcpu_info_placement) {
872                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
873
874                 pv_irq_ops.save_fl = xen_save_fl_direct;
875                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
876                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
877                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
878                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
879                 pv_cpu_ops.iret = xen_iret_direct;
880         }
881 }
882
883 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
884                           unsigned long addr, unsigned len)
885 {
886         char *start, *end, *reloc;
887         unsigned ret;
888
889         start = end = reloc = NULL;
890
891 #define SITE(op, x)                                                     \
892         case PARAVIRT_PATCH(op.x):                                      \
893         if (have_vcpu_info_placement) {                                 \
894                 start = (char *)xen_##x##_direct;                       \
895                 end = xen_##x##_direct_end;                             \
896                 reloc = xen_##x##_direct_reloc;                         \
897         }                                                               \
898         goto patch_site
899
900         switch (type) {
901                 SITE(pv_irq_ops, irq_enable);
902                 SITE(pv_irq_ops, irq_disable);
903                 SITE(pv_irq_ops, save_fl);
904                 SITE(pv_irq_ops, restore_fl);
905 #undef SITE
906
907         patch_site:
908                 if (start == NULL || (end-start) > len)
909                         goto default_patch;
910
911                 ret = paravirt_patch_insns(insnbuf, len, start, end);
912
913                 /* Note: because reloc is assigned from something that
914                    appears to be an array, gcc assumes it's non-null,
915                    but doesn't know its relationship with start and
916                    end. */
917                 if (reloc > start && reloc < end) {
918                         int reloc_off = reloc - start;
919                         long *relocp = (long *)(insnbuf + reloc_off);
920                         long delta = start - (char *)addr;
921
922                         *relocp += delta;
923                 }
924                 break;
925
926         default_patch:
927         default:
928                 ret = paravirt_patch_default(type, clobbers, insnbuf,
929                                              addr, len);
930                 break;
931         }
932
933         return ret;
934 }
935
936 static const struct pv_info xen_info __initdata = {
937         .paravirt_enabled = 1,
938         .shared_kernel_pmd = 0,
939
940         .name = "Xen",
941 };
942
943 static const struct pv_init_ops xen_init_ops __initdata = {
944         .patch = xen_patch,
945
946         .banner = xen_banner,
947         .memory_setup = xen_memory_setup,
948         .arch_setup = xen_arch_setup,
949         .post_allocator_init = xen_mark_init_mm_pinned,
950 };
951
952 static const struct pv_time_ops xen_time_ops __initdata = {
953         .time_init = xen_time_init,
954
955         .set_wallclock = xen_set_wallclock,
956         .get_wallclock = xen_get_wallclock,
957         .get_cpu_khz = xen_cpu_khz,
958         .sched_clock = xen_sched_clock,
959 };
960
961 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
962         .cpuid = xen_cpuid,
963
964         .set_debugreg = xen_set_debugreg,
965         .get_debugreg = xen_get_debugreg,
966
967         .clts = native_clts,
968
969         .read_cr0 = native_read_cr0,
970         .write_cr0 = native_write_cr0,
971
972         .read_cr4 = native_read_cr4,
973         .read_cr4_safe = native_read_cr4_safe,
974         .write_cr4 = xen_write_cr4,
975
976         .wbinvd = native_wbinvd,
977
978         .read_msr = native_read_msr_safe,
979         .write_msr = native_write_msr_safe,
980         .read_tsc = native_read_tsc,
981         .read_pmc = native_read_pmc,
982
983         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
984         .irq_enable_syscall_ret = NULL,  /* never called */
985
986         .load_tr_desc = paravirt_nop,
987         .set_ldt = xen_set_ldt,
988         .load_gdt = xen_load_gdt,
989         .load_idt = xen_load_idt,
990         .load_tls = xen_load_tls,
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         .lazy_mode = {
1005                 .enter = paravirt_enter_lazy_cpu,
1006                 .leave = xen_leave_lazy,
1007         },
1008 };
1009
1010 static const struct pv_irq_ops xen_irq_ops __initdata = {
1011         .init_IRQ = xen_init_IRQ,
1012         .save_fl = xen_save_fl,
1013         .restore_fl = xen_restore_fl,
1014         .irq_disable = xen_irq_disable,
1015         .irq_enable = xen_irq_enable,
1016         .safe_halt = xen_safe_halt,
1017         .halt = xen_halt,
1018 };
1019
1020 static const struct pv_apic_ops xen_apic_ops __initdata = {
1021 #ifdef CONFIG_X86_LOCAL_APIC
1022         .apic_write = xen_apic_write,
1023         .apic_write_atomic = xen_apic_write,
1024         .apic_read = xen_apic_read,
1025         .setup_boot_clock = paravirt_nop,
1026         .setup_secondary_clock = paravirt_nop,
1027         .startup_ipi_hook = paravirt_nop,
1028 #endif
1029 };
1030
1031 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1032         .pagetable_setup_start = xen_pagetable_setup_start,
1033         .pagetable_setup_done = xen_pagetable_setup_done,
1034
1035         .read_cr2 = xen_read_cr2,
1036         .write_cr2 = xen_write_cr2,
1037
1038         .read_cr3 = xen_read_cr3,
1039         .write_cr3 = xen_write_cr3,
1040
1041         .flush_tlb_user = xen_flush_tlb,
1042         .flush_tlb_kernel = xen_flush_tlb,
1043         .flush_tlb_single = xen_flush_tlb_single,
1044         .flush_tlb_others = xen_flush_tlb_others,
1045
1046         .pte_update = paravirt_nop,
1047         .pte_update_defer = paravirt_nop,
1048
1049         .alloc_pt = xen_alloc_pt_init,
1050         .release_pt = xen_release_pt_init,
1051         .alloc_pd = xen_alloc_pt_init,
1052         .alloc_pd_clone = paravirt_nop,
1053         .release_pd = xen_release_pt_init,
1054
1055 #ifdef CONFIG_HIGHPTE
1056         .kmap_atomic_pte = xen_kmap_atomic_pte,
1057 #endif
1058
1059         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1060         .set_pte_at = xen_set_pte_at,
1061         .set_pmd = xen_set_pmd,
1062
1063         .pte_val = xen_pte_val,
1064         .pgd_val = xen_pgd_val,
1065
1066         .make_pte = xen_make_pte,
1067         .make_pgd = xen_make_pgd,
1068
1069 #ifdef CONFIG_X86_PAE
1070         .set_pte_atomic = xen_set_pte_atomic,
1071         .set_pte_present = xen_set_pte_at,
1072         .set_pud = xen_set_pud,
1073         .pte_clear = xen_pte_clear,
1074         .pmd_clear = xen_pmd_clear,
1075
1076         .make_pmd = xen_make_pmd,
1077         .pmd_val = xen_pmd_val,
1078 #endif  /* PAE */
1079
1080         .activate_mm = xen_activate_mm,
1081         .dup_mmap = xen_dup_mmap,
1082         .exit_mmap = xen_exit_mmap,
1083
1084         .lazy_mode = {
1085                 .enter = paravirt_enter_lazy_mmu,
1086                 .leave = xen_leave_lazy,
1087         },
1088 };
1089
1090 #ifdef CONFIG_SMP
1091 static const struct smp_ops xen_smp_ops __initdata = {
1092         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
1093         .smp_prepare_cpus = xen_smp_prepare_cpus,
1094         .cpu_up = xen_cpu_up,
1095         .smp_cpus_done = xen_smp_cpus_done,
1096
1097         .smp_send_stop = xen_smp_send_stop,
1098         .smp_send_reschedule = xen_smp_send_reschedule,
1099         .smp_call_function_mask = xen_smp_call_function_mask,
1100 };
1101 #endif  /* CONFIG_SMP */
1102
1103 static void xen_reboot(int reason)
1104 {
1105 #ifdef CONFIG_SMP
1106         smp_send_stop();
1107 #endif
1108
1109         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, reason))
1110                 BUG();
1111 }
1112
1113 static void xen_restart(char *msg)
1114 {
1115         xen_reboot(SHUTDOWN_reboot);
1116 }
1117
1118 static void xen_emergency_restart(void)
1119 {
1120         xen_reboot(SHUTDOWN_reboot);
1121 }
1122
1123 static void xen_machine_halt(void)
1124 {
1125         xen_reboot(SHUTDOWN_poweroff);
1126 }
1127
1128 static void xen_crash_shutdown(struct pt_regs *regs)
1129 {
1130         xen_reboot(SHUTDOWN_crash);
1131 }
1132
1133 static const struct machine_ops __initdata xen_machine_ops = {
1134         .restart = xen_restart,
1135         .halt = xen_machine_halt,
1136         .power_off = xen_machine_halt,
1137         .shutdown = xen_machine_halt,
1138         .crash_shutdown = xen_crash_shutdown,
1139         .emergency_restart = xen_emergency_restart,
1140 };
1141
1142
1143 static void __init xen_reserve_top(void)
1144 {
1145         unsigned long top = HYPERVISOR_VIRT_START;
1146         struct xen_platform_parameters pp;
1147
1148         if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1149                 top = pp.virt_start;
1150
1151         reserve_top_address(-top + 2 * PAGE_SIZE);
1152 }
1153
1154 /* First C function to be called on Xen boot */
1155 asmlinkage void __init xen_start_kernel(void)
1156 {
1157         pgd_t *pgd;
1158
1159         if (!xen_start_info)
1160                 return;
1161
1162         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
1163
1164         /* Install Xen paravirt ops */
1165         pv_info = xen_info;
1166         pv_init_ops = xen_init_ops;
1167         pv_time_ops = xen_time_ops;
1168         pv_cpu_ops = xen_cpu_ops;
1169         pv_irq_ops = xen_irq_ops;
1170         pv_apic_ops = xen_apic_ops;
1171         pv_mmu_ops = xen_mmu_ops;
1172
1173         machine_ops = xen_machine_ops;
1174
1175 #ifdef CONFIG_SMP
1176         smp_ops = xen_smp_ops;
1177 #endif
1178
1179         xen_setup_features();
1180
1181         /* Get mfn list */
1182         if (!xen_feature(XENFEAT_auto_translated_physmap))
1183                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1184
1185         pgd = (pgd_t *)xen_start_info->pt_base;
1186
1187         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1188
1189         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1190
1191         /* keep using Xen gdt for now; no urgent need to change it */
1192
1193         x86_write_percpu(xen_cr3, __pa(pgd));
1194         x86_write_percpu(xen_current_cr3, __pa(pgd));
1195
1196         /* Don't do the full vcpu_info placement stuff until we have a
1197            possible map and a non-dummy shared_info. */
1198         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1199
1200         pv_info.kernel_rpl = 1;
1201         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1202                 pv_info.kernel_rpl = 0;
1203
1204         /* set the limit of our address space */
1205         xen_reserve_top();
1206
1207         /* set up basic CPUID stuff */
1208         cpu_detect(&new_cpu_data);
1209         new_cpu_data.hard_math = 1;
1210         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1211
1212         /* Poke various useful things into boot_params */
1213         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1214         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1215                 ? __pa(xen_start_info->mod_start) : 0;
1216         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1217
1218         /* Start the world */
1219         start_kernel();
1220 }