xref: /openbmc/linux/arch/x86/xen/enlighten.c (revision 7e035230)
1 /*
2  * Core of Xen paravirt_ops implementation.
3  *
4  * This file contains the xen_paravirt_ops structure itself, and the
5  * implementations for:
6  * - privileged instructions
7  * - interrupt flags
8  * - segment operations
9  * - booting and setup
10  *
11  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
12  */
13 
14 #include <linux/cpu.h>
15 #include <linux/kernel.h>
16 #include <linux/init.h>
17 #include <linux/smp.h>
18 #include <linux/preempt.h>
19 #include <linux/hardirq.h>
20 #include <linux/percpu.h>
21 #include <linux/delay.h>
22 #include <linux/start_kernel.h>
23 #include <linux/sched.h>
24 #include <linux/kprobes.h>
25 #include <linux/bootmem.h>
26 #include <linux/module.h>
27 #include <linux/mm.h>
28 #include <linux/page-flags.h>
29 #include <linux/highmem.h>
30 #include <linux/console.h>
31 #include <linux/pci.h>
32 #include <linux/gfp.h>
33 #include <linux/memblock.h>
34 #include <linux/syscore_ops.h>
35 
36 #include <xen/xen.h>
37 #include <xen/interface/xen.h>
38 #include <xen/interface/version.h>
39 #include <xen/interface/physdev.h>
40 #include <xen/interface/vcpu.h>
41 #include <xen/interface/memory.h>
42 #include <xen/interface/xen-mca.h>
43 #include <xen/features.h>
44 #include <xen/page.h>
45 #include <xen/hvm.h>
46 #include <xen/hvc-console.h>
47 #include <xen/acpi.h>
48 
49 #include <asm/paravirt.h>
50 #include <asm/apic.h>
51 #include <asm/page.h>
52 #include <asm/xen/pci.h>
53 #include <asm/xen/hypercall.h>
54 #include <asm/xen/hypervisor.h>
55 #include <asm/fixmap.h>
56 #include <asm/processor.h>
57 #include <asm/proto.h>
58 #include <asm/msr-index.h>
59 #include <asm/traps.h>
60 #include <asm/setup.h>
61 #include <asm/desc.h>
62 #include <asm/pgalloc.h>
63 #include <asm/pgtable.h>
64 #include <asm/tlbflush.h>
65 #include <asm/reboot.h>
66 #include <asm/stackprotector.h>
67 #include <asm/hypervisor.h>
68 #include <asm/mwait.h>
69 #include <asm/pci_x86.h>
70 
71 #ifdef CONFIG_ACPI
72 #include <linux/acpi.h>
73 #include <asm/acpi.h>
74 #include <acpi/pdc_intel.h>
75 #include <acpi/processor.h>
76 #include <xen/interface/platform.h>
77 #endif
78 
79 #include "xen-ops.h"
80 #include "mmu.h"
81 #include "smp.h"
82 #include "multicalls.h"
83 
84 EXPORT_SYMBOL_GPL(hypercall_page);
85 
86 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
87 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
88 
89 enum xen_domain_type xen_domain_type = XEN_NATIVE;
90 EXPORT_SYMBOL_GPL(xen_domain_type);
91 
92 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
93 EXPORT_SYMBOL(machine_to_phys_mapping);
94 unsigned long  machine_to_phys_nr;
95 EXPORT_SYMBOL(machine_to_phys_nr);
96 
97 struct start_info *xen_start_info;
98 EXPORT_SYMBOL_GPL(xen_start_info);
99 
100 struct shared_info xen_dummy_shared_info;
101 
102 void *xen_initial_gdt;
103 
104 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
105 __read_mostly int xen_have_vector_callback;
106 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
107 
108 /*
109  * Point at some empty memory to start with. We map the real shared_info
110  * page as soon as fixmap is up and running.
111  */
112 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
113 
114 /*
115  * Flag to determine whether vcpu info placement is available on all
116  * VCPUs.  We assume it is to start with, and then set it to zero on
117  * the first failure.  This is because it can succeed on some VCPUs
118  * and not others, since it can involve hypervisor memory allocation,
119  * or because the guest failed to guarantee all the appropriate
120  * constraints on all VCPUs (ie buffer can't cross a page boundary).
121  *
122  * Note that any particular CPU may be using a placed vcpu structure,
123  * but we can only optimise if the all are.
124  *
125  * 0: not available, 1: available
126  */
127 static int have_vcpu_info_placement = 1;
128 
129 struct tls_descs {
130 	struct desc_struct desc[3];
131 };
132 
133 /*
134  * Updating the 3 TLS descriptors in the GDT on every task switch is
135  * surprisingly expensive so we avoid updating them if they haven't
136  * changed.  Since Xen writes different descriptors than the one
137  * passed in the update_descriptor hypercall we keep shadow copies to
138  * compare against.
139  */
140 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
141 
142 static void clamp_max_cpus(void)
143 {
144 #ifdef CONFIG_SMP
145 	if (setup_max_cpus > MAX_VIRT_CPUS)
146 		setup_max_cpus = MAX_VIRT_CPUS;
147 #endif
148 }
149 
150 static void xen_vcpu_setup(int cpu)
151 {
152 	struct vcpu_register_vcpu_info info;
153 	int err;
154 	struct vcpu_info *vcpup;
155 
156 	BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
157 
158 	if (cpu < MAX_VIRT_CPUS)
159 		per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
160 
161 	if (!have_vcpu_info_placement) {
162 		if (cpu >= MAX_VIRT_CPUS)
163 			clamp_max_cpus();
164 		return;
165 	}
166 
167 	vcpup = &per_cpu(xen_vcpu_info, cpu);
168 	info.mfn = arbitrary_virt_to_mfn(vcpup);
169 	info.offset = offset_in_page(vcpup);
170 
171 	/* Check to see if the hypervisor will put the vcpu_info
172 	   structure where we want it, which allows direct access via
173 	   a percpu-variable. */
174 	err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
175 
176 	if (err) {
177 		printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
178 		have_vcpu_info_placement = 0;
179 		clamp_max_cpus();
180 	} else {
181 		/* This cpu is using the registered vcpu info, even if
182 		   later ones fail to. */
183 		per_cpu(xen_vcpu, cpu) = vcpup;
184 	}
185 }
186 
187 /*
188  * On restore, set the vcpu placement up again.
189  * If it fails, then we're in a bad state, since
190  * we can't back out from using it...
191  */
192 void xen_vcpu_restore(void)
193 {
194 	int cpu;
195 
196 	for_each_online_cpu(cpu) {
197 		bool other_cpu = (cpu != smp_processor_id());
198 
199 		if (other_cpu &&
200 		    HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
201 			BUG();
202 
203 		xen_setup_runstate_info(cpu);
204 
205 		if (have_vcpu_info_placement)
206 			xen_vcpu_setup(cpu);
207 
208 		if (other_cpu &&
209 		    HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
210 			BUG();
211 	}
212 }
213 
214 static void __init xen_banner(void)
215 {
216 	unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
217 	struct xen_extraversion extra;
218 	HYPERVISOR_xen_version(XENVER_extraversion, &extra);
219 
220 	printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
221 	       pv_info.name);
222 	printk(KERN_INFO "Xen version: %d.%d%s%s\n",
223 	       version >> 16, version & 0xffff, extra.extraversion,
224 	       xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
225 }
226 
227 #define CPUID_THERM_POWER_LEAF 6
228 #define APERFMPERF_PRESENT 0
229 
230 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
231 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
232 
233 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
234 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
235 static __read_mostly unsigned int cpuid_leaf5_edx_val;
236 
237 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
238 		      unsigned int *cx, unsigned int *dx)
239 {
240 	unsigned maskebx = ~0;
241 	unsigned maskecx = ~0;
242 	unsigned maskedx = ~0;
243 	unsigned setecx = 0;
244 	/*
245 	 * Mask out inconvenient features, to try and disable as many
246 	 * unsupported kernel subsystems as possible.
247 	 */
248 	switch (*ax) {
249 	case 1:
250 		maskecx = cpuid_leaf1_ecx_mask;
251 		setecx = cpuid_leaf1_ecx_set_mask;
252 		maskedx = cpuid_leaf1_edx_mask;
253 		break;
254 
255 	case CPUID_MWAIT_LEAF:
256 		/* Synthesize the values.. */
257 		*ax = 0;
258 		*bx = 0;
259 		*cx = cpuid_leaf5_ecx_val;
260 		*dx = cpuid_leaf5_edx_val;
261 		return;
262 
263 	case CPUID_THERM_POWER_LEAF:
264 		/* Disabling APERFMPERF for kernel usage */
265 		maskecx = ~(1 << APERFMPERF_PRESENT);
266 		break;
267 
268 	case 0xb:
269 		/* Suppress extended topology stuff */
270 		maskebx = 0;
271 		break;
272 	}
273 
274 	asm(XEN_EMULATE_PREFIX "cpuid"
275 		: "=a" (*ax),
276 		  "=b" (*bx),
277 		  "=c" (*cx),
278 		  "=d" (*dx)
279 		: "0" (*ax), "2" (*cx));
280 
281 	*bx &= maskebx;
282 	*cx &= maskecx;
283 	*cx |= setecx;
284 	*dx &= maskedx;
285 
286 }
287 
288 static bool __init xen_check_mwait(void)
289 {
290 #if defined(CONFIG_ACPI) && !defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR) && \
291 	!defined(CONFIG_ACPI_PROCESSOR_AGGREGATOR_MODULE)
292 	struct xen_platform_op op = {
293 		.cmd			= XENPF_set_processor_pminfo,
294 		.u.set_pminfo.id	= -1,
295 		.u.set_pminfo.type	= XEN_PM_PDC,
296 	};
297 	uint32_t buf[3];
298 	unsigned int ax, bx, cx, dx;
299 	unsigned int mwait_mask;
300 
301 	/* We need to determine whether it is OK to expose the MWAIT
302 	 * capability to the kernel to harvest deeper than C3 states from ACPI
303 	 * _CST using the processor_harvest_xen.c module. For this to work, we
304 	 * need to gather the MWAIT_LEAF values (which the cstate.c code
305 	 * checks against). The hypervisor won't expose the MWAIT flag because
306 	 * it would break backwards compatibility; so we will find out directly
307 	 * from the hardware and hypercall.
308 	 */
309 	if (!xen_initial_domain())
310 		return false;
311 
312 	ax = 1;
313 	cx = 0;
314 
315 	native_cpuid(&ax, &bx, &cx, &dx);
316 
317 	mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
318 		     (1 << (X86_FEATURE_MWAIT % 32));
319 
320 	if ((cx & mwait_mask) != mwait_mask)
321 		return false;
322 
323 	/* We need to emulate the MWAIT_LEAF and for that we need both
324 	 * ecx and edx. The hypercall provides only partial information.
325 	 */
326 
327 	ax = CPUID_MWAIT_LEAF;
328 	bx = 0;
329 	cx = 0;
330 	dx = 0;
331 
332 	native_cpuid(&ax, &bx, &cx, &dx);
333 
334 	/* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
335 	 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
336 	 */
337 	buf[0] = ACPI_PDC_REVISION_ID;
338 	buf[1] = 1;
339 	buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
340 
341 	set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
342 
343 	if ((HYPERVISOR_dom0_op(&op) == 0) &&
344 	    (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
345 		cpuid_leaf5_ecx_val = cx;
346 		cpuid_leaf5_edx_val = dx;
347 	}
348 	return true;
349 #else
350 	return false;
351 #endif
352 }
353 static void __init xen_init_cpuid_mask(void)
354 {
355 	unsigned int ax, bx, cx, dx;
356 	unsigned int xsave_mask;
357 
358 	cpuid_leaf1_edx_mask =
359 		~((1 << X86_FEATURE_MTRR) |  /* disable MTRR */
360 		  (1 << X86_FEATURE_ACC));   /* thermal monitoring */
361 
362 	if (!xen_initial_domain())
363 		cpuid_leaf1_edx_mask &=
364 			~((1 << X86_FEATURE_APIC) |  /* disable local APIC */
365 			  (1 << X86_FEATURE_ACPI));  /* disable ACPI */
366 	ax = 1;
367 	cx = 0;
368 	xen_cpuid(&ax, &bx, &cx, &dx);
369 
370 	xsave_mask =
371 		(1 << (X86_FEATURE_XSAVE % 32)) |
372 		(1 << (X86_FEATURE_OSXSAVE % 32));
373 
374 	/* Xen will set CR4.OSXSAVE if supported and not disabled by force */
375 	if ((cx & xsave_mask) != xsave_mask)
376 		cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
377 	if (xen_check_mwait())
378 		cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
379 }
380 
381 static void xen_set_debugreg(int reg, unsigned long val)
382 {
383 	HYPERVISOR_set_debugreg(reg, val);
384 }
385 
386 static unsigned long xen_get_debugreg(int reg)
387 {
388 	return HYPERVISOR_get_debugreg(reg);
389 }
390 
391 static void xen_end_context_switch(struct task_struct *next)
392 {
393 	xen_mc_flush();
394 	paravirt_end_context_switch(next);
395 }
396 
397 static unsigned long xen_store_tr(void)
398 {
399 	return 0;
400 }
401 
402 /*
403  * Set the page permissions for a particular virtual address.  If the
404  * address is a vmalloc mapping (or other non-linear mapping), then
405  * find the linear mapping of the page and also set its protections to
406  * match.
407  */
408 static void set_aliased_prot(void *v, pgprot_t prot)
409 {
410 	int level;
411 	pte_t *ptep;
412 	pte_t pte;
413 	unsigned long pfn;
414 	struct page *page;
415 
416 	ptep = lookup_address((unsigned long)v, &level);
417 	BUG_ON(ptep == NULL);
418 
419 	pfn = pte_pfn(*ptep);
420 	page = pfn_to_page(pfn);
421 
422 	pte = pfn_pte(pfn, prot);
423 
424 	if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
425 		BUG();
426 
427 	if (!PageHighMem(page)) {
428 		void *av = __va(PFN_PHYS(pfn));
429 
430 		if (av != v)
431 			if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
432 				BUG();
433 	} else
434 		kmap_flush_unused();
435 }
436 
437 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
438 {
439 	const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
440 	int i;
441 
442 	for(i = 0; i < entries; i += entries_per_page)
443 		set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
444 }
445 
446 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
447 {
448 	const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
449 	int i;
450 
451 	for(i = 0; i < entries; i += entries_per_page)
452 		set_aliased_prot(ldt + i, PAGE_KERNEL);
453 }
454 
455 static void xen_set_ldt(const void *addr, unsigned entries)
456 {
457 	struct mmuext_op *op;
458 	struct multicall_space mcs = xen_mc_entry(sizeof(*op));
459 
460 	trace_xen_cpu_set_ldt(addr, entries);
461 
462 	op = mcs.args;
463 	op->cmd = MMUEXT_SET_LDT;
464 	op->arg1.linear_addr = (unsigned long)addr;
465 	op->arg2.nr_ents = entries;
466 
467 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
468 
469 	xen_mc_issue(PARAVIRT_LAZY_CPU);
470 }
471 
472 static void xen_load_gdt(const struct desc_ptr *dtr)
473 {
474 	unsigned long va = dtr->address;
475 	unsigned int size = dtr->size + 1;
476 	unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
477 	unsigned long frames[pages];
478 	int f;
479 
480 	/*
481 	 * A GDT can be up to 64k in size, which corresponds to 8192
482 	 * 8-byte entries, or 16 4k pages..
483 	 */
484 
485 	BUG_ON(size > 65536);
486 	BUG_ON(va & ~PAGE_MASK);
487 
488 	for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
489 		int level;
490 		pte_t *ptep;
491 		unsigned long pfn, mfn;
492 		void *virt;
493 
494 		/*
495 		 * The GDT is per-cpu and is in the percpu data area.
496 		 * That can be virtually mapped, so we need to do a
497 		 * page-walk to get the underlying MFN for the
498 		 * hypercall.  The page can also be in the kernel's
499 		 * linear range, so we need to RO that mapping too.
500 		 */
501 		ptep = lookup_address(va, &level);
502 		BUG_ON(ptep == NULL);
503 
504 		pfn = pte_pfn(*ptep);
505 		mfn = pfn_to_mfn(pfn);
506 		virt = __va(PFN_PHYS(pfn));
507 
508 		frames[f] = mfn;
509 
510 		make_lowmem_page_readonly((void *)va);
511 		make_lowmem_page_readonly(virt);
512 	}
513 
514 	if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
515 		BUG();
516 }
517 
518 /*
519  * load_gdt for early boot, when the gdt is only mapped once
520  */
521 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
522 {
523 	unsigned long va = dtr->address;
524 	unsigned int size = dtr->size + 1;
525 	unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
526 	unsigned long frames[pages];
527 	int f;
528 
529 	/*
530 	 * A GDT can be up to 64k in size, which corresponds to 8192
531 	 * 8-byte entries, or 16 4k pages..
532 	 */
533 
534 	BUG_ON(size > 65536);
535 	BUG_ON(va & ~PAGE_MASK);
536 
537 	for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
538 		pte_t pte;
539 		unsigned long pfn, mfn;
540 
541 		pfn = virt_to_pfn(va);
542 		mfn = pfn_to_mfn(pfn);
543 
544 		pte = pfn_pte(pfn, PAGE_KERNEL_RO);
545 
546 		if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
547 			BUG();
548 
549 		frames[f] = mfn;
550 	}
551 
552 	if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
553 		BUG();
554 }
555 
556 static inline bool desc_equal(const struct desc_struct *d1,
557 			      const struct desc_struct *d2)
558 {
559 	return d1->a == d2->a && d1->b == d2->b;
560 }
561 
562 static void load_TLS_descriptor(struct thread_struct *t,
563 				unsigned int cpu, unsigned int i)
564 {
565 	struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
566 	struct desc_struct *gdt;
567 	xmaddr_t maddr;
568 	struct multicall_space mc;
569 
570 	if (desc_equal(shadow, &t->tls_array[i]))
571 		return;
572 
573 	*shadow = t->tls_array[i];
574 
575 	gdt = get_cpu_gdt_table(cpu);
576 	maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
577 	mc = __xen_mc_entry(0);
578 
579 	MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
580 }
581 
582 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
583 {
584 	/*
585 	 * XXX sleazy hack: If we're being called in a lazy-cpu zone
586 	 * and lazy gs handling is enabled, it means we're in a
587 	 * context switch, and %gs has just been saved.  This means we
588 	 * can zero it out to prevent faults on exit from the
589 	 * hypervisor if the next process has no %gs.  Either way, it
590 	 * has been saved, and the new value will get loaded properly.
591 	 * This will go away as soon as Xen has been modified to not
592 	 * save/restore %gs for normal hypercalls.
593 	 *
594 	 * On x86_64, this hack is not used for %gs, because gs points
595 	 * to KERNEL_GS_BASE (and uses it for PDA references), so we
596 	 * must not zero %gs on x86_64
597 	 *
598 	 * For x86_64, we need to zero %fs, otherwise we may get an
599 	 * exception between the new %fs descriptor being loaded and
600 	 * %fs being effectively cleared at __switch_to().
601 	 */
602 	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
603 #ifdef CONFIG_X86_32
604 		lazy_load_gs(0);
605 #else
606 		loadsegment(fs, 0);
607 #endif
608 	}
609 
610 	xen_mc_batch();
611 
612 	load_TLS_descriptor(t, cpu, 0);
613 	load_TLS_descriptor(t, cpu, 1);
614 	load_TLS_descriptor(t, cpu, 2);
615 
616 	xen_mc_issue(PARAVIRT_LAZY_CPU);
617 }
618 
619 #ifdef CONFIG_X86_64
620 static void xen_load_gs_index(unsigned int idx)
621 {
622 	if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
623 		BUG();
624 }
625 #endif
626 
627 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
628 				const void *ptr)
629 {
630 	xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
631 	u64 entry = *(u64 *)ptr;
632 
633 	trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
634 
635 	preempt_disable();
636 
637 	xen_mc_flush();
638 	if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
639 		BUG();
640 
641 	preempt_enable();
642 }
643 
644 static int cvt_gate_to_trap(int vector, const gate_desc *val,
645 			    struct trap_info *info)
646 {
647 	unsigned long addr;
648 
649 	if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
650 		return 0;
651 
652 	info->vector = vector;
653 
654 	addr = gate_offset(*val);
655 #ifdef CONFIG_X86_64
656 	/*
657 	 * Look for known traps using IST, and substitute them
658 	 * appropriately.  The debugger ones are the only ones we care
659 	 * about.  Xen will handle faults like double_fault,
660 	 * so we should never see them.  Warn if
661 	 * there's an unexpected IST-using fault handler.
662 	 */
663 	if (addr == (unsigned long)debug)
664 		addr = (unsigned long)xen_debug;
665 	else if (addr == (unsigned long)int3)
666 		addr = (unsigned long)xen_int3;
667 	else if (addr == (unsigned long)stack_segment)
668 		addr = (unsigned long)xen_stack_segment;
669 	else if (addr == (unsigned long)double_fault ||
670 		 addr == (unsigned long)nmi) {
671 		/* Don't need to handle these */
672 		return 0;
673 #ifdef CONFIG_X86_MCE
674 	} else if (addr == (unsigned long)machine_check) {
675 		/*
676 		 * when xen hypervisor inject vMCE to guest,
677 		 * use native mce handler to handle it
678 		 */
679 		;
680 #endif
681 	} else {
682 		/* Some other trap using IST? */
683 		if (WARN_ON(val->ist != 0))
684 			return 0;
685 	}
686 #endif	/* CONFIG_X86_64 */
687 	info->address = addr;
688 
689 	info->cs = gate_segment(*val);
690 	info->flags = val->dpl;
691 	/* interrupt gates clear IF */
692 	if (val->type == GATE_INTERRUPT)
693 		info->flags |= 1 << 2;
694 
695 	return 1;
696 }
697 
698 /* Locations of each CPU's IDT */
699 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
700 
701 /* Set an IDT entry.  If the entry is part of the current IDT, then
702    also update Xen. */
703 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
704 {
705 	unsigned long p = (unsigned long)&dt[entrynum];
706 	unsigned long start, end;
707 
708 	trace_xen_cpu_write_idt_entry(dt, entrynum, g);
709 
710 	preempt_disable();
711 
712 	start = __this_cpu_read(idt_desc.address);
713 	end = start + __this_cpu_read(idt_desc.size) + 1;
714 
715 	xen_mc_flush();
716 
717 	native_write_idt_entry(dt, entrynum, g);
718 
719 	if (p >= start && (p + 8) <= end) {
720 		struct trap_info info[2];
721 
722 		info[1].address = 0;
723 
724 		if (cvt_gate_to_trap(entrynum, g, &info[0]))
725 			if (HYPERVISOR_set_trap_table(info))
726 				BUG();
727 	}
728 
729 	preempt_enable();
730 }
731 
732 static void xen_convert_trap_info(const struct desc_ptr *desc,
733 				  struct trap_info *traps)
734 {
735 	unsigned in, out, count;
736 
737 	count = (desc->size+1) / sizeof(gate_desc);
738 	BUG_ON(count > 256);
739 
740 	for (in = out = 0; in < count; in++) {
741 		gate_desc *entry = (gate_desc*)(desc->address) + in;
742 
743 		if (cvt_gate_to_trap(in, entry, &traps[out]))
744 			out++;
745 	}
746 	traps[out].address = 0;
747 }
748 
749 void xen_copy_trap_info(struct trap_info *traps)
750 {
751 	const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
752 
753 	xen_convert_trap_info(desc, traps);
754 }
755 
756 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
757    hold a spinlock to protect the static traps[] array (static because
758    it avoids allocation, and saves stack space). */
759 static void xen_load_idt(const struct desc_ptr *desc)
760 {
761 	static DEFINE_SPINLOCK(lock);
762 	static struct trap_info traps[257];
763 
764 	trace_xen_cpu_load_idt(desc);
765 
766 	spin_lock(&lock);
767 
768 	__get_cpu_var(idt_desc) = *desc;
769 
770 	xen_convert_trap_info(desc, traps);
771 
772 	xen_mc_flush();
773 	if (HYPERVISOR_set_trap_table(traps))
774 		BUG();
775 
776 	spin_unlock(&lock);
777 }
778 
779 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
780    they're handled differently. */
781 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
782 				const void *desc, int type)
783 {
784 	trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
785 
786 	preempt_disable();
787 
788 	switch (type) {
789 	case DESC_LDT:
790 	case DESC_TSS:
791 		/* ignore */
792 		break;
793 
794 	default: {
795 		xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
796 
797 		xen_mc_flush();
798 		if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
799 			BUG();
800 	}
801 
802 	}
803 
804 	preempt_enable();
805 }
806 
807 /*
808  * Version of write_gdt_entry for use at early boot-time needed to
809  * update an entry as simply as possible.
810  */
811 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
812 					    const void *desc, int type)
813 {
814 	trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
815 
816 	switch (type) {
817 	case DESC_LDT:
818 	case DESC_TSS:
819 		/* ignore */
820 		break;
821 
822 	default: {
823 		xmaddr_t maddr = virt_to_machine(&dt[entry]);
824 
825 		if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
826 			dt[entry] = *(struct desc_struct *)desc;
827 	}
828 
829 	}
830 }
831 
832 static void xen_load_sp0(struct tss_struct *tss,
833 			 struct thread_struct *thread)
834 {
835 	struct multicall_space mcs;
836 
837 	mcs = xen_mc_entry(0);
838 	MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
839 	xen_mc_issue(PARAVIRT_LAZY_CPU);
840 }
841 
842 static void xen_set_iopl_mask(unsigned mask)
843 {
844 	struct physdev_set_iopl set_iopl;
845 
846 	/* Force the change at ring 0. */
847 	set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
848 	HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
849 }
850 
851 static void xen_io_delay(void)
852 {
853 }
854 
855 #ifdef CONFIG_X86_LOCAL_APIC
856 static unsigned long xen_set_apic_id(unsigned int x)
857 {
858 	WARN_ON(1);
859 	return x;
860 }
861 static unsigned int xen_get_apic_id(unsigned long x)
862 {
863 	return ((x)>>24) & 0xFFu;
864 }
865 static u32 xen_apic_read(u32 reg)
866 {
867 	struct xen_platform_op op = {
868 		.cmd = XENPF_get_cpuinfo,
869 		.interface_version = XENPF_INTERFACE_VERSION,
870 		.u.pcpu_info.xen_cpuid = 0,
871 	};
872 	int ret = 0;
873 
874 	/* Shouldn't need this as APIC is turned off for PV, and we only
875 	 * get called on the bootup processor. But just in case. */
876 	if (!xen_initial_domain() || smp_processor_id())
877 		return 0;
878 
879 	if (reg == APIC_LVR)
880 		return 0x10;
881 
882 	if (reg != APIC_ID)
883 		return 0;
884 
885 	ret = HYPERVISOR_dom0_op(&op);
886 	if (ret)
887 		return 0;
888 
889 	return op.u.pcpu_info.apic_id << 24;
890 }
891 
892 static void xen_apic_write(u32 reg, u32 val)
893 {
894 	/* Warn to see if there's any stray references */
895 	WARN_ON(1);
896 }
897 
898 static u64 xen_apic_icr_read(void)
899 {
900 	return 0;
901 }
902 
903 static void xen_apic_icr_write(u32 low, u32 id)
904 {
905 	/* Warn to see if there's any stray references */
906 	WARN_ON(1);
907 }
908 
909 static void xen_apic_wait_icr_idle(void)
910 {
911         return;
912 }
913 
914 static u32 xen_safe_apic_wait_icr_idle(void)
915 {
916         return 0;
917 }
918 
919 static void set_xen_basic_apic_ops(void)
920 {
921 	apic->read = xen_apic_read;
922 	apic->write = xen_apic_write;
923 	apic->icr_read = xen_apic_icr_read;
924 	apic->icr_write = xen_apic_icr_write;
925 	apic->wait_icr_idle = xen_apic_wait_icr_idle;
926 	apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
927 	apic->set_apic_id = xen_set_apic_id;
928 	apic->get_apic_id = xen_get_apic_id;
929 
930 #ifdef CONFIG_SMP
931 	apic->send_IPI_allbutself = xen_send_IPI_allbutself;
932 	apic->send_IPI_mask_allbutself = xen_send_IPI_mask_allbutself;
933 	apic->send_IPI_mask = xen_send_IPI_mask;
934 	apic->send_IPI_all = xen_send_IPI_all;
935 	apic->send_IPI_self = xen_send_IPI_self;
936 #endif
937 }
938 
939 #endif
940 
941 static void xen_clts(void)
942 {
943 	struct multicall_space mcs;
944 
945 	mcs = xen_mc_entry(0);
946 
947 	MULTI_fpu_taskswitch(mcs.mc, 0);
948 
949 	xen_mc_issue(PARAVIRT_LAZY_CPU);
950 }
951 
952 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
953 
954 static unsigned long xen_read_cr0(void)
955 {
956 	unsigned long cr0 = this_cpu_read(xen_cr0_value);
957 
958 	if (unlikely(cr0 == 0)) {
959 		cr0 = native_read_cr0();
960 		this_cpu_write(xen_cr0_value, cr0);
961 	}
962 
963 	return cr0;
964 }
965 
966 static void xen_write_cr0(unsigned long cr0)
967 {
968 	struct multicall_space mcs;
969 
970 	this_cpu_write(xen_cr0_value, cr0);
971 
972 	/* Only pay attention to cr0.TS; everything else is
973 	   ignored. */
974 	mcs = xen_mc_entry(0);
975 
976 	MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
977 
978 	xen_mc_issue(PARAVIRT_LAZY_CPU);
979 }
980 
981 static void xen_write_cr4(unsigned long cr4)
982 {
983 	cr4 &= ~X86_CR4_PGE;
984 	cr4 &= ~X86_CR4_PSE;
985 
986 	native_write_cr4(cr4);
987 }
988 
989 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
990 {
991 	int ret;
992 
993 	ret = 0;
994 
995 	switch (msr) {
996 #ifdef CONFIG_X86_64
997 		unsigned which;
998 		u64 base;
999 
1000 	case MSR_FS_BASE:		which = SEGBASE_FS; goto set;
1001 	case MSR_KERNEL_GS_BASE:	which = SEGBASE_GS_USER; goto set;
1002 	case MSR_GS_BASE:		which = SEGBASE_GS_KERNEL; goto set;
1003 
1004 	set:
1005 		base = ((u64)high << 32) | low;
1006 		if (HYPERVISOR_set_segment_base(which, base) != 0)
1007 			ret = -EIO;
1008 		break;
1009 #endif
1010 
1011 	case MSR_STAR:
1012 	case MSR_CSTAR:
1013 	case MSR_LSTAR:
1014 	case MSR_SYSCALL_MASK:
1015 	case MSR_IA32_SYSENTER_CS:
1016 	case MSR_IA32_SYSENTER_ESP:
1017 	case MSR_IA32_SYSENTER_EIP:
1018 		/* Fast syscall setup is all done in hypercalls, so
1019 		   these are all ignored.  Stub them out here to stop
1020 		   Xen console noise. */
1021 		break;
1022 
1023 	case MSR_IA32_CR_PAT:
1024 		if (smp_processor_id() == 0)
1025 			xen_set_pat(((u64)high << 32) | low);
1026 		break;
1027 
1028 	default:
1029 		ret = native_write_msr_safe(msr, low, high);
1030 	}
1031 
1032 	return ret;
1033 }
1034 
1035 void xen_setup_shared_info(void)
1036 {
1037 	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1038 		set_fixmap(FIX_PARAVIRT_BOOTMAP,
1039 			   xen_start_info->shared_info);
1040 
1041 		HYPERVISOR_shared_info =
1042 			(struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1043 	} else
1044 		HYPERVISOR_shared_info =
1045 			(struct shared_info *)__va(xen_start_info->shared_info);
1046 
1047 #ifndef CONFIG_SMP
1048 	/* In UP this is as good a place as any to set up shared info */
1049 	xen_setup_vcpu_info_placement();
1050 #endif
1051 
1052 	xen_setup_mfn_list_list();
1053 }
1054 
1055 /* This is called once we have the cpu_possible_mask */
1056 void xen_setup_vcpu_info_placement(void)
1057 {
1058 	int cpu;
1059 
1060 	for_each_possible_cpu(cpu)
1061 		xen_vcpu_setup(cpu);
1062 
1063 	/* xen_vcpu_setup managed to place the vcpu_info within the
1064 	   percpu area for all cpus, so make use of it */
1065 	if (have_vcpu_info_placement) {
1066 		pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1067 		pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1068 		pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1069 		pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1070 		pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1071 	}
1072 }
1073 
1074 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1075 			  unsigned long addr, unsigned len)
1076 {
1077 	char *start, *end, *reloc;
1078 	unsigned ret;
1079 
1080 	start = end = reloc = NULL;
1081 
1082 #define SITE(op, x)							\
1083 	case PARAVIRT_PATCH(op.x):					\
1084 	if (have_vcpu_info_placement) {					\
1085 		start = (char *)xen_##x##_direct;			\
1086 		end = xen_##x##_direct_end;				\
1087 		reloc = xen_##x##_direct_reloc;				\
1088 	}								\
1089 	goto patch_site
1090 
1091 	switch (type) {
1092 		SITE(pv_irq_ops, irq_enable);
1093 		SITE(pv_irq_ops, irq_disable);
1094 		SITE(pv_irq_ops, save_fl);
1095 		SITE(pv_irq_ops, restore_fl);
1096 #undef SITE
1097 
1098 	patch_site:
1099 		if (start == NULL || (end-start) > len)
1100 			goto default_patch;
1101 
1102 		ret = paravirt_patch_insns(insnbuf, len, start, end);
1103 
1104 		/* Note: because reloc is assigned from something that
1105 		   appears to be an array, gcc assumes it's non-null,
1106 		   but doesn't know its relationship with start and
1107 		   end. */
1108 		if (reloc > start && reloc < end) {
1109 			int reloc_off = reloc - start;
1110 			long *relocp = (long *)(insnbuf + reloc_off);
1111 			long delta = start - (char *)addr;
1112 
1113 			*relocp += delta;
1114 		}
1115 		break;
1116 
1117 	default_patch:
1118 	default:
1119 		ret = paravirt_patch_default(type, clobbers, insnbuf,
1120 					     addr, len);
1121 		break;
1122 	}
1123 
1124 	return ret;
1125 }
1126 
1127 static const struct pv_info xen_info __initconst = {
1128 	.paravirt_enabled = 1,
1129 	.shared_kernel_pmd = 0,
1130 
1131 #ifdef CONFIG_X86_64
1132 	.extra_user_64bit_cs = FLAT_USER_CS64,
1133 #endif
1134 
1135 	.name = "Xen",
1136 };
1137 
1138 static const struct pv_init_ops xen_init_ops __initconst = {
1139 	.patch = xen_patch,
1140 };
1141 
1142 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1143 	.cpuid = xen_cpuid,
1144 
1145 	.set_debugreg = xen_set_debugreg,
1146 	.get_debugreg = xen_get_debugreg,
1147 
1148 	.clts = xen_clts,
1149 
1150 	.read_cr0 = xen_read_cr0,
1151 	.write_cr0 = xen_write_cr0,
1152 
1153 	.read_cr4 = native_read_cr4,
1154 	.read_cr4_safe = native_read_cr4_safe,
1155 	.write_cr4 = xen_write_cr4,
1156 
1157 	.wbinvd = native_wbinvd,
1158 
1159 	.read_msr = native_read_msr_safe,
1160 	.write_msr = xen_write_msr_safe,
1161 
1162 	.read_tsc = native_read_tsc,
1163 	.read_pmc = native_read_pmc,
1164 
1165 	.iret = xen_iret,
1166 	.irq_enable_sysexit = xen_sysexit,
1167 #ifdef CONFIG_X86_64
1168 	.usergs_sysret32 = xen_sysret32,
1169 	.usergs_sysret64 = xen_sysret64,
1170 #endif
1171 
1172 	.load_tr_desc = paravirt_nop,
1173 	.set_ldt = xen_set_ldt,
1174 	.load_gdt = xen_load_gdt,
1175 	.load_idt = xen_load_idt,
1176 	.load_tls = xen_load_tls,
1177 #ifdef CONFIG_X86_64
1178 	.load_gs_index = xen_load_gs_index,
1179 #endif
1180 
1181 	.alloc_ldt = xen_alloc_ldt,
1182 	.free_ldt = xen_free_ldt,
1183 
1184 	.store_gdt = native_store_gdt,
1185 	.store_idt = native_store_idt,
1186 	.store_tr = xen_store_tr,
1187 
1188 	.write_ldt_entry = xen_write_ldt_entry,
1189 	.write_gdt_entry = xen_write_gdt_entry,
1190 	.write_idt_entry = xen_write_idt_entry,
1191 	.load_sp0 = xen_load_sp0,
1192 
1193 	.set_iopl_mask = xen_set_iopl_mask,
1194 	.io_delay = xen_io_delay,
1195 
1196 	/* Xen takes care of %gs when switching to usermode for us */
1197 	.swapgs = paravirt_nop,
1198 
1199 	.start_context_switch = paravirt_start_context_switch,
1200 	.end_context_switch = xen_end_context_switch,
1201 };
1202 
1203 static const struct pv_apic_ops xen_apic_ops __initconst = {
1204 #ifdef CONFIG_X86_LOCAL_APIC
1205 	.startup_ipi_hook = paravirt_nop,
1206 #endif
1207 };
1208 
1209 static void xen_reboot(int reason)
1210 {
1211 	struct sched_shutdown r = { .reason = reason };
1212 
1213 	if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1214 		BUG();
1215 }
1216 
1217 static void xen_restart(char *msg)
1218 {
1219 	xen_reboot(SHUTDOWN_reboot);
1220 }
1221 
1222 static void xen_emergency_restart(void)
1223 {
1224 	xen_reboot(SHUTDOWN_reboot);
1225 }
1226 
1227 static void xen_machine_halt(void)
1228 {
1229 	xen_reboot(SHUTDOWN_poweroff);
1230 }
1231 
1232 static void xen_machine_power_off(void)
1233 {
1234 	if (pm_power_off)
1235 		pm_power_off();
1236 	xen_reboot(SHUTDOWN_poweroff);
1237 }
1238 
1239 static void xen_crash_shutdown(struct pt_regs *regs)
1240 {
1241 	xen_reboot(SHUTDOWN_crash);
1242 }
1243 
1244 static int
1245 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1246 {
1247 	xen_reboot(SHUTDOWN_crash);
1248 	return NOTIFY_DONE;
1249 }
1250 
1251 static struct notifier_block xen_panic_block = {
1252 	.notifier_call= xen_panic_event,
1253 };
1254 
1255 int xen_panic_handler_init(void)
1256 {
1257 	atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1258 	return 0;
1259 }
1260 
1261 static const struct machine_ops xen_machine_ops __initconst = {
1262 	.restart = xen_restart,
1263 	.halt = xen_machine_halt,
1264 	.power_off = xen_machine_power_off,
1265 	.shutdown = xen_machine_halt,
1266 	.crash_shutdown = xen_crash_shutdown,
1267 	.emergency_restart = xen_emergency_restart,
1268 };
1269 
1270 /*
1271  * Set up the GDT and segment registers for -fstack-protector.  Until
1272  * we do this, we have to be careful not to call any stack-protected
1273  * function, which is most of the kernel.
1274  */
1275 static void __init xen_setup_stackprotector(void)
1276 {
1277 	pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1278 	pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1279 
1280 	setup_stack_canary_segment(0);
1281 	switch_to_new_gdt(0);
1282 
1283 	pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1284 	pv_cpu_ops.load_gdt = xen_load_gdt;
1285 }
1286 
1287 /* First C function to be called on Xen boot */
1288 asmlinkage void __init xen_start_kernel(void)
1289 {
1290 	struct physdev_set_iopl set_iopl;
1291 	int rc;
1292 	pgd_t *pgd;
1293 
1294 	if (!xen_start_info)
1295 		return;
1296 
1297 	xen_domain_type = XEN_PV_DOMAIN;
1298 
1299 	xen_setup_machphys_mapping();
1300 
1301 	/* Install Xen paravirt ops */
1302 	pv_info = xen_info;
1303 	pv_init_ops = xen_init_ops;
1304 	pv_cpu_ops = xen_cpu_ops;
1305 	pv_apic_ops = xen_apic_ops;
1306 
1307 	x86_init.resources.memory_setup = xen_memory_setup;
1308 	x86_init.oem.arch_setup = xen_arch_setup;
1309 	x86_init.oem.banner = xen_banner;
1310 
1311 	xen_init_time_ops();
1312 
1313 	/*
1314 	 * Set up some pagetable state before starting to set any ptes.
1315 	 */
1316 
1317 	xen_init_mmu_ops();
1318 
1319 	/* Prevent unwanted bits from being set in PTEs. */
1320 	__supported_pte_mask &= ~_PAGE_GLOBAL;
1321 #if 0
1322 	if (!xen_initial_domain())
1323 #endif
1324 		__supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1325 
1326 	__supported_pte_mask |= _PAGE_IOMAP;
1327 
1328 	/*
1329 	 * Prevent page tables from being allocated in highmem, even
1330 	 * if CONFIG_HIGHPTE is enabled.
1331 	 */
1332 	__userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1333 
1334 	/* Work out if we support NX */
1335 	x86_configure_nx();
1336 
1337 	xen_setup_features();
1338 
1339 	/* Get mfn list */
1340 	if (!xen_feature(XENFEAT_auto_translated_physmap))
1341 		xen_build_dynamic_phys_to_machine();
1342 
1343 	/*
1344 	 * Set up kernel GDT and segment registers, mainly so that
1345 	 * -fstack-protector code can be executed.
1346 	 */
1347 	xen_setup_stackprotector();
1348 
1349 	xen_init_irq_ops();
1350 	xen_init_cpuid_mask();
1351 
1352 #ifdef CONFIG_X86_LOCAL_APIC
1353 	/*
1354 	 * set up the basic apic ops.
1355 	 */
1356 	set_xen_basic_apic_ops();
1357 #endif
1358 
1359 	if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1360 		pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1361 		pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1362 	}
1363 
1364 	machine_ops = xen_machine_ops;
1365 
1366 	/*
1367 	 * The only reliable way to retain the initial address of the
1368 	 * percpu gdt_page is to remember it here, so we can go and
1369 	 * mark it RW later, when the initial percpu area is freed.
1370 	 */
1371 	xen_initial_gdt = &per_cpu(gdt_page, 0);
1372 
1373 	xen_smp_init();
1374 
1375 #ifdef CONFIG_ACPI_NUMA
1376 	/*
1377 	 * The pages we from Xen are not related to machine pages, so
1378 	 * any NUMA information the kernel tries to get from ACPI will
1379 	 * be meaningless.  Prevent it from trying.
1380 	 */
1381 	acpi_numa = -1;
1382 #endif
1383 
1384 	pgd = (pgd_t *)xen_start_info->pt_base;
1385 
1386 	/* Don't do the full vcpu_info placement stuff until we have a
1387 	   possible map and a non-dummy shared_info. */
1388 	per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1389 
1390 	local_irq_disable();
1391 	early_boot_irqs_disabled = true;
1392 
1393 	xen_raw_console_write("mapping kernel into physical memory\n");
1394 	pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1395 
1396 	/* Allocate and initialize top and mid mfn levels for p2m structure */
1397 	xen_build_mfn_list_list();
1398 
1399 	/* keep using Xen gdt for now; no urgent need to change it */
1400 
1401 #ifdef CONFIG_X86_32
1402 	pv_info.kernel_rpl = 1;
1403 	if (xen_feature(XENFEAT_supervisor_mode_kernel))
1404 		pv_info.kernel_rpl = 0;
1405 #else
1406 	pv_info.kernel_rpl = 0;
1407 #endif
1408 	/* set the limit of our address space */
1409 	xen_reserve_top();
1410 
1411 	/* We used to do this in xen_arch_setup, but that is too late on AMD
1412 	 * were early_cpu_init (run before ->arch_setup()) calls early_amd_init
1413 	 * which pokes 0xcf8 port.
1414 	 */
1415 	set_iopl.iopl = 1;
1416 	rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1417 	if (rc != 0)
1418 		xen_raw_printk("physdev_op failed %d\n", rc);
1419 
1420 #ifdef CONFIG_X86_32
1421 	/* set up basic CPUID stuff */
1422 	cpu_detect(&new_cpu_data);
1423 	new_cpu_data.hard_math = 1;
1424 	new_cpu_data.wp_works_ok = 1;
1425 	new_cpu_data.x86_capability[0] = cpuid_edx(1);
1426 #endif
1427 
1428 	/* Poke various useful things into boot_params */
1429 	boot_params.hdr.type_of_loader = (9 << 4) | 0;
1430 	boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1431 		? __pa(xen_start_info->mod_start) : 0;
1432 	boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1433 	boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1434 
1435 	if (!xen_initial_domain()) {
1436 		add_preferred_console("xenboot", 0, NULL);
1437 		add_preferred_console("tty", 0, NULL);
1438 		add_preferred_console("hvc", 0, NULL);
1439 		if (pci_xen)
1440 			x86_init.pci.arch_init = pci_xen_init;
1441 	} else {
1442 		const struct dom0_vga_console_info *info =
1443 			(void *)((char *)xen_start_info +
1444 				 xen_start_info->console.dom0.info_off);
1445 
1446 		xen_init_vga(info, xen_start_info->console.dom0.info_size);
1447 		xen_start_info->console.domU.mfn = 0;
1448 		xen_start_info->console.domU.evtchn = 0;
1449 
1450 		xen_init_apic();
1451 
1452 		/* Make sure ACS will be enabled */
1453 		pci_request_acs();
1454 
1455 		xen_acpi_sleep_register();
1456 	}
1457 #ifdef CONFIG_PCI
1458 	/* PCI BIOS service won't work from a PV guest. */
1459 	pci_probe &= ~PCI_PROBE_BIOS;
1460 #endif
1461 	xen_raw_console_write("about to get started...\n");
1462 
1463 	xen_setup_runstate_info(0);
1464 
1465 	/* Start the world */
1466 #ifdef CONFIG_X86_32
1467 	i386_start_kernel();
1468 #else
1469 	x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1470 #endif
1471 }
1472 
1473 #ifdef CONFIG_XEN_PVHVM
1474 /*
1475  * The pfn containing the shared_info is located somewhere in RAM. This
1476  * will cause trouble if the current kernel is doing a kexec boot into a
1477  * new kernel. The new kernel (and its startup code) can not know where
1478  * the pfn is, so it can not reserve the page. The hypervisor will
1479  * continue to update the pfn, and as a result memory corruption occours
1480  * in the new kernel.
1481  *
1482  * One way to work around this issue is to allocate a page in the
1483  * xen-platform pci device's BAR memory range. But pci init is done very
1484  * late and the shared_info page is already in use very early to read
1485  * the pvclock. So moving the pfn from RAM to MMIO is racy because some
1486  * code paths on other vcpus could access the pfn during the small
1487  * window when the old pfn is moved to the new pfn. There is even a
1488  * small window were the old pfn is not backed by a mfn, and during that
1489  * time all reads return -1.
1490  *
1491  * Because it is not known upfront where the MMIO region is located it
1492  * can not be used right from the start in xen_hvm_init_shared_info.
1493  *
1494  * To minimise trouble the move of the pfn is done shortly before kexec.
1495  * This does not eliminate the race because all vcpus are still online
1496  * when the syscore_ops will be called. But hopefully there is no work
1497  * pending at this point in time. Also the syscore_op is run last which
1498  * reduces the risk further.
1499  */
1500 
1501 static struct shared_info *xen_hvm_shared_info;
1502 
1503 static void xen_hvm_connect_shared_info(unsigned long pfn)
1504 {
1505 	struct xen_add_to_physmap xatp;
1506 
1507 	xatp.domid = DOMID_SELF;
1508 	xatp.idx = 0;
1509 	xatp.space = XENMAPSPACE_shared_info;
1510 	xatp.gpfn = pfn;
1511 	if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1512 		BUG();
1513 
1514 }
1515 static void xen_hvm_set_shared_info(struct shared_info *sip)
1516 {
1517 	int cpu;
1518 
1519 	HYPERVISOR_shared_info = sip;
1520 
1521 	/* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1522 	 * page, we use it in the event channel upcall and in some pvclock
1523 	 * related functions. We don't need the vcpu_info placement
1524 	 * optimizations because we don't use any pv_mmu or pv_irq op on
1525 	 * HVM.
1526 	 * When xen_hvm_set_shared_info is run at boot time only vcpu 0 is
1527 	 * online but xen_hvm_set_shared_info is run at resume time too and
1528 	 * in that case multiple vcpus might be online. */
1529 	for_each_online_cpu(cpu) {
1530 		per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1531 	}
1532 }
1533 
1534 /* Reconnect the shared_info pfn to a mfn */
1535 void xen_hvm_resume_shared_info(void)
1536 {
1537 	xen_hvm_connect_shared_info(__pa(xen_hvm_shared_info) >> PAGE_SHIFT);
1538 }
1539 
1540 #ifdef CONFIG_KEXEC
1541 static struct shared_info *xen_hvm_shared_info_kexec;
1542 static unsigned long xen_hvm_shared_info_pfn_kexec;
1543 
1544 /* Remember a pfn in MMIO space for kexec reboot */
1545 void __devinit xen_hvm_prepare_kexec(struct shared_info *sip, unsigned long pfn)
1546 {
1547 	xen_hvm_shared_info_kexec = sip;
1548 	xen_hvm_shared_info_pfn_kexec = pfn;
1549 }
1550 
1551 static void xen_hvm_syscore_shutdown(void)
1552 {
1553 	struct xen_memory_reservation reservation = {
1554 		.domid = DOMID_SELF,
1555 		.nr_extents = 1,
1556 	};
1557 	unsigned long prev_pfn;
1558 	int rc;
1559 
1560 	if (!xen_hvm_shared_info_kexec)
1561 		return;
1562 
1563 	prev_pfn = __pa(xen_hvm_shared_info) >> PAGE_SHIFT;
1564 	set_xen_guest_handle(reservation.extent_start, &prev_pfn);
1565 
1566 	/* Move pfn to MMIO, disconnects previous pfn from mfn */
1567 	xen_hvm_connect_shared_info(xen_hvm_shared_info_pfn_kexec);
1568 
1569 	/* Update pointers, following hypercall is also a memory barrier */
1570 	xen_hvm_set_shared_info(xen_hvm_shared_info_kexec);
1571 
1572 	/* Allocate new mfn for previous pfn */
1573 	do {
1574 		rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
1575 		if (rc == 0)
1576 			msleep(123);
1577 	} while (rc == 0);
1578 
1579 	/* Make sure the previous pfn is really connected to a (new) mfn */
1580 	BUG_ON(rc != 1);
1581 }
1582 
1583 static struct syscore_ops xen_hvm_syscore_ops = {
1584 	.shutdown = xen_hvm_syscore_shutdown,
1585 };
1586 #endif
1587 
1588 /* Use a pfn in RAM, may move to MMIO before kexec. */
1589 static void __init xen_hvm_init_shared_info(void)
1590 {
1591 	/* Remember pointer for resume */
1592 	xen_hvm_shared_info = extend_brk(PAGE_SIZE, PAGE_SIZE);
1593 	xen_hvm_connect_shared_info(__pa(xen_hvm_shared_info) >> PAGE_SHIFT);
1594 	xen_hvm_set_shared_info(xen_hvm_shared_info);
1595 }
1596 
1597 static void __init init_hvm_pv_info(void)
1598 {
1599 	int major, minor;
1600 	uint32_t eax, ebx, ecx, edx, pages, msr, base;
1601 	u64 pfn;
1602 
1603 	base = xen_cpuid_base();
1604 	cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1605 
1606 	major = eax >> 16;
1607 	minor = eax & 0xffff;
1608 	printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1609 
1610 	cpuid(base + 2, &pages, &msr, &ecx, &edx);
1611 
1612 	pfn = __pa(hypercall_page);
1613 	wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1614 
1615 	xen_setup_features();
1616 
1617 	pv_info.name = "Xen HVM";
1618 
1619 	xen_domain_type = XEN_HVM_DOMAIN;
1620 }
1621 
1622 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self,
1623 				    unsigned long action, void *hcpu)
1624 {
1625 	int cpu = (long)hcpu;
1626 	switch (action) {
1627 	case CPU_UP_PREPARE:
1628 		xen_vcpu_setup(cpu);
1629 		if (xen_have_vector_callback)
1630 			xen_init_lock_cpu(cpu);
1631 		break;
1632 	default:
1633 		break;
1634 	}
1635 	return NOTIFY_OK;
1636 }
1637 
1638 static struct notifier_block xen_hvm_cpu_notifier __cpuinitdata = {
1639 	.notifier_call	= xen_hvm_cpu_notify,
1640 };
1641 
1642 static void __init xen_hvm_guest_init(void)
1643 {
1644 	init_hvm_pv_info();
1645 
1646 	xen_hvm_init_shared_info();
1647 #ifdef CONFIG_KEXEC
1648 	register_syscore_ops(&xen_hvm_syscore_ops);
1649 #endif
1650 
1651 	if (xen_feature(XENFEAT_hvm_callback_vector))
1652 		xen_have_vector_callback = 1;
1653 	xen_hvm_smp_init();
1654 	register_cpu_notifier(&xen_hvm_cpu_notifier);
1655 	xen_unplug_emulated_devices();
1656 	x86_init.irqs.intr_init = xen_init_IRQ;
1657 	xen_hvm_init_time_ops();
1658 	xen_hvm_init_mmu_ops();
1659 }
1660 
1661 static bool __init xen_hvm_platform(void)
1662 {
1663 	if (xen_pv_domain())
1664 		return false;
1665 
1666 	if (!xen_cpuid_base())
1667 		return false;
1668 
1669 	return true;
1670 }
1671 
1672 bool xen_hvm_need_lapic(void)
1673 {
1674 	if (xen_pv_domain())
1675 		return false;
1676 	if (!xen_hvm_domain())
1677 		return false;
1678 	if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1679 		return false;
1680 	return true;
1681 }
1682 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1683 
1684 const struct hypervisor_x86 x86_hyper_xen_hvm __refconst = {
1685 	.name			= "Xen HVM",
1686 	.detect			= xen_hvm_platform,
1687 	.init_platform		= xen_hvm_guest_init,
1688 };
1689 EXPORT_SYMBOL(x86_hyper_xen_hvm);
1690 #endif
1691