xref: /openbmc/linux/arch/x86/xen/enlighten.c (revision a2cce7a9)
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/edd.h>
35 
36 #include <xen/xen.h>
37 #include <xen/events.h>
38 #include <xen/interface/xen.h>
39 #include <xen/interface/version.h>
40 #include <xen/interface/physdev.h>
41 #include <xen/interface/vcpu.h>
42 #include <xen/interface/memory.h>
43 #include <xen/interface/nmi.h>
44 #include <xen/interface/xen-mca.h>
45 #include <xen/features.h>
46 #include <xen/page.h>
47 #include <xen/hvm.h>
48 #include <xen/hvc-console.h>
49 #include <xen/acpi.h>
50 
51 #include <asm/paravirt.h>
52 #include <asm/apic.h>
53 #include <asm/page.h>
54 #include <asm/xen/pci.h>
55 #include <asm/xen/hypercall.h>
56 #include <asm/xen/hypervisor.h>
57 #include <asm/fixmap.h>
58 #include <asm/processor.h>
59 #include <asm/proto.h>
60 #include <asm/msr-index.h>
61 #include <asm/traps.h>
62 #include <asm/setup.h>
63 #include <asm/desc.h>
64 #include <asm/pgalloc.h>
65 #include <asm/pgtable.h>
66 #include <asm/tlbflush.h>
67 #include <asm/reboot.h>
68 #include <asm/stackprotector.h>
69 #include <asm/hypervisor.h>
70 #include <asm/mach_traps.h>
71 #include <asm/mwait.h>
72 #include <asm/pci_x86.h>
73 #include <asm/pat.h>
74 
75 #ifdef CONFIG_ACPI
76 #include <linux/acpi.h>
77 #include <asm/acpi.h>
78 #include <acpi/pdc_intel.h>
79 #include <acpi/processor.h>
80 #include <xen/interface/platform.h>
81 #endif
82 
83 #include "xen-ops.h"
84 #include "mmu.h"
85 #include "smp.h"
86 #include "multicalls.h"
87 #include "pmu.h"
88 
89 EXPORT_SYMBOL_GPL(hypercall_page);
90 
91 /*
92  * Pointer to the xen_vcpu_info structure or
93  * &HYPERVISOR_shared_info->vcpu_info[cpu]. See xen_hvm_init_shared_info
94  * and xen_vcpu_setup for details. By default it points to share_info->vcpu_info
95  * but if the hypervisor supports VCPUOP_register_vcpu_info then it can point
96  * to xen_vcpu_info. The pointer is used in __xen_evtchn_do_upcall to
97  * acknowledge pending events.
98  * Also more subtly it is used by the patched version of irq enable/disable
99  * e.g. xen_irq_enable_direct and xen_iret in PV mode.
100  *
101  * The desire to be able to do those mask/unmask operations as a single
102  * instruction by using the per-cpu offset held in %gs is the real reason
103  * vcpu info is in a per-cpu pointer and the original reason for this
104  * hypercall.
105  *
106  */
107 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
108 
109 /*
110  * Per CPU pages used if hypervisor supports VCPUOP_register_vcpu_info
111  * hypercall. This can be used both in PV and PVHVM mode. The structure
112  * overrides the default per_cpu(xen_vcpu, cpu) value.
113  */
114 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
115 
116 enum xen_domain_type xen_domain_type = XEN_NATIVE;
117 EXPORT_SYMBOL_GPL(xen_domain_type);
118 
119 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
120 EXPORT_SYMBOL(machine_to_phys_mapping);
121 unsigned long  machine_to_phys_nr;
122 EXPORT_SYMBOL(machine_to_phys_nr);
123 
124 struct start_info *xen_start_info;
125 EXPORT_SYMBOL_GPL(xen_start_info);
126 
127 struct shared_info xen_dummy_shared_info;
128 
129 void *xen_initial_gdt;
130 
131 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
132 __read_mostly int xen_have_vector_callback;
133 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
134 
135 /*
136  * Point at some empty memory to start with. We map the real shared_info
137  * page as soon as fixmap is up and running.
138  */
139 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
140 
141 /*
142  * Flag to determine whether vcpu info placement is available on all
143  * VCPUs.  We assume it is to start with, and then set it to zero on
144  * the first failure.  This is because it can succeed on some VCPUs
145  * and not others, since it can involve hypervisor memory allocation,
146  * or because the guest failed to guarantee all the appropriate
147  * constraints on all VCPUs (ie buffer can't cross a page boundary).
148  *
149  * Note that any particular CPU may be using a placed vcpu structure,
150  * but we can only optimise if the all are.
151  *
152  * 0: not available, 1: available
153  */
154 static int have_vcpu_info_placement = 1;
155 
156 struct tls_descs {
157 	struct desc_struct desc[3];
158 };
159 
160 /*
161  * Updating the 3 TLS descriptors in the GDT on every task switch is
162  * surprisingly expensive so we avoid updating them if they haven't
163  * changed.  Since Xen writes different descriptors than the one
164  * passed in the update_descriptor hypercall we keep shadow copies to
165  * compare against.
166  */
167 static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
168 
169 static void clamp_max_cpus(void)
170 {
171 #ifdef CONFIG_SMP
172 	if (setup_max_cpus > MAX_VIRT_CPUS)
173 		setup_max_cpus = MAX_VIRT_CPUS;
174 #endif
175 }
176 
177 static void xen_vcpu_setup(int cpu)
178 {
179 	struct vcpu_register_vcpu_info info;
180 	int err;
181 	struct vcpu_info *vcpup;
182 
183 	BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
184 
185 	/*
186 	 * This path is called twice on PVHVM - first during bootup via
187 	 * smp_init -> xen_hvm_cpu_notify, and then if the VCPU is being
188 	 * hotplugged: cpu_up -> xen_hvm_cpu_notify.
189 	 * As we can only do the VCPUOP_register_vcpu_info once lets
190 	 * not over-write its result.
191 	 *
192 	 * For PV it is called during restore (xen_vcpu_restore) and bootup
193 	 * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
194 	 * use this function.
195 	 */
196 	if (xen_hvm_domain()) {
197 		if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
198 			return;
199 	}
200 	if (cpu < MAX_VIRT_CPUS)
201 		per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
202 
203 	if (!have_vcpu_info_placement) {
204 		if (cpu >= MAX_VIRT_CPUS)
205 			clamp_max_cpus();
206 		return;
207 	}
208 
209 	vcpup = &per_cpu(xen_vcpu_info, cpu);
210 	info.mfn = arbitrary_virt_to_mfn(vcpup);
211 	info.offset = offset_in_page(vcpup);
212 
213 	/* Check to see if the hypervisor will put the vcpu_info
214 	   structure where we want it, which allows direct access via
215 	   a percpu-variable.
216 	   N.B. This hypercall can _only_ be called once per CPU. Subsequent
217 	   calls will error out with -EINVAL. This is due to the fact that
218 	   hypervisor has no unregister variant and this hypercall does not
219 	   allow to over-write info.mfn and info.offset.
220 	 */
221 	err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
222 
223 	if (err) {
224 		printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
225 		have_vcpu_info_placement = 0;
226 		clamp_max_cpus();
227 	} else {
228 		/* This cpu is using the registered vcpu info, even if
229 		   later ones fail to. */
230 		per_cpu(xen_vcpu, cpu) = vcpup;
231 	}
232 }
233 
234 /*
235  * On restore, set the vcpu placement up again.
236  * If it fails, then we're in a bad state, since
237  * we can't back out from using it...
238  */
239 void xen_vcpu_restore(void)
240 {
241 	int cpu;
242 
243 	for_each_possible_cpu(cpu) {
244 		bool other_cpu = (cpu != smp_processor_id());
245 		bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL);
246 
247 		if (other_cpu && is_up &&
248 		    HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
249 			BUG();
250 
251 		xen_setup_runstate_info(cpu);
252 
253 		if (have_vcpu_info_placement)
254 			xen_vcpu_setup(cpu);
255 
256 		if (other_cpu && is_up &&
257 		    HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
258 			BUG();
259 	}
260 }
261 
262 static void __init xen_banner(void)
263 {
264 	unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
265 	struct xen_extraversion extra;
266 	HYPERVISOR_xen_version(XENVER_extraversion, &extra);
267 
268 	pr_info("Booting paravirtualized kernel %son %s\n",
269 		xen_feature(XENFEAT_auto_translated_physmap) ?
270 			"with PVH extensions " : "", pv_info.name);
271 	printk(KERN_INFO "Xen version: %d.%d%s%s\n",
272 	       version >> 16, version & 0xffff, extra.extraversion,
273 	       xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
274 }
275 /* Check if running on Xen version (major, minor) or later */
276 bool
277 xen_running_on_version_or_later(unsigned int major, unsigned int minor)
278 {
279 	unsigned int version;
280 
281 	if (!xen_domain())
282 		return false;
283 
284 	version = HYPERVISOR_xen_version(XENVER_version, NULL);
285 	if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
286 		((version >> 16) > major))
287 		return true;
288 	return false;
289 }
290 
291 #define CPUID_THERM_POWER_LEAF 6
292 #define APERFMPERF_PRESENT 0
293 
294 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
295 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
296 
297 static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
298 static __read_mostly unsigned int cpuid_leaf5_ecx_val;
299 static __read_mostly unsigned int cpuid_leaf5_edx_val;
300 
301 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
302 		      unsigned int *cx, unsigned int *dx)
303 {
304 	unsigned maskebx = ~0;
305 	unsigned maskecx = ~0;
306 	unsigned maskedx = ~0;
307 	unsigned setecx = 0;
308 	/*
309 	 * Mask out inconvenient features, to try and disable as many
310 	 * unsupported kernel subsystems as possible.
311 	 */
312 	switch (*ax) {
313 	case 1:
314 		maskecx = cpuid_leaf1_ecx_mask;
315 		setecx = cpuid_leaf1_ecx_set_mask;
316 		maskedx = cpuid_leaf1_edx_mask;
317 		break;
318 
319 	case CPUID_MWAIT_LEAF:
320 		/* Synthesize the values.. */
321 		*ax = 0;
322 		*bx = 0;
323 		*cx = cpuid_leaf5_ecx_val;
324 		*dx = cpuid_leaf5_edx_val;
325 		return;
326 
327 	case CPUID_THERM_POWER_LEAF:
328 		/* Disabling APERFMPERF for kernel usage */
329 		maskecx = ~(1 << APERFMPERF_PRESENT);
330 		break;
331 
332 	case 0xb:
333 		/* Suppress extended topology stuff */
334 		maskebx = 0;
335 		break;
336 	}
337 
338 	asm(XEN_EMULATE_PREFIX "cpuid"
339 		: "=a" (*ax),
340 		  "=b" (*bx),
341 		  "=c" (*cx),
342 		  "=d" (*dx)
343 		: "0" (*ax), "2" (*cx));
344 
345 	*bx &= maskebx;
346 	*cx &= maskecx;
347 	*cx |= setecx;
348 	*dx &= maskedx;
349 
350 }
351 
352 static bool __init xen_check_mwait(void)
353 {
354 #ifdef CONFIG_ACPI
355 	struct xen_platform_op op = {
356 		.cmd			= XENPF_set_processor_pminfo,
357 		.u.set_pminfo.id	= -1,
358 		.u.set_pminfo.type	= XEN_PM_PDC,
359 	};
360 	uint32_t buf[3];
361 	unsigned int ax, bx, cx, dx;
362 	unsigned int mwait_mask;
363 
364 	/* We need to determine whether it is OK to expose the MWAIT
365 	 * capability to the kernel to harvest deeper than C3 states from ACPI
366 	 * _CST using the processor_harvest_xen.c module. For this to work, we
367 	 * need to gather the MWAIT_LEAF values (which the cstate.c code
368 	 * checks against). The hypervisor won't expose the MWAIT flag because
369 	 * it would break backwards compatibility; so we will find out directly
370 	 * from the hardware and hypercall.
371 	 */
372 	if (!xen_initial_domain())
373 		return false;
374 
375 	/*
376 	 * When running under platform earlier than Xen4.2, do not expose
377 	 * mwait, to avoid the risk of loading native acpi pad driver
378 	 */
379 	if (!xen_running_on_version_or_later(4, 2))
380 		return false;
381 
382 	ax = 1;
383 	cx = 0;
384 
385 	native_cpuid(&ax, &bx, &cx, &dx);
386 
387 	mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
388 		     (1 << (X86_FEATURE_MWAIT % 32));
389 
390 	if ((cx & mwait_mask) != mwait_mask)
391 		return false;
392 
393 	/* We need to emulate the MWAIT_LEAF and for that we need both
394 	 * ecx and edx. The hypercall provides only partial information.
395 	 */
396 
397 	ax = CPUID_MWAIT_LEAF;
398 	bx = 0;
399 	cx = 0;
400 	dx = 0;
401 
402 	native_cpuid(&ax, &bx, &cx, &dx);
403 
404 	/* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
405 	 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
406 	 */
407 	buf[0] = ACPI_PDC_REVISION_ID;
408 	buf[1] = 1;
409 	buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
410 
411 	set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
412 
413 	if ((HYPERVISOR_dom0_op(&op) == 0) &&
414 	    (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
415 		cpuid_leaf5_ecx_val = cx;
416 		cpuid_leaf5_edx_val = dx;
417 	}
418 	return true;
419 #else
420 	return false;
421 #endif
422 }
423 static void __init xen_init_cpuid_mask(void)
424 {
425 	unsigned int ax, bx, cx, dx;
426 	unsigned int xsave_mask;
427 
428 	cpuid_leaf1_edx_mask =
429 		~((1 << X86_FEATURE_MTRR) |  /* disable MTRR */
430 		  (1 << X86_FEATURE_ACC));   /* thermal monitoring */
431 
432 	if (!xen_initial_domain())
433 		cpuid_leaf1_edx_mask &=
434 			~((1 << X86_FEATURE_ACPI));  /* disable ACPI */
435 
436 	cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
437 
438 	ax = 1;
439 	cx = 0;
440 	cpuid(1, &ax, &bx, &cx, &dx);
441 
442 	xsave_mask =
443 		(1 << (X86_FEATURE_XSAVE % 32)) |
444 		(1 << (X86_FEATURE_OSXSAVE % 32));
445 
446 	/* Xen will set CR4.OSXSAVE if supported and not disabled by force */
447 	if ((cx & xsave_mask) != xsave_mask)
448 		cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
449 	if (xen_check_mwait())
450 		cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
451 }
452 
453 static void xen_set_debugreg(int reg, unsigned long val)
454 {
455 	HYPERVISOR_set_debugreg(reg, val);
456 }
457 
458 static unsigned long xen_get_debugreg(int reg)
459 {
460 	return HYPERVISOR_get_debugreg(reg);
461 }
462 
463 static void xen_end_context_switch(struct task_struct *next)
464 {
465 	xen_mc_flush();
466 	paravirt_end_context_switch(next);
467 }
468 
469 static unsigned long xen_store_tr(void)
470 {
471 	return 0;
472 }
473 
474 /*
475  * Set the page permissions for a particular virtual address.  If the
476  * address is a vmalloc mapping (or other non-linear mapping), then
477  * find the linear mapping of the page and also set its protections to
478  * match.
479  */
480 static void set_aliased_prot(void *v, pgprot_t prot)
481 {
482 	int level;
483 	pte_t *ptep;
484 	pte_t pte;
485 	unsigned long pfn;
486 	struct page *page;
487 	unsigned char dummy;
488 
489 	ptep = lookup_address((unsigned long)v, &level);
490 	BUG_ON(ptep == NULL);
491 
492 	pfn = pte_pfn(*ptep);
493 	page = pfn_to_page(pfn);
494 
495 	pte = pfn_pte(pfn, prot);
496 
497 	/*
498 	 * Careful: update_va_mapping() will fail if the virtual address
499 	 * we're poking isn't populated in the page tables.  We don't
500 	 * need to worry about the direct map (that's always in the page
501 	 * tables), but we need to be careful about vmap space.  In
502 	 * particular, the top level page table can lazily propagate
503 	 * entries between processes, so if we've switched mms since we
504 	 * vmapped the target in the first place, we might not have the
505 	 * top-level page table entry populated.
506 	 *
507 	 * We disable preemption because we want the same mm active when
508 	 * we probe the target and when we issue the hypercall.  We'll
509 	 * have the same nominal mm, but if we're a kernel thread, lazy
510 	 * mm dropping could change our pgd.
511 	 *
512 	 * Out of an abundance of caution, this uses __get_user() to fault
513 	 * in the target address just in case there's some obscure case
514 	 * in which the target address isn't readable.
515 	 */
516 
517 	preempt_disable();
518 
519 	pagefault_disable();	/* Avoid warnings due to being atomic. */
520 	__get_user(dummy, (unsigned char __user __force *)v);
521 	pagefault_enable();
522 
523 	if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
524 		BUG();
525 
526 	if (!PageHighMem(page)) {
527 		void *av = __va(PFN_PHYS(pfn));
528 
529 		if (av != v)
530 			if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
531 				BUG();
532 	} else
533 		kmap_flush_unused();
534 
535 	preempt_enable();
536 }
537 
538 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
539 {
540 	const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
541 	int i;
542 
543 	/*
544 	 * We need to mark the all aliases of the LDT pages RO.  We
545 	 * don't need to call vm_flush_aliases(), though, since that's
546 	 * only responsible for flushing aliases out the TLBs, not the
547 	 * page tables, and Xen will flush the TLB for us if needed.
548 	 *
549 	 * To avoid confusing future readers: none of this is necessary
550 	 * to load the LDT.  The hypervisor only checks this when the
551 	 * LDT is faulted in due to subsequent descriptor access.
552 	 */
553 
554 	for(i = 0; i < entries; i += entries_per_page)
555 		set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
556 }
557 
558 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
559 {
560 	const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
561 	int i;
562 
563 	for(i = 0; i < entries; i += entries_per_page)
564 		set_aliased_prot(ldt + i, PAGE_KERNEL);
565 }
566 
567 static void xen_set_ldt(const void *addr, unsigned entries)
568 {
569 	struct mmuext_op *op;
570 	struct multicall_space mcs = xen_mc_entry(sizeof(*op));
571 
572 	trace_xen_cpu_set_ldt(addr, entries);
573 
574 	op = mcs.args;
575 	op->cmd = MMUEXT_SET_LDT;
576 	op->arg1.linear_addr = (unsigned long)addr;
577 	op->arg2.nr_ents = entries;
578 
579 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
580 
581 	xen_mc_issue(PARAVIRT_LAZY_CPU);
582 }
583 
584 static void xen_load_gdt(const struct desc_ptr *dtr)
585 {
586 	unsigned long va = dtr->address;
587 	unsigned int size = dtr->size + 1;
588 	unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
589 	unsigned long frames[pages];
590 	int f;
591 
592 	/*
593 	 * A GDT can be up to 64k in size, which corresponds to 8192
594 	 * 8-byte entries, or 16 4k pages..
595 	 */
596 
597 	BUG_ON(size > 65536);
598 	BUG_ON(va & ~PAGE_MASK);
599 
600 	for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
601 		int level;
602 		pte_t *ptep;
603 		unsigned long pfn, mfn;
604 		void *virt;
605 
606 		/*
607 		 * The GDT is per-cpu and is in the percpu data area.
608 		 * That can be virtually mapped, so we need to do a
609 		 * page-walk to get the underlying MFN for the
610 		 * hypercall.  The page can also be in the kernel's
611 		 * linear range, so we need to RO that mapping too.
612 		 */
613 		ptep = lookup_address(va, &level);
614 		BUG_ON(ptep == NULL);
615 
616 		pfn = pte_pfn(*ptep);
617 		mfn = pfn_to_mfn(pfn);
618 		virt = __va(PFN_PHYS(pfn));
619 
620 		frames[f] = mfn;
621 
622 		make_lowmem_page_readonly((void *)va);
623 		make_lowmem_page_readonly(virt);
624 	}
625 
626 	if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
627 		BUG();
628 }
629 
630 /*
631  * load_gdt for early boot, when the gdt is only mapped once
632  */
633 static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
634 {
635 	unsigned long va = dtr->address;
636 	unsigned int size = dtr->size + 1;
637 	unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
638 	unsigned long frames[pages];
639 	int f;
640 
641 	/*
642 	 * A GDT can be up to 64k in size, which corresponds to 8192
643 	 * 8-byte entries, or 16 4k pages..
644 	 */
645 
646 	BUG_ON(size > 65536);
647 	BUG_ON(va & ~PAGE_MASK);
648 
649 	for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
650 		pte_t pte;
651 		unsigned long pfn, mfn;
652 
653 		pfn = virt_to_pfn(va);
654 		mfn = pfn_to_mfn(pfn);
655 
656 		pte = pfn_pte(pfn, PAGE_KERNEL_RO);
657 
658 		if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
659 			BUG();
660 
661 		frames[f] = mfn;
662 	}
663 
664 	if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
665 		BUG();
666 }
667 
668 static inline bool desc_equal(const struct desc_struct *d1,
669 			      const struct desc_struct *d2)
670 {
671 	return d1->a == d2->a && d1->b == d2->b;
672 }
673 
674 static void load_TLS_descriptor(struct thread_struct *t,
675 				unsigned int cpu, unsigned int i)
676 {
677 	struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
678 	struct desc_struct *gdt;
679 	xmaddr_t maddr;
680 	struct multicall_space mc;
681 
682 	if (desc_equal(shadow, &t->tls_array[i]))
683 		return;
684 
685 	*shadow = t->tls_array[i];
686 
687 	gdt = get_cpu_gdt_table(cpu);
688 	maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
689 	mc = __xen_mc_entry(0);
690 
691 	MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
692 }
693 
694 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
695 {
696 	/*
697 	 * XXX sleazy hack: If we're being called in a lazy-cpu zone
698 	 * and lazy gs handling is enabled, it means we're in a
699 	 * context switch, and %gs has just been saved.  This means we
700 	 * can zero it out to prevent faults on exit from the
701 	 * hypervisor if the next process has no %gs.  Either way, it
702 	 * has been saved, and the new value will get loaded properly.
703 	 * This will go away as soon as Xen has been modified to not
704 	 * save/restore %gs for normal hypercalls.
705 	 *
706 	 * On x86_64, this hack is not used for %gs, because gs points
707 	 * to KERNEL_GS_BASE (and uses it for PDA references), so we
708 	 * must not zero %gs on x86_64
709 	 *
710 	 * For x86_64, we need to zero %fs, otherwise we may get an
711 	 * exception between the new %fs descriptor being loaded and
712 	 * %fs being effectively cleared at __switch_to().
713 	 */
714 	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
715 #ifdef CONFIG_X86_32
716 		lazy_load_gs(0);
717 #else
718 		loadsegment(fs, 0);
719 #endif
720 	}
721 
722 	xen_mc_batch();
723 
724 	load_TLS_descriptor(t, cpu, 0);
725 	load_TLS_descriptor(t, cpu, 1);
726 	load_TLS_descriptor(t, cpu, 2);
727 
728 	xen_mc_issue(PARAVIRT_LAZY_CPU);
729 }
730 
731 #ifdef CONFIG_X86_64
732 static void xen_load_gs_index(unsigned int idx)
733 {
734 	if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
735 		BUG();
736 }
737 #endif
738 
739 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
740 				const void *ptr)
741 {
742 	xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
743 	u64 entry = *(u64 *)ptr;
744 
745 	trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
746 
747 	preempt_disable();
748 
749 	xen_mc_flush();
750 	if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
751 		BUG();
752 
753 	preempt_enable();
754 }
755 
756 static int cvt_gate_to_trap(int vector, const gate_desc *val,
757 			    struct trap_info *info)
758 {
759 	unsigned long addr;
760 
761 	if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
762 		return 0;
763 
764 	info->vector = vector;
765 
766 	addr = gate_offset(*val);
767 #ifdef CONFIG_X86_64
768 	/*
769 	 * Look for known traps using IST, and substitute them
770 	 * appropriately.  The debugger ones are the only ones we care
771 	 * about.  Xen will handle faults like double_fault,
772 	 * so we should never see them.  Warn if
773 	 * there's an unexpected IST-using fault handler.
774 	 */
775 	if (addr == (unsigned long)debug)
776 		addr = (unsigned long)xen_debug;
777 	else if (addr == (unsigned long)int3)
778 		addr = (unsigned long)xen_int3;
779 	else if (addr == (unsigned long)stack_segment)
780 		addr = (unsigned long)xen_stack_segment;
781 	else if (addr == (unsigned long)double_fault) {
782 		/* Don't need to handle these */
783 		return 0;
784 #ifdef CONFIG_X86_MCE
785 	} else if (addr == (unsigned long)machine_check) {
786 		/*
787 		 * when xen hypervisor inject vMCE to guest,
788 		 * use native mce handler to handle it
789 		 */
790 		;
791 #endif
792 	} else if (addr == (unsigned long)nmi)
793 		/*
794 		 * Use the native version as well.
795 		 */
796 		;
797 	else {
798 		/* Some other trap using IST? */
799 		if (WARN_ON(val->ist != 0))
800 			return 0;
801 	}
802 #endif	/* CONFIG_X86_64 */
803 	info->address = addr;
804 
805 	info->cs = gate_segment(*val);
806 	info->flags = val->dpl;
807 	/* interrupt gates clear IF */
808 	if (val->type == GATE_INTERRUPT)
809 		info->flags |= 1 << 2;
810 
811 	return 1;
812 }
813 
814 /* Locations of each CPU's IDT */
815 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
816 
817 /* Set an IDT entry.  If the entry is part of the current IDT, then
818    also update Xen. */
819 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
820 {
821 	unsigned long p = (unsigned long)&dt[entrynum];
822 	unsigned long start, end;
823 
824 	trace_xen_cpu_write_idt_entry(dt, entrynum, g);
825 
826 	preempt_disable();
827 
828 	start = __this_cpu_read(idt_desc.address);
829 	end = start + __this_cpu_read(idt_desc.size) + 1;
830 
831 	xen_mc_flush();
832 
833 	native_write_idt_entry(dt, entrynum, g);
834 
835 	if (p >= start && (p + 8) <= end) {
836 		struct trap_info info[2];
837 
838 		info[1].address = 0;
839 
840 		if (cvt_gate_to_trap(entrynum, g, &info[0]))
841 			if (HYPERVISOR_set_trap_table(info))
842 				BUG();
843 	}
844 
845 	preempt_enable();
846 }
847 
848 static void xen_convert_trap_info(const struct desc_ptr *desc,
849 				  struct trap_info *traps)
850 {
851 	unsigned in, out, count;
852 
853 	count = (desc->size+1) / sizeof(gate_desc);
854 	BUG_ON(count > 256);
855 
856 	for (in = out = 0; in < count; in++) {
857 		gate_desc *entry = (gate_desc*)(desc->address) + in;
858 
859 		if (cvt_gate_to_trap(in, entry, &traps[out]))
860 			out++;
861 	}
862 	traps[out].address = 0;
863 }
864 
865 void xen_copy_trap_info(struct trap_info *traps)
866 {
867 	const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
868 
869 	xen_convert_trap_info(desc, traps);
870 }
871 
872 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
873    hold a spinlock to protect the static traps[] array (static because
874    it avoids allocation, and saves stack space). */
875 static void xen_load_idt(const struct desc_ptr *desc)
876 {
877 	static DEFINE_SPINLOCK(lock);
878 	static struct trap_info traps[257];
879 
880 	trace_xen_cpu_load_idt(desc);
881 
882 	spin_lock(&lock);
883 
884 	memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
885 
886 	xen_convert_trap_info(desc, traps);
887 
888 	xen_mc_flush();
889 	if (HYPERVISOR_set_trap_table(traps))
890 		BUG();
891 
892 	spin_unlock(&lock);
893 }
894 
895 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
896    they're handled differently. */
897 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
898 				const void *desc, int type)
899 {
900 	trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
901 
902 	preempt_disable();
903 
904 	switch (type) {
905 	case DESC_LDT:
906 	case DESC_TSS:
907 		/* ignore */
908 		break;
909 
910 	default: {
911 		xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
912 
913 		xen_mc_flush();
914 		if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
915 			BUG();
916 	}
917 
918 	}
919 
920 	preempt_enable();
921 }
922 
923 /*
924  * Version of write_gdt_entry for use at early boot-time needed to
925  * update an entry as simply as possible.
926  */
927 static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
928 					    const void *desc, int type)
929 {
930 	trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
931 
932 	switch (type) {
933 	case DESC_LDT:
934 	case DESC_TSS:
935 		/* ignore */
936 		break;
937 
938 	default: {
939 		xmaddr_t maddr = virt_to_machine(&dt[entry]);
940 
941 		if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
942 			dt[entry] = *(struct desc_struct *)desc;
943 	}
944 
945 	}
946 }
947 
948 static void xen_load_sp0(struct tss_struct *tss,
949 			 struct thread_struct *thread)
950 {
951 	struct multicall_space mcs;
952 
953 	mcs = xen_mc_entry(0);
954 	MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
955 	xen_mc_issue(PARAVIRT_LAZY_CPU);
956 	tss->x86_tss.sp0 = thread->sp0;
957 }
958 
959 static void xen_set_iopl_mask(unsigned mask)
960 {
961 	struct physdev_set_iopl set_iopl;
962 
963 	/* Force the change at ring 0. */
964 	set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
965 	HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
966 }
967 
968 static void xen_io_delay(void)
969 {
970 }
971 
972 static void xen_clts(void)
973 {
974 	struct multicall_space mcs;
975 
976 	mcs = xen_mc_entry(0);
977 
978 	MULTI_fpu_taskswitch(mcs.mc, 0);
979 
980 	xen_mc_issue(PARAVIRT_LAZY_CPU);
981 }
982 
983 static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
984 
985 static unsigned long xen_read_cr0(void)
986 {
987 	unsigned long cr0 = this_cpu_read(xen_cr0_value);
988 
989 	if (unlikely(cr0 == 0)) {
990 		cr0 = native_read_cr0();
991 		this_cpu_write(xen_cr0_value, cr0);
992 	}
993 
994 	return cr0;
995 }
996 
997 static void xen_write_cr0(unsigned long cr0)
998 {
999 	struct multicall_space mcs;
1000 
1001 	this_cpu_write(xen_cr0_value, cr0);
1002 
1003 	/* Only pay attention to cr0.TS; everything else is
1004 	   ignored. */
1005 	mcs = xen_mc_entry(0);
1006 
1007 	MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
1008 
1009 	xen_mc_issue(PARAVIRT_LAZY_CPU);
1010 }
1011 
1012 static void xen_write_cr4(unsigned long cr4)
1013 {
1014 	cr4 &= ~(X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PCE);
1015 
1016 	native_write_cr4(cr4);
1017 }
1018 #ifdef CONFIG_X86_64
1019 static inline unsigned long xen_read_cr8(void)
1020 {
1021 	return 0;
1022 }
1023 static inline void xen_write_cr8(unsigned long val)
1024 {
1025 	BUG_ON(val);
1026 }
1027 #endif
1028 
1029 static u64 xen_read_msr_safe(unsigned int msr, int *err)
1030 {
1031 	u64 val;
1032 
1033 	if (pmu_msr_read(msr, &val, err))
1034 		return val;
1035 
1036 	val = native_read_msr_safe(msr, err);
1037 	switch (msr) {
1038 	case MSR_IA32_APICBASE:
1039 #ifdef CONFIG_X86_X2APIC
1040 		if (!(cpuid_ecx(1) & (1 << (X86_FEATURE_X2APIC & 31))))
1041 #endif
1042 			val &= ~X2APIC_ENABLE;
1043 		break;
1044 	}
1045 	return val;
1046 }
1047 
1048 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
1049 {
1050 	int ret;
1051 
1052 	ret = 0;
1053 
1054 	switch (msr) {
1055 #ifdef CONFIG_X86_64
1056 		unsigned which;
1057 		u64 base;
1058 
1059 	case MSR_FS_BASE:		which = SEGBASE_FS; goto set;
1060 	case MSR_KERNEL_GS_BASE:	which = SEGBASE_GS_USER; goto set;
1061 	case MSR_GS_BASE:		which = SEGBASE_GS_KERNEL; goto set;
1062 
1063 	set:
1064 		base = ((u64)high << 32) | low;
1065 		if (HYPERVISOR_set_segment_base(which, base) != 0)
1066 			ret = -EIO;
1067 		break;
1068 #endif
1069 
1070 	case MSR_STAR:
1071 	case MSR_CSTAR:
1072 	case MSR_LSTAR:
1073 	case MSR_SYSCALL_MASK:
1074 	case MSR_IA32_SYSENTER_CS:
1075 	case MSR_IA32_SYSENTER_ESP:
1076 	case MSR_IA32_SYSENTER_EIP:
1077 		/* Fast syscall setup is all done in hypercalls, so
1078 		   these are all ignored.  Stub them out here to stop
1079 		   Xen console noise. */
1080 
1081 	default:
1082 		if (!pmu_msr_write(msr, low, high, &ret))
1083 			ret = native_write_msr_safe(msr, low, high);
1084 	}
1085 
1086 	return ret;
1087 }
1088 
1089 void xen_setup_shared_info(void)
1090 {
1091 	if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1092 		set_fixmap(FIX_PARAVIRT_BOOTMAP,
1093 			   xen_start_info->shared_info);
1094 
1095 		HYPERVISOR_shared_info =
1096 			(struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1097 	} else
1098 		HYPERVISOR_shared_info =
1099 			(struct shared_info *)__va(xen_start_info->shared_info);
1100 
1101 #ifndef CONFIG_SMP
1102 	/* In UP this is as good a place as any to set up shared info */
1103 	xen_setup_vcpu_info_placement();
1104 #endif
1105 
1106 	xen_setup_mfn_list_list();
1107 }
1108 
1109 /* This is called once we have the cpu_possible_mask */
1110 void xen_setup_vcpu_info_placement(void)
1111 {
1112 	int cpu;
1113 
1114 	for_each_possible_cpu(cpu)
1115 		xen_vcpu_setup(cpu);
1116 
1117 	/* xen_vcpu_setup managed to place the vcpu_info within the
1118 	 * percpu area for all cpus, so make use of it. Note that for
1119 	 * PVH we want to use native IRQ mechanism. */
1120 	if (have_vcpu_info_placement && !xen_pvh_domain()) {
1121 		pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1122 		pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1123 		pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1124 		pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1125 		pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1126 	}
1127 }
1128 
1129 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1130 			  unsigned long addr, unsigned len)
1131 {
1132 	char *start, *end, *reloc;
1133 	unsigned ret;
1134 
1135 	start = end = reloc = NULL;
1136 
1137 #define SITE(op, x)							\
1138 	case PARAVIRT_PATCH(op.x):					\
1139 	if (have_vcpu_info_placement) {					\
1140 		start = (char *)xen_##x##_direct;			\
1141 		end = xen_##x##_direct_end;				\
1142 		reloc = xen_##x##_direct_reloc;				\
1143 	}								\
1144 	goto patch_site
1145 
1146 	switch (type) {
1147 		SITE(pv_irq_ops, irq_enable);
1148 		SITE(pv_irq_ops, irq_disable);
1149 		SITE(pv_irq_ops, save_fl);
1150 		SITE(pv_irq_ops, restore_fl);
1151 #undef SITE
1152 
1153 	patch_site:
1154 		if (start == NULL || (end-start) > len)
1155 			goto default_patch;
1156 
1157 		ret = paravirt_patch_insns(insnbuf, len, start, end);
1158 
1159 		/* Note: because reloc is assigned from something that
1160 		   appears to be an array, gcc assumes it's non-null,
1161 		   but doesn't know its relationship with start and
1162 		   end. */
1163 		if (reloc > start && reloc < end) {
1164 			int reloc_off = reloc - start;
1165 			long *relocp = (long *)(insnbuf + reloc_off);
1166 			long delta = start - (char *)addr;
1167 
1168 			*relocp += delta;
1169 		}
1170 		break;
1171 
1172 	default_patch:
1173 	default:
1174 		ret = paravirt_patch_default(type, clobbers, insnbuf,
1175 					     addr, len);
1176 		break;
1177 	}
1178 
1179 	return ret;
1180 }
1181 
1182 static const struct pv_info xen_info __initconst = {
1183 	.paravirt_enabled = 1,
1184 	.shared_kernel_pmd = 0,
1185 
1186 #ifdef CONFIG_X86_64
1187 	.extra_user_64bit_cs = FLAT_USER_CS64,
1188 #endif
1189 
1190 	.name = "Xen",
1191 };
1192 
1193 static const struct pv_init_ops xen_init_ops __initconst = {
1194 	.patch = xen_patch,
1195 };
1196 
1197 static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1198 	.cpuid = xen_cpuid,
1199 
1200 	.set_debugreg = xen_set_debugreg,
1201 	.get_debugreg = xen_get_debugreg,
1202 
1203 	.clts = xen_clts,
1204 
1205 	.read_cr0 = xen_read_cr0,
1206 	.write_cr0 = xen_write_cr0,
1207 
1208 	.read_cr4 = native_read_cr4,
1209 	.read_cr4_safe = native_read_cr4_safe,
1210 	.write_cr4 = xen_write_cr4,
1211 
1212 #ifdef CONFIG_X86_64
1213 	.read_cr8 = xen_read_cr8,
1214 	.write_cr8 = xen_write_cr8,
1215 #endif
1216 
1217 	.wbinvd = native_wbinvd,
1218 
1219 	.read_msr = xen_read_msr_safe,
1220 	.write_msr = xen_write_msr_safe,
1221 
1222 	.read_pmc = xen_read_pmc,
1223 
1224 	.iret = xen_iret,
1225 #ifdef CONFIG_X86_64
1226 	.usergs_sysret32 = xen_sysret32,
1227 	.usergs_sysret64 = xen_sysret64,
1228 #else
1229 	.irq_enable_sysexit = xen_sysexit,
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_idt = native_store_idt,
1245 	.store_tr = xen_store_tr,
1246 
1247 	.write_ldt_entry = xen_write_ldt_entry,
1248 	.write_gdt_entry = xen_write_gdt_entry,
1249 	.write_idt_entry = xen_write_idt_entry,
1250 	.load_sp0 = xen_load_sp0,
1251 
1252 	.set_iopl_mask = xen_set_iopl_mask,
1253 	.io_delay = xen_io_delay,
1254 
1255 	/* Xen takes care of %gs when switching to usermode for us */
1256 	.swapgs = paravirt_nop,
1257 
1258 	.start_context_switch = paravirt_start_context_switch,
1259 	.end_context_switch = xen_end_context_switch,
1260 };
1261 
1262 static const struct pv_apic_ops xen_apic_ops __initconst = {
1263 #ifdef CONFIG_X86_LOCAL_APIC
1264 	.startup_ipi_hook = paravirt_nop,
1265 #endif
1266 };
1267 
1268 static void xen_reboot(int reason)
1269 {
1270 	struct sched_shutdown r = { .reason = reason };
1271 	int cpu;
1272 
1273 	for_each_online_cpu(cpu)
1274 		xen_pmu_finish(cpu);
1275 
1276 	if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1277 		BUG();
1278 }
1279 
1280 static void xen_restart(char *msg)
1281 {
1282 	xen_reboot(SHUTDOWN_reboot);
1283 }
1284 
1285 static void xen_emergency_restart(void)
1286 {
1287 	xen_reboot(SHUTDOWN_reboot);
1288 }
1289 
1290 static void xen_machine_halt(void)
1291 {
1292 	xen_reboot(SHUTDOWN_poweroff);
1293 }
1294 
1295 static void xen_machine_power_off(void)
1296 {
1297 	if (pm_power_off)
1298 		pm_power_off();
1299 	xen_reboot(SHUTDOWN_poweroff);
1300 }
1301 
1302 static void xen_crash_shutdown(struct pt_regs *regs)
1303 {
1304 	xen_reboot(SHUTDOWN_crash);
1305 }
1306 
1307 static int
1308 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1309 {
1310 	xen_reboot(SHUTDOWN_crash);
1311 	return NOTIFY_DONE;
1312 }
1313 
1314 static struct notifier_block xen_panic_block = {
1315 	.notifier_call= xen_panic_event,
1316 	.priority = INT_MIN
1317 };
1318 
1319 int xen_panic_handler_init(void)
1320 {
1321 	atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1322 	return 0;
1323 }
1324 
1325 static const struct machine_ops xen_machine_ops __initconst = {
1326 	.restart = xen_restart,
1327 	.halt = xen_machine_halt,
1328 	.power_off = xen_machine_power_off,
1329 	.shutdown = xen_machine_halt,
1330 	.crash_shutdown = xen_crash_shutdown,
1331 	.emergency_restart = xen_emergency_restart,
1332 };
1333 
1334 static unsigned char xen_get_nmi_reason(void)
1335 {
1336 	unsigned char reason = 0;
1337 
1338 	/* Construct a value which looks like it came from port 0x61. */
1339 	if (test_bit(_XEN_NMIREASON_io_error,
1340 		     &HYPERVISOR_shared_info->arch.nmi_reason))
1341 		reason |= NMI_REASON_IOCHK;
1342 	if (test_bit(_XEN_NMIREASON_pci_serr,
1343 		     &HYPERVISOR_shared_info->arch.nmi_reason))
1344 		reason |= NMI_REASON_SERR;
1345 
1346 	return reason;
1347 }
1348 
1349 static void __init xen_boot_params_init_edd(void)
1350 {
1351 #if IS_ENABLED(CONFIG_EDD)
1352 	struct xen_platform_op op;
1353 	struct edd_info *edd_info;
1354 	u32 *mbr_signature;
1355 	unsigned nr;
1356 	int ret;
1357 
1358 	edd_info = boot_params.eddbuf;
1359 	mbr_signature = boot_params.edd_mbr_sig_buffer;
1360 
1361 	op.cmd = XENPF_firmware_info;
1362 
1363 	op.u.firmware_info.type = XEN_FW_DISK_INFO;
1364 	for (nr = 0; nr < EDDMAXNR; nr++) {
1365 		struct edd_info *info = edd_info + nr;
1366 
1367 		op.u.firmware_info.index = nr;
1368 		info->params.length = sizeof(info->params);
1369 		set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1370 				     &info->params);
1371 		ret = HYPERVISOR_dom0_op(&op);
1372 		if (ret)
1373 			break;
1374 
1375 #define C(x) info->x = op.u.firmware_info.u.disk_info.x
1376 		C(device);
1377 		C(version);
1378 		C(interface_support);
1379 		C(legacy_max_cylinder);
1380 		C(legacy_max_head);
1381 		C(legacy_sectors_per_track);
1382 #undef C
1383 	}
1384 	boot_params.eddbuf_entries = nr;
1385 
1386 	op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1387 	for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1388 		op.u.firmware_info.index = nr;
1389 		ret = HYPERVISOR_dom0_op(&op);
1390 		if (ret)
1391 			break;
1392 		mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1393 	}
1394 	boot_params.edd_mbr_sig_buf_entries = nr;
1395 #endif
1396 }
1397 
1398 /*
1399  * Set up the GDT and segment registers for -fstack-protector.  Until
1400  * we do this, we have to be careful not to call any stack-protected
1401  * function, which is most of the kernel.
1402  *
1403  * Note, that it is __ref because the only caller of this after init
1404  * is PVH which is not going to use xen_load_gdt_boot or other
1405  * __init functions.
1406  */
1407 static void __ref xen_setup_gdt(int cpu)
1408 {
1409 	if (xen_feature(XENFEAT_auto_translated_physmap)) {
1410 #ifdef CONFIG_X86_64
1411 		unsigned long dummy;
1412 
1413 		load_percpu_segment(cpu); /* We need to access per-cpu area */
1414 		switch_to_new_gdt(cpu); /* GDT and GS set */
1415 
1416 		/* We are switching of the Xen provided GDT to our HVM mode
1417 		 * GDT. The new GDT has  __KERNEL_CS with CS.L = 1
1418 		 * and we are jumping to reload it.
1419 		 */
1420 		asm volatile ("pushq %0\n"
1421 			      "leaq 1f(%%rip),%0\n"
1422 			      "pushq %0\n"
1423 			      "lretq\n"
1424 			      "1:\n"
1425 			      : "=&r" (dummy) : "0" (__KERNEL_CS));
1426 
1427 		/*
1428 		 * While not needed, we also set the %es, %ds, and %fs
1429 		 * to zero. We don't care about %ss as it is NULL.
1430 		 * Strictly speaking this is not needed as Xen zeros those
1431 		 * out (and also MSR_FS_BASE, MSR_GS_BASE, MSR_KERNEL_GS_BASE)
1432 		 *
1433 		 * Linux zeros them in cpu_init() and in secondary_startup_64
1434 		 * (for BSP).
1435 		 */
1436 		loadsegment(es, 0);
1437 		loadsegment(ds, 0);
1438 		loadsegment(fs, 0);
1439 #else
1440 		/* PVH: TODO Implement. */
1441 		BUG();
1442 #endif
1443 		return; /* PVH does not need any PV GDT ops. */
1444 	}
1445 	pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1446 	pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1447 
1448 	setup_stack_canary_segment(0);
1449 	switch_to_new_gdt(0);
1450 
1451 	pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1452 	pv_cpu_ops.load_gdt = xen_load_gdt;
1453 }
1454 
1455 #ifdef CONFIG_XEN_PVH
1456 /*
1457  * A PV guest starts with default flags that are not set for PVH, set them
1458  * here asap.
1459  */
1460 static void xen_pvh_set_cr_flags(int cpu)
1461 {
1462 
1463 	/* Some of these are setup in 'secondary_startup_64'. The others:
1464 	 * X86_CR0_TS, X86_CR0_PE, X86_CR0_ET are set by Xen for HVM guests
1465 	 * (which PVH shared codepaths), while X86_CR0_PG is for PVH. */
1466 	write_cr0(read_cr0() | X86_CR0_MP | X86_CR0_NE | X86_CR0_WP | X86_CR0_AM);
1467 
1468 	if (!cpu)
1469 		return;
1470 	/*
1471 	 * For BSP, PSE PGE are set in probe_page_size_mask(), for APs
1472 	 * set them here. For all, OSFXSR OSXMMEXCPT are set in fpu__init_cpu().
1473 	*/
1474 	if (cpu_has_pse)
1475 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
1476 
1477 	if (cpu_has_pge)
1478 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
1479 }
1480 
1481 /*
1482  * Note, that it is ref - because the only caller of this after init
1483  * is PVH which is not going to use xen_load_gdt_boot or other
1484  * __init functions.
1485  */
1486 void __ref xen_pvh_secondary_vcpu_init(int cpu)
1487 {
1488 	xen_setup_gdt(cpu);
1489 	xen_pvh_set_cr_flags(cpu);
1490 }
1491 
1492 static void __init xen_pvh_early_guest_init(void)
1493 {
1494 	if (!xen_feature(XENFEAT_auto_translated_physmap))
1495 		return;
1496 
1497 	if (!xen_feature(XENFEAT_hvm_callback_vector))
1498 		return;
1499 
1500 	xen_have_vector_callback = 1;
1501 
1502 	xen_pvh_early_cpu_init(0, false);
1503 	xen_pvh_set_cr_flags(0);
1504 
1505 #ifdef CONFIG_X86_32
1506 	BUG(); /* PVH: Implement proper support. */
1507 #endif
1508 }
1509 #endif    /* CONFIG_XEN_PVH */
1510 
1511 /* First C function to be called on Xen boot */
1512 asmlinkage __visible void __init xen_start_kernel(void)
1513 {
1514 	struct physdev_set_iopl set_iopl;
1515 	unsigned long initrd_start = 0;
1516 	u64 pat;
1517 	int rc;
1518 
1519 	if (!xen_start_info)
1520 		return;
1521 
1522 	xen_domain_type = XEN_PV_DOMAIN;
1523 
1524 	xen_setup_features();
1525 #ifdef CONFIG_XEN_PVH
1526 	xen_pvh_early_guest_init();
1527 #endif
1528 	xen_setup_machphys_mapping();
1529 
1530 	/* Install Xen paravirt ops */
1531 	pv_info = xen_info;
1532 	pv_init_ops = xen_init_ops;
1533 	pv_apic_ops = xen_apic_ops;
1534 	if (!xen_pvh_domain()) {
1535 		pv_cpu_ops = xen_cpu_ops;
1536 
1537 		x86_platform.get_nmi_reason = xen_get_nmi_reason;
1538 	}
1539 
1540 	if (xen_feature(XENFEAT_auto_translated_physmap))
1541 		x86_init.resources.memory_setup = xen_auto_xlated_memory_setup;
1542 	else
1543 		x86_init.resources.memory_setup = xen_memory_setup;
1544 	x86_init.oem.arch_setup = xen_arch_setup;
1545 	x86_init.oem.banner = xen_banner;
1546 
1547 	xen_init_time_ops();
1548 
1549 	/*
1550 	 * Set up some pagetable state before starting to set any ptes.
1551 	 */
1552 
1553 	xen_init_mmu_ops();
1554 
1555 	/* Prevent unwanted bits from being set in PTEs. */
1556 	__supported_pte_mask &= ~_PAGE_GLOBAL;
1557 
1558 	/*
1559 	 * Prevent page tables from being allocated in highmem, even
1560 	 * if CONFIG_HIGHPTE is enabled.
1561 	 */
1562 	__userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1563 
1564 	/* Work out if we support NX */
1565 	x86_configure_nx();
1566 
1567 	/* Get mfn list */
1568 	xen_build_dynamic_phys_to_machine();
1569 
1570 	/*
1571 	 * Set up kernel GDT and segment registers, mainly so that
1572 	 * -fstack-protector code can be executed.
1573 	 */
1574 	xen_setup_gdt(0);
1575 
1576 	xen_init_irq_ops();
1577 	xen_init_cpuid_mask();
1578 
1579 #ifdef CONFIG_X86_LOCAL_APIC
1580 	/*
1581 	 * set up the basic apic ops.
1582 	 */
1583 	xen_init_apic();
1584 #endif
1585 
1586 	if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1587 		pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1588 		pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1589 	}
1590 
1591 	machine_ops = xen_machine_ops;
1592 
1593 	/*
1594 	 * The only reliable way to retain the initial address of the
1595 	 * percpu gdt_page is to remember it here, so we can go and
1596 	 * mark it RW later, when the initial percpu area is freed.
1597 	 */
1598 	xen_initial_gdt = &per_cpu(gdt_page, 0);
1599 
1600 	xen_smp_init();
1601 
1602 #ifdef CONFIG_ACPI_NUMA
1603 	/*
1604 	 * The pages we from Xen are not related to machine pages, so
1605 	 * any NUMA information the kernel tries to get from ACPI will
1606 	 * be meaningless.  Prevent it from trying.
1607 	 */
1608 	acpi_numa = -1;
1609 #endif
1610 	/* Don't do the full vcpu_info placement stuff until we have a
1611 	   possible map and a non-dummy shared_info. */
1612 	per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1613 
1614 	local_irq_disable();
1615 	early_boot_irqs_disabled = true;
1616 
1617 	xen_raw_console_write("mapping kernel into physical memory\n");
1618 	xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1619 				   xen_start_info->nr_pages);
1620 	xen_reserve_special_pages();
1621 
1622 	/*
1623 	 * Modify the cache mode translation tables to match Xen's PAT
1624 	 * configuration.
1625 	 */
1626 	rdmsrl(MSR_IA32_CR_PAT, pat);
1627 	pat_init_cache_modes(pat);
1628 
1629 	/* keep using Xen gdt for now; no urgent need to change it */
1630 
1631 #ifdef CONFIG_X86_32
1632 	pv_info.kernel_rpl = 1;
1633 	if (xen_feature(XENFEAT_supervisor_mode_kernel))
1634 		pv_info.kernel_rpl = 0;
1635 #else
1636 	pv_info.kernel_rpl = 0;
1637 #endif
1638 	/* set the limit of our address space */
1639 	xen_reserve_top();
1640 
1641 	/* PVH: runs at default kernel iopl of 0 */
1642 	if (!xen_pvh_domain()) {
1643 		/*
1644 		 * We used to do this in xen_arch_setup, but that is too late
1645 		 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1646 		 * early_amd_init which pokes 0xcf8 port.
1647 		 */
1648 		set_iopl.iopl = 1;
1649 		rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1650 		if (rc != 0)
1651 			xen_raw_printk("physdev_op failed %d\n", rc);
1652 	}
1653 
1654 #ifdef CONFIG_X86_32
1655 	/* set up basic CPUID stuff */
1656 	cpu_detect(&new_cpu_data);
1657 	set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1658 	new_cpu_data.wp_works_ok = 1;
1659 	new_cpu_data.x86_capability[0] = cpuid_edx(1);
1660 #endif
1661 
1662 	if (xen_start_info->mod_start) {
1663 	    if (xen_start_info->flags & SIF_MOD_START_PFN)
1664 		initrd_start = PFN_PHYS(xen_start_info->mod_start);
1665 	    else
1666 		initrd_start = __pa(xen_start_info->mod_start);
1667 	}
1668 
1669 	/* Poke various useful things into boot_params */
1670 	boot_params.hdr.type_of_loader = (9 << 4) | 0;
1671 	boot_params.hdr.ramdisk_image = initrd_start;
1672 	boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1673 	boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1674 
1675 	if (!xen_initial_domain()) {
1676 		add_preferred_console("xenboot", 0, NULL);
1677 		add_preferred_console("tty", 0, NULL);
1678 		add_preferred_console("hvc", 0, NULL);
1679 		if (pci_xen)
1680 			x86_init.pci.arch_init = pci_xen_init;
1681 	} else {
1682 		const struct dom0_vga_console_info *info =
1683 			(void *)((char *)xen_start_info +
1684 				 xen_start_info->console.dom0.info_off);
1685 		struct xen_platform_op op = {
1686 			.cmd = XENPF_firmware_info,
1687 			.interface_version = XENPF_INTERFACE_VERSION,
1688 			.u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1689 		};
1690 
1691 		xen_init_vga(info, xen_start_info->console.dom0.info_size);
1692 		xen_start_info->console.domU.mfn = 0;
1693 		xen_start_info->console.domU.evtchn = 0;
1694 
1695 		if (HYPERVISOR_dom0_op(&op) == 0)
1696 			boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1697 
1698 		/* Make sure ACS will be enabled */
1699 		pci_request_acs();
1700 
1701 		xen_acpi_sleep_register();
1702 
1703 		/* Avoid searching for BIOS MP tables */
1704 		x86_init.mpparse.find_smp_config = x86_init_noop;
1705 		x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1706 
1707 		xen_boot_params_init_edd();
1708 	}
1709 #ifdef CONFIG_PCI
1710 	/* PCI BIOS service won't work from a PV guest. */
1711 	pci_probe &= ~PCI_PROBE_BIOS;
1712 #endif
1713 	xen_raw_console_write("about to get started...\n");
1714 
1715 	xen_setup_runstate_info(0);
1716 
1717 	xen_efi_init();
1718 
1719 	/* Start the world */
1720 #ifdef CONFIG_X86_32
1721 	i386_start_kernel();
1722 #else
1723 	cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1724 	x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1725 #endif
1726 }
1727 
1728 void __ref xen_hvm_init_shared_info(void)
1729 {
1730 	int cpu;
1731 	struct xen_add_to_physmap xatp;
1732 	static struct shared_info *shared_info_page = 0;
1733 
1734 	if (!shared_info_page)
1735 		shared_info_page = (struct shared_info *)
1736 			extend_brk(PAGE_SIZE, PAGE_SIZE);
1737 	xatp.domid = DOMID_SELF;
1738 	xatp.idx = 0;
1739 	xatp.space = XENMAPSPACE_shared_info;
1740 	xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1741 	if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1742 		BUG();
1743 
1744 	HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1745 
1746 	/* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1747 	 * page, we use it in the event channel upcall and in some pvclock
1748 	 * related functions. We don't need the vcpu_info placement
1749 	 * optimizations because we don't use any pv_mmu or pv_irq op on
1750 	 * HVM.
1751 	 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1752 	 * online but xen_hvm_init_shared_info is run at resume time too and
1753 	 * in that case multiple vcpus might be online. */
1754 	for_each_online_cpu(cpu) {
1755 		/* Leave it to be NULL. */
1756 		if (cpu >= MAX_VIRT_CPUS)
1757 			continue;
1758 		per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
1759 	}
1760 }
1761 
1762 #ifdef CONFIG_XEN_PVHVM
1763 static void __init init_hvm_pv_info(void)
1764 {
1765 	int major, minor;
1766 	uint32_t eax, ebx, ecx, edx, pages, msr, base;
1767 	u64 pfn;
1768 
1769 	base = xen_cpuid_base();
1770 	cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1771 
1772 	major = eax >> 16;
1773 	minor = eax & 0xffff;
1774 	printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1775 
1776 	cpuid(base + 2, &pages, &msr, &ecx, &edx);
1777 
1778 	pfn = __pa(hypercall_page);
1779 	wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1780 
1781 	xen_setup_features();
1782 
1783 	pv_info.name = "Xen HVM";
1784 
1785 	xen_domain_type = XEN_HVM_DOMAIN;
1786 }
1787 
1788 static int xen_hvm_cpu_notify(struct notifier_block *self, unsigned long action,
1789 			      void *hcpu)
1790 {
1791 	int cpu = (long)hcpu;
1792 	switch (action) {
1793 	case CPU_UP_PREPARE:
1794 		xen_vcpu_setup(cpu);
1795 		if (xen_have_vector_callback) {
1796 			if (xen_feature(XENFEAT_hvm_safe_pvclock))
1797 				xen_setup_timer(cpu);
1798 		}
1799 		break;
1800 	default:
1801 		break;
1802 	}
1803 	return NOTIFY_OK;
1804 }
1805 
1806 static struct notifier_block xen_hvm_cpu_notifier = {
1807 	.notifier_call	= xen_hvm_cpu_notify,
1808 };
1809 
1810 static void __init xen_hvm_guest_init(void)
1811 {
1812 	if (xen_pv_domain())
1813 		return;
1814 
1815 	init_hvm_pv_info();
1816 
1817 	xen_hvm_init_shared_info();
1818 
1819 	xen_panic_handler_init();
1820 
1821 	if (xen_feature(XENFEAT_hvm_callback_vector))
1822 		xen_have_vector_callback = 1;
1823 	xen_hvm_smp_init();
1824 	register_cpu_notifier(&xen_hvm_cpu_notifier);
1825 	xen_unplug_emulated_devices();
1826 	x86_init.irqs.intr_init = xen_init_IRQ;
1827 	xen_hvm_init_time_ops();
1828 	xen_hvm_init_mmu_ops();
1829 }
1830 #endif
1831 
1832 static bool xen_nopv = false;
1833 static __init int xen_parse_nopv(char *arg)
1834 {
1835        xen_nopv = true;
1836        return 0;
1837 }
1838 early_param("xen_nopv", xen_parse_nopv);
1839 
1840 static uint32_t __init xen_platform(void)
1841 {
1842 	if (xen_nopv)
1843 		return 0;
1844 
1845 	return xen_cpuid_base();
1846 }
1847 
1848 bool xen_hvm_need_lapic(void)
1849 {
1850 	if (xen_nopv)
1851 		return false;
1852 	if (xen_pv_domain())
1853 		return false;
1854 	if (!xen_hvm_domain())
1855 		return false;
1856 	if (xen_feature(XENFEAT_hvm_pirqs) && xen_have_vector_callback)
1857 		return false;
1858 	return true;
1859 }
1860 EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1861 
1862 static void xen_set_cpu_features(struct cpuinfo_x86 *c)
1863 {
1864 	if (xen_pv_domain())
1865 		clear_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1866 }
1867 
1868 const struct hypervisor_x86 x86_hyper_xen = {
1869 	.name			= "Xen",
1870 	.detect			= xen_platform,
1871 #ifdef CONFIG_XEN_PVHVM
1872 	.init_platform		= xen_hvm_guest_init,
1873 #endif
1874 	.x2apic_available	= xen_x2apic_para_available,
1875 	.set_cpu_features       = xen_set_cpu_features,
1876 };
1877 EXPORT_SYMBOL(x86_hyper_xen);
1878