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/export.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#include <linux/frame.h>
36
37#include <linux/kexec.h>
38
39#include <xen/xen.h>
40#include <xen/events.h>
41#include <xen/interface/xen.h>
42#include <xen/interface/version.h>
43#include <xen/interface/physdev.h>
44#include <xen/interface/vcpu.h>
45#include <xen/interface/memory.h>
46#include <xen/interface/nmi.h>
47#include <xen/interface/xen-mca.h>
48#include <xen/features.h>
49#include <xen/page.h>
50#include <xen/hvm.h>
51#include <xen/hvc-console.h>
52#include <xen/acpi.h>
53
54#include <asm/paravirt.h>
55#include <asm/apic.h>
56#include <asm/page.h>
57#include <asm/xen/pci.h>
58#include <asm/xen/hypercall.h>
59#include <asm/xen/hypervisor.h>
60#include <asm/xen/cpuid.h>
61#include <asm/fixmap.h>
62#include <asm/processor.h>
63#include <asm/proto.h>
64#include <asm/msr-index.h>
65#include <asm/traps.h>
66#include <asm/setup.h>
67#include <asm/desc.h>
68#include <asm/pgalloc.h>
69#include <asm/pgtable.h>
70#include <asm/tlbflush.h>
71#include <asm/reboot.h>
72#include <asm/stackprotector.h>
73#include <asm/hypervisor.h>
74#include <asm/mach_traps.h>
75#include <asm/mwait.h>
76#include <asm/pci_x86.h>
77#include <asm/cpu.h>
78
79#ifdef CONFIG_ACPI
80#include <linux/acpi.h>
81#include <asm/acpi.h>
82#include <acpi/pdc_intel.h>
83#include <acpi/processor.h>
84#include <xen/interface/platform.h>
85#endif
86
87#include "xen-ops.h"
88#include "mmu.h"
89#include "smp.h"
90#include "multicalls.h"
91#include "pmu.h"
92
93EXPORT_SYMBOL_GPL(hypercall_page);
94
95/*
96 * Pointer to the xen_vcpu_info structure or
97 * &HYPERVISOR_shared_info->vcpu_info[cpu]. See xen_hvm_init_shared_info
98 * and xen_vcpu_setup for details. By default it points to share_info->vcpu_info
99 * but if the hypervisor supports VCPUOP_register_vcpu_info then it can point
100 * to xen_vcpu_info. The pointer is used in __xen_evtchn_do_upcall to
101 * acknowledge pending events.
102 * Also more subtly it is used by the patched version of irq enable/disable
103 * e.g. xen_irq_enable_direct and xen_iret in PV mode.
104 *
105 * The desire to be able to do those mask/unmask operations as a single
106 * instruction by using the per-cpu offset held in %gs is the real reason
107 * vcpu info is in a per-cpu pointer and the original reason for this
108 * hypercall.
109 *
110 */
111DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
112
113/*
114 * Per CPU pages used if hypervisor supports VCPUOP_register_vcpu_info
115 * hypercall. This can be used both in PV and PVHVM mode. The structure
116 * overrides the default per_cpu(xen_vcpu, cpu) value.
117 */
118DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
119
120/* Linux <-> Xen vCPU id mapping */
121DEFINE_PER_CPU(uint32_t, xen_vcpu_id);
122EXPORT_PER_CPU_SYMBOL(xen_vcpu_id);
123
124enum xen_domain_type xen_domain_type = XEN_NATIVE;
125EXPORT_SYMBOL_GPL(xen_domain_type);
126
127unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
128EXPORT_SYMBOL(machine_to_phys_mapping);
129unsigned long machine_to_phys_nr;
130EXPORT_SYMBOL(machine_to_phys_nr);
131
132struct start_info *xen_start_info;
133EXPORT_SYMBOL_GPL(xen_start_info);
134
135struct shared_info xen_dummy_shared_info;
136
137void *xen_initial_gdt;
138
139RESERVE_BRK(shared_info_page_brk, PAGE_SIZE);
140
141static int xen_cpu_up_prepare(unsigned int cpu);
142static int xen_cpu_up_online(unsigned int cpu);
143static int xen_cpu_dead(unsigned int cpu);
144
145/*
146 * Point at some empty memory to start with. We map the real shared_info
147 * page as soon as fixmap is up and running.
148 */
149struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
150
151/*
152 * Flag to determine whether vcpu info placement is available on all
153 * VCPUs. We assume it is to start with, and then set it to zero on
154 * the first failure. This is because it can succeed on some VCPUs
155 * and not others, since it can involve hypervisor memory allocation,
156 * or because the guest failed to guarantee all the appropriate
157 * constraints on all VCPUs (ie buffer can't cross a page boundary).
158 *
159 * Note that any particular CPU may be using a placed vcpu structure,
160 * but we can only optimise if the all are.
161 *
162 * 0: not available, 1: available
163 */
164static int have_vcpu_info_placement = 1;
165
166struct tls_descs {
167 struct desc_struct desc[3];
168};
169
170/*
171 * Updating the 3 TLS descriptors in the GDT on every task switch is
172 * surprisingly expensive so we avoid updating them if they haven't
173 * changed. Since Xen writes different descriptors than the one
174 * passed in the update_descriptor hypercall we keep shadow copies to
175 * compare against.
176 */
177static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
178
179static void clamp_max_cpus(void)
180{
181#ifdef CONFIG_SMP
182 if (setup_max_cpus > MAX_VIRT_CPUS)
183 setup_max_cpus = MAX_VIRT_CPUS;
184#endif
185}
186
187void xen_vcpu_setup(int cpu)
188{
189 struct vcpu_register_vcpu_info info;
190 int err;
191 struct vcpu_info *vcpup;
192
193 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
194
195 /*
196 * This path is called twice on PVHVM - first during bootup via
197 * smp_init -> xen_hvm_cpu_notify, and then if the VCPU is being
198 * hotplugged: cpu_up -> xen_hvm_cpu_notify.
199 * As we can only do the VCPUOP_register_vcpu_info once lets
200 * not over-write its result.
201 *
202 * For PV it is called during restore (xen_vcpu_restore) and bootup
203 * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
204 * use this function.
205 */
206 if (xen_hvm_domain()) {
207 if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
208 return;
209 }
210 if (xen_vcpu_nr(cpu) < MAX_VIRT_CPUS)
211 per_cpu(xen_vcpu, cpu) =
212 &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
213
214 if (!have_vcpu_info_placement) {
215 if (cpu >= MAX_VIRT_CPUS)
216 clamp_max_cpus();
217 return;
218 }
219
220 vcpup = &per_cpu(xen_vcpu_info, cpu);
221 info.mfn = arbitrary_virt_to_mfn(vcpup);
222 info.offset = offset_in_page(vcpup);
223
224 /* Check to see if the hypervisor will put the vcpu_info
225 structure where we want it, which allows direct access via
226 a percpu-variable.
227 N.B. This hypercall can _only_ be called once per CPU. Subsequent
228 calls will error out with -EINVAL. This is due to the fact that
229 hypervisor has no unregister variant and this hypercall does not
230 allow to over-write info.mfn and info.offset.
231 */
232 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, xen_vcpu_nr(cpu),
233 &info);
234
235 if (err) {
236 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
237 have_vcpu_info_placement = 0;
238 clamp_max_cpus();
239 } else {
240 /* This cpu is using the registered vcpu info, even if
241 later ones fail to. */
242 per_cpu(xen_vcpu, cpu) = vcpup;
243 }
244}
245
246/*
247 * On restore, set the vcpu placement up again.
248 * If it fails, then we're in a bad state, since
249 * we can't back out from using it...
250 */
251void xen_vcpu_restore(void)
252{
253 int cpu;
254
255 for_each_possible_cpu(cpu) {
256 bool other_cpu = (cpu != smp_processor_id());
257 bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu),
258 NULL);
259
260 if (other_cpu && is_up &&
261 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
262 BUG();
263
264 xen_setup_runstate_info(cpu);
265
266 if (have_vcpu_info_placement)
267 xen_vcpu_setup(cpu);
268
269 if (other_cpu && is_up &&
270 HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
271 BUG();
272 }
273}
274
275static void __init xen_banner(void)
276{
277 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
278 struct xen_extraversion extra;
279 HYPERVISOR_xen_version(XENVER_extraversion, &extra);
280
281 pr_info("Booting paravirtualized kernel %son %s\n",
282 xen_feature(XENFEAT_auto_translated_physmap) ?
283 "with PVH extensions " : "", pv_info.name);
284 printk(KERN_INFO "Xen version: %d.%d%s%s\n",
285 version >> 16, version & 0xffff, extra.extraversion,
286 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
287}
288/* Check if running on Xen version (major, minor) or later */
289bool
290xen_running_on_version_or_later(unsigned int major, unsigned int minor)
291{
292 unsigned int version;
293
294 if (!xen_domain())
295 return false;
296
297 version = HYPERVISOR_xen_version(XENVER_version, NULL);
298 if ((((version >> 16) == major) && ((version & 0xffff) >= minor)) ||
299 ((version >> 16) > major))
300 return true;
301 return false;
302}
303
304#define CPUID_THERM_POWER_LEAF 6
305#define APERFMPERF_PRESENT 0
306
307static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
308static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
309
310static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
311static __read_mostly unsigned int cpuid_leaf5_ecx_val;
312static __read_mostly unsigned int cpuid_leaf5_edx_val;
313
314static void xen_cpuid(unsigned int *ax, unsigned int *bx,
315 unsigned int *cx, unsigned int *dx)
316{
317 unsigned maskebx = ~0;
318 unsigned maskecx = ~0;
319 unsigned maskedx = ~0;
320 unsigned setecx = 0;
321 /*
322 * Mask out inconvenient features, to try and disable as many
323 * unsupported kernel subsystems as possible.
324 */
325 switch (*ax) {
326 case 1:
327 maskecx = cpuid_leaf1_ecx_mask;
328 setecx = cpuid_leaf1_ecx_set_mask;
329 maskedx = cpuid_leaf1_edx_mask;
330 break;
331
332 case CPUID_MWAIT_LEAF:
333 /* Synthesize the values.. */
334 *ax = 0;
335 *bx = 0;
336 *cx = cpuid_leaf5_ecx_val;
337 *dx = cpuid_leaf5_edx_val;
338 return;
339
340 case CPUID_THERM_POWER_LEAF:
341 /* Disabling APERFMPERF for kernel usage */
342 maskecx = ~(1 << APERFMPERF_PRESENT);
343 break;
344
345 case 0xb:
346 /* Suppress extended topology stuff */
347 maskebx = 0;
348 break;
349 }
350
351 asm(XEN_EMULATE_PREFIX "cpuid"
352 : "=a" (*ax),
353 "=b" (*bx),
354 "=c" (*cx),
355 "=d" (*dx)
356 : "0" (*ax), "2" (*cx));
357
358 *bx &= maskebx;
359 *cx &= maskecx;
360 *cx |= setecx;
361 *dx &= maskedx;
362}
363STACK_FRAME_NON_STANDARD(xen_cpuid); /* XEN_EMULATE_PREFIX */
364
365static bool __init xen_check_mwait(void)
366{
367#ifdef CONFIG_ACPI
368 struct xen_platform_op op = {
369 .cmd = XENPF_set_processor_pminfo,
370 .u.set_pminfo.id = -1,
371 .u.set_pminfo.type = XEN_PM_PDC,
372 };
373 uint32_t buf[3];
374 unsigned int ax, bx, cx, dx;
375 unsigned int mwait_mask;
376
377 /* We need to determine whether it is OK to expose the MWAIT
378 * capability to the kernel to harvest deeper than C3 states from ACPI
379 * _CST using the processor_harvest_xen.c module. For this to work, we
380 * need to gather the MWAIT_LEAF values (which the cstate.c code
381 * checks against). The hypervisor won't expose the MWAIT flag because
382 * it would break backwards compatibility; so we will find out directly
383 * from the hardware and hypercall.
384 */
385 if (!xen_initial_domain())
386 return false;
387
388 /*
389 * When running under platform earlier than Xen4.2, do not expose
390 * mwait, to avoid the risk of loading native acpi pad driver
391 */
392 if (!xen_running_on_version_or_later(4, 2))
393 return false;
394
395 ax = 1;
396 cx = 0;
397
398 native_cpuid(&ax, &bx, &cx, &dx);
399
400 mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
401 (1 << (X86_FEATURE_MWAIT % 32));
402
403 if ((cx & mwait_mask) != mwait_mask)
404 return false;
405
406 /* We need to emulate the MWAIT_LEAF and for that we need both
407 * ecx and edx. The hypercall provides only partial information.
408 */
409
410 ax = CPUID_MWAIT_LEAF;
411 bx = 0;
412 cx = 0;
413 dx = 0;
414
415 native_cpuid(&ax, &bx, &cx, &dx);
416
417 /* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
418 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
419 */
420 buf[0] = ACPI_PDC_REVISION_ID;
421 buf[1] = 1;
422 buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
423
424 set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
425
426 if ((HYPERVISOR_platform_op(&op) == 0) &&
427 (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
428 cpuid_leaf5_ecx_val = cx;
429 cpuid_leaf5_edx_val = dx;
430 }
431 return true;
432#else
433 return false;
434#endif
435}
436static void __init xen_init_cpuid_mask(void)
437{
438 unsigned int ax, bx, cx, dx;
439 unsigned int xsave_mask;
440
441 cpuid_leaf1_edx_mask =
442 ~((1 << X86_FEATURE_MTRR) | /* disable MTRR */
443 (1 << X86_FEATURE_ACC)); /* thermal monitoring */
444
445 if (!xen_initial_domain())
446 cpuid_leaf1_edx_mask &=
447 ~((1 << X86_FEATURE_ACPI)); /* disable ACPI */
448
449 cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_X2APIC % 32));
450
451 ax = 1;
452 cx = 0;
453 cpuid(1, &ax, &bx, &cx, &dx);
454
455 xsave_mask =
456 (1 << (X86_FEATURE_XSAVE % 32)) |
457 (1 << (X86_FEATURE_OSXSAVE % 32));
458
459 /* Xen will set CR4.OSXSAVE if supported and not disabled by force */
460 if ((cx & xsave_mask) != xsave_mask)
461 cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
462 if (xen_check_mwait())
463 cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
464}
465
466static void xen_set_debugreg(int reg, unsigned long val)
467{
468 HYPERVISOR_set_debugreg(reg, val);
469}
470
471static unsigned long xen_get_debugreg(int reg)
472{
473 return HYPERVISOR_get_debugreg(reg);
474}
475
476static void xen_end_context_switch(struct task_struct *next)
477{
478 xen_mc_flush();
479 paravirt_end_context_switch(next);
480}
481
482static unsigned long xen_store_tr(void)
483{
484 return 0;
485}
486
487/*
488 * Set the page permissions for a particular virtual address. If the
489 * address is a vmalloc mapping (or other non-linear mapping), then
490 * find the linear mapping of the page and also set its protections to
491 * match.
492 */
493static void set_aliased_prot(void *v, pgprot_t prot)
494{
495 int level;
496 pte_t *ptep;
497 pte_t pte;
498 unsigned long pfn;
499 struct page *page;
500 unsigned char dummy;
501
502 ptep = lookup_address((unsigned long)v, &level);
503 BUG_ON(ptep == NULL);
504
505 pfn = pte_pfn(*ptep);
506 page = pfn_to_page(pfn);
507
508 pte = pfn_pte(pfn, prot);
509
510 /*
511 * Careful: update_va_mapping() will fail if the virtual address
512 * we're poking isn't populated in the page tables. We don't
513 * need to worry about the direct map (that's always in the page
514 * tables), but we need to be careful about vmap space. In
515 * particular, the top level page table can lazily propagate
516 * entries between processes, so if we've switched mms since we
517 * vmapped the target in the first place, we might not have the
518 * top-level page table entry populated.
519 *
520 * We disable preemption because we want the same mm active when
521 * we probe the target and when we issue the hypercall. We'll
522 * have the same nominal mm, but if we're a kernel thread, lazy
523 * mm dropping could change our pgd.
524 *
525 * Out of an abundance of caution, this uses __get_user() to fault
526 * in the target address just in case there's some obscure case
527 * in which the target address isn't readable.
528 */
529
530 preempt_disable();
531
532 probe_kernel_read(&dummy, v, 1);
533
534 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
535 BUG();
536
537 if (!PageHighMem(page)) {
538 void *av = __va(PFN_PHYS(pfn));
539
540 if (av != v)
541 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
542 BUG();
543 } else
544 kmap_flush_unused();
545
546 preempt_enable();
547}
548
549static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
550{
551 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
552 int i;
553
554 /*
555 * We need to mark the all aliases of the LDT pages RO. We
556 * don't need to call vm_flush_aliases(), though, since that's
557 * only responsible for flushing aliases out the TLBs, not the
558 * page tables, and Xen will flush the TLB for us if needed.
559 *
560 * To avoid confusing future readers: none of this is necessary
561 * to load the LDT. The hypervisor only checks this when the
562 * LDT is faulted in due to subsequent descriptor access.
563 */
564
565 for(i = 0; i < entries; i += entries_per_page)
566 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
567}
568
569static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
570{
571 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
572 int i;
573
574 for(i = 0; i < entries; i += entries_per_page)
575 set_aliased_prot(ldt + i, PAGE_KERNEL);
576}
577
578static void xen_set_ldt(const void *addr, unsigned entries)
579{
580 struct mmuext_op *op;
581 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
582
583 trace_xen_cpu_set_ldt(addr, entries);
584
585 op = mcs.args;
586 op->cmd = MMUEXT_SET_LDT;
587 op->arg1.linear_addr = (unsigned long)addr;
588 op->arg2.nr_ents = entries;
589
590 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
591
592 xen_mc_issue(PARAVIRT_LAZY_CPU);
593}
594
595static void xen_load_gdt(const struct desc_ptr *dtr)
596{
597 unsigned long va = dtr->address;
598 unsigned int size = dtr->size + 1;
599 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
600 unsigned long frames[pages];
601 int f;
602
603 /*
604 * A GDT can be up to 64k in size, which corresponds to 8192
605 * 8-byte entries, or 16 4k pages..
606 */
607
608 BUG_ON(size > 65536);
609 BUG_ON(va & ~PAGE_MASK);
610
611 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
612 int level;
613 pte_t *ptep;
614 unsigned long pfn, mfn;
615 void *virt;
616
617 /*
618 * The GDT is per-cpu and is in the percpu data area.
619 * That can be virtually mapped, so we need to do a
620 * page-walk to get the underlying MFN for the
621 * hypercall. The page can also be in the kernel's
622 * linear range, so we need to RO that mapping too.
623 */
624 ptep = lookup_address(va, &level);
625 BUG_ON(ptep == NULL);
626
627 pfn = pte_pfn(*ptep);
628 mfn = pfn_to_mfn(pfn);
629 virt = __va(PFN_PHYS(pfn));
630
631 frames[f] = mfn;
632
633 make_lowmem_page_readonly((void *)va);
634 make_lowmem_page_readonly(virt);
635 }
636
637 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
638 BUG();
639}
640
641/*
642 * load_gdt for early boot, when the gdt is only mapped once
643 */
644static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
645{
646 unsigned long va = dtr->address;
647 unsigned int size = dtr->size + 1;
648 unsigned pages = DIV_ROUND_UP(size, PAGE_SIZE);
649 unsigned long frames[pages];
650 int f;
651
652 /*
653 * A GDT can be up to 64k in size, which corresponds to 8192
654 * 8-byte entries, or 16 4k pages..
655 */
656
657 BUG_ON(size > 65536);
658 BUG_ON(va & ~PAGE_MASK);
659
660 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
661 pte_t pte;
662 unsigned long pfn, mfn;
663
664 pfn = virt_to_pfn(va);
665 mfn = pfn_to_mfn(pfn);
666
667 pte = pfn_pte(pfn, PAGE_KERNEL_RO);
668
669 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0))
670 BUG();
671
672 frames[f] = mfn;
673 }
674
675 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct)))
676 BUG();
677}
678
679static inline bool desc_equal(const struct desc_struct *d1,
680 const struct desc_struct *d2)
681{
682 return d1->a == d2->a && d1->b == d2->b;
683}
684
685static void load_TLS_descriptor(struct thread_struct *t,
686 unsigned int cpu, unsigned int i)
687{
688 struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
689 struct desc_struct *gdt;
690 xmaddr_t maddr;
691 struct multicall_space mc;
692
693 if (desc_equal(shadow, &t->tls_array[i]))
694 return;
695
696 *shadow = t->tls_array[i];
697
698 gdt = get_cpu_gdt_table(cpu);
699 maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
700 mc = __xen_mc_entry(0);
701
702 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
703}
704
705static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
706{
707 /*
708 * XXX sleazy hack: If we're being called in a lazy-cpu zone
709 * and lazy gs handling is enabled, it means we're in a
710 * context switch, and %gs has just been saved. This means we
711 * can zero it out to prevent faults on exit from the
712 * hypervisor if the next process has no %gs. Either way, it
713 * has been saved, and the new value will get loaded properly.
714 * This will go away as soon as Xen has been modified to not
715 * save/restore %gs for normal hypercalls.
716 *
717 * On x86_64, this hack is not used for %gs, because gs points
718 * to KERNEL_GS_BASE (and uses it for PDA references), so we
719 * must not zero %gs on x86_64
720 *
721 * For x86_64, we need to zero %fs, otherwise we may get an
722 * exception between the new %fs descriptor being loaded and
723 * %fs being effectively cleared at __switch_to().
724 */
725 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
726#ifdef CONFIG_X86_32
727 lazy_load_gs(0);
728#else
729 loadsegment(fs, 0);
730#endif
731 }
732
733 xen_mc_batch();
734
735 load_TLS_descriptor(t, cpu, 0);
736 load_TLS_descriptor(t, cpu, 1);
737 load_TLS_descriptor(t, cpu, 2);
738
739 xen_mc_issue(PARAVIRT_LAZY_CPU);
740}
741
742#ifdef CONFIG_X86_64
743static void xen_load_gs_index(unsigned int idx)
744{
745 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
746 BUG();
747}
748#endif
749
750static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
751 const void *ptr)
752{
753 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
754 u64 entry = *(u64 *)ptr;
755
756 trace_xen_cpu_write_ldt_entry(dt, entrynum, entry);
757
758 preempt_disable();
759
760 xen_mc_flush();
761 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
762 BUG();
763
764 preempt_enable();
765}
766
767static int cvt_gate_to_trap(int vector, const gate_desc *val,
768 struct trap_info *info)
769{
770 unsigned long addr;
771
772 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT)
773 return 0;
774
775 info->vector = vector;
776
777 addr = gate_offset(*val);
778#ifdef CONFIG_X86_64
779 /*
780 * Look for known traps using IST, and substitute them
781 * appropriately. The debugger ones are the only ones we care
782 * about. Xen will handle faults like double_fault,
783 * so we should never see them. Warn if
784 * there's an unexpected IST-using fault handler.
785 */
786 if (addr == (unsigned long)debug)
787 addr = (unsigned long)xen_debug;
788 else if (addr == (unsigned long)int3)
789 addr = (unsigned long)xen_int3;
790 else if (addr == (unsigned long)stack_segment)
791 addr = (unsigned long)xen_stack_segment;
792 else if (addr == (unsigned long)double_fault) {
793 /* Don't need to handle these */
794 return 0;
795#ifdef CONFIG_X86_MCE
796 } else if (addr == (unsigned long)machine_check) {
797 /*
798 * when xen hypervisor inject vMCE to guest,
799 * use native mce handler to handle it
800 */
801 ;
802#endif
803 } else if (addr == (unsigned long)nmi)
804 /*
805 * Use the native version as well.
806 */
807 ;
808 else {
809 /* Some other trap using IST? */
810 if (WARN_ON(val->ist != 0))
811 return 0;
812 }
813#endif /* CONFIG_X86_64 */
814 info->address = addr;
815
816 info->cs = gate_segment(*val);
817 info->flags = val->dpl;
818 /* interrupt gates clear IF */
819 if (val->type == GATE_INTERRUPT)
820 info->flags |= 1 << 2;
821
822 return 1;
823}
824
825/* Locations of each CPU's IDT */
826static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
827
828/* Set an IDT entry. If the entry is part of the current IDT, then
829 also update Xen. */
830static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
831{
832 unsigned long p = (unsigned long)&dt[entrynum];
833 unsigned long start, end;
834
835 trace_xen_cpu_write_idt_entry(dt, entrynum, g);
836
837 preempt_disable();
838
839 start = __this_cpu_read(idt_desc.address);
840 end = start + __this_cpu_read(idt_desc.size) + 1;
841
842 xen_mc_flush();
843
844 native_write_idt_entry(dt, entrynum, g);
845
846 if (p >= start && (p + 8) <= end) {
847 struct trap_info info[2];
848
849 info[1].address = 0;
850
851 if (cvt_gate_to_trap(entrynum, g, &info[0]))
852 if (HYPERVISOR_set_trap_table(info))
853 BUG();
854 }
855
856 preempt_enable();
857}
858
859static void xen_convert_trap_info(const struct desc_ptr *desc,
860 struct trap_info *traps)
861{
862 unsigned in, out, count;
863
864 count = (desc->size+1) / sizeof(gate_desc);
865 BUG_ON(count > 256);
866
867 for (in = out = 0; in < count; in++) {
868 gate_desc *entry = (gate_desc*)(desc->address) + in;
869
870 if (cvt_gate_to_trap(in, entry, &traps[out]))
871 out++;
872 }
873 traps[out].address = 0;
874}
875
876void xen_copy_trap_info(struct trap_info *traps)
877{
878 const struct desc_ptr *desc = this_cpu_ptr(&idt_desc);
879
880 xen_convert_trap_info(desc, traps);
881}
882
883/* Load a new IDT into Xen. In principle this can be per-CPU, so we
884 hold a spinlock to protect the static traps[] array (static because
885 it avoids allocation, and saves stack space). */
886static void xen_load_idt(const struct desc_ptr *desc)
887{
888 static DEFINE_SPINLOCK(lock);
889 static struct trap_info traps[257];
890
891 trace_xen_cpu_load_idt(desc);
892
893 spin_lock(&lock);
894
895 memcpy(this_cpu_ptr(&idt_desc), desc, sizeof(idt_desc));
896
897 xen_convert_trap_info(desc, traps);
898
899 xen_mc_flush();
900 if (HYPERVISOR_set_trap_table(traps))
901 BUG();
902
903 spin_unlock(&lock);
904}
905
906/* Write a GDT descriptor entry. Ignore LDT descriptors, since
907 they're handled differently. */
908static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
909 const void *desc, int type)
910{
911 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
912
913 preempt_disable();
914
915 switch (type) {
916 case DESC_LDT:
917 case DESC_TSS:
918 /* ignore */
919 break;
920
921 default: {
922 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
923
924 xen_mc_flush();
925 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
926 BUG();
927 }
928
929 }
930
931 preempt_enable();
932}
933
934/*
935 * Version of write_gdt_entry for use at early boot-time needed to
936 * update an entry as simply as possible.
937 */
938static void __init xen_write_gdt_entry_boot(struct desc_struct *dt, int entry,
939 const void *desc, int type)
940{
941 trace_xen_cpu_write_gdt_entry(dt, entry, desc, type);
942
943 switch (type) {
944 case DESC_LDT:
945 case DESC_TSS:
946 /* ignore */
947 break;
948
949 default: {
950 xmaddr_t maddr = virt_to_machine(&dt[entry]);
951
952 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
953 dt[entry] = *(struct desc_struct *)desc;
954 }
955
956 }
957}
958
959static void xen_load_sp0(struct tss_struct *tss,
960 struct thread_struct *thread)
961{
962 struct multicall_space mcs;
963
964 mcs = xen_mc_entry(0);
965 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
966 xen_mc_issue(PARAVIRT_LAZY_CPU);
967 tss->x86_tss.sp0 = thread->sp0;
968}
969
970void xen_set_iopl_mask(unsigned mask)
971{
972 struct physdev_set_iopl set_iopl;
973
974 /* Force the change at ring 0. */
975 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
976 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
977}
978
979static void xen_io_delay(void)
980{
981}
982
983static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
984
985static 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
997static 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
1012static 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
1019static inline unsigned long xen_read_cr8(void)
1020{
1021 return 0;
1022}
1023static inline void xen_write_cr8(unsigned long val)
1024{
1025 BUG_ON(val);
1026}
1027#endif
1028
1029static 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
1048static 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 break;
1081
1082 default:
1083 if (!pmu_msr_write(msr, low, high, &ret))
1084 ret = native_write_msr_safe(msr, low, high);
1085 }
1086
1087 return ret;
1088}
1089
1090static u64 xen_read_msr(unsigned int msr)
1091{
1092 /*
1093 * This will silently swallow a #GP from RDMSR. It may be worth
1094 * changing that.
1095 */
1096 int err;
1097
1098 return xen_read_msr_safe(msr, &err);
1099}
1100
1101static void xen_write_msr(unsigned int msr, unsigned low, unsigned high)
1102{
1103 /*
1104 * This will silently swallow a #GP from WRMSR. It may be worth
1105 * changing that.
1106 */
1107 xen_write_msr_safe(msr, low, high);
1108}
1109
1110void xen_setup_shared_info(void)
1111{
1112 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1113 set_fixmap(FIX_PARAVIRT_BOOTMAP,
1114 xen_start_info->shared_info);
1115
1116 HYPERVISOR_shared_info =
1117 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
1118 } else
1119 HYPERVISOR_shared_info =
1120 (struct shared_info *)__va(xen_start_info->shared_info);
1121
1122#ifndef CONFIG_SMP
1123 /* In UP this is as good a place as any to set up shared info */
1124 xen_setup_vcpu_info_placement();
1125#endif
1126
1127 xen_setup_mfn_list_list();
1128}
1129
1130/* This is called once we have the cpu_possible_mask */
1131void xen_setup_vcpu_info_placement(void)
1132{
1133 int cpu;
1134
1135 for_each_possible_cpu(cpu) {
1136 /* Set up direct vCPU id mapping for PV guests. */
1137 per_cpu(xen_vcpu_id, cpu) = cpu;
1138 xen_vcpu_setup(cpu);
1139 }
1140
1141 /* xen_vcpu_setup managed to place the vcpu_info within the
1142 * percpu area for all cpus, so make use of it. Note that for
1143 * PVH we want to use native IRQ mechanism. */
1144 if (have_vcpu_info_placement && !xen_pvh_domain()) {
1145 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
1146 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
1147 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
1148 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
1149 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
1150 }
1151}
1152
1153static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
1154 unsigned long addr, unsigned len)
1155{
1156 char *start, *end, *reloc;
1157 unsigned ret;
1158
1159 start = end = reloc = NULL;
1160
1161#define SITE(op, x) \
1162 case PARAVIRT_PATCH(op.x): \
1163 if (have_vcpu_info_placement) { \
1164 start = (char *)xen_##x##_direct; \
1165 end = xen_##x##_direct_end; \
1166 reloc = xen_##x##_direct_reloc; \
1167 } \
1168 goto patch_site
1169
1170 switch (type) {
1171 SITE(pv_irq_ops, irq_enable);
1172 SITE(pv_irq_ops, irq_disable);
1173 SITE(pv_irq_ops, save_fl);
1174 SITE(pv_irq_ops, restore_fl);
1175#undef SITE
1176
1177 patch_site:
1178 if (start == NULL || (end-start) > len)
1179 goto default_patch;
1180
1181 ret = paravirt_patch_insns(insnbuf, len, start, end);
1182
1183 /* Note: because reloc is assigned from something that
1184 appears to be an array, gcc assumes it's non-null,
1185 but doesn't know its relationship with start and
1186 end. */
1187 if (reloc > start && reloc < end) {
1188 int reloc_off = reloc - start;
1189 long *relocp = (long *)(insnbuf + reloc_off);
1190 long delta = start - (char *)addr;
1191
1192 *relocp += delta;
1193 }
1194 break;
1195
1196 default_patch:
1197 default:
1198 ret = paravirt_patch_default(type, clobbers, insnbuf,
1199 addr, len);
1200 break;
1201 }
1202
1203 return ret;
1204}
1205
1206static const struct pv_info xen_info __initconst = {
1207 .shared_kernel_pmd = 0,
1208
1209#ifdef CONFIG_X86_64
1210 .extra_user_64bit_cs = FLAT_USER_CS64,
1211#endif
1212 .name = "Xen",
1213};
1214
1215static const struct pv_init_ops xen_init_ops __initconst = {
1216 .patch = xen_patch,
1217};
1218
1219static const struct pv_cpu_ops xen_cpu_ops __initconst = {
1220 .cpuid = xen_cpuid,
1221
1222 .set_debugreg = xen_set_debugreg,
1223 .get_debugreg = xen_get_debugreg,
1224
1225 .read_cr0 = xen_read_cr0,
1226 .write_cr0 = xen_write_cr0,
1227
1228 .read_cr4 = native_read_cr4,
1229 .write_cr4 = xen_write_cr4,
1230
1231#ifdef CONFIG_X86_64
1232 .read_cr8 = xen_read_cr8,
1233 .write_cr8 = xen_write_cr8,
1234#endif
1235
1236 .wbinvd = native_wbinvd,
1237
1238 .read_msr = xen_read_msr,
1239 .write_msr = xen_write_msr,
1240
1241 .read_msr_safe = xen_read_msr_safe,
1242 .write_msr_safe = xen_write_msr_safe,
1243
1244 .read_pmc = xen_read_pmc,
1245
1246 .iret = xen_iret,
1247#ifdef CONFIG_X86_64
1248 .usergs_sysret64 = xen_sysret64,
1249#endif
1250
1251 .load_tr_desc = paravirt_nop,
1252 .set_ldt = xen_set_ldt,
1253 .load_gdt = xen_load_gdt,
1254 .load_idt = xen_load_idt,
1255 .load_tls = xen_load_tls,
1256#ifdef CONFIG_X86_64
1257 .load_gs_index = xen_load_gs_index,
1258#endif
1259
1260 .alloc_ldt = xen_alloc_ldt,
1261 .free_ldt = xen_free_ldt,
1262
1263 .store_idt = native_store_idt,
1264 .store_tr = xen_store_tr,
1265
1266 .write_ldt_entry = xen_write_ldt_entry,
1267 .write_gdt_entry = xen_write_gdt_entry,
1268 .write_idt_entry = xen_write_idt_entry,
1269 .load_sp0 = xen_load_sp0,
1270
1271 .set_iopl_mask = xen_set_iopl_mask,
1272 .io_delay = xen_io_delay,
1273
1274 /* Xen takes care of %gs when switching to usermode for us */
1275 .swapgs = paravirt_nop,
1276
1277 .start_context_switch = paravirt_start_context_switch,
1278 .end_context_switch = xen_end_context_switch,
1279};
1280
1281static void xen_reboot(int reason)
1282{
1283 struct sched_shutdown r = { .reason = reason };
1284 int cpu;
1285
1286 for_each_online_cpu(cpu)
1287 xen_pmu_finish(cpu);
1288
1289 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1290 BUG();
1291}
1292
1293static void xen_restart(char *msg)
1294{
1295 xen_reboot(SHUTDOWN_reboot);
1296}
1297
1298static void xen_emergency_restart(void)
1299{
1300 xen_reboot(SHUTDOWN_reboot);
1301}
1302
1303static void xen_machine_halt(void)
1304{
1305 xen_reboot(SHUTDOWN_poweroff);
1306}
1307
1308static void xen_machine_power_off(void)
1309{
1310 if (pm_power_off)
1311 pm_power_off();
1312 xen_reboot(SHUTDOWN_poweroff);
1313}
1314
1315static void xen_crash_shutdown(struct pt_regs *regs)
1316{
1317 xen_reboot(SHUTDOWN_crash);
1318}
1319
1320static int
1321xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
1322{
1323 if (!kexec_crash_loaded())
1324 xen_reboot(SHUTDOWN_crash);
1325 return NOTIFY_DONE;
1326}
1327
1328static struct notifier_block xen_panic_block = {
1329 .notifier_call= xen_panic_event,
1330 .priority = INT_MIN
1331};
1332
1333int xen_panic_handler_init(void)
1334{
1335 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
1336 return 0;
1337}
1338
1339static const struct machine_ops xen_machine_ops __initconst = {
1340 .restart = xen_restart,
1341 .halt = xen_machine_halt,
1342 .power_off = xen_machine_power_off,
1343 .shutdown = xen_machine_halt,
1344 .crash_shutdown = xen_crash_shutdown,
1345 .emergency_restart = xen_emergency_restart,
1346};
1347
1348static unsigned char xen_get_nmi_reason(void)
1349{
1350 unsigned char reason = 0;
1351
1352 /* Construct a value which looks like it came from port 0x61. */
1353 if (test_bit(_XEN_NMIREASON_io_error,
1354 &HYPERVISOR_shared_info->arch.nmi_reason))
1355 reason |= NMI_REASON_IOCHK;
1356 if (test_bit(_XEN_NMIREASON_pci_serr,
1357 &HYPERVISOR_shared_info->arch.nmi_reason))
1358 reason |= NMI_REASON_SERR;
1359
1360 return reason;
1361}
1362
1363static void __init xen_boot_params_init_edd(void)
1364{
1365#if IS_ENABLED(CONFIG_EDD)
1366 struct xen_platform_op op;
1367 struct edd_info *edd_info;
1368 u32 *mbr_signature;
1369 unsigned nr;
1370 int ret;
1371
1372 edd_info = boot_params.eddbuf;
1373 mbr_signature = boot_params.edd_mbr_sig_buffer;
1374
1375 op.cmd = XENPF_firmware_info;
1376
1377 op.u.firmware_info.type = XEN_FW_DISK_INFO;
1378 for (nr = 0; nr < EDDMAXNR; nr++) {
1379 struct edd_info *info = edd_info + nr;
1380
1381 op.u.firmware_info.index = nr;
1382 info->params.length = sizeof(info->params);
1383 set_xen_guest_handle(op.u.firmware_info.u.disk_info.edd_params,
1384 &info->params);
1385 ret = HYPERVISOR_platform_op(&op);
1386 if (ret)
1387 break;
1388
1389#define C(x) info->x = op.u.firmware_info.u.disk_info.x
1390 C(device);
1391 C(version);
1392 C(interface_support);
1393 C(legacy_max_cylinder);
1394 C(legacy_max_head);
1395 C(legacy_sectors_per_track);
1396#undef C
1397 }
1398 boot_params.eddbuf_entries = nr;
1399
1400 op.u.firmware_info.type = XEN_FW_DISK_MBR_SIGNATURE;
1401 for (nr = 0; nr < EDD_MBR_SIG_MAX; nr++) {
1402 op.u.firmware_info.index = nr;
1403 ret = HYPERVISOR_platform_op(&op);
1404 if (ret)
1405 break;
1406 mbr_signature[nr] = op.u.firmware_info.u.disk_mbr_signature.mbr_signature;
1407 }
1408 boot_params.edd_mbr_sig_buf_entries = nr;
1409#endif
1410}
1411
1412/*
1413 * Set up the GDT and segment registers for -fstack-protector. Until
1414 * we do this, we have to be careful not to call any stack-protected
1415 * function, which is most of the kernel.
1416 *
1417 * Note, that it is __ref because the only caller of this after init
1418 * is PVH which is not going to use xen_load_gdt_boot or other
1419 * __init functions.
1420 */
1421static void __ref xen_setup_gdt(int cpu)
1422{
1423 if (xen_feature(XENFEAT_auto_translated_physmap)) {
1424#ifdef CONFIG_X86_64
1425 unsigned long dummy;
1426
1427 load_percpu_segment(cpu); /* We need to access per-cpu area */
1428 switch_to_new_gdt(cpu); /* GDT and GS set */
1429
1430 /* We are switching of the Xen provided GDT to our HVM mode
1431 * GDT. The new GDT has __KERNEL_CS with CS.L = 1
1432 * and we are jumping to reload it.
1433 */
1434 asm volatile ("pushq %0\n"
1435 "leaq 1f(%%rip),%0\n"
1436 "pushq %0\n"
1437 "lretq\n"
1438 "1:\n"
1439 : "=&r" (dummy) : "0" (__KERNEL_CS));
1440
1441 /*
1442 * While not needed, we also set the %es, %ds, and %fs
1443 * to zero. We don't care about %ss as it is NULL.
1444 * Strictly speaking this is not needed as Xen zeros those
1445 * out (and also MSR_FS_BASE, MSR_GS_BASE, MSR_KERNEL_GS_BASE)
1446 *
1447 * Linux zeros them in cpu_init() and in secondary_startup_64
1448 * (for BSP).
1449 */
1450 loadsegment(es, 0);
1451 loadsegment(ds, 0);
1452 loadsegment(fs, 0);
1453#else
1454 /* PVH: TODO Implement. */
1455 BUG();
1456#endif
1457 return; /* PVH does not need any PV GDT ops. */
1458 }
1459 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot;
1460 pv_cpu_ops.load_gdt = xen_load_gdt_boot;
1461
1462 setup_stack_canary_segment(0);
1463 switch_to_new_gdt(0);
1464
1465 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry;
1466 pv_cpu_ops.load_gdt = xen_load_gdt;
1467}
1468
1469#ifdef CONFIG_XEN_PVH
1470/*
1471 * A PV guest starts with default flags that are not set for PVH, set them
1472 * here asap.
1473 */
1474static void xen_pvh_set_cr_flags(int cpu)
1475{
1476
1477 /* Some of these are setup in 'secondary_startup_64'. The others:
1478 * X86_CR0_TS, X86_CR0_PE, X86_CR0_ET are set by Xen for HVM guests
1479 * (which PVH shared codepaths), while X86_CR0_PG is for PVH. */
1480 write_cr0(read_cr0() | X86_CR0_MP | X86_CR0_NE | X86_CR0_WP | X86_CR0_AM);
1481
1482 if (!cpu)
1483 return;
1484 /*
1485 * For BSP, PSE PGE are set in probe_page_size_mask(), for APs
1486 * set them here. For all, OSFXSR OSXMMEXCPT are set in fpu__init_cpu().
1487 */
1488 if (boot_cpu_has(X86_FEATURE_PSE))
1489 cr4_set_bits_and_update_boot(X86_CR4_PSE);
1490
1491 if (boot_cpu_has(X86_FEATURE_PGE))
1492 cr4_set_bits_and_update_boot(X86_CR4_PGE);
1493}
1494
1495/*
1496 * Note, that it is ref - because the only caller of this after init
1497 * is PVH which is not going to use xen_load_gdt_boot or other
1498 * __init functions.
1499 */
1500void __ref xen_pvh_secondary_vcpu_init(int cpu)
1501{
1502 xen_setup_gdt(cpu);
1503 xen_pvh_set_cr_flags(cpu);
1504}
1505
1506static void __init xen_pvh_early_guest_init(void)
1507{
1508 if (!xen_feature(XENFEAT_auto_translated_physmap))
1509 return;
1510
1511 BUG_ON(!xen_feature(XENFEAT_hvm_callback_vector));
1512
1513 xen_pvh_early_cpu_init(0, false);
1514 xen_pvh_set_cr_flags(0);
1515
1516#ifdef CONFIG_X86_32
1517 BUG(); /* PVH: Implement proper support. */
1518#endif
1519}
1520#endif /* CONFIG_XEN_PVH */
1521
1522static void __init xen_dom0_set_legacy_features(void)
1523{
1524 x86_platform.legacy.rtc = 1;
1525}
1526
1527static int xen_cpuhp_setup(void)
1528{
1529 int rc;
1530
1531 rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
1532 "x86/xen/hvm_guest:prepare",
1533 xen_cpu_up_prepare, xen_cpu_dead);
1534 if (rc >= 0) {
1535 rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
1536 "x86/xen/hvm_guest:online",
1537 xen_cpu_up_online, NULL);
1538 if (rc < 0)
1539 cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
1540 }
1541
1542 return rc >= 0 ? 0 : rc;
1543}
1544
1545/* First C function to be called on Xen boot */
1546asmlinkage __visible void __init xen_start_kernel(void)
1547{
1548 struct physdev_set_iopl set_iopl;
1549 unsigned long initrd_start = 0;
1550 int rc;
1551
1552 if (!xen_start_info)
1553 return;
1554
1555 xen_domain_type = XEN_PV_DOMAIN;
1556
1557 xen_setup_features();
1558#ifdef CONFIG_XEN_PVH
1559 xen_pvh_early_guest_init();
1560#endif
1561 xen_setup_machphys_mapping();
1562
1563 /* Install Xen paravirt ops */
1564 pv_info = xen_info;
1565 pv_init_ops = xen_init_ops;
1566 if (!xen_pvh_domain()) {
1567 pv_cpu_ops = xen_cpu_ops;
1568
1569 x86_platform.get_nmi_reason = xen_get_nmi_reason;
1570 }
1571
1572 if (xen_feature(XENFEAT_auto_translated_physmap))
1573 x86_init.resources.memory_setup = xen_auto_xlated_memory_setup;
1574 else
1575 x86_init.resources.memory_setup = xen_memory_setup;
1576 x86_init.oem.arch_setup = xen_arch_setup;
1577 x86_init.oem.banner = xen_banner;
1578
1579 xen_init_time_ops();
1580
1581 /*
1582 * Set up some pagetable state before starting to set any ptes.
1583 */
1584
1585 xen_init_mmu_ops();
1586
1587 /* Prevent unwanted bits from being set in PTEs. */
1588 __supported_pte_mask &= ~_PAGE_GLOBAL;
1589
1590 /*
1591 * Prevent page tables from being allocated in highmem, even
1592 * if CONFIG_HIGHPTE is enabled.
1593 */
1594 __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
1595
1596 /* Work out if we support NX */
1597 x86_configure_nx();
1598
1599 /* Get mfn list */
1600 xen_build_dynamic_phys_to_machine();
1601
1602 /*
1603 * Set up kernel GDT and segment registers, mainly so that
1604 * -fstack-protector code can be executed.
1605 */
1606 xen_setup_gdt(0);
1607
1608 xen_init_irq_ops();
1609 xen_init_cpuid_mask();
1610
1611#ifdef CONFIG_X86_LOCAL_APIC
1612 /*
1613 * set up the basic apic ops.
1614 */
1615 xen_init_apic();
1616#endif
1617
1618 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1619 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1620 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1621 }
1622
1623 machine_ops = xen_machine_ops;
1624
1625 /*
1626 * The only reliable way to retain the initial address of the
1627 * percpu gdt_page is to remember it here, so we can go and
1628 * mark it RW later, when the initial percpu area is freed.
1629 */
1630 xen_initial_gdt = &per_cpu(gdt_page, 0);
1631
1632 xen_smp_init();
1633
1634#ifdef CONFIG_ACPI_NUMA
1635 /*
1636 * The pages we from Xen are not related to machine pages, so
1637 * any NUMA information the kernel tries to get from ACPI will
1638 * be meaningless. Prevent it from trying.
1639 */
1640 acpi_numa = -1;
1641#endif
1642 /* Don't do the full vcpu_info placement stuff until we have a
1643 possible map and a non-dummy shared_info. */
1644 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1645
1646 WARN_ON(xen_cpuhp_setup());
1647
1648 local_irq_disable();
1649 early_boot_irqs_disabled = true;
1650
1651 xen_raw_console_write("mapping kernel into physical memory\n");
1652 xen_setup_kernel_pagetable((pgd_t *)xen_start_info->pt_base,
1653 xen_start_info->nr_pages);
1654 xen_reserve_special_pages();
1655
1656 /* keep using Xen gdt for now; no urgent need to change it */
1657
1658#ifdef CONFIG_X86_32
1659 pv_info.kernel_rpl = 1;
1660 if (xen_feature(XENFEAT_supervisor_mode_kernel))
1661 pv_info.kernel_rpl = 0;
1662#else
1663 pv_info.kernel_rpl = 0;
1664#endif
1665 /* set the limit of our address space */
1666 xen_reserve_top();
1667
1668 /* PVH: runs at default kernel iopl of 0 */
1669 if (!xen_pvh_domain()) {
1670 /*
1671 * We used to do this in xen_arch_setup, but that is too late
1672 * on AMD were early_cpu_init (run before ->arch_setup()) calls
1673 * early_amd_init which pokes 0xcf8 port.
1674 */
1675 set_iopl.iopl = 1;
1676 rc = HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
1677 if (rc != 0)
1678 xen_raw_printk("physdev_op failed %d\n", rc);
1679 }
1680
1681#ifdef CONFIG_X86_32
1682 /* set up basic CPUID stuff */
1683 cpu_detect(&new_cpu_data);
1684 set_cpu_cap(&new_cpu_data, X86_FEATURE_FPU);
1685 new_cpu_data.wp_works_ok = 1;
1686 new_cpu_data.x86_capability[CPUID_1_EDX] = cpuid_edx(1);
1687#endif
1688
1689 if (xen_start_info->mod_start) {
1690 if (xen_start_info->flags & SIF_MOD_START_PFN)
1691 initrd_start = PFN_PHYS(xen_start_info->mod_start);
1692 else
1693 initrd_start = __pa(xen_start_info->mod_start);
1694 }
1695
1696 /* Poke various useful things into boot_params */
1697 boot_params.hdr.type_of_loader = (9 << 4) | 0;
1698 boot_params.hdr.ramdisk_image = initrd_start;
1699 boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1700 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1701 boot_params.hdr.hardware_subarch = X86_SUBARCH_XEN;
1702
1703 if (!xen_initial_domain()) {
1704 add_preferred_console("xenboot", 0, NULL);
1705 add_preferred_console("tty", 0, NULL);
1706 add_preferred_console("hvc", 0, NULL);
1707 if (pci_xen)
1708 x86_init.pci.arch_init = pci_xen_init;
1709 } else {
1710 const struct dom0_vga_console_info *info =
1711 (void *)((char *)xen_start_info +
1712 xen_start_info->console.dom0.info_off);
1713 struct xen_platform_op op = {
1714 .cmd = XENPF_firmware_info,
1715 .interface_version = XENPF_INTERFACE_VERSION,
1716 .u.firmware_info.type = XEN_FW_KBD_SHIFT_FLAGS,
1717 };
1718
1719 x86_platform.set_legacy_features =
1720 xen_dom0_set_legacy_features;
1721 xen_init_vga(info, xen_start_info->console.dom0.info_size);
1722 xen_start_info->console.domU.mfn = 0;
1723 xen_start_info->console.domU.evtchn = 0;
1724
1725 if (HYPERVISOR_platform_op(&op) == 0)
1726 boot_params.kbd_status = op.u.firmware_info.u.kbd_shift_flags;
1727
1728 /* Make sure ACS will be enabled */
1729 pci_request_acs();
1730
1731 xen_acpi_sleep_register();
1732
1733 /* Avoid searching for BIOS MP tables */
1734 x86_init.mpparse.find_smp_config = x86_init_noop;
1735 x86_init.mpparse.get_smp_config = x86_init_uint_noop;
1736
1737 xen_boot_params_init_edd();
1738 }
1739#ifdef CONFIG_PCI
1740 /* PCI BIOS service won't work from a PV guest. */
1741 pci_probe &= ~PCI_PROBE_BIOS;
1742#endif
1743 xen_raw_console_write("about to get started...\n");
1744
1745 /* Let's presume PV guests always boot on vCPU with id 0. */
1746 per_cpu(xen_vcpu_id, 0) = 0;
1747
1748 xen_setup_runstate_info(0);
1749
1750 xen_efi_init();
1751
1752 /* Start the world */
1753#ifdef CONFIG_X86_32
1754 i386_start_kernel();
1755#else
1756 cr4_init_shadow(); /* 32b kernel does this in i386_start_kernel() */
1757 x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1758#endif
1759}
1760
1761void __ref xen_hvm_init_shared_info(void)
1762{
1763 int cpu;
1764 struct xen_add_to_physmap xatp;
1765 static struct shared_info *shared_info_page = 0;
1766
1767 if (!shared_info_page)
1768 shared_info_page = (struct shared_info *)
1769 extend_brk(PAGE_SIZE, PAGE_SIZE);
1770 xatp.domid = DOMID_SELF;
1771 xatp.idx = 0;
1772 xatp.space = XENMAPSPACE_shared_info;
1773 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT;
1774 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
1775 BUG();
1776
1777 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page;
1778
1779 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info
1780 * page, we use it in the event channel upcall and in some pvclock
1781 * related functions. We don't need the vcpu_info placement
1782 * optimizations because we don't use any pv_mmu or pv_irq op on
1783 * HVM.
1784 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is
1785 * online but xen_hvm_init_shared_info is run at resume time too and
1786 * in that case multiple vcpus might be online. */
1787 for_each_online_cpu(cpu) {
1788 /* Leave it to be NULL. */
1789 if (xen_vcpu_nr(cpu) >= MAX_VIRT_CPUS)
1790 continue;
1791 per_cpu(xen_vcpu, cpu) =
1792 &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
1793 }
1794}
1795
1796#ifdef CONFIG_XEN_PVHVM
1797static void __init init_hvm_pv_info(void)
1798{
1799 int major, minor;
1800 uint32_t eax, ebx, ecx, edx, pages, msr, base;
1801 u64 pfn;
1802
1803 base = xen_cpuid_base();
1804 cpuid(base + 1, &eax, &ebx, &ecx, &edx);
1805
1806 major = eax >> 16;
1807 minor = eax & 0xffff;
1808 printk(KERN_INFO "Xen version %d.%d.\n", major, minor);
1809
1810 cpuid(base + 2, &pages, &msr, &ecx, &edx);
1811
1812 pfn = __pa(hypercall_page);
1813 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
1814
1815 xen_setup_features();
1816
1817 cpuid(base + 4, &eax, &ebx, &ecx, &edx);
1818 if (eax & XEN_HVM_CPUID_VCPU_ID_PRESENT)
1819 this_cpu_write(xen_vcpu_id, ebx);
1820 else
1821 this_cpu_write(xen_vcpu_id, smp_processor_id());
1822
1823 pv_info.name = "Xen HVM";
1824
1825 xen_domain_type = XEN_HVM_DOMAIN;
1826}
1827#endif
1828
1829static int xen_cpu_up_prepare(unsigned int cpu)
1830{
1831 int rc;
1832
1833 if (xen_hvm_domain()) {
1834 /*
1835 * This can happen if CPU was offlined earlier and
1836 * offlining timed out in common_cpu_die().
1837 */
1838 if (cpu_report_state(cpu) == CPU_DEAD_FROZEN) {
1839 xen_smp_intr_free(cpu);
1840 xen_uninit_lock_cpu(cpu);
1841 }
1842
1843 if (cpu_acpi_id(cpu) != U32_MAX)
1844 per_cpu(xen_vcpu_id, cpu) = cpu_acpi_id(cpu);
1845 else
1846 per_cpu(xen_vcpu_id, cpu) = cpu;
1847 xen_vcpu_setup(cpu);
1848 }
1849
1850 if (xen_pv_domain() || xen_feature(XENFEAT_hvm_safe_pvclock))
1851 xen_setup_timer(cpu);
1852
1853 rc = xen_smp_intr_init(cpu);
1854 if (rc) {
1855 WARN(1, "xen_smp_intr_init() for CPU %d failed: %d\n",
1856 cpu, rc);
1857 return rc;
1858 }
1859 return 0;
1860}
1861
1862static int xen_cpu_dead(unsigned int cpu)
1863{
1864 xen_smp_intr_free(cpu);
1865
1866 if (xen_pv_domain() || xen_feature(XENFEAT_hvm_safe_pvclock))
1867 xen_teardown_timer(cpu);
1868
1869 return 0;
1870}
1871
1872static int xen_cpu_up_online(unsigned int cpu)
1873{
1874 xen_init_lock_cpu(cpu);
1875 return 0;
1876}
1877
1878#ifdef CONFIG_XEN_PVHVM
1879#ifdef CONFIG_KEXEC_CORE
1880static void xen_hvm_shutdown(void)
1881{
1882 native_machine_shutdown();
1883 if (kexec_in_progress)
1884 xen_reboot(SHUTDOWN_soft_reset);
1885}
1886
1887static void xen_hvm_crash_shutdown(struct pt_regs *regs)
1888{
1889 native_machine_crash_shutdown(regs);
1890 xen_reboot(SHUTDOWN_soft_reset);
1891}
1892#endif
1893
1894static void __init xen_hvm_guest_init(void)
1895{
1896 if (xen_pv_domain())
1897 return;
1898
1899 init_hvm_pv_info();
1900
1901 xen_hvm_init_shared_info();
1902
1903 xen_panic_handler_init();
1904
1905 BUG_ON(!xen_feature(XENFEAT_hvm_callback_vector));
1906
1907 xen_hvm_smp_init();
1908 WARN_ON(xen_cpuhp_setup());
1909 xen_unplug_emulated_devices();
1910 x86_init.irqs.intr_init = xen_init_IRQ;
1911 xen_hvm_init_time_ops();
1912 xen_hvm_init_mmu_ops();
1913#ifdef CONFIG_KEXEC_CORE
1914 machine_ops.shutdown = xen_hvm_shutdown;
1915 machine_ops.crash_shutdown = xen_hvm_crash_shutdown;
1916#endif
1917}
1918#endif
1919
1920static bool xen_nopv = false;
1921static __init int xen_parse_nopv(char *arg)
1922{
1923 xen_nopv = true;
1924 return 0;
1925}
1926early_param("xen_nopv", xen_parse_nopv);
1927
1928static uint32_t __init xen_platform(void)
1929{
1930 if (xen_nopv)
1931 return 0;
1932
1933 return xen_cpuid_base();
1934}
1935
1936bool xen_hvm_need_lapic(void)
1937{
1938 if (xen_nopv)
1939 return false;
1940 if (xen_pv_domain())
1941 return false;
1942 if (!xen_hvm_domain())
1943 return false;
1944 if (xen_feature(XENFEAT_hvm_pirqs))
1945 return false;
1946 return true;
1947}
1948EXPORT_SYMBOL_GPL(xen_hvm_need_lapic);
1949
1950static void xen_set_cpu_features(struct cpuinfo_x86 *c)
1951{
1952 if (xen_pv_domain()) {
1953 clear_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1954 set_cpu_cap(c, X86_FEATURE_XENPV);
1955 }
1956}
1957
1958static void xen_pin_vcpu(int cpu)
1959{
1960 static bool disable_pinning;
1961 struct sched_pin_override pin_override;
1962 int ret;
1963
1964 if (disable_pinning)
1965 return;
1966
1967 pin_override.pcpu = cpu;
1968 ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
1969
1970 /* Ignore errors when removing override. */
1971 if (cpu < 0)
1972 return;
1973
1974 switch (ret) {
1975 case -ENOSYS:
1976 pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
1977 cpu);
1978 disable_pinning = true;
1979 break;
1980 case -EPERM:
1981 WARN(1, "Trying to pin vcpu without having privilege to do so\n");
1982 disable_pinning = true;
1983 break;
1984 case -EINVAL:
1985 case -EBUSY:
1986 pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
1987 cpu);
1988 break;
1989 case 0:
1990 break;
1991 default:
1992 WARN(1, "rc %d while trying to pin vcpu\n", ret);
1993 disable_pinning = true;
1994 }
1995}
1996
1997const struct hypervisor_x86 x86_hyper_xen = {
1998 .name = "Xen",
1999 .detect = xen_platform,
2000#ifdef CONFIG_XEN_PVHVM
2001 .init_platform = xen_hvm_guest_init,
2002#endif
2003 .x2apic_available = xen_x2apic_para_available,
2004 .set_cpu_features = xen_set_cpu_features,
2005 .pin_vcpu = xen_pin_vcpu,
2006};
2007EXPORT_SYMBOL(x86_hyper_xen);
2008
2009#ifdef CONFIG_HOTPLUG_CPU
2010void xen_arch_register_cpu(int num)
2011{
2012 arch_register_cpu(num);
2013}
2014EXPORT_SYMBOL(xen_arch_register_cpu);
2015
2016void xen_arch_unregister_cpu(int num)
2017{
2018 arch_unregister_cpu(num);
2019}
2020EXPORT_SYMBOL(xen_arch_unregister_cpu);
2021#endif
2022