2 * Jailhouse, a Linux-based partitioning hypervisor
4 * Copyright (c) Siemens AG, 2013
5 * Copyright (c) Valentine Sinitsyn, 2014
8 * Jan Kiszka <jan.kiszka@siemens.com>
9 * Valentine Sinitsyn <valentine.sinitsyn@gmail.com>
11 * Based on vmx.c written by Jan Kiszka.
13 * This work is licensed under the terms of the GNU GPL, version 2. See
14 * the COPYING file in the top-level directory.
17 #include <jailhouse/entry.h>
18 #include <jailhouse/cell-config.h>
19 #include <jailhouse/control.h>
20 #include <jailhouse/paging.h>
21 #include <jailhouse/printk.h>
22 #include <jailhouse/processor.h>
23 #include <jailhouse/string.h>
24 #include <jailhouse/utils.h>
27 #include <asm/control.h>
28 #include <asm/iommu.h>
29 #include <asm/paging.h>
30 #include <asm/percpu.h>
31 #include <asm/processor.h>
36 * NW bit is ignored by all modern processors, however some
37 * combinations of NW and CD bits are prohibited by SVM (see APMv2,
38 * Sect. 15.5). To handle this, we always keep the NW bit off.
40 #define SVM_CR0_ALLOWED_BITS (~X86_CR0_NW)
42 static bool has_avic, has_assists, has_flush_by_asid;
44 static const struct segment invalid_seg;
46 static struct paging npt_paging[NPT_PAGE_DIR_LEVELS];
48 /* bit cleared: direct access allowed */
49 // TODO: convert to whitelist
50 static u8 __attribute__((aligned(PAGE_SIZE))) msrpm[][0x2000/4] = {
51 [ SVM_MSRPM_0000 ] = {
52 [ 0/4 ... 0x017/4 ] = 0,
53 [ 0x018/4 ... 0x01b/4 ] = 0x80, /* 0x01b (w) */
54 [ 0x01c/4 ... 0x1ff/4 ] = 0,
55 [ 0x200/4 ... 0x273/4 ] = 0xaa, /* 0x200 - 0x273 (w) */
56 [ 0x274/4 ... 0x277/4 ] = 0xea, /* 0x274 - 0x276 (w), 0x277 (rw) */
57 [ 0x278/4 ... 0x2fb/4 ] = 0,
58 [ 0x2fc/4 ... 0x2ff/4 ] = 0x80, /* 0x2ff (w) */
59 [ 0x300/4 ... 0x7ff/4 ] = 0,
60 /* x2APIC MSRs - emulated if not present */
61 [ 0x800/4 ... 0x803/4 ] = 0x90, /* 0x802 (r), 0x803 (r) */
62 [ 0x804/4 ... 0x807/4 ] = 0,
63 [ 0x808/4 ... 0x80b/4 ] = 0x93, /* 0x808 (rw), 0x80a (r), 0x80b (w) */
64 [ 0x80c/4 ... 0x80f/4 ] = 0xc8, /* 0x80d (w), 0x80f (rw) */
65 [ 0x810/4 ... 0x813/4 ] = 0x55, /* 0x810 - 0x813 (r) */
66 [ 0x814/4 ... 0x817/4 ] = 0x55, /* 0x814 - 0x817 (r) */
67 [ 0x818/4 ... 0x81b/4 ] = 0x55, /* 0x818 - 0x81b (r) */
68 [ 0x81c/4 ... 0x81f/4 ] = 0x55, /* 0x81c - 0x81f (r) */
69 [ 0x820/4 ... 0x823/4 ] = 0x55, /* 0x820 - 0x823 (r) */
70 [ 0x824/4 ... 0x827/4 ] = 0x55, /* 0x823 - 0x827 (r) */
71 [ 0x828/4 ... 0x82b/4 ] = 0x03, /* 0x828 (rw) */
72 [ 0x82c/4 ... 0x82f/4 ] = 0xc0, /* 0x82f (rw) */
73 [ 0x830/4 ... 0x833/4 ] = 0xf3, /* 0x830 (rw), 0x832 (rw), 0x833 (rw) */
74 [ 0x834/4 ... 0x837/4 ] = 0xff, /* 0x834 - 0x837 (rw) */
75 [ 0x838/4 ... 0x83b/4 ] = 0x07, /* 0x838 (rw), 0x839 (r) */
76 [ 0x83c/4 ... 0x83f/4 ] = 0x70, /* 0x83e (rw), 0x83f (r) */
77 [ 0x840/4 ... 0x1fff/4 ] = 0,
79 [ SVM_MSRPM_C000 ] = {
80 [ 0/4 ... 0x07f/4 ] = 0,
81 [ 0x080/4 ... 0x083/4 ] = 0x02, /* 0x080 (w) */
82 [ 0x084/4 ... 0x1fff/4 ] = 0
84 [ SVM_MSRPM_C001 ] = {
85 [ 0/4 ... 0x1fff/4 ] = 0,
87 [ SVM_MSRPM_RESV ] = {
88 [ 0/4 ... 0x1fff/4 ] = 0,
92 /* This page is mapped so the code begins at 0x000ffff0 */
93 static u8 __attribute__((aligned(PAGE_SIZE))) parking_code[PAGE_SIZE] = {
94 [0xff0] = 0xfa, /* 1: cli */
95 [0xff1] = 0xf4, /* hlt */
97 [0xff3] = 0xfc /* jmp 1b */
100 static void *parked_mode_npt;
102 static void *avic_page;
104 static int svm_check_features(void)
106 /* SVM is available */
107 if (!(cpuid_ecx(0x80000001) & X86_FEATURE_SVM))
108 return trace_error(-ENODEV);
111 if (!(cpuid_edx(0x8000000A) & X86_FEATURE_NP))
112 return trace_error(-EIO);
115 if ((cpuid_edx(0x8000000A) & X86_FEATURE_DECODE_ASSISTS))
119 if (cpuid_edx(0x8000000A) & X86_FEATURE_AVIC)
122 /* TLB Flush by ASID support */
123 if (cpuid_edx(0x8000000A) & X86_FEATURE_FLUSH_BY_ASID)
124 has_flush_by_asid = true;
129 static void set_svm_segment_from_dtr(struct svm_segment *svm_segment,
130 const struct desc_table_reg *dtr)
132 svm_segment->base = dtr->base;
133 svm_segment->limit = dtr->limit & 0xffff;
136 static void set_svm_segment_from_segment(struct svm_segment *svm_segment,
137 const struct segment *segment)
139 svm_segment->selector = segment->selector;
140 svm_segment->access_rights = ((segment->access_rights & 0xf000) >> 4) |
141 (segment->access_rights & 0x00ff);
142 svm_segment->limit = segment->limit;
143 svm_segment->base = segment->base;
146 static void svm_set_cell_config(struct cell *cell, struct vmcb *vmcb)
148 vmcb->iopm_base_pa = paging_hvirt2phys(cell->svm.iopm);
149 vmcb->n_cr3 = paging_hvirt2phys(cell->svm.npt_structs.root_table);
152 static void vmcb_setup(struct per_cpu *cpu_data)
154 struct vmcb *vmcb = &cpu_data->vmcb;
156 memset(vmcb, 0, sizeof(struct vmcb));
158 vmcb->cr0 = cpu_data->linux_cr0 & SVM_CR0_ALLOWED_BITS;
159 vmcb->cr3 = cpu_data->linux_cr3;
160 vmcb->cr4 = cpu_data->linux_cr4;
162 set_svm_segment_from_segment(&vmcb->cs, &cpu_data->linux_cs);
163 set_svm_segment_from_segment(&vmcb->ds, &cpu_data->linux_ds);
164 set_svm_segment_from_segment(&vmcb->es, &cpu_data->linux_es);
165 set_svm_segment_from_segment(&vmcb->fs, &cpu_data->linux_fs);
166 set_svm_segment_from_segment(&vmcb->gs, &cpu_data->linux_gs);
167 set_svm_segment_from_segment(&vmcb->ss, &invalid_seg);
168 set_svm_segment_from_segment(&vmcb->tr, &cpu_data->linux_tss);
169 set_svm_segment_from_segment(&vmcb->ldtr, &invalid_seg);
171 set_svm_segment_from_dtr(&vmcb->gdtr, &cpu_data->linux_gdtr);
172 set_svm_segment_from_dtr(&vmcb->idtr, &cpu_data->linux_idtr);
174 vmcb->cpl = 0; /* Linux runs in ring 0 before migration */
177 /* Indicate success to the caller of arch_entry */
179 vmcb->rsp = cpu_data->linux_sp +
180 (NUM_ENTRY_REGS + 1) * sizeof(unsigned long);
181 vmcb->rip = cpu_data->linux_ip;
183 vmcb->sysenter_cs = read_msr(MSR_IA32_SYSENTER_CS);
184 vmcb->sysenter_eip = read_msr(MSR_IA32_SYSENTER_EIP);
185 vmcb->sysenter_esp = read_msr(MSR_IA32_SYSENTER_ESP);
186 vmcb->star = read_msr(MSR_STAR);
187 vmcb->lstar = read_msr(MSR_LSTAR);
188 vmcb->cstar = read_msr(MSR_CSTAR);
189 vmcb->sfmask = read_msr(MSR_SFMASK);
190 vmcb->kerngsbase = read_msr(MSR_KERNGS_BASE);
192 vmcb->dr6 = 0x00000ff0;
193 vmcb->dr7 = 0x00000400;
195 /* Make the hypervisor visible */
196 vmcb->efer = (cpu_data->linux_efer | EFER_SVME);
198 vmcb->g_pat = cpu_data->pat;
200 vmcb->general1_intercepts |= GENERAL1_INTERCEPT_NMI;
201 vmcb->general1_intercepts |= GENERAL1_INTERCEPT_CR0_SEL_WRITE;
202 vmcb->general1_intercepts |= GENERAL1_INTERCEPT_IOIO_PROT;
203 vmcb->general1_intercepts |= GENERAL1_INTERCEPT_MSR_PROT;
204 vmcb->general1_intercepts |= GENERAL1_INTERCEPT_SHUTDOWN_EVT;
206 vmcb->general2_intercepts |= GENERAL2_INTERCEPT_VMRUN; /* Required */
207 vmcb->general2_intercepts |= GENERAL2_INTERCEPT_VMMCALL;
209 vmcb->msrpm_base_pa = paging_hvirt2phys(msrpm);
212 /* No more than one guest owns the CPU */
213 vmcb->guest_asid = 1;
215 /* TODO: Setup AVIC */
217 /* Explicitly mark all of the state as new */
218 vmcb->clean_bits = 0;
220 svm_set_cell_config(cpu_data->cell, vmcb);
223 unsigned long arch_paging_gphys2phys(struct per_cpu *cpu_data,
227 return paging_virt2phys(&cpu_data->cell->svm.npt_structs,
231 static void npt_set_next_pt(pt_entry_t pte, unsigned long next_pt)
233 /* See APMv2, Section 15.25.5 */
234 *pte = (next_pt & 0x000ffffffffff000UL) |
235 (PAGE_DEFAULT_FLAGS | PAGE_FLAG_US);
238 int vcpu_vendor_init(void)
240 struct paging_structures parking_pt;
244 err = svm_check_features();
248 vm_cr = read_msr(MSR_VM_CR);
249 if (vm_cr & VM_CR_SVMDIS)
250 /* SVM disabled in BIOS */
251 return trace_error(-EPERM);
253 /* Nested paging is the same as the native one */
254 memcpy(npt_paging, x86_64_paging, sizeof(npt_paging));
255 for (n = 0; n < NPT_PAGE_DIR_LEVELS; n++)
256 npt_paging[n].set_next_pt = npt_set_next_pt;
258 /* Map guest parking code (shared between cells and CPUs) */
259 parking_pt.root_paging = npt_paging;
260 parking_pt.root_table = parked_mode_npt = page_alloc(&mem_pool, 1);
261 if (!parked_mode_npt)
263 err = paging_create(&parking_pt, paging_hvirt2phys(parking_code),
264 PAGE_SIZE, 0x000ff000,
265 PAGE_READONLY_FLAGS | PAGE_FLAG_US,
266 PAGING_NON_COHERENT);
270 /* This is always false for AMD now (except in nested SVM);
271 see Sect. 16.3.1 in APMv2 */
273 /* allow direct x2APIC access except for ICR writes */
274 memset(&msrpm[SVM_MSRPM_0000][MSR_X2APIC_BASE/4], 0,
275 (MSR_X2APIC_END - MSR_X2APIC_BASE + 1)/4);
276 msrpm[SVM_MSRPM_0000][MSR_X2APIC_ICR/4] = 0x02;
279 avic_page = page_alloc(&remap_pool, 1);
281 return trace_error(-ENOMEM);
285 return vcpu_cell_init(&root_cell);
288 int vcpu_vendor_cell_init(struct cell *cell)
293 /* allocate iopm (two 4-K pages + 3 bits) */
294 cell->svm.iopm = page_alloc(&mem_pool, 3);
298 /* build root NPT of cell */
299 cell->svm.npt_structs.root_paging = npt_paging;
300 cell->svm.npt_structs.root_table = page_alloc(&mem_pool, 1);
301 if (!cell->svm.npt_structs.root_table)
306 * Map xAPIC as is; reads are passed, writes are trapped.
308 flags = PAGE_READONLY_FLAGS | PAGE_FLAG_US | PAGE_FLAG_DEVICE;
309 err = paging_create(&cell->svm.npt_structs, XAPIC_BASE,
310 PAGE_SIZE, XAPIC_BASE,
312 PAGING_NON_COHERENT);
314 flags = PAGE_DEFAULT_FLAGS | PAGE_FLAG_DEVICE;
315 err = paging_create(&cell->svm.npt_structs,
316 paging_hvirt2phys(avic_page),
317 PAGE_SIZE, XAPIC_BASE,
319 PAGING_NON_COHERENT);
325 int vcpu_map_memory_region(struct cell *cell,
326 const struct jailhouse_memory *mem)
328 u64 phys_start = mem->phys_start;
329 u64 flags = PAGE_FLAG_US; /* See APMv2, Section 15.25.5 */
331 if (mem->flags & JAILHOUSE_MEM_READ)
332 flags |= PAGE_FLAG_PRESENT;
333 if (mem->flags & JAILHOUSE_MEM_WRITE)
334 flags |= PAGE_FLAG_RW;
335 if (!(mem->flags & JAILHOUSE_MEM_EXECUTE))
336 flags |= PAGE_FLAG_NOEXECUTE;
337 if (mem->flags & JAILHOUSE_MEM_COMM_REGION)
338 phys_start = paging_hvirt2phys(&cell->comm_page);
340 return paging_create(&cell->svm.npt_structs, phys_start, mem->size,
341 mem->virt_start, flags, PAGING_NON_COHERENT);
344 int vcpu_unmap_memory_region(struct cell *cell,
345 const struct jailhouse_memory *mem)
347 return paging_destroy(&cell->svm.npt_structs, mem->virt_start,
348 mem->size, PAGING_NON_COHERENT);
351 void vcpu_vendor_cell_exit(struct cell *cell)
353 paging_destroy(&cell->svm.npt_structs, XAPIC_BASE, PAGE_SIZE,
354 PAGING_NON_COHERENT);
355 page_free(&mem_pool, cell->svm.npt_structs.root_table, 1);
358 int vcpu_init(struct per_cpu *cpu_data)
363 err = svm_check_features();
367 efer = read_msr(MSR_EFER);
368 if (efer & EFER_SVME)
369 return trace_error(-EBUSY);
372 write_msr(MSR_EFER, efer);
374 cpu_data->svm_state = SVMON;
376 vmcb_setup(cpu_data);
379 * APM Volume 2, 3.1.1: "When writing the CR0 register, software should
380 * set the values of reserved bits to the values found during the
381 * previous CR0 read."
382 * But we want to avoid surprises with new features unknown to us but
383 * set by Linux. So check if any assumed revered bit was set and bail
385 * Note that the APM defines all reserved CR4 bits as must-be-zero.
387 if (cpu_data->linux_cr0 & X86_CR0_RESERVED)
390 /* bring CR0 and CR4 into well-defined states */
391 write_cr0(X86_CR0_HOST_STATE);
392 write_cr4(X86_CR4_HOST_STATE);
394 write_msr(MSR_VM_HSAVE_PA, paging_hvirt2phys(cpu_data->host_state));
399 void vcpu_exit(struct per_cpu *cpu_data)
403 if (cpu_data->svm_state == SVMOFF)
406 cpu_data->svm_state = SVMOFF;
408 /* We are leaving - set the GIF */
409 asm volatile ("stgi" : : : "memory");
411 efer = read_msr(MSR_EFER);
413 write_msr(MSR_EFER, efer);
415 write_msr(MSR_VM_HSAVE_PA, 0);
418 void __attribute__((noreturn)) vcpu_activate_vmm(struct per_cpu *cpu_data)
420 unsigned long vmcb_pa, host_stack;
422 vmcb_pa = paging_hvirt2phys(&cpu_data->vmcb);
423 host_stack = (unsigned long)cpu_data->stack + sizeof(cpu_data->stack);
425 /* We enter Linux at the point arch_entry would return to as well.
426 * rax is cleared to signal success to the caller. */
429 "mov (%%rdi),%%r15\n\t"
430 "mov 0x8(%%rdi),%%r14\n\t"
431 "mov 0x10(%%rdi),%%r13\n\t"
432 "mov 0x18(%%rdi),%%r12\n\t"
433 "mov 0x20(%%rdi),%%rbx\n\t"
434 "mov 0x28(%%rdi),%%rbp\n\t"
439 : "D" (cpu_data->linux_reg), "a" (vmcb_pa), "m" (host_stack));
440 __builtin_unreachable();
443 void __attribute__((noreturn)) vcpu_deactivate_vmm(void)
445 struct per_cpu *cpu_data = this_cpu_data();
446 struct vmcb *vmcb = &cpu_data->vmcb;
447 unsigned long *stack = (unsigned long *)vmcb->rsp;
448 unsigned long linux_ip = vmcb->rip;
450 cpu_data->linux_cr0 = vmcb->cr0;
451 cpu_data->linux_cr3 = vmcb->cr3;
453 cpu_data->linux_gdtr.base = vmcb->gdtr.base;
454 cpu_data->linux_gdtr.limit = vmcb->gdtr.limit;
455 cpu_data->linux_idtr.base = vmcb->idtr.base;
456 cpu_data->linux_idtr.limit = vmcb->idtr.limit;
458 cpu_data->linux_cs.selector = vmcb->cs.selector;
460 asm volatile("str %0" : "=m" (cpu_data->linux_tss.selector));
462 cpu_data->linux_efer = vmcb->efer & (~EFER_SVME);
463 cpu_data->linux_gs.base = vmcb->gs.base;
465 cpu_data->linux_ds.selector = vmcb->ds.selector;
466 cpu_data->linux_es.selector = vmcb->es.selector;
468 asm volatile("mov %%fs,%0" : "=m" (cpu_data->linux_fs.selector));
469 asm volatile("mov %%gs,%0" : "=m" (cpu_data->linux_gs.selector));
471 arch_cpu_restore(cpu_data, 0);
477 "mov %%rbx,%%rsp\n\t"
493 "mov %%rax,%%rsp\n\t"
494 "xor %%rax,%%rax\n\t"
496 : : "a" (stack), "b" (&cpu_data->guest_regs));
497 __builtin_unreachable();
500 static void svm_vcpu_reset(struct per_cpu *cpu_data, unsigned int sipi_vector)
502 static const struct svm_segment dataseg_reset_state = {
506 .access_rights = 0x0093,
508 static const struct svm_segment dtr_reset_state = {
514 struct vmcb *vmcb = &cpu_data->vmcb;
517 vmcb->cr0 = X86_CR0_NW | X86_CR0_CD | X86_CR0_ET;
524 if (sipi_vector == APIC_BSP_PSEUDO_SIPI) {
531 vmcb->cs.selector = sipi_vector << 8;
532 vmcb->cs.base = sipi_vector << 12;
533 vmcb->cs.limit = 0xffff;
534 vmcb->cs.access_rights = 0x009b;
536 vmcb->ds = dataseg_reset_state;
537 vmcb->es = dataseg_reset_state;
538 vmcb->fs = dataseg_reset_state;
539 vmcb->gs = dataseg_reset_state;
540 vmcb->ss = dataseg_reset_state;
542 vmcb->tr.selector = 0;
544 vmcb->tr.limit = 0xffff;
545 vmcb->tr.access_rights = 0x008b;
547 vmcb->ldtr.selector = 0;
549 vmcb->ldtr.limit = 0xffff;
550 vmcb->ldtr.access_rights = 0x0082;
552 vmcb->gdtr = dtr_reset_state;
553 vmcb->idtr = dtr_reset_state;
555 vmcb->efer = EFER_SVME;
557 /* These MSRs are undefined on reset */
562 vmcb->sysenter_cs = 0;
563 vmcb->sysenter_eip = 0;
564 vmcb->sysenter_esp = 0;
565 vmcb->kerngsbase = 0;
567 vmcb->dr7 = 0x00000400;
569 /* Almost all of the guest state changed */
570 vmcb->clean_bits = 0;
572 svm_set_cell_config(cpu_data->cell, vmcb);
576 : : "a" (paging_hvirt2phys(vmcb)) : "memory");
577 /* vmload overwrites GS_BASE - restore the host state */
578 write_msr(MSR_GS_BASE, (unsigned long)cpu_data);
581 void vcpu_skip_emulated_instruction(unsigned int inst_len)
583 this_cpu_data()->vmcb.rip += inst_len;
586 static void update_efer(struct vmcb *vmcb)
588 unsigned long efer = vmcb->efer;
590 if ((efer & (EFER_LME | EFER_LMA)) != EFER_LME)
595 /* Flush TLB on LMA/LME change: See APMv2, Sect. 15.16 */
596 if ((vmcb->efer ^ efer) & EFER_LMA)
600 vmcb->clean_bits &= ~CLEAN_BITS_CRX;
603 bool vcpu_get_guest_paging_structs(struct guest_paging_structures *pg_structs)
605 struct vmcb *vmcb = &this_cpu_data()->vmcb;
607 if (vmcb->efer & EFER_LMA) {
608 pg_structs->root_paging = x86_64_paging;
609 pg_structs->root_table_gphys =
610 vmcb->cr3 & 0x000ffffffffff000UL;
611 } else if ((vmcb->cr0 & X86_CR0_PG) &&
612 !(vmcb->cr4 & X86_CR4_PAE)) {
613 pg_structs->root_paging = i386_paging;
614 pg_structs->root_table_gphys =
615 vmcb->cr3 & 0xfffff000UL;
616 } else if (!(vmcb->cr0 & X86_CR0_PG)) {
618 * Can be in non-paged protected mode as well, but
619 * the translation mechanism will stay the same ayway.
621 pg_structs->root_paging = realmode_paging;
623 * This will make paging_get_guest_pages map the page
624 * that also contains the bootstrap code and, thus, is
625 * always present in a cell.
627 pg_structs->root_table_gphys = 0xff000;
629 printk("FATAL: Unsupported paging mode\n");
635 void vcpu_vendor_set_guest_pat(unsigned long val)
637 struct vmcb *vmcb = &this_cpu_data()->vmcb;
640 vmcb->clean_bits &= ~CLEAN_BITS_NP;
643 struct parse_context {
644 unsigned int remaining;
646 unsigned long cs_base;
650 static bool ctx_advance(struct parse_context *ctx,
652 struct guest_paging_structures *pg_structs)
655 ctx->size = ctx->remaining;
656 ctx->inst = vcpu_map_inst(pg_structs, ctx->cs_base + *pc,
660 ctx->remaining -= ctx->size;
666 static bool svm_parse_mov_to_cr(struct vmcb *vmcb, unsigned long pc,
667 unsigned char reg, unsigned long *gpr)
669 struct guest_paging_structures pg_structs;
670 struct parse_context ctx = {};
671 /* No prefixes are supported yet */
672 u8 opcodes[] = {0x0f, 0x22}, modrm;
675 ctx.remaining = ARRAY_SIZE(opcodes);
676 if (!vcpu_get_guest_paging_structs(&pg_structs))
678 ctx.cs_base = (vmcb->efer & EFER_LMA) ? 0 : vmcb->cs.base;
680 if (!ctx_advance(&ctx, &pc, &pg_structs))
683 for (n = 0; n < ARRAY_SIZE(opcodes); n++, ctx.inst++)
684 if (*(ctx.inst) != opcodes[n] ||
685 !ctx_advance(&ctx, &pc, &pg_structs))
688 if (!ctx_advance(&ctx, &pc, &pg_structs))
693 if (((modrm & 0x38) >> 3) != reg)
697 *gpr = (modrm & 0x7);
703 * XXX: The only visible reason to have this function (vmx.c consistency
704 * aside) is to prevent cells from setting invalid CD+NW combinations that
705 * result in no more than VMEXIT_INVALID. Maybe we can get along without it
708 static bool svm_handle_cr(struct per_cpu *cpu_data)
710 struct vmcb *vmcb = &cpu_data->vmcb;
711 /* Workaround GCC 4.8 warning on uninitialized variable 'reg' */
712 unsigned long reg = -1, val, bits;
715 if (!(vmcb->exitinfo1 & (1UL << 63))) {
716 panic_printk("FATAL: Unsupported CR access (LMSW or CLTS)\n");
719 reg = vmcb->exitinfo1 & 0x07;
721 if (!svm_parse_mov_to_cr(vmcb, vmcb->rip, 0, ®)) {
722 panic_printk("FATAL: Unable to parse MOV-to-CR instruction\n");
730 val = cpu_data->guest_regs.by_index[15 - reg];
732 vcpu_skip_emulated_instruction(X86_INST_LEN_MOV_TO_CR);
733 /* Flush TLB on PG/WP/CD/NW change: See APMv2, Sect. 15.16 */
734 bits = (X86_CR0_PG | X86_CR0_WP | X86_CR0_CD | X86_CR0_NW);
735 if ((val ^ vmcb->cr0) & bits)
737 /* TODO: better check for #GP reasons */
738 vmcb->cr0 = val & SVM_CR0_ALLOWED_BITS;
739 if (val & X86_CR0_PG)
741 vmcb->clean_bits &= ~CLEAN_BITS_CRX;
746 static bool svm_handle_msr_write(struct per_cpu *cpu_data)
748 struct vmcb *vmcb = &cpu_data->vmcb;
751 if (cpu_data->guest_regs.rcx == MSR_EFER) {
752 /* Never let a guest to disable SVME; see APMv2, Sect. 3.1.7 */
753 efer = get_wrmsr_value(&cpu_data->guest_regs) | EFER_SVME;
754 /* Flush TLB on LME/NXE change: See APMv2, Sect. 15.16 */
755 if ((efer ^ vmcb->efer) & (EFER_LME | EFER_NXE))
758 vmcb->clean_bits &= ~CLEAN_BITS_CRX;
759 vcpu_skip_emulated_instruction(X86_INST_LEN_WRMSR);
763 return vcpu_handle_msr_write();
767 * TODO: This handles unaccelerated (non-AVIC) access. AVIC should
768 * be treated separately in svm_handle_avic_access().
770 static bool svm_handle_apic_access(struct vmcb *vmcb)
772 struct guest_paging_structures pg_structs;
773 unsigned int inst_len, offset;
776 /* The caller is responsible for sanity checks */
777 is_write = !!(vmcb->exitinfo1 & 0x2);
778 offset = vmcb->exitinfo2 - XAPIC_BASE;
783 if (!vcpu_get_guest_paging_structs(&pg_structs))
786 inst_len = apic_mmio_access(vmcb->rip, &pg_structs, offset >> 4,
791 vcpu_skip_emulated_instruction(inst_len);
795 panic_printk("FATAL: Unhandled APIC access, offset %d, is_write: %d\n",
800 static void dump_guest_regs(union registers *guest_regs, struct vmcb *vmcb)
802 panic_printk("RIP: %p RSP: %p FLAGS: %x\n", vmcb->rip,
803 vmcb->rsp, vmcb->rflags);
804 panic_printk("RAX: %p RBX: %p RCX: %p\n", guest_regs->rax,
805 guest_regs->rbx, guest_regs->rcx);
806 panic_printk("RDX: %p RSI: %p RDI: %p\n", guest_regs->rdx,
807 guest_regs->rsi, guest_regs->rdi);
808 panic_printk("CS: %x BASE: %p AR-BYTES: %x EFER.LMA %d\n",
809 vmcb->cs.selector, vmcb->cs.base, vmcb->cs.access_rights,
810 !!(vmcb->efer & EFER_LMA));
811 panic_printk("CR0: %p CR3: %p CR4: %p\n", vmcb->cr0,
812 vmcb->cr3, vmcb->cr4);
813 panic_printk("EFER: %p\n", vmcb->efer);
816 void vcpu_vendor_get_io_intercept(struct vcpu_io_intercept *io)
818 struct vmcb *vmcb = &this_cpu_data()->vmcb;
819 u64 exitinfo = vmcb->exitinfo1;
821 /* parse exit info for I/O instructions (see APM, 15.10.2 ) */
822 io->port = (exitinfo >> 16) & 0xFFFF;
823 io->size = (exitinfo >> 4) & 0x7;
824 io->in = !!(exitinfo & 0x1);
825 io->inst_len = vmcb->exitinfo2 - vmcb->rip;
826 io->rep_or_str = !!(exitinfo & 0x0c);
829 void vcpu_vendor_get_mmio_intercept(struct vcpu_mmio_intercept *mmio)
831 struct vmcb *vmcb = &this_cpu_data()->vmcb;
833 mmio->phys_addr = vmcb->exitinfo2;
834 mmio->is_write = !!(vmcb->exitinfo1 & 0x2);
837 void vcpu_handle_exit(struct per_cpu *cpu_data)
839 struct vmcb *vmcb = &cpu_data->vmcb;
843 vmcb->gs.base = read_msr(MSR_GS_BASE);
845 /* Restore GS value expected by per_cpu data accessors */
846 write_msr(MSR_GS_BASE, (unsigned long)cpu_data);
848 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_TOTAL]++;
850 * All guest state is marked unmodified; individual handlers must clear
851 * the bits as needed.
853 vmcb->clean_bits = 0xffffffff;
855 switch (vmcb->exitcode) {
857 panic_printk("FATAL: VM-Entry failure, error %d\n",
861 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_MANAGEMENT]++;
862 /* Temporarily enable GIF to consume pending NMI */
863 asm volatile("stgi; clgi" : : : "memory");
864 sipi_vector = x86_handle_events(cpu_data);
865 if (sipi_vector >= 0) {
866 printk("CPU %d received SIPI, vector %x\n",
867 cpu_data->cpu_id, sipi_vector);
868 svm_vcpu_reset(cpu_data, sipi_vector);
871 iommu_check_pending_faults(cpu_data);
874 vcpu_handle_hypercall();
876 case VMEXIT_CR0_SEL_WRITE:
877 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_CR]++;
878 if (svm_handle_cr(cpu_data))
882 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_MSR]++;
883 if (!vmcb->exitinfo1)
884 res = vcpu_handle_msr_read();
886 res = svm_handle_msr_write(cpu_data);
891 if ((vmcb->exitinfo1 & 0x7) == 0x7 &&
892 vmcb->exitinfo2 >= XAPIC_BASE &&
893 vmcb->exitinfo2 < XAPIC_BASE + PAGE_SIZE) {
894 /* APIC access in non-AVIC mode */
895 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_XAPIC]++;
896 if (svm_handle_apic_access(vmcb))
899 /* General MMIO (IOAPIC, PCI etc) */
900 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_MMIO]++;
901 if (vcpu_handle_mmio_access())
905 panic_printk("FATAL: Unhandled Nested Page Fault for (%p), "
906 "error code is %x\n", vmcb->exitinfo2,
907 vmcb->exitinfo1 & 0xf);
910 if (vcpu_handle_xsetbv())
914 cpu_data->stats[JAILHOUSE_CPU_STAT_VMEXITS_PIO]++;
915 if (vcpu_handle_io_access())
918 /* TODO: Handle VMEXIT_AVIC_NOACCEL and VMEXIT_AVIC_INCOMPLETE_IPI */
920 panic_printk("FATAL: Unexpected #VMEXIT, exitcode %x, "
921 "exitinfo1 %p exitinfo2 %p\n",
922 vmcb->exitcode, vmcb->exitinfo1, vmcb->exitinfo2);
924 dump_guest_regs(&cpu_data->guest_regs, vmcb);
928 write_msr(MSR_GS_BASE, vmcb->gs.base);
933 svm_vcpu_reset(this_cpu_data(), APIC_BSP_PSEUDO_SIPI);
934 /* No need to clear VMCB Clean bit: vcpu_reset() already does this */
935 this_cpu_data()->vmcb.n_cr3 = paging_hvirt2phys(parked_mode_npt);
940 void vcpu_nmi_handler(void)
944 void vcpu_tlb_flush(void)
946 struct vmcb *vmcb = &this_cpu_data()->vmcb;
948 if (has_flush_by_asid)
949 vmcb->tlb_control = SVM_TLB_FLUSH_GUEST;
951 vmcb->tlb_control = SVM_TLB_FLUSH_ALL;
954 const u8 *vcpu_get_inst_bytes(const struct guest_paging_structures *pg_structs,
955 unsigned long pc, unsigned int *size)
957 struct vmcb *vmcb = &this_cpu_data()->vmcb;
963 start = vmcb->rip - pc;
964 if (start < vmcb->bytes_fetched) {
965 *size = vmcb->bytes_fetched - start;
966 return &vmcb->guest_bytes[start];
971 return vcpu_map_inst(pg_structs, pc, size);
975 void vcpu_vendor_get_cell_io_bitmap(struct cell *cell,
976 struct vcpu_io_bitmap *iobm)
978 iobm->data = cell->svm.iopm;
979 iobm->size = sizeof(cell->svm.iopm);
982 void vcpu_vendor_get_execution_state(struct vcpu_execution_state *x_state)
984 struct vmcb *vmcb = &this_cpu_data()->vmcb;
986 x_state->efer = vmcb->efer;
987 x_state->rflags = vmcb->rflags;
988 x_state->cs = vmcb->cs.selector;
989 x_state->rip = vmcb->rip;
992 /* GIF must be set for interrupts to be delivered (APMv2, Sect. 15.17) */
993 void enable_irq(void)
995 asm volatile("stgi; sti" : : : "memory");
998 /* Jailhouse runs with GIF cleared, so we need to restore this state */
999 void disable_irq(void)
1001 asm volatile("cli; clgi" : : : "memory");