?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
get rid of GET_GUESTVM
[freebsd-src/fkvm-freebsd.git] / sys / kern / kern_fkvm.c
blobdb0f68582cf797e510a8b7e4a00a90555184fe53
1 /*-
2 * Copyright (c) 2008 The FreeBSD Project
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27 #include <sys/cdefs.h>
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/kernel.h>
31 #include <sys/malloc.h>
32 #include <sys/sysproto.h>
33 #include <sys/file.h>
34 #include <sys/mman.h>
35 #include <sys/proc.h>
36 #include <vm/vm.h>
37 #include <vm/pmap.h>
38 #include <vm/vm_extern.h>
39 #include <vm/vm_map.h>
40 #include <vm/vm_object.h>
41 #include <vm/vm_param.h>
42 #include <machine/_inttypes.h>
43 #include <machine/specialreg.h>
44 #include <machine/segments.h>
45 #include <machine/vmcb.h>
46
47 #define IOPM_SIZE (8*1024 + 1) /* TODO: ensure that this need not be 12Kbtes, not just 8Kb+1 */
48 #define MSRPM_SIZE (8*1024)
49
50 /* fkvm data */
51 static void *iopm = NULL; /* Should I allocate a vm_object_t instead? */
52 static void *msrpm = NULL; /* Should I allocate a vm_object_t instead? */
53
54 static void *hsave_area = NULL;
55
56 /* per-guest data */
57
58 enum {
59 VCPU_REGS_RAX = 0,
60 VCPU_REGS_RCX = 1,
61 VCPU_REGS_RDX = 2,
62 VCPU_REGS_RBX = 3,
63 VCPU_REGS_RSP = 4,
64 VCPU_REGS_RBP = 5,
65 VCPU_REGS_RSI = 6,
66 VCPU_REGS_RDI = 7,
67 VCPU_REGS_R8 = 8,
68 VCPU_REGS_R9 = 9,
69 VCPU_REGS_R10 = 10,
70 VCPU_REGS_R11 = 11,
71 VCPU_REGS_R12 = 12,
72 VCPU_REGS_R13 = 13,
73 VCPU_REGS_R14 = 14,
74 VCPU_REGS_R15 = 15,
75 VCPU_REGS_RIP,
76 NR_VCPU_REGS
77 };
78
79 struct vcpu {
80 /* VCPU data */
81 struct vmcb *vmcb;
82 unsigned long vmcb_pa;
83
84 unsigned long regs[NR_VCPU_REGS];
85 u_int64_t host_gs_base;
86 u_int64_t cr2;
87 u_int64_t cr3;
88
89 struct guestvm *guest_vm;
90 };
91
92 #define MAX_VCPUS 8
93
94 struct guestvm {
95 struct vcpu *vcpus[MAX_VCPUS];
96 int nr_vcpus;
97
98 struct vmspace *sp;
99 u_int64_t nested_cr3;
100
101 };
102
103 #define SVM_VMLOAD ".byte 0x0f, 0x01, 0xda"
104 #define SVM_VMRUN ".byte 0x0f, 0x01, 0xd8"
105 #define SVM_VMSAVE ".byte 0x0f, 0x01, 0xdb"
106 #define SVM_CLGI ".byte 0x0f, 0x01, 0xdd"
107 #define SVM_STGI ".byte 0x0f, 0x01, 0xdc"
108 #define SVM_INVLPGA ".byte 0x0f, 0x01, 0xdf"
109
110 #define TD_GET_VCPU(td) td->vcpu
111
112 #define TD_SET_VCPU(td, vcpu) do { \
113 td->vcpu = vcpu; \
114 } while (0)
115
116 #define PROC_GET_GUESTVM(p) p->p_guestvm
117
118 #define PROC_SET_GUESTVM(p, guestvm) do { \
119 p->p_guestvm = guestvm; \
120 } while (0)
121
122 static void
123 print_vmcb_seg(struct vmcb_seg* vmcb_seg, const char* name)
124 {
125 printf("%s Selector\n", name);
126 printf("Selector : %" PRIx16 "\n", vmcb_seg->selector);
127 printf("Attributes : %" PRIx16 "\n", vmcb_seg->attrib);
128 printf("Limit : %" PRIx32 "\n", vmcb_seg->limit);
129 printf("Base Address : %" PRIx64 "\n", vmcb_seg->base);
130 printf("\n");
131 }
132
133 static void
134 print_vmcb(struct vmcb *vmcb)
135 {
136 printf("VMCB Control Area\n");
137 printf("Intercept CR Reads : %" PRIx16 "\n", vmcb->control.intercept_cr_reads);
138 printf("Intercept CR Writes : %" PRIx16 "\n", vmcb->control.intercept_cr_writes);
139 printf("Intercept DR Reads : %" PRIx16 "\n", vmcb->control.intercept_dr_reads);
140 printf("Intercept DR Writes : %" PRIx16 "\n", vmcb->control.intercept_dr_writes);
141 printf("Intercept Exceptions : %" PRIx32 "\n", vmcb->control.intercept_exceptions);
142 printf("Intercepts : %" PRIx64 "\n", vmcb->control.intercepts);
143 printf("Reserved 1: \n");
144 for(int i=0; i < 44; i++) {
145 printf("%" PRIx8 "", vmcb->control.reserved_1[i]); /* Should be Zero */
146 }
147 printf("\n");
148 printf("IOPM Base PA : %" PRIx64 "\n", vmcb->control.iopm_base_pa);
149 printf("MSRPM Base PA : %" PRIx64 "\n", vmcb->control.msrpm_base_pa);
150 printf("TSC Offset : %" PRIx64 "\n", vmcb->control.tsc_offset);
151 printf("Guest ASID : %" PRIx32 "\n", vmcb->control.guest_asid);
152 printf("TLB Control : %" PRIx8 "\n", vmcb->control.tlb_control);
153 printf("Reserved 2 : \n");
154 for(int i=0; i < 3; i++) {
155 printf("%" PRIx8 "", vmcb->control.reserved_1[i]); /* Should be Zero */
156 }
157 printf("\n");
158 printf("Virtual TPR : %" PRIx8 "\n", vmcb->control.v_tpr);
159 printf("Virtual IRQ : %" PRIx8 "\n", vmcb->control.v_irq);
160 printf("Virtual Interrupt : %" PRIx8 "\n", vmcb->control.v_intr);
161 printf("Virtual Interrupt Masking: %" PRIx8 "\n", vmcb->control.v_intr_masking);
162 printf("Virtual Interrupt Vector : %" PRIx8 "\n", vmcb->control.v_intr_vector);
163 printf("Reserved 6 : \n");
164 for(int i=0; i < 3; i++) {
165 printf("%" PRIx8 "", vmcb->control.reserved_6[i]); /* Should be Zero */
166 }
167 printf("\n");
168 printf("Interrupt Shadow : %" PRIx8 "\n", vmcb->control.intr_shadow);
169 printf("Reserved 7 : \n");
170 for(int i=0; i < 7; i++) {
171 printf("%" PRIx8 "", vmcb->control.reserved_7[i]); /* Should be Zero */
172 }
173 printf("\n");
174 printf("Exit Code : %" PRIx64 "\n", vmcb->control.exit_code);
175 printf("Exit Info 1 : %" PRIx64 "\n", vmcb->control.exit_info_1);
176 printf("Exit Info 2 : %" PRIx64 "\n", vmcb->control.exit_info_2);
177 printf("Exit Interrupt Info : %" PRIx32 "\n", vmcb->control.exit_int_info);
178 printf("Exit Interrupt Info Err Code: %" PRIx32 "\n", vmcb->control.exit_int_info_err_code);
179 printf("Nested Control : %" PRIx64 "\n", vmcb->control.nested_ctl);
180 printf("Reserved 8 : \n");
181 for(int i=0; i < 16; i++) {
182 printf("%" PRIx8 "", vmcb->control.reserved_8[i]); /* Should be Zero */
183 }
184 printf("\n");
185 printf("Event Injection : %" PRIx64 "\n", vmcb->control.event_inj);
186 printf("Nested CR3 : %" PRIx64 "\n", vmcb->control.nested_cr3);
187 printf("LBR Virtualization Enable: %" PRIx64 "\n", vmcb->control.lbr_virt_enable);
188 printf("Reserved 9 : \n");
189 for(int i=0; i < 832; i++) {
190 printf("%" PRIx8 "", vmcb->control.reserved_9[i]); /* Should be Zero */
191 }
192 printf("\n");
193
194 printf("\n");
195
196 printf("VMCB Save Area\n");
197 print_vmcb_seg(&(vmcb->save.es), "ES");
198 print_vmcb_seg(&(vmcb->save.es), "CS");
199 print_vmcb_seg(&(vmcb->save.es), "SS");
200 print_vmcb_seg(&(vmcb->save.es), "DS");
201 print_vmcb_seg(&(vmcb->save.es), "FS");
202 print_vmcb_seg(&(vmcb->save.es), "GS");
203 print_vmcb_seg(&(vmcb->save.es), "GDTR");
204 print_vmcb_seg(&(vmcb->save.es), "LDTR");
205 print_vmcb_seg(&(vmcb->save.es), "IDTR");
206 print_vmcb_seg(&(vmcb->save.es), "TR");
207 printf("Reserved 1 : \n");
208 for(int i=0; i < 43; i++) {
209 printf("%" PRIx8 "", vmcb->save.reserved_1[i]); /* Should be Zero */
210 }
211 printf("\n");
212 printf("Current Processor Level : %" PRIx8 "\n", vmcb->save.cpl);
213 printf("Reserved 2 : \n");
214 for(int i=0; i < 4; i++) {
215 printf("%" PRIx8 "", vmcb->save.reserved_2[i]); /* Should be Zero */
216 }
217 printf("\n");
218 printf("EFER : %" PRIx64 "\n", vmcb->save.efer);
219 printf("Reserved 3 : \n");
220 for(int i=0; i < 112; i++) {
221 printf("%" PRIx8 "", vmcb->save.reserved_3[i]); /* Should be Zero */
222 }
223 printf("\n");
224 printf("Control Register 4 : %" PRIx64 "\n", vmcb->save.cr4);
225 printf("Control Register 3 : %" PRIx64 "\n", vmcb->save.cr3);
226 printf("Control Register 0 : %" PRIx64 "\n", vmcb->save.cr0);
227 printf("Debug Register 7 : %" PRIx64 "\n", vmcb->save.dr7);
228 printf("Debug Register 6 : %" PRIx64 "\n", vmcb->save.dr6);
229 printf("RFlags : %" PRIx64 "\n", vmcb->save.rflags);
230 printf("RIP : %" PRIx64 "\n", vmcb->save.rip);
231 printf("Reserved 4 : \n");
232 for(int i=0; i < 88; i++) {
233 printf("%" PRIx8 "", vmcb->save.reserved_4[i]); /* Should be Zero */
234 }
235 printf("\n");
236 printf("RSP : %" PRIx64 "\n", vmcb->save.rsp);
237 printf("Reserved 5 : \n");
238 for(int i=0; i < 24; i++) {
239 printf("%" PRIx8 "", vmcb->save.reserved_5[i]); /* Should be Zero */
240 }
241 printf("\n");
242 printf("RAX : %" PRIx64 "\n", vmcb->save.rax);
243 printf("STAR : %" PRIx64 "\n", vmcb->save.star);
244 printf("LSTAR : %" PRIx64 "\n", vmcb->save.lstar);
245 printf("CSTAR : %" PRIx64 "\n", vmcb->save.cstar);
246 printf("SFMASK : %" PRIx64 "\n", vmcb->save.sfmask);
247 printf("Kernel GS Base : %" PRIx64 "\n", vmcb->save.kernel_gs_base);
248 printf("SYSENTER CS : %" PRIx64 "\n", vmcb->save.sysenter_cs);
249 printf("SYSENTER ESP : %" PRIx64 "\n", vmcb->save.sysenter_esp);
250 printf("SYSENTER EIP : %" PRIx64 "\n", vmcb->save.sysenter_eip);
251 printf("Control Register 2 : %" PRIx64 "\n", vmcb->save.cr2);
252 printf("Reserved 6 : \n");
253 for(int i=0; i < 32; i++) {
254 printf("%" PRIx8 "", vmcb->save.reserved_6[i]); /* Should be Zero */
255 }
256 printf("\n");
257 printf("Global PAT : %" PRIx64 "\n", vmcb->save.g_pat);
258 printf("Debug Control : %" PRIx64 "\n", vmcb->save.dbg_ctl);
259 printf("BR From : %" PRIx64 "\n", vmcb->save.br_from);
260 printf("BR To : %" PRIx64 "\n", vmcb->save.br_to);
261 printf("Last Exception From : %" PRIx64 "\n", vmcb->save.last_excp_from);
262 printf("Last Exception To : %" PRIx64 "\n", vmcb->save.last_excp_to);
263
264 printf("\n\n");
265 }
266
267 #if 0
268 static void
269 print_tss_desc(struct system_segment_descriptor *tss_desc)
270 {
271 printf("TSS desc @ %p:\n", tss_desc);
272 printf("sd_lolimit: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_lolimit);
273 printf("sd_lobase: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_lobase);
274 printf("sd_type: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_type);
275 printf("sd_dpl: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_dpl);
276 printf("sd_p: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_p);
277 printf("sd_hilimit: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_hilimit);
278 printf("sd_xx0: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_xx0);
279 printf("sd_gran: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_gran);
280 printf("sd_hibase: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_hibase);
281 printf("sd_xx1: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_xx1);
282 printf("sd_mbz: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_mbz);
283 printf("sd_xx2: 0x%" PRIx64 "\n", (u_int64_t) tss_desc->sd_xx2);
284 printf("\n\n");
285 }
286
287 static void
288 print_tss(struct system_segment_descriptor *tss_desc)
289 {
290 u_int32_t *base;
291 int limit;
292 int i;
293
294 base = (u_int32_t*) ((((u_int64_t) tss_desc->sd_hibase) << 24) | ((u_int64_t) tss_desc->sd_lobase));
295 limit = ((tss_desc->sd_hilimit << 16) | tss_desc->sd_lolimit) / 4;
296
297 printf("TSS: @ %p\n", base);
298 for (i = 0; i <= limit; i++)
299 printf("%x: 0x%" PRIx32 "\n", i, base[i]);
300 printf("\n\n");
301 }
302 #endif
303
304 static void
305 print_vmcb_save_area(struct vmcb *vmcb)
306 {
307 printf("VMCB save area:\n");
308 printf(" cs: [selector %" PRIx16 ", attrib %" PRIx16 ", limit %" PRIx32 ", base %" PRIx64 "]\n",
309 vmcb->save.cs.selector,
310 vmcb->save.cs.attrib,
311 vmcb->save.cs.limit,
312 vmcb->save.cs.base);
313 printf(" fs: [selector %" PRIx16 ", attrib %" PRIx16 ", limit %" PRIx32 ", base %" PRIx64 "]\n",
314 vmcb->save.fs.selector,
315 vmcb->save.fs.attrib,
316 vmcb->save.fs.limit,
317 vmcb->save.fs.base);
318 printf(" gs: [selector %" PRIx16 ", attrib %" PRIx16 ", limit %" PRIx32 ", base %" PRIx64 "]\n",
319 vmcb->save.gs.selector,
320 vmcb->save.gs.attrib,
321 vmcb->save.gs.limit,
322 vmcb->save.gs.base);
323 printf(" tr: [selector %" PRIx16 ", attrib %" PRIx16 ", limit %" PRIx32 ", base %" PRIx64 "]\n",
324 vmcb->save.tr.selector,
325 vmcb->save.tr.attrib,
326 vmcb->save.tr.limit,
327 vmcb->save.tr.base);
328 printf(" ldtr: [selector %" PRIx16 ", attrib %" PRIx16 ", limit %" PRIx32 ", base %" PRIx64 "]\n",
329 vmcb->save.ldtr.selector,
330 vmcb->save.ldtr.attrib,
331 vmcb->save.ldtr.limit,
332 vmcb->save.ldtr.base);
333 printf(" rip: %" PRIx64 "\n", vmcb->save.rip);
334 printf(" kernel_gs_base: %" PRIx64 "\n", vmcb->save.kernel_gs_base);
335 printf(" star: %" PRIx64 "\n", vmcb->save.star);
336 printf(" lstar: %" PRIx64 "\n", vmcb->save.lstar);
337 printf(" cstar: %" PRIx64 "\n", vmcb->save.cstar);
338 printf(" sfmask: %" PRIx64 "\n", vmcb->save.sfmask);
339 printf(" sysenter_cs: %" PRIx64 "\n", vmcb->save.sysenter_cs);
340 printf(" sysenter_esp: %" PRIx64 "\n", vmcb->save.sysenter_esp);
341 printf(" sysenter_eip: %" PRIx64 "\n", vmcb->save.sysenter_eip);
342 printf("\n\n");
343 }
344
345 static int
346 vmrun_assert(struct vmcb *vmcb)
347 {
348 #define A(cond) do { if ((cond)) { printf("Error: assertion not met on line %d\n", __LINE__); bad = 1; } } while (0)
349
350 int bad;
351
352 bad = 0;
353
354 // The following are illegal:
355
356 //EFER.SVME is zero.
357 A((vmcb->save.efer & 0x0000000000001000) == 0);
358
359 // CR0.CD is zero and CR0.NW is set
360 A( ((vmcb->save.cr0 & 0x0000000040000000) == 0) &&
361 ((vmcb->save.cr0 & 0x0000000020000000) != 0));
362
363 // CR0[63:32] are not zero.
364 A((vmcb->save.cr0 & 0xFFFFFFFF00000000) == 0xFFFFFFFF00000000);
365
366 // Any MBZ bit of CR3 is set.
367 A((vmcb->save.cr3 & 0xFFF0000000000000) != 0);
368
369 // CR4[63:11] are not zero.
370 A((vmcb->save.cr4 & 0xFFFFFFFFFFFFF800) == 0xFFFFFFFFFFFFF800);
371
372 // DR6[63:32] are not zero.
373 A((vmcb->save.dr6 & 0xFFFFFFFF00000000) == 0xFFFFFFFF00000000);
374
375 // DR7[63:32] are not zero.
376 A((vmcb->save.dr7 & 0xFFFFFFFF00000000) == 0xFFFFFFFF00000000);
377
378 // EFER[63:15] are not zero.
379 A((vmcb->save.efer & 0xFFFFFFFFFFFF8000) == 0xFFFFFFFFFFF8000);
380
381 // EFER.LMA or EFER.LME is non-zero and this processor does not support long mode.
382 //// A((vmcb->save.efer & 0x0000000000000500) != 0);
383
384 // EFER.LME and CR0.PG are both set and CR4.PAE is zero.
385 A( ((vmcb->save.efer & 0x0000000000000100) != 0) &&
386 ((vmcb->save.cr0 & 0x0000000080000000) != 0) &&
387 ((vmcb->save.cr4 & 0x0000000000000020) != 0));
388
389 // EFER.LME and CR0.PG are both non-zero and CR0.PE is zero.
390 A( ((vmcb->save.efer & 0x0000000000000100) != 0) &&
391 ((vmcb->save.cr0 & 0x0000000080000000) != 0) &&
392 ((vmcb->save.cr0 & 0x0000000000000001) == 0));
393
394 // EFER.LME, CR0.PG, CR4.PAE, CS.L, and CS.D are all non-zero.
395 // cs.attrib = concat 55-52 and 47-40 (p372 v2)
396 A( ((vmcb->save.efer & 0x0000000000000100) != 0) &&
397 ((vmcb->save.cr0 & 0x0000000080000000) != 0) &&
398 ((vmcb->save.cr4 & 0x0000000000000020) != 0) &&
399 ((vmcb->save.cs.attrib & 0x0200) != 0) &&
400 ((vmcb->save.cs.attrib & 0x0400) != 0));
401
402 // The VMRUN intercept bit is clear.
403 A((vmcb->control.intercepts & 0x0000000100000000) == 0);
404
405 // The MSR or IOIO intercept tables extend to a physical address that is
406 // greater than or equal to the maximum supported physical address.
407
408 // Illegal event injection (see Section 15.19 on page 391).
409
410 // ASID is equal to zero.
411 A(vmcb->control.guest_asid == 0);
412
413 // VMRUN can load a guest value of CR0 with PE = 0 but PG = 1, a
414 // combination that is otherwise illegal (see Section 15.18).
415
416 // In addition to consistency checks, VMRUN and #VMEXIT canonicalize (i.e.,
417 // sign-extend to 63 bits) all base addresses in the segment registers
418 // that have been loaded.
419
420 return bad;
421
422 #undef A
423 }
424
425 static void
426 fkvm_vcpu_run(struct vcpu *vcpu)
427 {
428 u_int64_t lstar;
429 u_int64_t cstar;
430 u_int64_t star;
431 u_int64_t sfmask;
432
433 u_short fs_selector;
434 u_short gs_selector;
435 u_short ldt_selector;
436
437 unsigned long host_cr2;
438 unsigned long host_dr6;
439 unsigned long host_dr7;
440
441 struct system_segment_descriptor *tss_desc;
442 u_int64_t sel;
443
444 struct vmcb *vmcb;
445
446 printf("begin fkvm_vcpu_run\n");
447
448 vmcb = vcpu->vmcb;
449
450 if (vmrun_assert(vmcb))
451 return;
452
453 tss_desc = (struct system_segment_descriptor*) (&gdt[GPROC0_SEL]);
454 sel = GSEL(GPROC0_SEL, SEL_KPL);
455
456 // printf("GSEL(GPROC0_SEL, SEL_KPL)=0x%" PRIx64 "\n", sel);
457 // print_tss_desc(tss_desc);
458 // print_tss(tss_desc);
459
460 print_vmcb_save_area(vmcb);
461 printf("vcpu->regs[VCPU_REGS_RIP]: 0x%lx\n", vcpu->regs[VCPU_REGS_RIP]);
462 // disable_intr();
463
464 vmcb->save.rax = vcpu->regs[VCPU_REGS_RAX];
465 vmcb->save.rsp = vcpu->regs[VCPU_REGS_RSP];
466 vmcb->save.rip = vcpu->regs[VCPU_REGS_RIP];
467
468 /* meh: kvm has pre_svm_run(svm); */
469
470 vcpu->host_gs_base = rdmsr(MSR_GSBASE);
471 // printf("host_gs_base: 0x%" PRIx64 "\n", vcpu->host_gs_base);
472
473 fs_selector = rfs();
474 gs_selector = rgs();
475 ldt_selector = rldt();
476 // printf("fs selector: %hx\n", fs_selector);
477 // printf("gs selector: %hx\n", gs_selector);
478 // printf("ldt selector: %hx\n", ldt_selector);
479
480 host_cr2 = rcr2();
481
482 host_dr6 = rdr6();
483 host_dr7 = rdr7();
484
485 vmcb->save.cr2 = vcpu->cr2;
486 /* meh: cr3? */
487
488 /* meh: dr7? db_regs? */
489
490 // printf("MSR_STAR: %" PRIx64 "\n", rdmsr(MSR_STAR));
491 // printf("MSR_LSTAR: %" PRIx64 "\n", rdmsr(MSR_LSTAR));
492 // printf("MSR_CSTAR: %" PRIx64 "\n", rdmsr(MSR_CSTAR));
493 // printf("MSR_SF_MASK: %" PRIx64 "\n", rdmsr(MSR_SF_MASK));
494
495 star = rdmsr(MSR_STAR);
496 lstar = rdmsr(MSR_LSTAR);
497 cstar = rdmsr(MSR_CSTAR);
498 sfmask = rdmsr(MSR_SF_MASK);
499
500 printf("CLGI...\n");
501
502 __asm __volatile (SVM_CLGI);
503
504
505 // enable_intr();
506
507 __asm __volatile (
508 "push %%rbp; \n\t"
509 "mov %c[rbx](%[svm]), %%rbx \n\t"
510 "mov %c[rcx](%[svm]), %%rcx \n\t"
511 "mov %c[rdx](%[svm]), %%rdx \n\t"
512 "mov %c[rsi](%[svm]), %%rsi \n\t"
513 "mov %c[rdi](%[svm]), %%rdi \n\t"
514 "mov %c[rbp](%[svm]), %%rbp \n\t"
515 "mov %c[r8](%[svm]), %%r8 \n\t"
516 "mov %c[r9](%[svm]), %%r9 \n\t"
517 "mov %c[r10](%[svm]), %%r10 \n\t"
518 "mov %c[r11](%[svm]), %%r11 \n\t"
519 "mov %c[r12](%[svm]), %%r12 \n\t"
520 "mov %c[r13](%[svm]), %%r13 \n\t"
521 "mov %c[r14](%[svm]), %%r14 \n\t"
522 "mov %c[r15](%[svm]), %%r15 \n\t"
523
524 /* Enter guest mode */
525 "push %%rax \n\t"
526 "mov %c[vmcb](%[svm]), %%rax \n\t"
527 SVM_VMLOAD "\n\t"
528 SVM_VMRUN "\n\t"
529 SVM_VMSAVE "\n\t"
530 "pop %%rax \n\t"
531
532 /* Save guest registers, load host registers */
533 "mov %%rbx, %c[rbx](%[svm]) \n\t"
534 "mov %%rcx, %c[rcx](%[svm]) \n\t"
535 "mov %%rdx, %c[rdx](%[svm]) \n\t"
536 "mov %%rsi, %c[rsi](%[svm]) \n\t"
537 "mov %%rdi, %c[rdi](%[svm]) \n\t"
538 "mov %%rbp, %c[rbp](%[svm]) \n\t"
539 "mov %%r8, %c[r8](%[svm]) \n\t"
540 "mov %%r9, %c[r9](%[svm]) \n\t"
541 "mov %%r10, %c[r10](%[svm]) \n\t"
542 "mov %%r11, %c[r11](%[svm]) \n\t"
543 "mov %%r12, %c[r12](%[svm]) \n\t"
544 "mov %%r13, %c[r13](%[svm]) \n\t"
545 "mov %%r14, %c[r14](%[svm]) \n\t"
546 "mov %%r15, %c[r15](%[svm]) \n\t"
547 "pop %%rbp"
548 :
549 : [svm]"a"(vcpu),
550 [vmcb]"i"(offsetof(struct vcpu, vmcb_pa)),
551 [rbx]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RBX])),
552 [rcx]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RCX])),
553 [rdx]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RDX])),
554 [rsi]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RSI])),
555 [rdi]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RDI])),
556 [rbp]"i"(offsetof(struct vcpu, regs[VCPU_REGS_RBP])),
557 [r8 ]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R8 ])),
558 [r9 ]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R9 ])),
559 [r10]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R10])),
560 [r11]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R11])),
561 [r12]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R12])),
562 [r13]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R13])),
563 [r14]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R14])),
564 [r15]"i"(offsetof(struct vcpu, regs[VCPU_REGS_R15]))
565 : "cc", "memory",
566 "rbx", "rcx", "rdx", "rsi", "rdi",
567 "r8", "r9", "r10", "r11" , "r12", "r13", "r14", "r15"
568 );
569
570
571 /* meh: dr7? db_regs? */
572
573 vcpu->cr2 = vmcb->save.cr2;
574
575 vcpu->regs[VCPU_REGS_RAX] = vmcb->save.rax;
576 vcpu->regs[VCPU_REGS_RSP] = vmcb->save.rsp;
577 vcpu->regs[VCPU_REGS_RIP] = vmcb->save.rip;
578
579 load_dr6(host_dr6);
580 load_dr7(host_dr7);
581
582 load_cr2(host_cr2);
583
584 load_fs(fs_selector);
585 load_gs(gs_selector);
586 lldt(ldt_selector);
587
588 wrmsr(MSR_GSBASE, vcpu->host_gs_base);
589
590 tss_desc->sd_type = SDT_SYSTSS;
591 ltr(sel);
592
593 wrmsr(MSR_STAR, star);
594 wrmsr(MSR_LSTAR, lstar);
595 wrmsr(MSR_CSTAR, cstar);
596 wrmsr(MSR_SF_MASK, sfmask);
597
598 // disable_intr();
599
600 __asm __volatile (SVM_STGI);
601
602 printf("STGI\n");
603
604 printf("exit_code: %" PRIx64 "\n", vmcb->control.exit_code);
605
606 // print_tss_desc(tss_desc);
607 // print_tss(tss_desc);
608
609 print_vmcb_save_area(vmcb);
610
611 // enable_intr();
612
613 /* meh: next_rip */
614 }
615
616 static void
617 _fkvm_init_seg(struct vmcb_seg *seg, uint16_t attrib)
618 {
619 seg->selector = 0;
620 seg->attrib = VMCB_SELECTOR_P_MASK | attrib;
621 seg->limit = 0xffff;
622 seg->base = 0;
623 }
624
625 static inline void
626 fkvm_init_seg(struct vmcb_seg *seg)
627 {
628 _fkvm_init_seg(seg, VMCB_SELECTOR_S_MASK | VMCB_SELECTOR_WRITE_MASK);
629 }
630
631 static inline void
632 fkvm_init_sys_seg(struct vmcb_seg *seg, uint16_t attrib)
633 {
634 _fkvm_init_seg(seg, attrib);
635 }
636
637 static void*
638 fkvm_iopm_alloc(void)
639 {
640 return contigmalloc(IOPM_SIZE, M_DEVBUF, 0, 0, -1UL, PAGE_SIZE, 0);
641 }
642
643 static void
644 fkvm_iopm_init(void *iopm)
645 {
646 memset(iopm, 0xff, IOPM_SIZE); /* TODO: we may want to allow access to PC debug port */
647 }
648
649 static void
650 fkvm_iopm_free(void *iopm)
651 {
652 contigfree(iopm, IOPM_SIZE, M_DEVBUF);
653 }
654
655 static void*
656 fkvm_msrpm_alloc(void)
657 {
658 return contigmalloc(MSRPM_SIZE, M_DEVBUF, 0, 0, -1UL, PAGE_SIZE, 0);
659 }
660
661 static void
662 fkvm_msrpm_init(void *msrpm)
663 {
664 memset(msrpm, 0xff, MSRPM_SIZE); /* TODO: we may want to allow some MSR accesses */
665 }
666
667 static void
668 fkvm_msrpm_free(void *msrpm)
669 {
670 contigfree(msrpm, MSRPM_SIZE, M_DEVBUF);
671 }
672
673 static void*
674 fkvm_hsave_area_alloc(void)
675 {
676 return contigmalloc(PAGE_SIZE, M_DEVBUF, 0, 0, -1UL, PAGE_SIZE, 0);
677 }
678
679 static void
680 fkvm_hsave_area_init(void *hsave_area)
681 {
682 }
683
684 static void
685 fkvm_hsave_area_free(void *hsave_area)
686 {
687 contigfree(hsave_area, PAGE_SIZE, M_DEVBUF);
688 }
689
690 static struct vmspace*
691 fkvm_make_vmspace(void)
692 {
693 struct vmspace *sp;
694
695 sp = vmspace_alloc(0, 0xffffffffffffffff);
696 if (sp == NULL) {
697 printf("vmspace_alloc failed\n");
698 return NULL;
699 }
700
701 return sp;
702 }
703
704 static void
705 fkvm_destroy_vmspace(struct vmspace* sp)
706 {
707 vmspace_free(sp);
708 }
709
710 static struct vmcb*
711 fkvm_vmcb_alloc(void)
712 {
713 return contigmalloc(PAGE_SIZE, M_DEVBUF, M_ZERO, 0, -1UL,
714 PAGE_SIZE, 0);
715 }
716
717 static void
718 fkvm_vmcb_init(struct vmcb *vmcb)
719 {
720 struct vmcb_control_area *control = &vmcb->control;
721 struct vmcb_save_area *save = &vmcb->save;
722
723 control->intercept_cr_reads = INTERCEPT_CR4_MASK;
724
725 control->intercept_cr_writes = INTERCEPT_CR4_MASK |
726 INTERCEPT_CR8_MASK;
727
728 control->intercept_dr_reads = INTERCEPT_DR0_MASK |
729 INTERCEPT_DR1_MASK |
730 INTERCEPT_DR2_MASK |
731 INTERCEPT_DR3_MASK;
732
733 control->intercept_dr_writes = INTERCEPT_DR0_MASK |
734 INTERCEPT_DR1_MASK |
735 INTERCEPT_DR2_MASK |
736 INTERCEPT_DR3_MASK |
737 INTERCEPT_DR5_MASK |
738 INTERCEPT_DR7_MASK;
739
740 control->intercept_exceptions = (1 << IDT_UD) | // Invalid Opcode
741 (1 << IDT_MC); // Machine Check
742
743 control->intercepts = INTERCEPT_INTR |
744 INTERCEPT_NMI |
745 INTERCEPT_SMI |
746 INTERCEPT_CPUID |
747 INTERCEPT_INVD |
748 INTERCEPT_HLT |
749 INTERCEPT_INVLPGA |
750 INTERCEPT_IOIO_PROT |
751 INTERCEPT_MSR_PROT |
752 INTERCEPT_SHUTDOWN |
753 INTERCEPT_VMRUN |
754 INTERCEPT_VMMCALL |
755 INTERCEPT_VMLOAD |
756 INTERCEPT_VMSAVE |
757 INTERCEPT_STGI |
758 INTERCEPT_CLGI |
759 INTERCEPT_SKINIT |
760 INTERCEPT_WBINVD |
761 INTERCEPT_MONITOR |
762 INTERCEPT_MWAIT_UNCOND;
763
764 control->iopm_base_pa = vtophys(iopm);
765 control->msrpm_base_pa = vtophys(msrpm);
766 control->tsc_offset = 0;
767
768 /* TODO: remove this once we assign asid's to distinct VM's */
769 control->guest_asid = 1;
770 control->tlb_control = VMCB_TLB_CONTROL_FLUSH_ALL;
771
772 /* let v_tpr default to 0 */
773 /* let v_irq default to 0 */
774 /* let v_intr default to 0 */
775
776 control->v_intr_masking = 1;
777
778 /* let v_intr_vector default to 0 */
779 /* let intr_shadow default to 0 */
780 /* let exit_code, exit_info_1, exit_info_2, exit_int_info,
781 exit_int_info_err_code default to 0 */
782
783 control->nested_ctl = 1;
784
785 /* let event_inj default to 0 */
786
787 // (nested_cr3 is later)
788
789 /* let lbr_virt_enable default to 0 */
790
791
792 fkvm_init_seg(&save->ds);
793 fkvm_init_seg(&save->es);
794 fkvm_init_seg(&save->fs);
795 fkvm_init_seg(&save->gs);
796 fkvm_init_seg(&save->ss);
797
798 _fkvm_init_seg(&save->cs, VMCB_SELECTOR_READ_MASK | VMCB_SELECTOR_S_MASK |
799 VMCB_SELECTOR_CODE_MASK);
800 save->cs.selector = 0xf000;
801 save->cs.base = 0xffff0000;
802
803 save->gdtr.limit = 0xffff;
804 save->idtr.limit = 0xffff;
805
806 fkvm_init_sys_seg(&save->ldtr, SDT_SYSLDT);
807 fkvm_init_sys_seg(&save->tr, SDT_SYS286BSY);
808
809 save->g_pat = PAT_VALUE(PAT_WRITE_BACK, 0) | PAT_VALUE(PAT_WRITE_THROUGH, 1) |
810 PAT_VALUE(PAT_UNCACHED, 2) | PAT_VALUE(PAT_UNCACHEABLE, 3) |
811 PAT_VALUE(PAT_WRITE_BACK, 4) | PAT_VALUE(PAT_WRITE_THROUGH, 5) |
812 PAT_VALUE(PAT_UNCACHED, 6) | PAT_VALUE(PAT_UNCACHEABLE, 7);
813
814 /* CR0 = 6000_0010h at boot */
815 save->cr0 = CR0_ET | CR0_NW | CR0_CD;
816 save->dr6 = 0xffff0ff0;
817 save->dr7 = 0x400;
818 save->rflags = 2;
819 save->rip = 0x0000fff0;
820
821 save->efer = EFER_SVME;
822 }
823
824 static void
825 fkvm_vmcb_free(struct vmcb *vmcb)
826 {
827 contigfree(vmcb, PAGE_SIZE, M_DEVBUF);
828 }
829
830 static struct vcpu*
831 fkvm_vcpu_create(struct guestvm *guest_vm)
832 {
833 struct vcpu *vcpu;
834 vcpu = malloc(sizeof(struct vcpu), M_DEVBUF, M_WAITOK|M_ZERO);
835
836 vcpu->vmcb = fkvm_vmcb_alloc();
837 vcpu->vmcb_pa = vtophys(vcpu->vmcb);
838 printf("vmcb = 0x%p\n", vcpu->vmcb);
839 printf("vcpu->vmcb_pa = 0x%lx\n", vcpu->vmcb_pa);
840
841 fkvm_vmcb_init(vcpu->vmcb);
842 vcpu->vmcb->control.nested_cr3 = guest_vm->nested_cr3;
843 vcpu->regs[VCPU_REGS_RIP] = vcpu->vmcb->save.rip;
844
845 vcpu->guest_vm = guest_vm;
846
847 return vcpu;
848 }
849
850 static void
851 fkvm_vcpu_destroy(struct vcpu *vcpu)
852 {
853 fkvm_vmcb_free(vcpu->vmcb);
854 free(vcpu, M_DEVBUF);
855 }
856
857 static struct guestvm*
858 fkvm_guestvm_alloc(void)
859 {
860 return malloc(sizeof(struct guestvm), M_DEVBUF, M_WAITOK|M_ZERO);
861 }
862
863 static void
864 fkvm_guestvm_free(struct guestvm* guest_vm)
865 {
866 free(guest_vm, M_DEVBUF);
867 }
868
869 static void
870 fkvm_guestvm_add_vcpu(struct guestvm *guest_vm, struct vcpu *vcpu)
871 {
872 guest_vm->vcpus[guest_vm->nr_vcpus] = vcpu;
873 guest_vm->nr_vcpus++; /* TODO: Probably not safe to increment */
874 /* How about a lock to protect all of this? */
875 }
876
877
878 int
879 fkvm_userpoke(struct thread *td, struct fkvm_userpoke_args *uap)
880 {
881 printf("fkvm_userpoke\n");
882 return ENOSYS;
883 }
884
885 static int
886 fkvm_mem_has_entry(vm_map_entry_t expected_entry, vm_map_t vm_map, vm_offset_t vaddr)
887 {
888 vm_map_entry_t lookup_entry;
889 vm_object_t throwaway_object;
890 vm_pindex_t throwaway_pindex;
891 vm_prot_t throwaway_prot;
892 boolean_t throwaway_wired;
893 int error;
894
895 error = vm_map_lookup(&vm_map, /* IN/OUT */
896 vaddr,
897 VM_PROT_READ|VM_PROT_WRITE,
898 &lookup_entry, /* OUT */
899 &throwaway_object, /* OUT */
900 &throwaway_pindex, /* OUT */
901 &throwaway_prot, /* OUT */
902 &throwaway_wired); /* OUT */
903 if (error != KERN_SUCCESS)
904 return 0;
905 vm_map_lookup_done(vm_map, lookup_entry);
906 return (lookup_entry == expected_entry);
907 }
908
909 static int
910 fkvm_guest_check_range(struct guestvm *guest_vm, uint64_t start, uint64_t end)
911 {
912 vm_map_t guest_vm_map;
913 vm_map_entry_t lookup_entry;
914 vm_object_t throwaway_object;
915 vm_pindex_t throwaway_pindex;
916 vm_prot_t throwaway_prot;
917 boolean_t throwaway_wired;
918 int ret;
919 int error;
920
921 guest_vm_map = &guest_vm->sp->vm_map;
922
923 error = vm_map_lookup(&guest_vm_map, /* IN/OUT */
924 start,
925 VM_PROT_READ|VM_PROT_WRITE,
926 &lookup_entry, /* OUT */
927 &throwaway_object, /* OUT */
928 &throwaway_pindex, /* OUT */
929 &throwaway_prot, /* OUT */
930 &throwaway_wired); /* OUT */
931 if (error != KERN_SUCCESS)
932 return EFAULT;
933
934 if (fkvm_mem_has_entry(lookup_entry, guest_vm_map, end))
935 ret = 0;
936 else
937 ret = EFAULT;
938
939 vm_map_lookup_done(guest_vm_map, lookup_entry);
940 return ret;
941 }
942
943 /* System Calls */
944 /* This function can only be called with multiples of page sizes */
945 /* vaddr as NULL overloads to fkvm_guest_check_range */
946 int
947 fkvm_set_user_mem_region(struct thread *td, struct fkvm_set_user_mem_region_args *uap)
948 {
949 struct guestvm *guest_vm = PROC_GET_GUESTVM(td->td_proc);
950
951 vm_offset_t start;
952 vm_offset_t end;
953
954 struct vmspace *user_vm_space;
955 vm_map_t user_vm_map;
956
957 vm_object_t vm_object;
958 vm_pindex_t vm_object_pindex;
959 vm_ooffset_t vm_object_offset;
960 vm_prot_t throwaway_prot;
961 boolean_t throwaway_wired;
962 vm_map_entry_t lookup_entry;
963
964 int error;
965
966 start = uap->guest_pa;
967 end = uap->guest_pa + uap->size - 1;
968 printf("start: 0x%" PRIx64 " bytes\n", start);
969 printf("end: 0x%" PRIx64 " bytes\n", end);
970
971 if (uap->vaddr == 0)
972 return fkvm_guest_check_range(guest_vm, start, end);
973
974 user_vm_space = td->td_proc->p_vmspace;
975 user_vm_map = &user_vm_space->vm_map;
976 printf("user vm space: %p\n", user_vm_space);
977 printf("user vm map: %p\n", user_vm_map);
978
979 error = vm_map_lookup(&user_vm_map, /* IN/OUT */
980 uap->vaddr,
981 VM_PROT_READ|VM_PROT_WRITE,
982 &lookup_entry, /* OUT */
983 &vm_object, /* OUT */
984 &vm_object_pindex, /* OUT */
985 &throwaway_prot, /* OUT */
986 &throwaway_wired); /* OUT */
987 if (error != KERN_SUCCESS) {
988 printf("vm_map_lookup failed: %d\n", error);
989 return EFAULT;
990 }
991
992 if (!fkvm_mem_has_entry(lookup_entry, user_vm_map, uap->vaddr + uap->size)) {
993 printf("end of range not contained in same vm map entry as start\n");
994 return EFAULT;
995 }
996
997 printf("vm object: %p\n", vm_object);
998 printf(" size: %d pages\n", (int) vm_object->size);
999
1000 vm_object_offset = IDX_TO_OFF(vm_object_pindex);
1001 printf("vm_ooffset: 0x%" PRIx64 "\n", vm_object_offset);
1002
1003 vm_object_reference(vm_object); // TODO: this might be a mem leak
1004
1005 vm_map_lookup_done(user_vm_map, lookup_entry);
1006
1007 error = vm_map_insert(&guest_vm->sp->vm_map,
1008 vm_object,
1009 vm_object_offset,
1010 start,
1011 end,
1012 VM_PROT_ALL, VM_PROT_ALL,
1013 0);
1014 if (error != KERN_SUCCESS) {
1015 printf("vm_map_insert failed: %d\n", error);
1016 switch (error) {
1017 case KERN_INVALID_ADDRESS:
1018 return EINVAL;
1019 case KERN_NO_SPACE:
1020 return ENOMEM;
1021 default:
1022 return 1;
1023 }
1024 }
1025
1026 return 0;
1027 }
1028
1029 int
1030 fkvm_unset_user_mem_region(struct thread *td, struct fkvm_unset_user_mem_region_args *uap)
1031 {
1032 struct guestvm *guest_vm = PROC_GET_GUESTVM(td->td_proc);
1033
1034 vm_offset_t start;
1035 vm_offset_t end;
1036
1037 vm_map_t guest_vm_map;
1038
1039 int error;
1040
1041 start = uap->guest_pa;
1042 end = uap->guest_pa + uap->size - 1;
1043 printf("start: 0x%" PRIx64 " bytes\n", start);
1044 printf("end: 0x%" PRIx64 " bytes\n", end);
1045
1046 guest_vm_map = &guest_vm->sp->vm_map;
1047
1048 error = vm_map_remove(guest_vm_map, start, end);
1049 if (error != KERN_SUCCESS)
1050 return -1;
1051
1052 return 0;
1053 }
1054
1055 int
1056 fkvm_create_vm(struct thread *td, struct fkvm_create_vm_args *uap)
1057 {
1058 struct vcpu *vcpu;
1059 struct guestvm *guest_vm;
1060
1061 printf("SYSCALL : fkvm_create_vm\n");
1062
1063 /* Allocate Guest VM */
1064 guest_vm = fkvm_guestvm_alloc();
1065
1066 /* Set up the vm address space */
1067 guest_vm->sp = fkvm_make_vmspace();
1068 if (guest_vm->sp == NULL) {
1069 fkvm_guestvm_free(guest_vm);
1070 return ENOMEM;
1071 }
1072 guest_vm->nested_cr3 = vtophys(vmspace_pmap(guest_vm->sp)->pm_pml4);
1073
1074 printf("guest:\n");
1075 printf(" vm space: %p\n", guest_vm->sp);
1076 printf(" vm map: %p\n", &guest_vm->sp->vm_map);
1077 printf(" ncr3: 0x%" PRIx64 "\n", guest_vm->nested_cr3);
1078
1079 /* Allocate VCPU0 */
1080 vcpu = fkvm_vcpu_create(guest_vm);
1081 fkvm_guestvm_add_vcpu(guest_vm, vcpu);
1082
1083 PROC_SET_GUESTVM(td->td_proc, guest_vm);
1084 TD_SET_VCPU(td, vcpu);
1085
1086 printf("fkvm_create_vm done\n");
1087 return 0;
1088 }
1089
1090 int
1091 fkvm_destroy_vm(struct thread *td, struct fkvm_destroy_vm_args *uap)
1092 {
1093 struct guestvm *guest_vm = PROC_GET_GUESTVM(td->td_proc);
1094
1095 /* Destroy the VCPUs */
1096 while (guest_vm->nr_vcpus > 0) {
1097 guest_vm->nr_vcpus--;
1098 fkvm_vcpu_destroy(guest_vm->vcpus[guest_vm->nr_vcpus]);
1099 guest_vm->vcpus[guest_vm->nr_vcpus] = NULL;
1100 }
1101
1102 /* Destroy the vmspace */
1103 fkvm_destroy_vmspace(guest_vm->sp);
1104
1105 /* Destroy the Guest VM itself */
1106 fkvm_guestvm_free(guest_vm);
1107
1108 return 0;
1109 }
1110
1111 int
1112 fkvm_vm_run(struct thread *td, struct fkvm_vm_run_args *uap)
1113 {
1114 struct vcpu *vcpu = TD_GET_VCPU(td);
1115 struct guestvm *guest_vm = vcpu->guest_vm;
1116 struct vmcb *vmcb = vcpu->vmcb;
1117 int ret = 0;
1118
1119 fkvm_vcpu_run(vcpu);
1120
1121 switch (vmcb->control.exit_code) {
1122
1123 case VMCB_EXIT_EXCP_BASE ... (VMCB_EXIT_EXCP_BASE + 31): {
1124 int excp_vector;
1125
1126 excp_vector = vmcb->control.exit_code - VMCB_EXIT_EXCP_BASE;
1127
1128 printf("VMCB_EXIT_EXCP_BASE, exception vector: 0x%x\n",
1129 excp_vector);
1130 ret = ENOSYS;
1131 break;
1132 }
1133
1134 case VMCB_EXIT_INTR: {
1135 printf("VMCB_EXIT_INTR - nothing to do\n");
1136 break;
1137 }
1138
1139 case VMCB_EXIT_NPF: {
1140 /* EXITINFO1 contains fault error code */
1141 /* EXITINFO2 contains the guest physical address causing the fault. */
1142
1143 u_int64_t fault_code;
1144 u_int64_t fault_gpa;
1145
1146 vm_prot_t fault_type;
1147 int fault_flags;
1148 int rc;
1149
1150 fault_code = vmcb->control.exit_info_1;
1151 fault_gpa = vmcb->control.exit_info_2;
1152
1153 printf("VMCB_EXIT_NPF:\n");
1154 printf("gpa=0x%" PRIx64 "\n", fault_gpa);
1155 printf("fault code=0x%" PRIx64 " [P=%x, R/W=%x, U/S=%x, I/D=%x]\n",
1156 fault_code,
1157 (fault_code & PGEX_P) != 0,
1158 (fault_code & PGEX_W) != 0,
1159 (fault_code & PGEX_U) != 0,
1160 (fault_code & PGEX_I) != 0);
1161
1162 if (fault_code & PGEX_W)
1163 fault_type = VM_PROT_WRITE;
1164 else if (fault_code & PGEX_I)
1165 fault_type = VM_PROT_EXECUTE;
1166 else
1167 fault_type = VM_PROT_READ;
1168
1169 fault_flags = 0; /* TODO: is that right? */
1170 rc = vm_fault(&guest_vm->sp->vm_map, (fault_gpa & (~PAGE_MASK)), fault_type, fault_flags);
1171 if (rc != KERN_SUCCESS) {
1172 printf("vm_fault failed: %d\n", rc);
1173 }
1174 break;
1175 }
1176 default:
1177 printf("Unhandled vmexit:\n"
1178 " code: 0x%" PRIx64 "\n"
1179 " info1: 0x%" PRIx64 "\n"
1180 " info2: 0x%" PRIx64 "\n",
1181 vmcb->control.exit_code,
1182 vmcb->control.exit_info_1,
1183 vmcb->control.exit_info_2);
1184 print_vmcb(vmcb);
1185 ret = ENOSYS;
1186 }
1187
1188 printf("\n\n");
1189
1190 return ret;
1191 }
1192
1193 int
1194 fkvm_create_vcpu(struct thread *td, struct fkvm_create_vcpu_args *uap)
1195 {
1196 struct guestvm *guest_vm = PROC_GET_GUESTVM(td->td_proc);
1197 struct vcpu *vcpu;
1198
1199 /* Allocate VCPU */
1200 vcpu = fkvm_vcpu_create(guest_vm);
1201 fkvm_guestvm_add_vcpu(guest_vm, vcpu);
1202
1203 TD_SET_VCPU(td, vcpu);
1204 return 0;
1205 }
1206
1207 static void
1208 fkvm_load(void *unused)
1209 {
1210 u_int64_t efer;
1211
1212 printf("fkvm_load\n");
1213 printf("sizeof(struct vmcb) = %" PRIx64 "\n", sizeof(struct vmcb));
1214
1215 hsave_area = NULL;
1216 iopm = NULL;
1217 msrpm = NULL;
1218
1219 /* TODO: check for the presense of extensions */
1220
1221 /* allocate structures */
1222 hsave_area = fkvm_hsave_area_alloc();
1223 iopm = fkvm_iopm_alloc();
1224 msrpm = fkvm_msrpm_alloc();
1225
1226 /* Initialize structures */
1227 fkvm_hsave_area_init(hsave_area);
1228 fkvm_iopm_init(iopm);
1229 fkvm_msrpm_init(msrpm);
1230
1231 /* Enable SVM in EFER */
1232 efer = rdmsr(MSR_EFER);
1233 printf("EFER = %" PRIx64 "\n", efer);
1234 wrmsr(MSR_EFER, efer | EFER_SVME);
1235 efer = rdmsr(MSR_EFER);
1236 printf("new EFER = %" PRIx64 "\n", efer);
1237
1238 /* Write Host save address in MSR_VM_HSAVE_PA */
1239 wrmsr(MSR_VM_HSAVE_PA, vtophys(hsave_area));
1240 }
1241 SYSINIT(fkvm, SI_SUB_PSEUDO, SI_ORDER_MIDDLE, fkvm_load, NULL);
1242
1243 static void
1244 fkvm_unload(void *unused)
1245 {
1246 printf("fkvm_unload\n");
1247
1248 if (msrpm != NULL) {
1249 fkvm_msrpm_free(iopm);
1250 msrpm = NULL;
1251 }
1252 if (iopm != NULL) {
1253 fkvm_iopm_free(iopm);
1254 iopm = NULL;
1255 }
1256 if (hsave_area != NULL) {
1257 fkvm_hsave_area_free(hsave_area);
1258 hsave_area = NULL;
1259 }
1260 }
1261 SYSUNINIT(fkvm, SI_SUB_PSEUDO, SI_ORDER_MIDDLE, fkvm_unload, NULL);
FreeBSD kvm: FreeBSD sources