xref: /freebsd/sys/amd64/vmm/x86.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright (c) 2011 NetApp, Inc.
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 NETAPP, INC ``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 NETAPP, INC 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  * $FreeBSD$
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/pcpu.h>
34 #include <sys/systm.h>
35 #include <sys/sysctl.h>
36 
37 #include <machine/clock.h>
38 #include <machine/cpufunc.h>
39 #include <machine/md_var.h>
40 #include <machine/segments.h>
41 #include <machine/specialreg.h>
42 
43 #include <machine/vmm.h>
44 
45 #include "vmm_host.h"
46 #include "vmm_ktr.h"
47 #include "vmm_util.h"
48 #include "x86.h"
49 
50 SYSCTL_DECL(_hw_vmm);
51 static SYSCTL_NODE(_hw_vmm, OID_AUTO, topology, CTLFLAG_RD, 0, NULL);
52 
53 #define	CPUID_VM_HIGH		0x40000000
54 
55 static const char bhyve_id[12] = "bhyve bhyve ";
56 
57 static uint64_t bhyve_xcpuids;
58 SYSCTL_ULONG(_hw_vmm, OID_AUTO, bhyve_xcpuids, CTLFLAG_RW, &bhyve_xcpuids, 0,
59     "Number of times an unknown cpuid leaf was accessed");
60 
61 /*
62  * The default CPU topology is a single thread per package.
63  */
64 static u_int threads_per_core = 1;
65 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, threads_per_core, CTLFLAG_RDTUN,
66     &threads_per_core, 0, NULL);
67 
68 static u_int cores_per_package = 1;
69 SYSCTL_UINT(_hw_vmm_topology, OID_AUTO, cores_per_package, CTLFLAG_RDTUN,
70     &cores_per_package, 0, NULL);
71 
72 static int cpuid_leaf_b = 1;
73 SYSCTL_INT(_hw_vmm_topology, OID_AUTO, cpuid_leaf_b, CTLFLAG_RDTUN,
74     &cpuid_leaf_b, 0, NULL);
75 
76 /*
77  * Round up to the next power of two, if necessary, and then take log2.
78  * Returns -1 if argument is zero.
79  */
80 static __inline int
81 log2(u_int x)
82 {
83 
84 	return (fls(x << (1 - powerof2(x))) - 1);
85 }
86 
87 int
88 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
89 		  uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
90 {
91 	const struct xsave_limits *limits;
92 	uint64_t cr4;
93 	int error, enable_invpcid, level, width, x2apic_id;
94 	unsigned int func, regs[4], logical_cpus;
95 	enum x2apic_state x2apic_state;
96 
97 	VCPU_CTR2(vm, vcpu_id, "cpuid %#x,%#x", *eax, *ecx);
98 
99 	/*
100 	 * Requests for invalid CPUID levels should map to the highest
101 	 * available level instead.
102 	 */
103 	if (cpu_exthigh != 0 && *eax >= 0x80000000) {
104 		if (*eax > cpu_exthigh)
105 			*eax = cpu_exthigh;
106 	} else if (*eax >= 0x40000000) {
107 		if (*eax > CPUID_VM_HIGH)
108 			*eax = CPUID_VM_HIGH;
109 	} else if (*eax > cpu_high) {
110 		*eax = cpu_high;
111 	}
112 
113 	func = *eax;
114 
115 	/*
116 	 * In general the approach used for CPU topology is to
117 	 * advertise a flat topology where all CPUs are packages with
118 	 * no multi-core or SMT.
119 	 */
120 	switch (func) {
121 		/*
122 		 * Pass these through to the guest
123 		 */
124 		case CPUID_0000_0000:
125 		case CPUID_0000_0002:
126 		case CPUID_0000_0003:
127 		case CPUID_8000_0000:
128 		case CPUID_8000_0002:
129 		case CPUID_8000_0003:
130 		case CPUID_8000_0004:
131 		case CPUID_8000_0006:
132 			cpuid_count(*eax, *ecx, regs);
133 			break;
134 		case CPUID_8000_0008:
135 			cpuid_count(*eax, *ecx, regs);
136 			if (vmm_is_amd()) {
137 				/*
138 				 * XXX this might appear silly because AMD
139 				 * cpus don't have threads.
140 				 *
141 				 * However this matches the logical cpus as
142 				 * advertised by leaf 0x1 and will work even
143 				 * if the 'threads_per_core' tunable is set
144 				 * incorrectly on an AMD host.
145 				 */
146 				logical_cpus = threads_per_core *
147 				    cores_per_package;
148 				regs[2] = logical_cpus - 1;
149 			}
150 			break;
151 
152 		case CPUID_8000_0001:
153 			cpuid_count(*eax, *ecx, regs);
154 
155 			/*
156 			 * Hide SVM and Topology Extension features from guest.
157 			 */
158 			regs[2] &= ~(AMDID2_SVM | AMDID2_TOPOLOGY);
159 
160 			/*
161 			 * Don't advertise extended performance counter MSRs
162 			 * to the guest.
163 			 */
164 			regs[2] &= ~AMDID2_PCXC;
165 			regs[2] &= ~AMDID2_PNXC;
166 			regs[2] &= ~AMDID2_PTSCEL2I;
167 
168 			/*
169 			 * Don't advertise Instruction Based Sampling feature.
170 			 */
171 			regs[2] &= ~AMDID2_IBS;
172 
173 			/* NodeID MSR not available */
174 			regs[2] &= ~AMDID2_NODE_ID;
175 
176 			/* Don't advertise the OS visible workaround feature */
177 			regs[2] &= ~AMDID2_OSVW;
178 
179 			/* Hide mwaitx/monitorx capability from the guest */
180 			regs[2] &= ~AMDID2_MWAITX;
181 
182 			/*
183 			 * Hide rdtscp/ia32_tsc_aux until we know how
184 			 * to deal with them.
185 			 */
186 			regs[3] &= ~AMDID_RDTSCP;
187 			break;
188 
189 		case CPUID_8000_0007:
190 			/*
191 			 * AMD uses this leaf to advertise the processor's
192 			 * power monitoring and RAS capabilities. These
193 			 * features are hardware-specific and exposing
194 			 * them to a guest doesn't make a lot of sense.
195 			 *
196 			 * Intel uses this leaf only to advertise the
197 			 * "Invariant TSC" feature with all other bits
198 			 * being reserved (set to zero).
199 			 */
200 			regs[0] = 0;
201 			regs[1] = 0;
202 			regs[2] = 0;
203 			regs[3] = 0;
204 
205 			/*
206 			 * "Invariant TSC" can be advertised to the guest if:
207 			 * - host TSC frequency is invariant
208 			 * - host TSCs are synchronized across physical cpus
209 			 *
210 			 * XXX This still falls short because the vcpu
211 			 * can observe the TSC moving backwards as it
212 			 * migrates across physical cpus. But at least
213 			 * it should discourage the guest from using the
214 			 * TSC to keep track of time.
215 			 */
216 			if (tsc_is_invariant && smp_tsc)
217 				regs[3] |= AMDPM_TSC_INVARIANT;
218 			break;
219 
220 		case CPUID_0000_0001:
221 			do_cpuid(1, regs);
222 
223 			error = vm_get_x2apic_state(vm, vcpu_id, &x2apic_state);
224 			if (error) {
225 				panic("x86_emulate_cpuid: error %d "
226 				      "fetching x2apic state", error);
227 			}
228 
229 			/*
230 			 * Override the APIC ID only in ebx
231 			 */
232 			regs[1] &= ~(CPUID_LOCAL_APIC_ID);
233 			regs[1] |= (vcpu_id << CPUID_0000_0001_APICID_SHIFT);
234 
235 			/*
236 			 * Don't expose VMX, SpeedStep, TME or SMX capability.
237 			 * Advertise x2APIC capability and Hypervisor guest.
238 			 */
239 			regs[2] &= ~(CPUID2_VMX | CPUID2_EST | CPUID2_TM2);
240 			regs[2] &= ~(CPUID2_SMX);
241 
242 			regs[2] |= CPUID2_HV;
243 
244 			if (x2apic_state != X2APIC_DISABLED)
245 				regs[2] |= CPUID2_X2APIC;
246 			else
247 				regs[2] &= ~CPUID2_X2APIC;
248 
249 			/*
250 			 * Only advertise CPUID2_XSAVE in the guest if
251 			 * the host is using XSAVE.
252 			 */
253 			if (!(regs[2] & CPUID2_OSXSAVE))
254 				regs[2] &= ~CPUID2_XSAVE;
255 
256 			/*
257 			 * If CPUID2_XSAVE is being advertised and the
258 			 * guest has set CR4_XSAVE, set
259 			 * CPUID2_OSXSAVE.
260 			 */
261 			regs[2] &= ~CPUID2_OSXSAVE;
262 			if (regs[2] & CPUID2_XSAVE) {
263 				error = vm_get_register(vm, vcpu_id,
264 				    VM_REG_GUEST_CR4, &cr4);
265 				if (error)
266 					panic("x86_emulate_cpuid: error %d "
267 					      "fetching %%cr4", error);
268 				if (cr4 & CR4_XSAVE)
269 					regs[2] |= CPUID2_OSXSAVE;
270 			}
271 
272 			/*
273 			 * Hide monitor/mwait until we know how to deal with
274 			 * these instructions.
275 			 */
276 			regs[2] &= ~CPUID2_MON;
277 
278                         /*
279 			 * Hide the performance and debug features.
280 			 */
281 			regs[2] &= ~CPUID2_PDCM;
282 
283 			/*
284 			 * No TSC deadline support in the APIC yet
285 			 */
286 			regs[2] &= ~CPUID2_TSCDLT;
287 
288 			/*
289 			 * Hide thermal monitoring
290 			 */
291 			regs[3] &= ~(CPUID_ACPI | CPUID_TM);
292 
293 			/*
294 			 * Hide the debug store capability.
295 			 */
296 			regs[3] &= ~CPUID_DS;
297 
298 			/*
299 			 * Advertise the Machine Check and MTRR capability.
300 			 *
301 			 * Some guest OSes (e.g. Windows) will not boot if
302 			 * these features are absent.
303 			 */
304 			regs[3] |= (CPUID_MCA | CPUID_MCE | CPUID_MTRR);
305 
306 			logical_cpus = threads_per_core * cores_per_package;
307 			regs[1] &= ~CPUID_HTT_CORES;
308 			regs[1] |= (logical_cpus & 0xff) << 16;
309 			regs[3] |= CPUID_HTT;
310 			break;
311 
312 		case CPUID_0000_0004:
313 			cpuid_count(*eax, *ecx, regs);
314 
315 			if (regs[0] || regs[1] || regs[2] || regs[3]) {
316 				regs[0] &= 0x3ff;
317 				regs[0] |= (cores_per_package - 1) << 26;
318 				/*
319 				 * Cache topology:
320 				 * - L1 and L2 are shared only by the logical
321 				 *   processors in a single core.
322 				 * - L3 and above are shared by all logical
323 				 *   processors in the package.
324 				 */
325 				logical_cpus = threads_per_core;
326 				level = (regs[0] >> 5) & 0x7;
327 				if (level >= 3)
328 					logical_cpus *= cores_per_package;
329 				regs[0] |= (logical_cpus - 1) << 14;
330 			}
331 			break;
332 
333 		case CPUID_0000_0007:
334 			regs[0] = 0;
335 			regs[1] = 0;
336 			regs[2] = 0;
337 			regs[3] = 0;
338 
339 			/* leaf 0 */
340 			if (*ecx == 0) {
341 				cpuid_count(*eax, *ecx, regs);
342 
343 				/* Only leaf 0 is supported */
344 				regs[0] = 0;
345 
346 				/*
347 				 * Expose known-safe features.
348 				 */
349 				regs[1] &= (CPUID_STDEXT_FSGSBASE |
350 				    CPUID_STDEXT_BMI1 | CPUID_STDEXT_HLE |
351 				    CPUID_STDEXT_AVX2 | CPUID_STDEXT_BMI2 |
352 				    CPUID_STDEXT_ERMS | CPUID_STDEXT_RTM |
353 				    CPUID_STDEXT_AVX512F |
354 				    CPUID_STDEXT_AVX512PF |
355 				    CPUID_STDEXT_AVX512ER |
356 				    CPUID_STDEXT_AVX512CD);
357 				regs[2] = 0;
358 				regs[3] = 0;
359 
360 				/* Advertise INVPCID if it is enabled. */
361 				error = vm_get_capability(vm, vcpu_id,
362 				    VM_CAP_ENABLE_INVPCID, &enable_invpcid);
363 				if (error == 0 && enable_invpcid)
364 					regs[1] |= CPUID_STDEXT_INVPCID;
365 			}
366 			break;
367 
368 		case CPUID_0000_0006:
369 			regs[0] = CPUTPM1_ARAT;
370 			regs[1] = 0;
371 			regs[2] = 0;
372 			regs[3] = 0;
373 			break;
374 
375 		case CPUID_0000_000A:
376 			/*
377 			 * Handle the access, but report 0 for
378 			 * all options
379 			 */
380 			regs[0] = 0;
381 			regs[1] = 0;
382 			regs[2] = 0;
383 			regs[3] = 0;
384 			break;
385 
386 		case CPUID_0000_000B:
387 			/*
388 			 * Processor topology enumeration
389 			 */
390 			if (*ecx == 0) {
391 				logical_cpus = threads_per_core;
392 				width = log2(logical_cpus);
393 				level = CPUID_TYPE_SMT;
394 				x2apic_id = vcpu_id;
395 			}
396 
397 			if (*ecx == 1) {
398 				logical_cpus = threads_per_core *
399 				    cores_per_package;
400 				width = log2(logical_cpus);
401 				level = CPUID_TYPE_CORE;
402 				x2apic_id = vcpu_id;
403 			}
404 
405 			if (!cpuid_leaf_b || *ecx >= 2) {
406 				width = 0;
407 				logical_cpus = 0;
408 				level = 0;
409 				x2apic_id = 0;
410 			}
411 
412 			regs[0] = width & 0x1f;
413 			regs[1] = logical_cpus & 0xffff;
414 			regs[2] = (level << 8) | (*ecx & 0xff);
415 			regs[3] = x2apic_id;
416 			break;
417 
418 		case CPUID_0000_000D:
419 			limits = vmm_get_xsave_limits();
420 			if (!limits->xsave_enabled) {
421 				regs[0] = 0;
422 				regs[1] = 0;
423 				regs[2] = 0;
424 				regs[3] = 0;
425 				break;
426 			}
427 
428 			cpuid_count(*eax, *ecx, regs);
429 			switch (*ecx) {
430 			case 0:
431 				/*
432 				 * Only permit the guest to use bits
433 				 * that are active in the host in
434 				 * %xcr0.  Also, claim that the
435 				 * maximum save area size is
436 				 * equivalent to the host's current
437 				 * save area size.  Since this runs
438 				 * "inside" of vmrun(), it runs with
439 				 * the guest's xcr0, so the current
440 				 * save area size is correct as-is.
441 				 */
442 				regs[0] &= limits->xcr0_allowed;
443 				regs[2] = limits->xsave_max_size;
444 				regs[3] &= (limits->xcr0_allowed >> 32);
445 				break;
446 			case 1:
447 				/* Only permit XSAVEOPT. */
448 				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
449 				regs[1] = 0;
450 				regs[2] = 0;
451 				regs[3] = 0;
452 				break;
453 			default:
454 				/*
455 				 * If the leaf is for a permitted feature,
456 				 * pass through as-is, otherwise return
457 				 * all zeroes.
458 				 */
459 				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
460 					regs[0] = 0;
461 					regs[1] = 0;
462 					regs[2] = 0;
463 					regs[3] = 0;
464 				}
465 				break;
466 			}
467 			break;
468 
469 		case 0x40000000:
470 			regs[0] = CPUID_VM_HIGH;
471 			bcopy(bhyve_id, &regs[1], 4);
472 			bcopy(bhyve_id + 4, &regs[2], 4);
473 			bcopy(bhyve_id + 8, &regs[3], 4);
474 			break;
475 
476 		default:
477 			/*
478 			 * The leaf value has already been clamped so
479 			 * simply pass this through, keeping count of
480 			 * how many unhandled leaf values have been seen.
481 			 */
482 			atomic_add_long(&bhyve_xcpuids, 1);
483 			cpuid_count(*eax, *ecx, regs);
484 			break;
485 	}
486 
487 	*eax = regs[0];
488 	*ebx = regs[1];
489 	*ecx = regs[2];
490 	*edx = regs[3];
491 
492 	return (1);
493 }
494 
495 bool
496 vm_cpuid_capability(struct vm *vm, int vcpuid, enum vm_cpuid_capability cap)
497 {
498 	bool rv;
499 
500 	KASSERT(cap > 0 && cap < VCC_LAST, ("%s: invalid vm_cpu_capability %d",
501 	    __func__, cap));
502 
503 	/*
504 	 * Simply passthrough the capabilities of the host cpu for now.
505 	 */
506 	rv = false;
507 	switch (cap) {
508 	case VCC_NO_EXECUTE:
509 		if (amd_feature & AMDID_NX)
510 			rv = true;
511 		break;
512 	case VCC_FFXSR:
513 		if (amd_feature & AMDID_FFXSR)
514 			rv = true;
515 		break;
516 	case VCC_TCE:
517 		if (amd_feature2 & AMDID2_TCE)
518 			rv = true;
519 		break;
520 	default:
521 		panic("%s: unknown vm_cpu_capability %d", __func__, cap);
522 	}
523 	return (rv);
524 }
525