1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Test for x86 KVM_CAP_SYNC_REGS
4 *
5 * Copyright (C) 2018, Google LLC.
6 *
7 * Verifies expected behavior of x86 KVM_CAP_SYNC_REGS functionality,
8 * including requesting an invalid register set, updates to/from values
9 * in kvm_run.s.regs when kvm_valid_regs and kvm_dirty_regs are toggled.
10 */
11 #include <fcntl.h>
12 #include <stdio.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include <sys/ioctl.h>
16 #include <pthread.h>
17
18 #include "kvm_test_harness.h"
19 #include "test_util.h"
20 #include "kvm_util.h"
21 #include "processor.h"
22
23 #define UCALL_PIO_PORT ((uint16_t)0x1000)
24
25 struct ucall uc_none = {
26 .cmd = UCALL_NONE,
27 };
28
29 /*
30 * ucall is embedded here to protect against compiler reshuffling registers
31 * before calling a function. In this test we only need to get KVM_EXIT_IO
32 * vmexit and preserve RBX, no additional information is needed.
33 */
guest_code(void)34 void guest_code(void)
35 {
36 asm volatile("1: in %[port], %%al\n"
37 "add $0x1, %%rbx\n"
38 "jmp 1b"
39 : : [port] "d" (UCALL_PIO_PORT), "D" (&uc_none)
40 : "rax", "rbx");
41 }
42
43 KVM_ONE_VCPU_TEST_SUITE(sync_regs_test);
44
compare_regs(struct kvm_regs * left,struct kvm_regs * right)45 static void compare_regs(struct kvm_regs *left, struct kvm_regs *right)
46 {
47 #define REG_COMPARE(reg) \
48 TEST_ASSERT(left->reg == right->reg, \
49 "Register " #reg \
50 " values did not match: 0x%llx, 0x%llx", \
51 left->reg, right->reg)
52 REG_COMPARE(rax);
53 REG_COMPARE(rbx);
54 REG_COMPARE(rcx);
55 REG_COMPARE(rdx);
56 REG_COMPARE(rsi);
57 REG_COMPARE(rdi);
58 REG_COMPARE(rsp);
59 REG_COMPARE(rbp);
60 REG_COMPARE(r8);
61 REG_COMPARE(r9);
62 REG_COMPARE(r10);
63 REG_COMPARE(r11);
64 REG_COMPARE(r12);
65 REG_COMPARE(r13);
66 REG_COMPARE(r14);
67 REG_COMPARE(r15);
68 REG_COMPARE(rip);
69 REG_COMPARE(rflags);
70 #undef REG_COMPARE
71 }
72
compare_sregs(struct kvm_sregs * left,struct kvm_sregs * right)73 static void compare_sregs(struct kvm_sregs *left, struct kvm_sregs *right)
74 {
75 }
76
compare_vcpu_events(struct kvm_vcpu_events * left,struct kvm_vcpu_events * right)77 static void compare_vcpu_events(struct kvm_vcpu_events *left,
78 struct kvm_vcpu_events *right)
79 {
80 }
81
82 #define TEST_SYNC_FIELDS (KVM_SYNC_X86_REGS|KVM_SYNC_X86_SREGS|KVM_SYNC_X86_EVENTS)
83 #define INVALID_SYNC_FIELD 0x80000000
84
85 /*
86 * Set an exception as pending *and* injected while KVM is processing events.
87 * KVM is supposed to ignore/drop pending exceptions if userspace is also
88 * requesting that an exception be injected.
89 */
race_events_inj_pen(void * arg)90 static void *race_events_inj_pen(void *arg)
91 {
92 struct kvm_run *run = (struct kvm_run *)arg;
93 struct kvm_vcpu_events *events = &run->s.regs.events;
94
95 WRITE_ONCE(events->exception.nr, UD_VECTOR);
96
97 for (;;) {
98 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_EVENTS);
99 WRITE_ONCE(events->flags, 0);
100 WRITE_ONCE(events->exception.injected, 1);
101 WRITE_ONCE(events->exception.pending, 1);
102
103 pthread_testcancel();
104 }
105
106 return NULL;
107 }
108
109 /*
110 * Set an invalid exception vector while KVM is processing events. KVM is
111 * supposed to reject any vector >= 32, as well as NMIs (vector 2).
112 */
race_events_exc(void * arg)113 static void *race_events_exc(void *arg)
114 {
115 struct kvm_run *run = (struct kvm_run *)arg;
116 struct kvm_vcpu_events *events = &run->s.regs.events;
117
118 for (;;) {
119 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_EVENTS);
120 WRITE_ONCE(events->flags, 0);
121 WRITE_ONCE(events->exception.nr, UD_VECTOR);
122 WRITE_ONCE(events->exception.pending, 1);
123 WRITE_ONCE(events->exception.nr, 255);
124
125 pthread_testcancel();
126 }
127
128 return NULL;
129 }
130
131 /*
132 * Toggle CR4.PAE while KVM is processing SREGS, EFER.LME=1 with CR4.PAE=0 is
133 * illegal, and KVM's MMU heavily relies on vCPU state being valid.
134 */
race_sregs_cr4(void * arg)135 static noinline void *race_sregs_cr4(void *arg)
136 {
137 struct kvm_run *run = (struct kvm_run *)arg;
138 __u64 *cr4 = &run->s.regs.sregs.cr4;
139 __u64 pae_enabled = *cr4;
140 __u64 pae_disabled = *cr4 & ~X86_CR4_PAE;
141
142 for (;;) {
143 WRITE_ONCE(run->kvm_dirty_regs, KVM_SYNC_X86_SREGS);
144 WRITE_ONCE(*cr4, pae_enabled);
145 asm volatile(".rept 512\n\t"
146 "nop\n\t"
147 ".endr");
148 WRITE_ONCE(*cr4, pae_disabled);
149
150 pthread_testcancel();
151 }
152
153 return NULL;
154 }
155
race_sync_regs(struct kvm_vcpu * vcpu,void * racer)156 static void race_sync_regs(struct kvm_vcpu *vcpu, void *racer)
157 {
158 const time_t TIMEOUT = 2; /* seconds, roughly */
159 struct kvm_x86_state *state;
160 struct kvm_translation tr;
161 struct kvm_run *run;
162 pthread_t thread;
163 time_t t;
164
165 run = vcpu->run;
166
167 run->kvm_valid_regs = KVM_SYNC_X86_SREGS;
168 vcpu_run(vcpu);
169 run->kvm_valid_regs = 0;
170
171 /* Save state *before* spawning the thread that mucks with vCPU state. */
172 state = vcpu_save_state(vcpu);
173
174 /*
175 * Selftests run 64-bit guests by default, both EFER.LME and CR4.PAE
176 * should already be set in guest state.
177 */
178 TEST_ASSERT((run->s.regs.sregs.cr4 & X86_CR4_PAE) &&
179 (run->s.regs.sregs.efer & EFER_LME),
180 "vCPU should be in long mode, CR4.PAE=%d, EFER.LME=%d",
181 !!(run->s.regs.sregs.cr4 & X86_CR4_PAE),
182 !!(run->s.regs.sregs.efer & EFER_LME));
183
184 TEST_ASSERT_EQ(pthread_create(&thread, NULL, racer, (void *)run), 0);
185
186 for (t = time(NULL) + TIMEOUT; time(NULL) < t;) {
187 /*
188 * Reload known good state if the vCPU triple faults, e.g. due
189 * to the unhandled #GPs being injected. VMX preserves state
190 * on shutdown, but SVM synthesizes an INIT as the VMCB state
191 * is architecturally undefined on triple fault.
192 */
193 if (!__vcpu_run(vcpu) && run->exit_reason == KVM_EXIT_SHUTDOWN)
194 vcpu_load_state(vcpu, state);
195
196 if (racer == race_sregs_cr4) {
197 tr = (struct kvm_translation) { .linear_address = 0 };
198 __vcpu_ioctl(vcpu, KVM_TRANSLATE, &tr);
199 }
200 }
201
202 TEST_ASSERT_EQ(pthread_cancel(thread), 0);
203 TEST_ASSERT_EQ(pthread_join(thread, NULL), 0);
204
205 kvm_x86_state_cleanup(state);
206 }
207
KVM_ONE_VCPU_TEST(sync_regs_test,read_invalid,guest_code)208 KVM_ONE_VCPU_TEST(sync_regs_test, read_invalid, guest_code)
209 {
210 struct kvm_run *run = vcpu->run;
211 int rv;
212
213 /* Request reading invalid register set from VCPU. */
214 run->kvm_valid_regs = INVALID_SYNC_FIELD;
215 rv = _vcpu_run(vcpu);
216 TEST_ASSERT(rv < 0 && errno == EINVAL,
217 "Invalid kvm_valid_regs did not cause expected KVM_RUN error: %d",
218 rv);
219 run->kvm_valid_regs = 0;
220
221 run->kvm_valid_regs = INVALID_SYNC_FIELD | TEST_SYNC_FIELDS;
222 rv = _vcpu_run(vcpu);
223 TEST_ASSERT(rv < 0 && errno == EINVAL,
224 "Invalid kvm_valid_regs did not cause expected KVM_RUN error: %d",
225 rv);
226 run->kvm_valid_regs = 0;
227 }
228
KVM_ONE_VCPU_TEST(sync_regs_test,set_invalid,guest_code)229 KVM_ONE_VCPU_TEST(sync_regs_test, set_invalid, guest_code)
230 {
231 struct kvm_run *run = vcpu->run;
232 int rv;
233
234 /* Request setting invalid register set into VCPU. */
235 run->kvm_dirty_regs = INVALID_SYNC_FIELD;
236 rv = _vcpu_run(vcpu);
237 TEST_ASSERT(rv < 0 && errno == EINVAL,
238 "Invalid kvm_dirty_regs did not cause expected KVM_RUN error: %d",
239 rv);
240 run->kvm_dirty_regs = 0;
241
242 run->kvm_dirty_regs = INVALID_SYNC_FIELD | TEST_SYNC_FIELDS;
243 rv = _vcpu_run(vcpu);
244 TEST_ASSERT(rv < 0 && errno == EINVAL,
245 "Invalid kvm_dirty_regs did not cause expected KVM_RUN error: %d",
246 rv);
247 run->kvm_dirty_regs = 0;
248 }
249
KVM_ONE_VCPU_TEST(sync_regs_test,req_and_verify_all_valid,guest_code)250 KVM_ONE_VCPU_TEST(sync_regs_test, req_and_verify_all_valid, guest_code)
251 {
252 struct kvm_run *run = vcpu->run;
253 struct kvm_vcpu_events events;
254 struct kvm_sregs sregs;
255 struct kvm_regs regs;
256
257 /* Request and verify all valid register sets. */
258 /* TODO: BUILD TIME CHECK: TEST_ASSERT(KVM_SYNC_X86_NUM_FIELDS != 3); */
259 run->kvm_valid_regs = TEST_SYNC_FIELDS;
260 vcpu_run(vcpu);
261 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
262
263 vcpu_regs_get(vcpu, ®s);
264 compare_regs(®s, &run->s.regs.regs);
265
266 vcpu_sregs_get(vcpu, &sregs);
267 compare_sregs(&sregs, &run->s.regs.sregs);
268
269 vcpu_events_get(vcpu, &events);
270 compare_vcpu_events(&events, &run->s.regs.events);
271 }
272
KVM_ONE_VCPU_TEST(sync_regs_test,set_and_verify_various,guest_code)273 KVM_ONE_VCPU_TEST(sync_regs_test, set_and_verify_various, guest_code)
274 {
275 struct kvm_run *run = vcpu->run;
276 struct kvm_vcpu_events events;
277 struct kvm_sregs sregs;
278 struct kvm_regs regs;
279
280 /* Run once to get register set */
281 run->kvm_valid_regs = TEST_SYNC_FIELDS;
282 vcpu_run(vcpu);
283 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
284
285 /* Set and verify various register values. */
286 run->s.regs.regs.rbx = 0xBAD1DEA;
287 run->s.regs.sregs.apic_base = 1 << 11;
288 /* TODO run->s.regs.events.XYZ = ABC; */
289
290 run->kvm_valid_regs = TEST_SYNC_FIELDS;
291 run->kvm_dirty_regs = KVM_SYNC_X86_REGS | KVM_SYNC_X86_SREGS;
292 vcpu_run(vcpu);
293 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
294 TEST_ASSERT(run->s.regs.regs.rbx == 0xBAD1DEA + 1,
295 "rbx sync regs value incorrect 0x%llx.",
296 run->s.regs.regs.rbx);
297 TEST_ASSERT(run->s.regs.sregs.apic_base == 1 << 11,
298 "apic_base sync regs value incorrect 0x%llx.",
299 run->s.regs.sregs.apic_base);
300
301 vcpu_regs_get(vcpu, ®s);
302 compare_regs(®s, &run->s.regs.regs);
303
304 vcpu_sregs_get(vcpu, &sregs);
305 compare_sregs(&sregs, &run->s.regs.sregs);
306
307 vcpu_events_get(vcpu, &events);
308 compare_vcpu_events(&events, &run->s.regs.events);
309 }
310
KVM_ONE_VCPU_TEST(sync_regs_test,clear_kvm_dirty_regs_bits,guest_code)311 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_dirty_regs_bits, guest_code)
312 {
313 struct kvm_run *run = vcpu->run;
314
315 /* Clear kvm_dirty_regs bits, verify new s.regs values are
316 * overwritten with existing guest values.
317 */
318 run->kvm_valid_regs = TEST_SYNC_FIELDS;
319 run->kvm_dirty_regs = 0;
320 run->s.regs.regs.rbx = 0xDEADBEEF;
321 vcpu_run(vcpu);
322 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
323 TEST_ASSERT(run->s.regs.regs.rbx != 0xDEADBEEF,
324 "rbx sync regs value incorrect 0x%llx.",
325 run->s.regs.regs.rbx);
326 }
327
KVM_ONE_VCPU_TEST(sync_regs_test,clear_kvm_valid_and_dirty_regs,guest_code)328 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_valid_and_dirty_regs, guest_code)
329 {
330 struct kvm_run *run = vcpu->run;
331 struct kvm_regs regs;
332
333 /* Run once to get register set */
334 run->kvm_valid_regs = TEST_SYNC_FIELDS;
335 vcpu_run(vcpu);
336 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
337
338 /* Clear kvm_valid_regs bits and kvm_dirty_bits.
339 * Verify s.regs values are not overwritten with existing guest values
340 * and that guest values are not overwritten with kvm_sync_regs values.
341 */
342 run->kvm_valid_regs = 0;
343 run->kvm_dirty_regs = 0;
344 run->s.regs.regs.rbx = 0xAAAA;
345 vcpu_regs_get(vcpu, ®s);
346 regs.rbx = 0xBAC0;
347 vcpu_regs_set(vcpu, ®s);
348 vcpu_run(vcpu);
349 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
350 TEST_ASSERT(run->s.regs.regs.rbx == 0xAAAA,
351 "rbx sync regs value incorrect 0x%llx.",
352 run->s.regs.regs.rbx);
353 vcpu_regs_get(vcpu, ®s);
354 TEST_ASSERT(regs.rbx == 0xBAC0 + 1,
355 "rbx guest value incorrect 0x%llx.",
356 regs.rbx);
357 }
358
KVM_ONE_VCPU_TEST(sync_regs_test,clear_kvm_valid_regs_bits,guest_code)359 KVM_ONE_VCPU_TEST(sync_regs_test, clear_kvm_valid_regs_bits, guest_code)
360 {
361 struct kvm_run *run = vcpu->run;
362 struct kvm_regs regs;
363
364 /* Run once to get register set */
365 run->kvm_valid_regs = TEST_SYNC_FIELDS;
366 vcpu_run(vcpu);
367 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
368
369 /* Clear kvm_valid_regs bits. Verify s.regs values are not overwritten
370 * with existing guest values but that guest values are overwritten
371 * with kvm_sync_regs values.
372 */
373 run->kvm_valid_regs = 0;
374 run->kvm_dirty_regs = TEST_SYNC_FIELDS;
375 run->s.regs.regs.rbx = 0xBBBB;
376 vcpu_run(vcpu);
377 TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
378 TEST_ASSERT(run->s.regs.regs.rbx == 0xBBBB,
379 "rbx sync regs value incorrect 0x%llx.",
380 run->s.regs.regs.rbx);
381 vcpu_regs_get(vcpu, ®s);
382 TEST_ASSERT(regs.rbx == 0xBBBB + 1,
383 "rbx guest value incorrect 0x%llx.",
384 regs.rbx);
385 }
386
KVM_ONE_VCPU_TEST(sync_regs_test,race_cr4,guest_code)387 KVM_ONE_VCPU_TEST(sync_regs_test, race_cr4, guest_code)
388 {
389 race_sync_regs(vcpu, race_sregs_cr4);
390 }
391
KVM_ONE_VCPU_TEST(sync_regs_test,race_exc,guest_code)392 KVM_ONE_VCPU_TEST(sync_regs_test, race_exc, guest_code)
393 {
394 race_sync_regs(vcpu, race_events_exc);
395 }
396
KVM_ONE_VCPU_TEST(sync_regs_test,race_inj_pen,guest_code)397 KVM_ONE_VCPU_TEST(sync_regs_test, race_inj_pen, guest_code)
398 {
399 race_sync_regs(vcpu, race_events_inj_pen);
400 }
401
main(int argc,char * argv[])402 int main(int argc, char *argv[])
403 {
404 int cap;
405
406 cap = kvm_check_cap(KVM_CAP_SYNC_REGS);
407 TEST_REQUIRE((cap & TEST_SYNC_FIELDS) == TEST_SYNC_FIELDS);
408 TEST_REQUIRE(!(cap & INVALID_SYNC_FIELD));
409
410 return test_harness_run(argc, argv);
411 }
412