1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * vmx_nested_tsc_scaling_test
4 *
5 * Copyright 2021 Amazon.com, Inc. or its affiliates. All Rights Reserved.
6 *
7 * This test case verifies that nested TSC scaling behaves as expected when
8 * both L1 and L2 are scaled using different ratios. For this test we scale
9 * L1 down and scale L2 up.
10 */
11
12 #include <time.h>
13
14 #include "kvm_util.h"
15 #include "vmx.h"
16 #include "kselftest.h"
17
18
19 #define VCPU_ID 0
20
21 /* L2 is scaled up (from L1's perspective) by this factor */
22 #define L2_SCALE_FACTOR 4ULL
23
24 #define TSC_OFFSET_L2 ((uint64_t) -33125236320908)
25 #define TSC_MULTIPLIER_L2 (L2_SCALE_FACTOR << 48)
26
27 #define L2_GUEST_STACK_SIZE 64
28
29 enum { USLEEP, UCHECK_L1, UCHECK_L2 };
30 #define GUEST_SLEEP(sec) ucall(UCALL_SYNC, 2, USLEEP, sec)
31 #define GUEST_CHECK(level, freq) ucall(UCALL_SYNC, 2, level, freq)
32
33
34 /*
35 * This function checks whether the "actual" TSC frequency of a guest matches
36 * its expected frequency. In order to account for delays in taking the TSC
37 * measurements, a difference of 1% between the actual and the expected value
38 * is tolerated.
39 */
compare_tsc_freq(uint64_t actual,uint64_t expected)40 static void compare_tsc_freq(uint64_t actual, uint64_t expected)
41 {
42 uint64_t tolerance, thresh_low, thresh_high;
43
44 tolerance = expected / 100;
45 thresh_low = expected - tolerance;
46 thresh_high = expected + tolerance;
47
48 TEST_ASSERT(thresh_low < actual,
49 "TSC freq is expected to be between %"PRIu64" and %"PRIu64
50 " but it actually is %"PRIu64,
51 thresh_low, thresh_high, actual);
52 TEST_ASSERT(thresh_high > actual,
53 "TSC freq is expected to be between %"PRIu64" and %"PRIu64
54 " but it actually is %"PRIu64,
55 thresh_low, thresh_high, actual);
56 }
57
check_tsc_freq(int level)58 static void check_tsc_freq(int level)
59 {
60 uint64_t tsc_start, tsc_end, tsc_freq;
61
62 /*
63 * Reading the TSC twice with about a second's difference should give
64 * us an approximation of the TSC frequency from the guest's
65 * perspective. Now, this won't be completely accurate, but it should
66 * be good enough for the purposes of this test.
67 */
68 tsc_start = rdmsr(MSR_IA32_TSC);
69 GUEST_SLEEP(1);
70 tsc_end = rdmsr(MSR_IA32_TSC);
71
72 tsc_freq = tsc_end - tsc_start;
73
74 GUEST_CHECK(level, tsc_freq);
75 }
76
l2_guest_code(void)77 static void l2_guest_code(void)
78 {
79 check_tsc_freq(UCHECK_L2);
80
81 /* exit to L1 */
82 __asm__ __volatile__("vmcall");
83 }
84
l1_guest_code(struct vmx_pages * vmx_pages)85 static void l1_guest_code(struct vmx_pages *vmx_pages)
86 {
87 unsigned long l2_guest_stack[L2_GUEST_STACK_SIZE];
88 uint32_t control;
89
90 /* check that L1's frequency looks alright before launching L2 */
91 check_tsc_freq(UCHECK_L1);
92
93 GUEST_ASSERT(prepare_for_vmx_operation(vmx_pages));
94 GUEST_ASSERT(load_vmcs(vmx_pages));
95
96 /* prepare the VMCS for L2 execution */
97 prepare_vmcs(vmx_pages, l2_guest_code, &l2_guest_stack[L2_GUEST_STACK_SIZE]);
98
99 /* enable TSC offsetting and TSC scaling for L2 */
100 control = vmreadz(CPU_BASED_VM_EXEC_CONTROL);
101 control |= CPU_BASED_USE_MSR_BITMAPS | CPU_BASED_USE_TSC_OFFSETTING;
102 vmwrite(CPU_BASED_VM_EXEC_CONTROL, control);
103
104 control = vmreadz(SECONDARY_VM_EXEC_CONTROL);
105 control |= SECONDARY_EXEC_TSC_SCALING;
106 vmwrite(SECONDARY_VM_EXEC_CONTROL, control);
107
108 vmwrite(TSC_OFFSET, TSC_OFFSET_L2);
109 vmwrite(TSC_MULTIPLIER, TSC_MULTIPLIER_L2);
110 vmwrite(TSC_MULTIPLIER_HIGH, TSC_MULTIPLIER_L2 >> 32);
111
112 /* launch L2 */
113 GUEST_ASSERT(!vmlaunch());
114 GUEST_ASSERT(vmreadz(VM_EXIT_REASON) == EXIT_REASON_VMCALL);
115
116 /* check that L1's frequency still looks good */
117 check_tsc_freq(UCHECK_L1);
118
119 GUEST_DONE();
120 }
121
tsc_scaling_check_supported(void)122 static void tsc_scaling_check_supported(void)
123 {
124 if (!kvm_check_cap(KVM_CAP_TSC_CONTROL)) {
125 print_skip("TSC scaling not supported by the HW");
126 exit(KSFT_SKIP);
127 }
128 }
129
stable_tsc_check_supported(void)130 static void stable_tsc_check_supported(void)
131 {
132 FILE *fp;
133 char buf[4];
134
135 fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r");
136 if (fp == NULL)
137 goto skip_test;
138
139 if (fgets(buf, sizeof(buf), fp) == NULL)
140 goto skip_test;
141
142 if (strncmp(buf, "tsc", sizeof(buf)))
143 goto skip_test;
144
145 return;
146 skip_test:
147 print_skip("Kernel does not use TSC clocksource - assuming that host TSC is not stable");
148 exit(KSFT_SKIP);
149 }
150
main(int argc,char * argv[])151 int main(int argc, char *argv[])
152 {
153 struct kvm_vm *vm;
154 vm_vaddr_t vmx_pages_gva;
155
156 uint64_t tsc_start, tsc_end;
157 uint64_t tsc_khz;
158 uint64_t l1_scale_factor;
159 uint64_t l0_tsc_freq = 0;
160 uint64_t l1_tsc_freq = 0;
161 uint64_t l2_tsc_freq = 0;
162
163 nested_vmx_check_supported();
164 tsc_scaling_check_supported();
165 stable_tsc_check_supported();
166
167 /*
168 * We set L1's scale factor to be a random number from 2 to 10.
169 * Ideally we would do the same for L2's factor but that one is
170 * referenced by both main() and l1_guest_code() and using a global
171 * variable does not work.
172 */
173 srand(time(NULL));
174 l1_scale_factor = (rand() % 9) + 2;
175 printf("L1's scale down factor is: %"PRIu64"\n", l1_scale_factor);
176 printf("L2's scale up factor is: %llu\n", L2_SCALE_FACTOR);
177
178 tsc_start = rdtsc();
179 sleep(1);
180 tsc_end = rdtsc();
181
182 l0_tsc_freq = tsc_end - tsc_start;
183 printf("real TSC frequency is around: %"PRIu64"\n", l0_tsc_freq);
184
185 vm = vm_create_default(VCPU_ID, 0, (void *) l1_guest_code);
186 vcpu_alloc_vmx(vm, &vmx_pages_gva);
187 vcpu_args_set(vm, VCPU_ID, 1, vmx_pages_gva);
188
189 tsc_khz = _vcpu_ioctl(vm, VCPU_ID, KVM_GET_TSC_KHZ, NULL);
190 TEST_ASSERT(tsc_khz != -1, "vcpu ioctl KVM_GET_TSC_KHZ failed");
191
192 /* scale down L1's TSC frequency */
193 vcpu_ioctl(vm, VCPU_ID, KVM_SET_TSC_KHZ,
194 (void *) (tsc_khz / l1_scale_factor));
195
196 for (;;) {
197 volatile struct kvm_run *run = vcpu_state(vm, VCPU_ID);
198 struct ucall uc;
199
200 vcpu_run(vm, VCPU_ID);
201 TEST_ASSERT(run->exit_reason == KVM_EXIT_IO,
202 "Got exit_reason other than KVM_EXIT_IO: %u (%s)\n",
203 run->exit_reason,
204 exit_reason_str(run->exit_reason));
205
206 switch (get_ucall(vm, VCPU_ID, &uc)) {
207 case UCALL_ABORT:
208 TEST_FAIL("%s", (const char *) uc.args[0]);
209 case UCALL_SYNC:
210 switch (uc.args[0]) {
211 case USLEEP:
212 sleep(uc.args[1]);
213 break;
214 case UCHECK_L1:
215 l1_tsc_freq = uc.args[1];
216 printf("L1's TSC frequency is around: %"PRIu64
217 "\n", l1_tsc_freq);
218
219 compare_tsc_freq(l1_tsc_freq,
220 l0_tsc_freq / l1_scale_factor);
221 break;
222 case UCHECK_L2:
223 l2_tsc_freq = uc.args[1];
224 printf("L2's TSC frequency is around: %"PRIu64
225 "\n", l2_tsc_freq);
226
227 compare_tsc_freq(l2_tsc_freq,
228 l1_tsc_freq * L2_SCALE_FACTOR);
229 break;
230 }
231 break;
232 case UCALL_DONE:
233 goto done;
234 default:
235 TEST_FAIL("Unknown ucall %lu", uc.cmd);
236 }
237 }
238
239 done:
240 kvm_vm_free(vm);
241 return 0;
242 }
243