1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* KVM paravirtual clock driver. A clocksource implementation 3 Copyright (C) 2008 Glauber de Oliveira Costa, Red Hat Inc. 4 */ 5 6 #include <linux/clocksource.h> 7 #include <linux/kvm_para.h> 8 #include <asm/pvclock.h> 9 #include <asm/msr.h> 10 #include <asm/apic.h> 11 #include <linux/percpu.h> 12 #include <linux/hardirq.h> 13 #include <linux/cpuhotplug.h> 14 #include <linux/sched.h> 15 #include <linux/sched/clock.h> 16 #include <linux/mm.h> 17 #include <linux/slab.h> 18 #include <linux/set_memory.h> 19 #include <linux/cc_platform.h> 20 21 #include <asm/hypervisor.h> 22 #include <asm/timer.h> 23 #include <asm/x86_init.h> 24 #include <asm/kvmclock.h> 25 26 static int kvmclock __initdata = 1; 27 static int kvmclock_vsyscall __initdata = 1; 28 static int msr_kvm_system_time __ro_after_init; 29 static int msr_kvm_wall_clock __ro_after_init; 30 static u64 kvm_sched_clock_offset __ro_after_init; 31 32 static int __init parse_no_kvmclock(char *arg) 33 { 34 kvmclock = 0; 35 return 0; 36 } 37 early_param("no-kvmclock", parse_no_kvmclock); 38 39 static int __init parse_no_kvmclock_vsyscall(char *arg) 40 { 41 kvmclock_vsyscall = 0; 42 return 0; 43 } 44 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall); 45 46 /* Aligned to page sizes to match what's mapped via vsyscalls to userspace */ 47 #define HVC_BOOT_ARRAY_SIZE \ 48 (PAGE_SIZE / sizeof(struct pvclock_vsyscall_time_info)) 49 50 static struct pvclock_vsyscall_time_info 51 hv_clock_boot[HVC_BOOT_ARRAY_SIZE] __bss_decrypted __aligned(PAGE_SIZE); 52 static struct pvclock_wall_clock wall_clock __bss_decrypted; 53 static struct pvclock_vsyscall_time_info *hvclock_mem; 54 DEFINE_PER_CPU(struct pvclock_vsyscall_time_info *, hv_clock_per_cpu); 55 EXPORT_PER_CPU_SYMBOL_GPL(hv_clock_per_cpu); 56 57 /* 58 * The wallclock is the time of day when we booted. Since then, some time may 59 * have elapsed since the hypervisor wrote the data. So we try to account for 60 * that with system time 61 */ 62 static void kvm_get_wallclock(struct timespec64 *now) 63 { 64 wrmsrq(msr_kvm_wall_clock, slow_virt_to_phys(&wall_clock)); 65 preempt_disable(); 66 pvclock_read_wallclock(&wall_clock, this_cpu_pvti(), now); 67 preempt_enable(); 68 } 69 70 static int kvm_set_wallclock(const struct timespec64 *now) 71 { 72 return -ENODEV; 73 } 74 75 static u64 kvm_clock_read(void) 76 { 77 u64 ret; 78 79 preempt_disable_notrace(); 80 ret = pvclock_clocksource_read_nowd(this_cpu_pvti()); 81 preempt_enable_notrace(); 82 return ret; 83 } 84 85 static u64 kvm_clock_get_cycles(struct clocksource *cs) 86 { 87 return kvm_clock_read(); 88 } 89 90 static noinstr u64 kvm_sched_clock_read(void) 91 { 92 return pvclock_clocksource_read_nowd(this_cpu_pvti()) - kvm_sched_clock_offset; 93 } 94 95 static inline void kvm_sched_clock_init(bool stable) 96 { 97 if (!stable) 98 clear_sched_clock_stable(); 99 kvm_sched_clock_offset = kvm_clock_read(); 100 paravirt_set_sched_clock(kvm_sched_clock_read); 101 102 pr_info("kvm-clock: using sched offset of %llu cycles", 103 kvm_sched_clock_offset); 104 105 BUILD_BUG_ON(sizeof(kvm_sched_clock_offset) > 106 sizeof(((struct pvclock_vcpu_time_info *)NULL)->system_time)); 107 } 108 109 /* 110 * If we don't do that, there is the possibility that the guest 111 * will calibrate under heavy load - thus, getting a lower lpj - 112 * and execute the delays themselves without load. This is wrong, 113 * because no delay loop can finish beforehand. 114 * Any heuristics is subject to fail, because ultimately, a large 115 * poll of guests can be running and trouble each other. So we preset 116 * lpj here 117 */ 118 static unsigned long kvm_get_tsc_khz(void) 119 { 120 setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ); 121 return pvclock_tsc_khz(this_cpu_pvti()); 122 } 123 124 static void __init kvm_get_preset_lpj(void) 125 { 126 unsigned long khz; 127 u64 lpj; 128 129 khz = kvm_get_tsc_khz(); 130 131 lpj = ((u64)khz * 1000); 132 do_div(lpj, HZ); 133 preset_lpj = lpj; 134 } 135 136 bool kvm_check_and_clear_guest_paused(void) 137 { 138 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock(); 139 bool ret = false; 140 141 if (!src) 142 return ret; 143 144 if ((src->pvti.flags & PVCLOCK_GUEST_STOPPED) != 0) { 145 src->pvti.flags &= ~PVCLOCK_GUEST_STOPPED; 146 pvclock_touch_watchdogs(); 147 ret = true; 148 } 149 return ret; 150 } 151 152 static int kvm_cs_enable(struct clocksource *cs) 153 { 154 vclocks_set_used(VDSO_CLOCKMODE_PVCLOCK); 155 return 0; 156 } 157 158 static struct clocksource kvm_clock = { 159 .name = "kvm-clock", 160 .read = kvm_clock_get_cycles, 161 .rating = 400, 162 .mask = CLOCKSOURCE_MASK(64), 163 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 164 .id = CSID_X86_KVM_CLK, 165 .enable = kvm_cs_enable, 166 }; 167 168 static void kvm_register_clock(char *txt) 169 { 170 struct pvclock_vsyscall_time_info *src = this_cpu_hvclock(); 171 u64 pa; 172 173 if (!src) 174 return; 175 176 pa = slow_virt_to_phys(&src->pvti) | 0x01ULL; 177 wrmsrq(msr_kvm_system_time, pa); 178 pr_debug("kvm-clock: cpu %d, msr %llx, %s", smp_processor_id(), pa, txt); 179 } 180 181 static void kvm_save_sched_clock_state(void) 182 { 183 } 184 185 static void kvm_restore_sched_clock_state(void) 186 { 187 kvm_register_clock("primary cpu clock, resume"); 188 } 189 190 #ifdef CONFIG_X86_LOCAL_APIC 191 static void kvm_setup_secondary_clock(void) 192 { 193 kvm_register_clock("secondary cpu clock"); 194 } 195 #endif 196 197 void kvmclock_disable(void) 198 { 199 if (msr_kvm_system_time) 200 native_write_msr(msr_kvm_system_time, 0); 201 } 202 203 static void __init kvmclock_init_mem(void) 204 { 205 unsigned long ncpus; 206 unsigned int order; 207 struct page *p; 208 int r; 209 210 if (HVC_BOOT_ARRAY_SIZE >= num_possible_cpus()) 211 return; 212 213 ncpus = num_possible_cpus() - HVC_BOOT_ARRAY_SIZE; 214 order = get_order(ncpus * sizeof(*hvclock_mem)); 215 216 p = alloc_pages(GFP_KERNEL, order); 217 if (!p) { 218 pr_warn("%s: failed to alloc %d pages", __func__, (1U << order)); 219 return; 220 } 221 222 hvclock_mem = page_address(p); 223 224 /* 225 * hvclock is shared between the guest and the hypervisor, must 226 * be mapped decrypted. 227 */ 228 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) { 229 r = set_memory_decrypted((unsigned long) hvclock_mem, 230 1UL << order); 231 if (r) { 232 __free_pages(p, order); 233 hvclock_mem = NULL; 234 pr_warn("kvmclock: set_memory_decrypted() failed. Disabling\n"); 235 return; 236 } 237 } 238 239 memset(hvclock_mem, 0, PAGE_SIZE << order); 240 } 241 242 static int __init kvm_setup_vsyscall_timeinfo(void) 243 { 244 if (!kvm_para_available() || !kvmclock || nopv) 245 return 0; 246 247 kvmclock_init_mem(); 248 249 #ifdef CONFIG_X86_64 250 if (per_cpu(hv_clock_per_cpu, 0) && kvmclock_vsyscall) { 251 u8 flags; 252 253 flags = pvclock_read_flags(&hv_clock_boot[0].pvti); 254 if (!(flags & PVCLOCK_TSC_STABLE_BIT)) 255 return 0; 256 257 kvm_clock.vdso_clock_mode = VDSO_CLOCKMODE_PVCLOCK; 258 } 259 #endif 260 261 return 0; 262 } 263 early_initcall(kvm_setup_vsyscall_timeinfo); 264 265 static int kvmclock_setup_percpu(unsigned int cpu) 266 { 267 struct pvclock_vsyscall_time_info *p = per_cpu(hv_clock_per_cpu, cpu); 268 269 /* 270 * The per cpu area setup replicates CPU0 data to all cpu 271 * pointers. So carefully check. CPU0 has been set up in init 272 * already. 273 */ 274 if (!cpu || (p && p != per_cpu(hv_clock_per_cpu, 0))) 275 return 0; 276 277 /* Use the static page for the first CPUs, allocate otherwise */ 278 if (cpu < HVC_BOOT_ARRAY_SIZE) 279 p = &hv_clock_boot[cpu]; 280 else if (hvclock_mem) 281 p = hvclock_mem + cpu - HVC_BOOT_ARRAY_SIZE; 282 else 283 return -ENOMEM; 284 285 per_cpu(hv_clock_per_cpu, cpu) = p; 286 return p ? 0 : -ENOMEM; 287 } 288 289 void __init kvmclock_init(void) 290 { 291 u8 flags; 292 293 if (!kvm_para_available() || !kvmclock) 294 return; 295 296 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE2)) { 297 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME_NEW; 298 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK_NEW; 299 } else if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) { 300 msr_kvm_system_time = MSR_KVM_SYSTEM_TIME; 301 msr_kvm_wall_clock = MSR_KVM_WALL_CLOCK; 302 } else { 303 return; 304 } 305 306 if (cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "kvmclock:setup_percpu", 307 kvmclock_setup_percpu, NULL) < 0) { 308 return; 309 } 310 311 pr_info("kvm-clock: Using msrs %x and %x", 312 msr_kvm_system_time, msr_kvm_wall_clock); 313 314 this_cpu_write(hv_clock_per_cpu, &hv_clock_boot[0]); 315 kvm_register_clock("primary cpu clock"); 316 pvclock_set_pvti_cpu0_va(hv_clock_boot); 317 318 if (kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE_STABLE_BIT)) 319 pvclock_set_flags(PVCLOCK_TSC_STABLE_BIT); 320 321 flags = pvclock_read_flags(&hv_clock_boot[0].pvti); 322 kvm_sched_clock_init(flags & PVCLOCK_TSC_STABLE_BIT); 323 324 x86_platform.calibrate_tsc = kvm_get_tsc_khz; 325 x86_platform.calibrate_cpu = kvm_get_tsc_khz; 326 x86_platform.get_wallclock = kvm_get_wallclock; 327 x86_platform.set_wallclock = kvm_set_wallclock; 328 #ifdef CONFIG_X86_LOCAL_APIC 329 x86_cpuinit.early_percpu_clock_init = kvm_setup_secondary_clock; 330 #endif 331 x86_platform.save_sched_clock_state = kvm_save_sched_clock_state; 332 x86_platform.restore_sched_clock_state = kvm_restore_sched_clock_state; 333 kvm_get_preset_lpj(); 334 335 /* 336 * X86_FEATURE_NONSTOP_TSC is TSC runs at constant rate 337 * with P/T states and does not stop in deep C-states. 338 * 339 * Invariant TSC exposed by host means kvmclock is not necessary: 340 * can use TSC as clocksource. 341 * 342 */ 343 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && 344 boot_cpu_has(X86_FEATURE_NONSTOP_TSC) && 345 !check_tsc_unstable()) 346 kvm_clock.rating = 299; 347 348 clocksource_register_hz(&kvm_clock, NSEC_PER_SEC); 349 pv_info.name = "KVM"; 350 } 351