| 1 | /* |
| 2 | * QEMU TCG Single Threaded vCPUs implementation |
| 3 | * |
| 4 | * Copyright (c) 2003-2008 Fabrice Bellard |
| 5 | * Copyright (c) 2014 Red Hat Inc. |
| 6 | * |
| 7 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 8 | * of this software and associated documentation files (the "Software"), to deal |
| 9 | * in the Software without restriction, including without limitation the rights |
| 10 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 11 | * copies of the Software, and to permit persons to whom the Software is |
| 12 | * furnished to do so, subject to the following conditions: |
| 13 | * |
| 14 | * The above copyright notice and this permission notice shall be included in |
| 15 | * all copies or substantial portions of the Software. |
| 16 | * |
| 17 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 18 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 19 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 20 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 21 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 22 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 23 | * THE SOFTWARE. |
| 24 | */ |
| 25 | |
| 26 | #include "qemu/osdep.h" |
| 27 | #include "qemu/lockable.h" |
| 28 | #include "system/tcg.h" |
| 29 | #include "system/replay.h" |
| 30 | #include "exec/icount.h" |
| 31 | #include "qemu/main-loop.h" |
| 32 | #include "qemu/notify.h" |
| 33 | #include "qemu/guest-random.h" |
| 34 | #include "exec/cpu-common.h" |
| 35 | #include "accel/tcg/cpu-loop.h" |
| 36 | #include "tcg/startup.h" |
| 37 | #include "tcg-accel-ops.h" |
| 38 | #include "tcg-accel-ops-rr.h" |
| 39 | #include "tcg-accel-ops-icount.h" |
| 40 | |
| 41 | /* Kick all RR vCPUs */ |
| 42 | void rr_kick_vcpu_thread(CPUState *unused) |
| 43 | { |
| 44 | CPUState *cpu; |
| 45 | |
| 46 | CPU_FOREACH(cpu) { |
| 47 | tcg_kick_vcpu_thread(cpu); |
| 48 | }; |
| 49 | } |
| 50 | |
| 51 | /* |
| 52 | * TCG vCPU kick timer |
| 53 | * |
| 54 | * The kick timer is responsible for moving single threaded vCPU |
| 55 | * emulation on to the next vCPU. If more than one vCPU is running a |
| 56 | * timer event we force a cpu->exit so the next vCPU can get |
| 57 | * scheduled. |
| 58 | * |
| 59 | * The timer is removed if all vCPUs are idle and restarted again once |
| 60 | * idleness is complete. |
| 61 | */ |
| 62 | |
| 63 | static QEMUTimer *rr_kick_vcpu_timer; |
| 64 | static CPUState *rr_current_cpu; |
| 65 | |
| 66 | static inline int64_t rr_next_kick_time(void) |
| 67 | { |
| 68 | return qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + TCG_KICK_PERIOD; |
| 69 | } |
| 70 | |
| 71 | /* Kick the currently round-robin scheduled vCPU to next */ |
| 72 | static void rr_kick_next_cpu(void) |
| 73 | { |
| 74 | CPUState *cpu; |
| 75 | do { |
| 76 | cpu = qatomic_read(&rr_current_cpu); |
| 77 | if (cpu) { |
| 78 | cpu_exit(cpu); |
| 79 | } |
| 80 | /* Finish kicking this cpu before reading again. */ |
| 81 | smp_mb(); |
| 82 | } while (cpu != qatomic_read(&rr_current_cpu)); |
| 83 | } |
| 84 | |
| 85 | static void rr_kick_thread(void *opaque) |
| 86 | { |
| 87 | timer_mod(rr_kick_vcpu_timer, rr_next_kick_time()); |
| 88 | rr_kick_next_cpu(); |
| 89 | } |
| 90 | |
| 91 | static void rr_start_kick_timer(void) |
| 92 | { |
| 93 | if (!rr_kick_vcpu_timer && CPU_NEXT(first_cpu)) { |
| 94 | rr_kick_vcpu_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, |
| 95 | rr_kick_thread, NULL); |
| 96 | } |
| 97 | if (rr_kick_vcpu_timer && !timer_pending(rr_kick_vcpu_timer)) { |
| 98 | timer_mod(rr_kick_vcpu_timer, rr_next_kick_time()); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | static void rr_stop_kick_timer(void) |
| 103 | { |
| 104 | if (rr_kick_vcpu_timer && timer_pending(rr_kick_vcpu_timer)) { |
| 105 | timer_del(rr_kick_vcpu_timer); |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | static void rr_wait_io_event(void) |
| 110 | { |
| 111 | CPUState *cpu; |
| 112 | |
| 113 | while (all_cpu_threads_idle()) { |
| 114 | rr_stop_kick_timer(); |
| 115 | qemu_cond_wait_bql(first_cpu->halt_cond); |
| 116 | } |
| 117 | |
| 118 | rr_start_kick_timer(); |
| 119 | |
| 120 | CPU_FOREACH(cpu) { |
| 121 | qemu_process_cpu_events_common(cpu); |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * Destroy any remaining vCPUs which have been unplugged and have |
| 127 | * finished running |
| 128 | */ |
| 129 | static void rr_deal_with_unplugged_cpus(void) |
| 130 | { |
| 131 | CPUState *cpu; |
| 132 | |
| 133 | CPU_FOREACH(cpu) { |
| 134 | if (cpu->unplug && !cpu_can_run(cpu)) { |
| 135 | tcg_cpu_destroy(cpu); |
| 136 | break; |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | static void rr_force_rcu(Notifier *notify, void *data) |
| 142 | { |
| 143 | rr_kick_next_cpu(); |
| 144 | } |
| 145 | |
| 146 | /* |
| 147 | * Calculate the number of CPUs that we will process in a single iteration of |
| 148 | * the main CPU thread loop so that we can fairly distribute the instruction |
| 149 | * count across CPUs. |
| 150 | * |
| 151 | * The CPU count is cached based on the CPU list generation ID to avoid |
| 152 | * iterating the list every time. |
| 153 | */ |
| 154 | static int rr_cpu_count(void) |
| 155 | { |
| 156 | static unsigned int last_gen_id = ~0; |
| 157 | static int cpu_count; |
| 158 | CPUState *cpu; |
| 159 | |
| 160 | QEMU_LOCK_GUARD(&qemu_cpu_list_lock); |
| 161 | |
| 162 | if (cpu_list_generation_id_get() != last_gen_id) { |
| 163 | cpu_count = 0; |
| 164 | CPU_FOREACH(cpu) { |
| 165 | ++cpu_count; |
| 166 | } |
| 167 | last_gen_id = cpu_list_generation_id_get(); |
| 168 | } |
| 169 | |
| 170 | return cpu_count; |
| 171 | } |
| 172 | |
| 173 | /* |
| 174 | * In the single-threaded case each vCPU is simulated in turn. If |
| 175 | * there is more than a single vCPU we create a simple timer to kick |
| 176 | * the vCPU and ensure we don't get stuck in a tight loop in one vCPU. |
| 177 | * This is done explicitly rather than relying on side-effects |
| 178 | * elsewhere. |
| 179 | */ |
| 180 | |
| 181 | static void *rr_cpu_thread_fn(void *arg) |
| 182 | { |
| 183 | Notifier force_rcu; |
| 184 | CPUState *cpu = arg; |
| 185 | |
| 186 | assert(tcg_enabled()); |
| 187 | rcu_register_thread(); |
| 188 | force_rcu.notify = rr_force_rcu; |
| 189 | rcu_add_force_rcu_notifier(&force_rcu); |
| 190 | tcg_register_thread(); |
| 191 | |
| 192 | bql_lock(); |
| 193 | qemu_thread_get_self(cpu->thread); |
| 194 | |
| 195 | cpu->thread_id = qemu_get_thread_id(); |
| 196 | cpu->neg.can_do_io = true; |
| 197 | cpu_thread_signal_created(cpu); |
| 198 | qemu_guest_random_seed_thread_part2(cpu->random_seed); |
| 199 | |
| 200 | /* wait for initial kick-off after machine start */ |
| 201 | while (cpu_is_stopped(first_cpu)) { |
| 202 | qemu_cond_wait_bql(first_cpu->halt_cond); |
| 203 | |
| 204 | /* process any pending work */ |
| 205 | CPU_FOREACH(cpu) { |
| 206 | current_cpu = cpu; |
| 207 | qemu_process_cpu_events_common(cpu); |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | rr_start_kick_timer(); |
| 212 | |
| 213 | cpu = first_cpu; |
| 214 | |
| 215 | while (1) { |
| 216 | /* Only used for icount_enabled() */ |
| 217 | int64_t cpu_budget = 0; |
| 218 | |
| 219 | if (cpu) { |
| 220 | /* |
| 221 | * This could even reset exit_request for all CPUs, but in practice |
| 222 | * races between CPU exits and changes to "cpu" are so rare that |
| 223 | * there's no advantage in doing so. |
| 224 | */ |
| 225 | qatomic_set(&cpu->exit_request, false); |
| 226 | } |
| 227 | |
| 228 | if (icount_enabled() && all_cpu_threads_idle()) { |
| 229 | /* |
| 230 | * When all cpus are sleeping (e.g in WFI), to avoid a deadlock |
| 231 | * in the main_loop, wake it up in order to start the warp timer. |
| 232 | */ |
| 233 | qemu_notify_event(); |
| 234 | } |
| 235 | |
| 236 | rr_wait_io_event(); |
| 237 | rr_deal_with_unplugged_cpus(); |
| 238 | |
| 239 | bql_unlock(); |
| 240 | replay_mutex_lock(); |
| 241 | bql_lock(); |
| 242 | |
| 243 | if (icount_enabled()) { |
| 244 | int cpu_count = rr_cpu_count(); |
| 245 | |
| 246 | /* Account partial waits to QEMU_CLOCK_VIRTUAL. */ |
| 247 | icount_account_warp_timer(); |
| 248 | /* |
| 249 | * Run the timers here. This is much more efficient than |
| 250 | * waking up the I/O thread and waiting for completion. |
| 251 | */ |
| 252 | icount_handle_deadline(); |
| 253 | |
| 254 | cpu_budget = icount_percpu_budget(cpu_count); |
| 255 | } |
| 256 | |
| 257 | replay_mutex_unlock(); |
| 258 | |
| 259 | if (!cpu) { |
| 260 | cpu = first_cpu; |
| 261 | } |
| 262 | |
| 263 | while (cpu && cpu_work_list_empty(cpu)) { |
| 264 | /* |
| 265 | * Store rr_current_cpu before evaluating cpu->exit_request. |
| 266 | * Pairs with rr_kick_next_cpu(). |
| 267 | */ |
| 268 | qatomic_set_mb(&rr_current_cpu, cpu); |
| 269 | |
| 270 | /* Pairs with store-release in cpu_exit. */ |
| 271 | if (qatomic_load_acquire(&cpu->exit_request)) { |
| 272 | break; |
| 273 | } |
| 274 | current_cpu = cpu; |
| 275 | |
| 276 | qemu_clock_enable(QEMU_CLOCK_VIRTUAL, |
| 277 | (cpu->singlestep_flags & SSTEP_NOTIMER) == 0); |
| 278 | |
| 279 | if (cpu_can_run(cpu)) { |
| 280 | int r; |
| 281 | |
| 282 | bql_unlock(); |
| 283 | if (icount_enabled()) { |
| 284 | icount_prepare_for_run(cpu, cpu_budget); |
| 285 | } |
| 286 | r = tcg_cpu_exec(cpu); |
| 287 | if (icount_enabled()) { |
| 288 | icount_process_data(cpu); |
| 289 | } |
| 290 | bql_lock(); |
| 291 | |
| 292 | if (r == EXCP_DEBUG) { |
| 293 | cpu_handle_guest_debug(cpu); |
| 294 | break; |
| 295 | } else if (r == EXCP_ATOMIC) { |
| 296 | bql_unlock(); |
| 297 | cpu_exec_step_atomic(cpu); |
| 298 | bql_lock(); |
| 299 | break; |
| 300 | } |
| 301 | } else if (cpu->stop) { |
| 302 | if (cpu->unplug) { |
| 303 | cpu = CPU_NEXT(cpu); |
| 304 | } |
| 305 | break; |
| 306 | } |
| 307 | |
| 308 | cpu = CPU_NEXT(cpu); |
| 309 | } /* while (cpu && !cpu->exit_request).. */ |
| 310 | |
| 311 | /* Does not need a memory barrier because a spurious wakeup is okay. */ |
| 312 | qatomic_set(&rr_current_cpu, NULL); |
| 313 | } |
| 314 | |
| 315 | g_assert_not_reached(); |
| 316 | } |
| 317 | |
| 318 | void rr_start_vcpu_thread(CPUState *cpu) |
| 319 | { |
| 320 | char thread_name[VCPU_THREAD_NAME_SIZE]; |
| 321 | static QemuCond *single_tcg_halt_cond; |
| 322 | static QemuThread *single_tcg_cpu_thread; |
| 323 | |
| 324 | g_assert(tcg_enabled()); |
| 325 | tcg_cpu_init_cflags(cpu, false); |
| 326 | |
| 327 | if (!single_tcg_cpu_thread) { |
| 328 | single_tcg_halt_cond = cpu->halt_cond; |
| 329 | single_tcg_cpu_thread = cpu->thread; |
| 330 | |
| 331 | /* share a single thread for all cpus with TCG */ |
| 332 | snprintf(thread_name, VCPU_THREAD_NAME_SIZE, "ALL CPUs/TCG"); |
| 333 | qemu_thread_create(cpu->thread, thread_name, |
| 334 | rr_cpu_thread_fn, |
| 335 | cpu, QEMU_THREAD_JOINABLE); |
| 336 | } else { |
| 337 | /* we share the thread, dump spare data */ |
| 338 | g_free(cpu->thread); |
| 339 | qemu_cond_destroy(cpu->halt_cond); |
| 340 | g_free(cpu->halt_cond); |
| 341 | cpu->thread = single_tcg_cpu_thread; |
| 342 | cpu->halt_cond = single_tcg_halt_cond; |
| 343 | |
| 344 | /* copy the stuff done at start of rr_cpu_thread_fn */ |
| 345 | cpu->thread_id = first_cpu->thread_id; |
| 346 | cpu->neg.can_do_io = 1; |
| 347 | cpu->created = true; |
| 348 | } |
| 349 | } |