1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Xen SMP support 4 * 5 * This file implements the Xen versions of smp_ops. SMP under Xen is 6 * very straightforward. Bringing a CPU up is simply a matter of 7 * loading its initial context and setting it running. 8 * 9 * IPIs are handled through the Xen event mechanism. 10 * 11 * Because virtual CPUs can be scheduled onto any real CPU, there's no 12 * useful topology information for the kernel to make use of. As a 13 * result, all CPUs are treated as if they're single-core and 14 * single-threaded. 15 */ 16 #include <linux/sched.h> 17 #include <linux/sched/task_stack.h> 18 #include <linux/err.h> 19 #include <linux/slab.h> 20 #include <linux/smp.h> 21 #include <linux/irq_work.h> 22 #include <linux/tick.h> 23 #include <linux/nmi.h> 24 #include <linux/cpuhotplug.h> 25 #include <linux/stackprotector.h> 26 #include <linux/pgtable.h> 27 28 #include <asm/paravirt.h> 29 #include <asm/idtentry.h> 30 #include <asm/desc.h> 31 #include <asm/cpu.h> 32 #include <asm/io_apic.h> 33 34 #include <xen/interface/xen.h> 35 #include <xen/interface/vcpu.h> 36 #include <xen/interface/xenpmu.h> 37 38 #include <asm/spec-ctrl.h> 39 #include <asm/xen/interface.h> 40 #include <asm/xen/hypercall.h> 41 42 #include <xen/xen.h> 43 #include <xen/page.h> 44 #include <xen/events.h> 45 46 #include <xen/hvc-console.h> 47 #include "xen-ops.h" 48 #include "mmu.h" 49 #include "smp.h" 50 #include "pmu.h" 51 52 cpumask_var_t xen_cpu_initialized_map; 53 54 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 }; 55 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 }; 56 57 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id); 58 59 static void cpu_bringup(void) 60 { 61 int cpu; 62 63 cr4_init(); 64 cpuhp_ap_sync_alive(); 65 cpu_init(); 66 fpu__init_cpu(); 67 touch_softlockup_watchdog(); 68 69 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */ 70 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) { 71 xen_enable_sysenter(); 72 xen_enable_syscall(); 73 } 74 cpu = smp_processor_id(); 75 smp_store_cpu_info(cpu); 76 set_cpu_sibling_map(cpu); 77 78 speculative_store_bypass_ht_init(); 79 80 xen_setup_cpu_clockevents(); 81 82 notify_cpu_starting(cpu); 83 84 set_cpu_online(cpu, true); 85 86 smp_mb(); 87 88 /* We can take interrupts now: we're officially "up". */ 89 local_irq_enable(); 90 } 91 92 asmlinkage __visible void cpu_bringup_and_idle(void) 93 { 94 cpu_bringup(); 95 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE); 96 } 97 98 void xen_smp_intr_free_pv(unsigned int cpu) 99 { 100 kfree(per_cpu(xen_irq_work, cpu).name); 101 per_cpu(xen_irq_work, cpu).name = NULL; 102 if (per_cpu(xen_irq_work, cpu).irq >= 0) { 103 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL); 104 per_cpu(xen_irq_work, cpu).irq = -1; 105 } 106 107 kfree(per_cpu(xen_pmu_irq, cpu).name); 108 per_cpu(xen_pmu_irq, cpu).name = NULL; 109 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) { 110 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL); 111 per_cpu(xen_pmu_irq, cpu).irq = -1; 112 } 113 } 114 115 int xen_smp_intr_init_pv(unsigned int cpu) 116 { 117 int rc; 118 char *callfunc_name, *pmu_name; 119 120 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu); 121 per_cpu(xen_irq_work, cpu).name = callfunc_name; 122 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR, 123 cpu, 124 xen_irq_work_interrupt, 125 IRQF_PERCPU|IRQF_NOBALANCING, 126 callfunc_name, 127 NULL); 128 if (rc < 0) 129 goto fail; 130 per_cpu(xen_irq_work, cpu).irq = rc; 131 132 if (is_xen_pmu) { 133 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu); 134 per_cpu(xen_pmu_irq, cpu).name = pmu_name; 135 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu, 136 xen_pmu_irq_handler, 137 IRQF_PERCPU|IRQF_NOBALANCING, 138 pmu_name, NULL); 139 if (rc < 0) 140 goto fail; 141 per_cpu(xen_pmu_irq, cpu).irq = rc; 142 } 143 144 return 0; 145 146 fail: 147 xen_smp_intr_free_pv(cpu); 148 return rc; 149 } 150 151 static void __init xen_pv_smp_config(void) 152 { 153 int i, rc; 154 unsigned int subtract = 0; 155 156 num_processors = 0; 157 disabled_cpus = 0; 158 for (i = 0; i < nr_cpu_ids; i++) { 159 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); 160 if (rc >= 0) { 161 num_processors++; 162 set_cpu_possible(i, true); 163 } else { 164 set_cpu_possible(i, false); 165 set_cpu_present(i, false); 166 subtract++; 167 } 168 } 169 #ifdef CONFIG_HOTPLUG_CPU 170 /* This is akin to using 'nr_cpus' on the Linux command line. 171 * Which is OK as when we use 'dom0_max_vcpus=X' we can only 172 * have up to X, while nr_cpu_ids is greater than X. This 173 * normally is not a problem, except when CPU hotplugging 174 * is involved and then there might be more than X CPUs 175 * in the guest - which will not work as there is no 176 * hypercall to expand the max number of VCPUs an already 177 * running guest has. So cap it up to X. */ 178 if (subtract) 179 set_nr_cpu_ids(nr_cpu_ids - subtract); 180 #endif 181 /* Pretend to be a proper enumerated system */ 182 smp_found_config = 1; 183 } 184 185 static void __init xen_pv_smp_prepare_boot_cpu(void) 186 { 187 BUG_ON(smp_processor_id() != 0); 188 native_smp_prepare_boot_cpu(); 189 190 if (!xen_feature(XENFEAT_writable_page_tables)) 191 /* We've switched to the "real" per-cpu gdt, so make 192 * sure the old memory can be recycled. */ 193 make_lowmem_page_readwrite(xen_initial_gdt); 194 195 xen_setup_vcpu_info_placement(); 196 197 /* 198 * The alternative logic (which patches the unlock/lock) runs before 199 * the smp bootup up code is activated. Hence we need to set this up 200 * the core kernel is being patched. Otherwise we will have only 201 * modules patched but not core code. 202 */ 203 xen_init_spinlocks(); 204 } 205 206 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus) 207 { 208 unsigned cpu; 209 210 if (ioapic_is_disabled) { 211 char *m = (max_cpus == 0) ? 212 "The nosmp parameter is incompatible with Xen; " \ 213 "use Xen dom0_max_vcpus=1 parameter" : 214 "The noapic parameter is incompatible with Xen"; 215 216 xen_raw_printk(m); 217 panic(m); 218 } 219 xen_init_lock_cpu(0); 220 221 smp_prepare_cpus_common(); 222 223 speculative_store_bypass_ht_init(); 224 225 xen_pmu_init(0); 226 227 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0)) 228 BUG(); 229 230 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL)) 231 panic("could not allocate xen_cpu_initialized_map\n"); 232 233 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0)); 234 235 /* Restrict the possible_map according to max_cpus. */ 236 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { 237 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--) 238 continue; 239 set_cpu_possible(cpu, false); 240 } 241 242 for_each_possible_cpu(cpu) 243 set_cpu_present(cpu, true); 244 } 245 246 static int 247 cpu_initialize_context(unsigned int cpu, struct task_struct *idle) 248 { 249 struct vcpu_guest_context *ctxt; 250 struct desc_struct *gdt; 251 unsigned long gdt_mfn; 252 253 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map)) 254 return 0; 255 256 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); 257 if (ctxt == NULL) { 258 cpumask_clear_cpu(cpu, xen_cpu_initialized_map); 259 return -ENOMEM; 260 } 261 262 gdt = get_cpu_gdt_rw(cpu); 263 264 /* 265 * Bring up the CPU in cpu_bringup_and_idle() with the stack 266 * pointing just below where pt_regs would be if it were a normal 267 * kernel entry. 268 */ 269 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle; 270 ctxt->flags = VGCF_IN_KERNEL; 271 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ 272 ctxt->user_regs.ds = __USER_DS; 273 ctxt->user_regs.es = __USER_DS; 274 ctxt->user_regs.ss = __KERNEL_DS; 275 ctxt->user_regs.cs = __KERNEL_CS; 276 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle); 277 278 xen_copy_trap_info(ctxt->trap_ctxt); 279 280 BUG_ON((unsigned long)gdt & ~PAGE_MASK); 281 282 gdt_mfn = arbitrary_virt_to_mfn(gdt); 283 make_lowmem_page_readonly(gdt); 284 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn)); 285 286 ctxt->gdt_frames[0] = gdt_mfn; 287 ctxt->gdt_ents = GDT_ENTRIES; 288 289 /* 290 * Set SS:SP that Xen will use when entering guest kernel mode 291 * from guest user mode. Subsequent calls to load_sp0() can 292 * change this value. 293 */ 294 ctxt->kernel_ss = __KERNEL_DS; 295 ctxt->kernel_sp = task_top_of_stack(idle); 296 297 ctxt->gs_base_kernel = per_cpu_offset(cpu); 298 ctxt->event_callback_eip = 299 (unsigned long)xen_asm_exc_xen_hypervisor_callback; 300 ctxt->failsafe_callback_eip = 301 (unsigned long)xen_failsafe_callback; 302 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); 303 304 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir)); 305 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt)) 306 BUG(); 307 308 kfree(ctxt); 309 return 0; 310 } 311 312 static int xen_pv_kick_ap(unsigned int cpu, struct task_struct *idle) 313 { 314 int rc; 315 316 rc = common_cpu_up(cpu, idle); 317 if (rc) 318 return rc; 319 320 xen_setup_runstate_info(cpu); 321 322 /* make sure interrupts start blocked */ 323 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; 324 325 rc = cpu_initialize_context(cpu, idle); 326 if (rc) 327 return rc; 328 329 xen_pmu_init(cpu); 330 331 /* 332 * Why is this a BUG? If the hypercall fails then everything can be 333 * rolled back, no? 334 */ 335 BUG_ON(HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL)); 336 337 return 0; 338 } 339 340 static void xen_pv_poll_sync_state(void) 341 { 342 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 343 } 344 345 #ifdef CONFIG_HOTPLUG_CPU 346 static int xen_pv_cpu_disable(void) 347 { 348 unsigned int cpu = smp_processor_id(); 349 if (cpu == 0) 350 return -EBUSY; 351 352 cpu_disable_common(); 353 354 load_cr3(swapper_pg_dir); 355 return 0; 356 } 357 358 static void xen_pv_cpu_die(unsigned int cpu) 359 { 360 while (HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu), NULL)) { 361 __set_current_state(TASK_UNINTERRUPTIBLE); 362 schedule_timeout(HZ/10); 363 } 364 } 365 366 static void xen_pv_cleanup_dead_cpu(unsigned int cpu) 367 { 368 xen_smp_intr_free(cpu); 369 xen_uninit_lock_cpu(cpu); 370 xen_teardown_timer(cpu); 371 xen_pmu_finish(cpu); 372 } 373 374 static void __noreturn xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */ 375 { 376 play_dead_common(); 377 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL); 378 xen_cpu_bringup_again((unsigned long)task_pt_regs(current)); 379 BUG(); 380 } 381 382 #else /* !CONFIG_HOTPLUG_CPU */ 383 static int xen_pv_cpu_disable(void) 384 { 385 return -ENOSYS; 386 } 387 388 static void xen_pv_cpu_die(unsigned int cpu) 389 { 390 BUG(); 391 } 392 393 static void xen_pv_cleanup_dead_cpu(unsigned int cpu) 394 { 395 BUG(); 396 } 397 398 static void __noreturn xen_pv_play_dead(void) 399 { 400 BUG(); 401 } 402 403 #endif 404 static void stop_self(void *v) 405 { 406 int cpu = smp_processor_id(); 407 408 /* make sure we're not pinning something down */ 409 load_cr3(swapper_pg_dir); 410 /* should set up a minimal gdt */ 411 412 set_cpu_online(cpu, false); 413 414 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL); 415 BUG(); 416 } 417 418 static void xen_pv_stop_other_cpus(int wait) 419 { 420 smp_call_function(stop_self, NULL, wait); 421 } 422 423 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id) 424 { 425 irq_work_run(); 426 inc_irq_stat(apic_irq_work_irqs); 427 428 return IRQ_HANDLED; 429 } 430 431 static const struct smp_ops xen_smp_ops __initconst = { 432 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu, 433 .smp_prepare_cpus = xen_pv_smp_prepare_cpus, 434 .smp_cpus_done = xen_smp_cpus_done, 435 436 .kick_ap_alive = xen_pv_kick_ap, 437 .cpu_die = xen_pv_cpu_die, 438 .cleanup_dead_cpu = xen_pv_cleanup_dead_cpu, 439 .poll_sync_state = xen_pv_poll_sync_state, 440 .cpu_disable = xen_pv_cpu_disable, 441 .play_dead = xen_pv_play_dead, 442 443 .stop_other_cpus = xen_pv_stop_other_cpus, 444 .smp_send_reschedule = xen_smp_send_reschedule, 445 446 .send_call_func_ipi = xen_smp_send_call_function_ipi, 447 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi, 448 }; 449 450 void __init xen_smp_init(void) 451 { 452 smp_ops = xen_smp_ops; 453 454 /* Avoid searching for BIOS MP tables */ 455 x86_init.mpparse.find_mptable = x86_init_noop; 456 x86_init.mpparse.early_parse_smp_cfg = x86_init_noop; 457 x86_init.mpparse.parse_smp_cfg = xen_pv_smp_config; 458 } 459