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 cpu_init(); 65 touch_softlockup_watchdog(); 66 67 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */ 68 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) { 69 xen_enable_sysenter(); 70 xen_enable_syscall(); 71 } 72 cpu = smp_processor_id(); 73 smp_store_cpu_info(cpu); 74 cpu_data(cpu).x86_max_cores = 1; 75 set_cpu_sibling_map(cpu); 76 77 speculative_store_bypass_ht_init(); 78 79 xen_setup_cpu_clockevents(); 80 81 notify_cpu_starting(cpu); 82 83 set_cpu_online(cpu, true); 84 85 cpu_set_state_online(cpu); /* Implies full memory barrier. */ 86 87 /* We can take interrupts now: we're officially "up". */ 88 local_irq_enable(); 89 } 90 91 asmlinkage __visible void cpu_bringup_and_idle(void) 92 { 93 cpu_bringup(); 94 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE); 95 } 96 97 void xen_smp_intr_free_pv(unsigned int cpu) 98 { 99 kfree(per_cpu(xen_irq_work, cpu).name); 100 per_cpu(xen_irq_work, cpu).name = NULL; 101 if (per_cpu(xen_irq_work, cpu).irq >= 0) { 102 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL); 103 per_cpu(xen_irq_work, cpu).irq = -1; 104 } 105 106 kfree(per_cpu(xen_pmu_irq, cpu).name); 107 per_cpu(xen_pmu_irq, cpu).name = NULL; 108 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) { 109 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL); 110 per_cpu(xen_pmu_irq, cpu).irq = -1; 111 } 112 } 113 114 int xen_smp_intr_init_pv(unsigned int cpu) 115 { 116 int rc; 117 char *callfunc_name, *pmu_name; 118 119 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu); 120 per_cpu(xen_irq_work, cpu).name = callfunc_name; 121 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR, 122 cpu, 123 xen_irq_work_interrupt, 124 IRQF_PERCPU|IRQF_NOBALANCING, 125 callfunc_name, 126 NULL); 127 if (rc < 0) 128 goto fail; 129 per_cpu(xen_irq_work, cpu).irq = rc; 130 131 if (is_xen_pmu) { 132 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu); 133 per_cpu(xen_pmu_irq, cpu).name = pmu_name; 134 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu, 135 xen_pmu_irq_handler, 136 IRQF_PERCPU|IRQF_NOBALANCING, 137 pmu_name, NULL); 138 if (rc < 0) 139 goto fail; 140 per_cpu(xen_pmu_irq, cpu).irq = rc; 141 } 142 143 return 0; 144 145 fail: 146 xen_smp_intr_free_pv(cpu); 147 return rc; 148 } 149 150 static void __init _get_smp_config(unsigned int early) 151 { 152 int i, rc; 153 unsigned int subtract = 0; 154 155 if (early) 156 return; 157 158 num_processors = 0; 159 disabled_cpus = 0; 160 for (i = 0; i < nr_cpu_ids; i++) { 161 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); 162 if (rc >= 0) { 163 num_processors++; 164 set_cpu_possible(i, true); 165 } else { 166 set_cpu_possible(i, false); 167 set_cpu_present(i, false); 168 subtract++; 169 } 170 } 171 #ifdef CONFIG_HOTPLUG_CPU 172 /* This is akin to using 'nr_cpus' on the Linux command line. 173 * Which is OK as when we use 'dom0_max_vcpus=X' we can only 174 * have up to X, while nr_cpu_ids is greater than X. This 175 * normally is not a problem, except when CPU hotplugging 176 * is involved and then there might be more than X CPUs 177 * in the guest - which will not work as there is no 178 * hypercall to expand the max number of VCPUs an already 179 * running guest has. So cap it up to X. */ 180 if (subtract) 181 set_nr_cpu_ids(nr_cpu_ids - subtract); 182 #endif 183 184 } 185 186 static void __init xen_pv_smp_prepare_boot_cpu(void) 187 { 188 BUG_ON(smp_processor_id() != 0); 189 native_smp_prepare_boot_cpu(); 190 191 if (!xen_feature(XENFEAT_writable_page_tables)) 192 /* We've switched to the "real" per-cpu gdt, so make 193 * sure the old memory can be recycled. */ 194 make_lowmem_page_readwrite(xen_initial_gdt); 195 196 xen_setup_vcpu_info_placement(); 197 198 /* 199 * The alternative logic (which patches the unlock/lock) runs before 200 * the smp bootup up code is activated. Hence we need to set this up 201 * the core kernel is being patched. Otherwise we will have only 202 * modules patched but not core code. 203 */ 204 xen_init_spinlocks(); 205 } 206 207 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus) 208 { 209 unsigned cpu; 210 211 if (skip_ioapic_setup) { 212 char *m = (max_cpus == 0) ? 213 "The nosmp parameter is incompatible with Xen; " \ 214 "use Xen dom0_max_vcpus=1 parameter" : 215 "The noapic parameter is incompatible with Xen"; 216 217 xen_raw_printk(m); 218 panic(m); 219 } 220 xen_init_lock_cpu(0); 221 222 smp_prepare_cpus_common(); 223 224 cpu_data(0).x86_max_cores = 1; 225 226 speculative_store_bypass_ht_init(); 227 228 xen_pmu_init(0); 229 230 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0)) 231 BUG(); 232 233 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL)) 234 panic("could not allocate xen_cpu_initialized_map\n"); 235 236 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0)); 237 238 /* Restrict the possible_map according to max_cpus. */ 239 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { 240 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--) 241 continue; 242 set_cpu_possible(cpu, false); 243 } 244 245 for_each_possible_cpu(cpu) 246 set_cpu_present(cpu, true); 247 } 248 249 static int 250 cpu_initialize_context(unsigned int cpu, struct task_struct *idle) 251 { 252 struct vcpu_guest_context *ctxt; 253 struct desc_struct *gdt; 254 unsigned long gdt_mfn; 255 256 /* used to tell cpu_init() that it can proceed with initialization */ 257 cpumask_set_cpu(cpu, cpu_callout_mask); 258 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map)) 259 return 0; 260 261 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); 262 if (ctxt == NULL) { 263 cpumask_clear_cpu(cpu, xen_cpu_initialized_map); 264 cpumask_clear_cpu(cpu, cpu_callout_mask); 265 return -ENOMEM; 266 } 267 268 gdt = get_cpu_gdt_rw(cpu); 269 270 /* 271 * Bring up the CPU in cpu_bringup_and_idle() with the stack 272 * pointing just below where pt_regs would be if it were a normal 273 * kernel entry. 274 */ 275 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle; 276 ctxt->flags = VGCF_IN_KERNEL; 277 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ 278 ctxt->user_regs.ds = __USER_DS; 279 ctxt->user_regs.es = __USER_DS; 280 ctxt->user_regs.ss = __KERNEL_DS; 281 ctxt->user_regs.cs = __KERNEL_CS; 282 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle); 283 284 xen_copy_trap_info(ctxt->trap_ctxt); 285 286 BUG_ON((unsigned long)gdt & ~PAGE_MASK); 287 288 gdt_mfn = arbitrary_virt_to_mfn(gdt); 289 make_lowmem_page_readonly(gdt); 290 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn)); 291 292 ctxt->gdt_frames[0] = gdt_mfn; 293 ctxt->gdt_ents = GDT_ENTRIES; 294 295 /* 296 * Set SS:SP that Xen will use when entering guest kernel mode 297 * from guest user mode. Subsequent calls to load_sp0() can 298 * change this value. 299 */ 300 ctxt->kernel_ss = __KERNEL_DS; 301 ctxt->kernel_sp = task_top_of_stack(idle); 302 303 ctxt->gs_base_kernel = per_cpu_offset(cpu); 304 ctxt->event_callback_eip = 305 (unsigned long)xen_asm_exc_xen_hypervisor_callback; 306 ctxt->failsafe_callback_eip = 307 (unsigned long)xen_failsafe_callback; 308 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); 309 310 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir)); 311 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt)) 312 BUG(); 313 314 kfree(ctxt); 315 return 0; 316 } 317 318 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle) 319 { 320 int rc; 321 322 rc = common_cpu_up(cpu, idle); 323 if (rc) 324 return rc; 325 326 xen_setup_runstate_info(cpu); 327 328 /* 329 * PV VCPUs are always successfully taken down (see 'while' loop 330 * in xen_cpu_die()), so -EBUSY is an error. 331 */ 332 rc = cpu_check_up_prepare(cpu); 333 if (rc) 334 return rc; 335 336 /* make sure interrupts start blocked */ 337 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; 338 339 rc = cpu_initialize_context(cpu, idle); 340 if (rc) 341 return rc; 342 343 xen_pmu_init(cpu); 344 345 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL); 346 BUG_ON(rc); 347 348 while (cpu_report_state(cpu) != CPU_ONLINE) 349 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 350 351 return 0; 352 } 353 354 #ifdef CONFIG_HOTPLUG_CPU 355 static int xen_pv_cpu_disable(void) 356 { 357 unsigned int cpu = smp_processor_id(); 358 if (cpu == 0) 359 return -EBUSY; 360 361 cpu_disable_common(); 362 363 load_cr3(swapper_pg_dir); 364 return 0; 365 } 366 367 static void xen_pv_cpu_die(unsigned int cpu) 368 { 369 while (HYPERVISOR_vcpu_op(VCPUOP_is_up, 370 xen_vcpu_nr(cpu), NULL)) { 371 __set_current_state(TASK_UNINTERRUPTIBLE); 372 schedule_timeout(HZ/10); 373 } 374 375 if (common_cpu_die(cpu) == 0) { 376 xen_smp_intr_free(cpu); 377 xen_uninit_lock_cpu(cpu); 378 xen_teardown_timer(cpu); 379 xen_pmu_finish(cpu); 380 } 381 } 382 383 static void __noreturn xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */ 384 { 385 play_dead_common(); 386 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL); 387 xen_cpu_bringup_again((unsigned long)task_pt_regs(current)); 388 BUG(); 389 } 390 391 #else /* !CONFIG_HOTPLUG_CPU */ 392 static int xen_pv_cpu_disable(void) 393 { 394 return -ENOSYS; 395 } 396 397 static void xen_pv_cpu_die(unsigned int cpu) 398 { 399 BUG(); 400 } 401 402 static void __noreturn xen_pv_play_dead(void) 403 { 404 BUG(); 405 } 406 407 #endif 408 static void stop_self(void *v) 409 { 410 int cpu = smp_processor_id(); 411 412 /* make sure we're not pinning something down */ 413 load_cr3(swapper_pg_dir); 414 /* should set up a minimal gdt */ 415 416 set_cpu_online(cpu, false); 417 418 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL); 419 BUG(); 420 } 421 422 static void xen_pv_stop_other_cpus(int wait) 423 { 424 smp_call_function(stop_self, NULL, wait); 425 } 426 427 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id) 428 { 429 irq_work_run(); 430 inc_irq_stat(apic_irq_work_irqs); 431 432 return IRQ_HANDLED; 433 } 434 435 static const struct smp_ops xen_smp_ops __initconst = { 436 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu, 437 .smp_prepare_cpus = xen_pv_smp_prepare_cpus, 438 .smp_cpus_done = xen_smp_cpus_done, 439 440 .cpu_up = xen_pv_cpu_up, 441 .cpu_die = xen_pv_cpu_die, 442 .cpu_disable = xen_pv_cpu_disable, 443 .play_dead = xen_pv_play_dead, 444 445 .stop_other_cpus = xen_pv_stop_other_cpus, 446 .smp_send_reschedule = xen_smp_send_reschedule, 447 448 .send_call_func_ipi = xen_smp_send_call_function_ipi, 449 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi, 450 }; 451 452 void __init xen_smp_init(void) 453 { 454 smp_ops = xen_smp_ops; 455 456 /* Avoid searching for BIOS MP tables */ 457 x86_init.mpparse.find_smp_config = x86_init_noop; 458 x86_init.mpparse.get_smp_config = _get_smp_config; 459 } 460