arch/x86/xen/smp_pv.c

   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 void asm_cpu_bringup_and_idle(void);
  59
  60 static void cpu_bringup(void)
  61 {
  62         int cpu;
  63
  64         cr4_init();
  65         cpu_init();
  66         touch_softlockup_watchdog();
  67         preempt_disable();
  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         cpu_data(cpu).x86_max_cores = 1;
  77         set_cpu_sibling_map(cpu);
  78
  79         speculative_store_bypass_ht_init();
  80
  81         xen_setup_cpu_clockevents();
  82
  83         notify_cpu_starting(cpu);
  84
  85         set_cpu_online(cpu, true);
  86
  87         cpu_set_state_online(cpu);  /* Implies full memory barrier. */
  88
  89         /* We can take interrupts now: we're officially "up". */
  90         local_irq_enable();
  91 }
  92
  93 asmlinkage __visible void cpu_bringup_and_idle(void)
  94 {
  95         cpu_bringup();
  96         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
  97 }
  98
  99 void xen_smp_intr_free_pv(unsigned int cpu)
 100 {
 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                 kfree(per_cpu(xen_irq_work, cpu).name);
 105                 per_cpu(xen_irq_work, cpu).name = NULL;
 106         }
 107
 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                 kfree(per_cpu(xen_pmu_irq, cpu).name);
 112                 per_cpu(xen_pmu_irq, cpu).name = NULL;
 113         }
 114 }
 115
 116 int xen_smp_intr_init_pv(unsigned int cpu)
 117 {
 118         int rc;
 119         char *callfunc_name, *pmu_name;
 120
 121         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
 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         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 132
 133         if (is_xen_pmu(cpu)) {
 134                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 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                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 143         }
 144
 145         return 0;
 146
 147  fail:
 148         xen_smp_intr_free_pv(cpu);
 149         return rc;
 150 }
 151
 152 static void __init xen_fill_possible_map(void)
 153 {
 154         int i, rc;
 155
 156         if (xen_initial_domain())
 157                 return;
 158
 159         for (i = 0; i < nr_cpu_ids; i++) {
 160                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 161                 if (rc >= 0) {
 162                         num_processors++;
 163                         set_cpu_possible(i, true);
 164                 }
 165         }
 166 }
 167
 168 static void __init xen_filter_cpu_maps(void)
 169 {
 170         int i, rc;
 171         unsigned int subtract = 0;
 172
 173         if (!xen_initial_domain())
 174                 return;
 175
 176         num_processors = 0;
 177         disabled_cpus = 0;
 178         for (i = 0; i < nr_cpu_ids; i++) {
 179                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 180                 if (rc >= 0) {
 181                         num_processors++;
 182                         set_cpu_possible(i, true);
 183                 } else {
 184                         set_cpu_possible(i, false);
 185                         set_cpu_present(i, false);
 186                         subtract++;
 187                 }
 188         }
 189 #ifdef CONFIG_HOTPLUG_CPU
 190         /* This is akin to using 'nr_cpus' on the Linux command line.
 191          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 192          * have up to X, while nr_cpu_ids is greater than X. This
 193          * normally is not a problem, except when CPU hotplugging
 194          * is involved and then there might be more than X CPUs
 195          * in the guest - which will not work as there is no
 196          * hypercall to expand the max number of VCPUs an already
 197          * running guest has. So cap it up to X. */
 198         if (subtract)
 199                 nr_cpu_ids = nr_cpu_ids - subtract;
 200 #endif
 201
 202 }
 203
 204 static void __init xen_pv_smp_prepare_boot_cpu(void)
 205 {
 206         BUG_ON(smp_processor_id() != 0);
 207         native_smp_prepare_boot_cpu();
 208
 209         if (!xen_feature(XENFEAT_writable_page_tables))
 210                 /* We've switched to the "real" per-cpu gdt, so make
 211                  * sure the old memory can be recycled. */
 212                 make_lowmem_page_readwrite(xen_initial_gdt);
 213
 214         xen_filter_cpu_maps();
 215         xen_setup_vcpu_info_placement();
 216
 217         /*
 218          * The alternative logic (which patches the unlock/lock) runs before
 219          * the smp bootup up code is activated. Hence we need to set this up
 220          * the core kernel is being patched. Otherwise we will have only
 221          * modules patched but not core code.
 222          */
 223         xen_init_spinlocks();
 224 }
 225
 226 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
 227 {
 228         unsigned cpu;
 229         unsigned int i;
 230
 231         if (skip_ioapic_setup) {
 232                 char *m = (max_cpus == 0) ?
 233                         "The nosmp parameter is incompatible with Xen; " \
 234                         "use Xen dom0_max_vcpus=1 parameter" :
 235                         "The noapic parameter is incompatible with Xen";
 236
 237                 xen_raw_printk(m);
 238                 panic(m);
 239         }
 240         xen_init_lock_cpu(0);
 241
 242         smp_store_boot_cpu_info();
 243         cpu_data(0).x86_max_cores = 1;
 244
 245         for_each_possible_cpu(i) {
 246                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 247                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 248                 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
 249                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 250         }
 251         set_cpu_sibling_map(0);
 252
 253         speculative_store_bypass_ht_init();
 254
 255         xen_pmu_init(0);
 256
 257         if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
 258                 BUG();
 259
 260         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 261                 panic("could not allocate xen_cpu_initialized_map\n");
 262
 263         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 264
 265         /* Restrict the possible_map according to max_cpus. */
 266         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 267                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 268                         continue;
 269                 set_cpu_possible(cpu, false);
 270         }
 271
 272         for_each_possible_cpu(cpu)
 273                 set_cpu_present(cpu, true);
 274 }
 275
 276 static int
 277 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 278 {
 279         struct vcpu_guest_context *ctxt;
 280         struct desc_struct *gdt;
 281         unsigned long gdt_mfn;
 282
 283         /* used to tell cpu_init() that it can proceed with initialization */
 284         cpumask_set_cpu(cpu, cpu_callout_mask);
 285         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 286                 return 0;
 287
 288         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 289         if (ctxt == NULL)
 290                 return -ENOMEM;
 291
 292         gdt = get_cpu_gdt_rw(cpu);
 293
 294         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 295
 296         /*
 297          * Bring up the CPU in cpu_bringup_and_idle() with the stack
 298          * pointing just below where pt_regs would be if it were a normal
 299          * kernel entry.
 300          */
 301         ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
 302         ctxt->flags = VGCF_IN_KERNEL;
 303         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 304         ctxt->user_regs.ds = __USER_DS;
 305         ctxt->user_regs.es = __USER_DS;
 306         ctxt->user_regs.ss = __KERNEL_DS;
 307         ctxt->user_regs.cs = __KERNEL_CS;
 308         ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
 309
 310         xen_copy_trap_info(ctxt->trap_ctxt);
 311
 312         ctxt->ldt_ents = 0;
 313
 314         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 315
 316         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 317         make_lowmem_page_readonly(gdt);
 318         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 319
 320         ctxt->gdt_frames[0] = gdt_mfn;
 321         ctxt->gdt_ents      = GDT_ENTRIES;
 322
 323         /*
 324          * Set SS:SP that Xen will use when entering guest kernel mode
 325          * from guest user mode.  Subsequent calls to load_sp0() can
 326          * change this value.
 327          */
 328         ctxt->kernel_ss = __KERNEL_DS;
 329         ctxt->kernel_sp = task_top_of_stack(idle);
 330
 331         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 332         ctxt->event_callback_eip    =
 333                 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
 334         ctxt->failsafe_callback_eip =
 335                 (unsigned long)xen_failsafe_callback;
 336         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 337
 338         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 339         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 340                 BUG();
 341
 342         kfree(ctxt);
 343         return 0;
 344 }
 345
 346 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
 347 {
 348         int rc;
 349
 350         rc = common_cpu_up(cpu, idle);
 351         if (rc)
 352                 return rc;
 353
 354         xen_setup_runstate_info(cpu);
 355
 356         /*
 357          * PV VCPUs are always successfully taken down (see 'while' loop
 358          * in xen_cpu_die()), so -EBUSY is an error.
 359          */
 360         rc = cpu_check_up_prepare(cpu);
 361         if (rc)
 362                 return rc;
 363
 364         /* make sure interrupts start blocked */
 365         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 366
 367         rc = cpu_initialize_context(cpu, idle);
 368         if (rc)
 369                 return rc;
 370
 371         xen_pmu_init(cpu);
 372
 373         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 374         BUG_ON(rc);
 375
 376         while (cpu_report_state(cpu) != CPU_ONLINE)
 377                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 378
 379         return 0;
 380 }
 381
 382 #ifdef CONFIG_HOTPLUG_CPU
 383 static int xen_pv_cpu_disable(void)
 384 {
 385         unsigned int cpu = smp_processor_id();
 386         if (cpu == 0)
 387                 return -EBUSY;
 388
 389         cpu_disable_common();
 390
 391         load_cr3(swapper_pg_dir);
 392         return 0;
 393 }
 394
 395 static void xen_pv_cpu_die(unsigned int cpu)
 396 {
 397         while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 398                                   xen_vcpu_nr(cpu), NULL)) {
 399                 __set_current_state(TASK_UNINTERRUPTIBLE);
 400                 schedule_timeout(HZ/10);
 401         }
 402
 403         if (common_cpu_die(cpu) == 0) {
 404                 xen_smp_intr_free(cpu);
 405                 xen_uninit_lock_cpu(cpu);
 406                 xen_teardown_timer(cpu);
 407                 xen_pmu_finish(cpu);
 408         }
 409 }
 410
 411 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
 412 {
 413         play_dead_common();
 414         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 415         cpu_bringup();
 416         /*
 417          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 418          * clears certain data that the cpu_idle loop (which called us
 419          * and that we return from) expects. The only way to get that
 420          * data back is to call:
 421          */
 422         tick_nohz_idle_enter();
 423         tick_nohz_idle_stop_tick_protected();
 424
 425         cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
 426 }
 427
 428 #else /* !CONFIG_HOTPLUG_CPU */
 429 static int xen_pv_cpu_disable(void)
 430 {
 431         return -ENOSYS;
 432 }
 433
 434 static void xen_pv_cpu_die(unsigned int cpu)
 435 {
 436         BUG();
 437 }
 438
 439 static void xen_pv_play_dead(void)
 440 {
 441         BUG();
 442 }
 443
 444 #endif
 445 static void stop_self(void *v)
 446 {
 447         int cpu = smp_processor_id();
 448
 449         /* make sure we're not pinning something down */
 450         load_cr3(swapper_pg_dir);
 451         /* should set up a minimal gdt */
 452
 453         set_cpu_online(cpu, false);
 454
 455         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 456         BUG();
 457 }
 458
 459 static void xen_pv_stop_other_cpus(int wait)
 460 {
 461         smp_call_function(stop_self, NULL, wait);
 462 }
 463
 464 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 465 {
 466         irq_enter();
 467         irq_work_run();
 468         inc_irq_stat(apic_irq_work_irqs);
 469         irq_exit();
 470
 471         return IRQ_HANDLED;
 472 }
 473
 474 static const struct smp_ops xen_smp_ops __initconst = {
 475         .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 476         .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
 477         .smp_cpus_done = xen_smp_cpus_done,
 478
 479         .cpu_up = xen_pv_cpu_up,
 480         .cpu_die = xen_pv_cpu_die,
 481         .cpu_disable = xen_pv_cpu_disable,
 482         .play_dead = xen_pv_play_dead,
 483
 484         .stop_other_cpus = xen_pv_stop_other_cpus,
 485         .smp_send_reschedule = xen_smp_send_reschedule,
 486
 487         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 488         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 489 };
 490
 491 void __init xen_smp_init(void)
 492 {
 493         smp_ops = xen_smp_ops;
 494         xen_fill_possible_map();
 495 }