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
  68         /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
  69         if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
  70                 xen_enable_sysenter();
  71                 xen_enable_syscall();
  72         }
  73         cpu = smp_processor_id();
  74         smp_store_cpu_info(cpu);
  75         cpu_data(cpu).x86_max_cores = 1;
  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         cpu_set_state_online(cpu);  /* Implies full memory barrier. */
  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         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 101                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 102                 per_cpu(xen_irq_work, cpu).irq = -1;
 103                 kfree(per_cpu(xen_irq_work, cpu).name);
 104                 per_cpu(xen_irq_work, cpu).name = NULL;
 105         }
 106
 107         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 108                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 109                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 110                 kfree(per_cpu(xen_pmu_irq, cpu).name);
 111                 per_cpu(xen_pmu_irq, cpu).name = NULL;
 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         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         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 131
 132         if (is_xen_pmu) {
 133                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 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                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 142         }
 143
 144         return 0;
 145
 146  fail:
 147         xen_smp_intr_free_pv(cpu);
 148         return rc;
 149 }
 150
 151 static void __init _get_smp_config(unsigned int early)
 152 {
 153         int i, rc;
 154         unsigned int subtract = 0;
 155
 156         if (early)
 157                 return;
 158
 159         num_processors = 0;
 160         disabled_cpus = 0;
 161         for (i = 0; i < nr_cpu_ids; i++) {
 162                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 163                 if (rc >= 0) {
 164                         num_processors++;
 165                         set_cpu_possible(i, true);
 166                 } else {
 167                         set_cpu_possible(i, false);
 168                         set_cpu_present(i, false);
 169                         subtract++;
 170                 }
 171         }
 172 #ifdef CONFIG_HOTPLUG_CPU
 173         /* This is akin to using 'nr_cpus' on the Linux command line.
 174          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 175          * have up to X, while nr_cpu_ids is greater than X. This
 176          * normally is not a problem, except when CPU hotplugging
 177          * is involved and then there might be more than X CPUs
 178          * in the guest - which will not work as there is no
 179          * hypercall to expand the max number of VCPUs an already
 180          * running guest has. So cap it up to X. */
 181         if (subtract)
 182                 nr_cpu_ids = nr_cpu_ids - subtract;
 183 #endif
 184
 185 }
 186
 187 static void __init xen_pv_smp_prepare_boot_cpu(void)
 188 {
 189         BUG_ON(smp_processor_id() != 0);
 190         native_smp_prepare_boot_cpu();
 191
 192         if (!xen_feature(XENFEAT_writable_page_tables))
 193                 /* We've switched to the "real" per-cpu gdt, so make
 194                  * sure the old memory can be recycled. */
 195                 make_lowmem_page_readwrite(xen_initial_gdt);
 196
 197         xen_setup_vcpu_info_placement();
 198
 199         /*
 200          * The alternative logic (which patches the unlock/lock) runs before
 201          * the smp bootup up code is activated. Hence we need to set this up
 202          * the core kernel is being patched. Otherwise we will have only
 203          * modules patched but not core code.
 204          */
 205         xen_init_spinlocks();
 206 }
 207
 208 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
 209 {
 210         unsigned cpu;
 211
 212         if (skip_ioapic_setup) {
 213                 char *m = (max_cpus == 0) ?
 214                         "The nosmp parameter is incompatible with Xen; " \
 215                         "use Xen dom0_max_vcpus=1 parameter" :
 216                         "The noapic parameter is incompatible with Xen";
 217
 218                 xen_raw_printk(m);
 219                 panic(m);
 220         }
 221         xen_init_lock_cpu(0);
 222
 223         smp_prepare_cpus_common();
 224
 225         cpu_data(0).x86_max_cores = 1;
 226
 227         speculative_store_bypass_ht_init();
 228
 229         xen_pmu_init(0);
 230
 231         if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
 232                 BUG();
 233
 234         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 235                 panic("could not allocate xen_cpu_initialized_map\n");
 236
 237         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 238
 239         /* Restrict the possible_map according to max_cpus. */
 240         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 241                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 242                         continue;
 243                 set_cpu_possible(cpu, false);
 244         }
 245
 246         for_each_possible_cpu(cpu)
 247                 set_cpu_present(cpu, true);
 248 }
 249
 250 static int
 251 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 252 {
 253         struct vcpu_guest_context *ctxt;
 254         struct desc_struct *gdt;
 255         unsigned long gdt_mfn;
 256
 257         /* used to tell cpu_init() that it can proceed with initialization */
 258         cpumask_set_cpu(cpu, cpu_callout_mask);
 259         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 260                 return 0;
 261
 262         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 263         if (ctxt == NULL)
 264                 return -ENOMEM;
 265
 266         gdt = get_cpu_gdt_rw(cpu);
 267
 268         /*
 269          * Bring up the CPU in cpu_bringup_and_idle() with the stack
 270          * pointing just below where pt_regs would be if it were a normal
 271          * kernel entry.
 272          */
 273         ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
 274         ctxt->flags = VGCF_IN_KERNEL;
 275         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 276         ctxt->user_regs.ds = __USER_DS;
 277         ctxt->user_regs.es = __USER_DS;
 278         ctxt->user_regs.ss = __KERNEL_DS;
 279         ctxt->user_regs.cs = __KERNEL_CS;
 280         ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
 281
 282         xen_copy_trap_info(ctxt->trap_ctxt);
 283
 284         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 285
 286         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 287         make_lowmem_page_readonly(gdt);
 288         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 289
 290         ctxt->gdt_frames[0] = gdt_mfn;
 291         ctxt->gdt_ents      = GDT_ENTRIES;
 292
 293         /*
 294          * Set SS:SP that Xen will use when entering guest kernel mode
 295          * from guest user mode.  Subsequent calls to load_sp0() can
 296          * change this value.
 297          */
 298         ctxt->kernel_ss = __KERNEL_DS;
 299         ctxt->kernel_sp = task_top_of_stack(idle);
 300
 301         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 302         ctxt->event_callback_eip    =
 303                 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
 304         ctxt->failsafe_callback_eip =
 305                 (unsigned long)xen_failsafe_callback;
 306         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 307
 308         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 309         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 310                 BUG();
 311
 312         kfree(ctxt);
 313         return 0;
 314 }
 315
 316 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
 317 {
 318         int rc;
 319
 320         rc = common_cpu_up(cpu, idle);
 321         if (rc)
 322                 return rc;
 323
 324         xen_setup_runstate_info(cpu);
 325
 326         /*
 327          * PV VCPUs are always successfully taken down (see 'while' loop
 328          * in xen_cpu_die()), so -EBUSY is an error.
 329          */
 330         rc = cpu_check_up_prepare(cpu);
 331         if (rc)
 332                 return rc;
 333
 334         /* make sure interrupts start blocked */
 335         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 336
 337         rc = cpu_initialize_context(cpu, idle);
 338         if (rc)
 339                 return rc;
 340
 341         xen_pmu_init(cpu);
 342
 343         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 344         BUG_ON(rc);
 345
 346         while (cpu_report_state(cpu) != CPU_ONLINE)
 347                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 348
 349         return 0;
 350 }
 351
 352 #ifdef CONFIG_HOTPLUG_CPU
 353 static int xen_pv_cpu_disable(void)
 354 {
 355         unsigned int cpu = smp_processor_id();
 356         if (cpu == 0)
 357                 return -EBUSY;
 358
 359         cpu_disable_common();
 360
 361         load_cr3(swapper_pg_dir);
 362         return 0;
 363 }
 364
 365 static void xen_pv_cpu_die(unsigned int cpu)
 366 {
 367         while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 368                                   xen_vcpu_nr(cpu), NULL)) {
 369                 __set_current_state(TASK_UNINTERRUPTIBLE);
 370                 schedule_timeout(HZ/10);
 371         }
 372
 373         if (common_cpu_die(cpu) == 0) {
 374                 xen_smp_intr_free(cpu);
 375                 xen_uninit_lock_cpu(cpu);
 376                 xen_teardown_timer(cpu);
 377                 xen_pmu_finish(cpu);
 378         }
 379 }
 380
 381 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
 382 {
 383         play_dead_common();
 384         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 385         cpu_bringup();
 386         /*
 387          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 388          * clears certain data that the cpu_idle loop (which called us
 389          * and that we return from) expects. The only way to get that
 390          * data back is to call:
 391          */
 392         tick_nohz_idle_enter();
 393         tick_nohz_idle_stop_tick_protected();
 394
 395         cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
 396 }
 397
 398 #else /* !CONFIG_HOTPLUG_CPU */
 399 static int xen_pv_cpu_disable(void)
 400 {
 401         return -ENOSYS;
 402 }
 403
 404 static void xen_pv_cpu_die(unsigned int cpu)
 405 {
 406         BUG();
 407 }
 408
 409 static void xen_pv_play_dead(void)
 410 {
 411         BUG();
 412 }
 413
 414 #endif
 415 static void stop_self(void *v)
 416 {
 417         int cpu = smp_processor_id();
 418
 419         /* make sure we're not pinning something down */
 420         load_cr3(swapper_pg_dir);
 421         /* should set up a minimal gdt */
 422
 423         set_cpu_online(cpu, false);
 424
 425         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 426         BUG();
 427 }
 428
 429 static void xen_pv_stop_other_cpus(int wait)
 430 {
 431         smp_call_function(stop_self, NULL, wait);
 432 }
 433
 434 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 435 {
 436         irq_work_run();
 437         inc_irq_stat(apic_irq_work_irqs);
 438
 439         return IRQ_HANDLED;
 440 }
 441
 442 static const struct smp_ops xen_smp_ops __initconst = {
 443         .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 444         .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
 445         .smp_cpus_done = xen_smp_cpus_done,
 446
 447         .cpu_up = xen_pv_cpu_up,
 448         .cpu_die = xen_pv_cpu_die,
 449         .cpu_disable = xen_pv_cpu_disable,
 450         .play_dead = xen_pv_play_dead,
 451
 452         .stop_other_cpus = xen_pv_stop_other_cpus,
 453         .smp_send_reschedule = xen_smp_send_reschedule,
 454
 455         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 456         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 457 };
 458
 459 void __init xen_smp_init(void)
 460 {
 461         smp_ops = xen_smp_ops;
 462
 463         /* Avoid searching for BIOS MP tables */
 464         x86_init.mpparse.find_smp_config = x86_init_noop;
 465         x86_init.mpparse.get_smp_config = _get_smp_config;
 466 }