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         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         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         kfree(per_cpu(xen_irq_work, cpu).name);
 102         per_cpu(xen_irq_work, cpu).name = NULL;
 103         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 104                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 105                 per_cpu(xen_irq_work, cpu).irq = -1;
 106         }
 107
 108         kfree(per_cpu(xen_pmu_irq, cpu).name);
 109         per_cpu(xen_pmu_irq, cpu).name = NULL;
 110         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 111                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 112                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 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         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 123         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
 124                                     cpu,
 125                                     xen_irq_work_interrupt,
 126                                     IRQF_PERCPU|IRQF_NOBALANCING,
 127                                     callfunc_name,
 128                                     NULL);
 129         if (rc < 0)
 130                 goto fail;
 131         per_cpu(xen_irq_work, cpu).irq = rc;
 132
 133         if (is_xen_pmu) {
 134                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 135                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 136                 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
 137                                              xen_pmu_irq_handler,
 138                                              IRQF_PERCPU|IRQF_NOBALANCING,
 139                                              pmu_name, NULL);
 140                 if (rc < 0)
 141                         goto fail;
 142                 per_cpu(xen_pmu_irq, cpu).irq = rc;
 143         }
 144
 145         return 0;
 146
 147  fail:
 148         xen_smp_intr_free_pv(cpu);
 149         return rc;
 150 }
 151
 152 static void __init _get_smp_config(unsigned int early)
 153 {
 154         int i, rc;
 155         unsigned int subtract = 0;
 156
 157         if (early)
 158                 return;
 159
 160         num_processors = 0;
 161         disabled_cpus = 0;
 162         for (i = 0; i < nr_cpu_ids; i++) {
 163                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 164                 if (rc >= 0) {
 165                         num_processors++;
 166                         set_cpu_possible(i, true);
 167                 } else {
 168                         set_cpu_possible(i, false);
 169                         set_cpu_present(i, false);
 170                         subtract++;
 171                 }
 172         }
 173 #ifdef CONFIG_HOTPLUG_CPU
 174         /* This is akin to using 'nr_cpus' on the Linux command line.
 175          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 176          * have up to X, while nr_cpu_ids is greater than X. This
 177          * normally is not a problem, except when CPU hotplugging
 178          * is involved and then there might be more than X CPUs
 179          * in the guest - which will not work as there is no
 180          * hypercall to expand the max number of VCPUs an already
 181          * running guest has. So cap it up to X. */
 182         if (subtract)
 183                 set_nr_cpu_ids(nr_cpu_ids - subtract);
 184 #endif
 185
 186 }
 187
 188 static void __init xen_pv_smp_prepare_boot_cpu(void)
 189 {
 190         BUG_ON(smp_processor_id() != 0);
 191         native_smp_prepare_boot_cpu();
 192
 193         if (!xen_feature(XENFEAT_writable_page_tables))
 194                 /* We've switched to the "real" per-cpu gdt, so make
 195                  * sure the old memory can be recycled. */
 196                 make_lowmem_page_readwrite(xen_initial_gdt);
 197
 198         xen_setup_vcpu_info_placement();
 199
 200         /*
 201          * The alternative logic (which patches the unlock/lock) runs before
 202          * the smp bootup up code is activated. Hence we need to set this up
 203          * the core kernel is being patched. Otherwise we will have only
 204          * modules patched but not core code.
 205          */
 206         xen_init_spinlocks();
 207 }
 208
 209 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
 210 {
 211         unsigned cpu;
 212
 213         if (skip_ioapic_setup) {
 214                 char *m = (max_cpus == 0) ?
 215                         "The nosmp parameter is incompatible with Xen; " \
 216                         "use Xen dom0_max_vcpus=1 parameter" :
 217                         "The noapic parameter is incompatible with Xen";
 218
 219                 xen_raw_printk(m);
 220                 panic(m);
 221         }
 222         xen_init_lock_cpu(0);
 223
 224         smp_prepare_cpus_common();
 225
 226         cpu_data(0).x86_max_cores = 1;
 227
 228         speculative_store_bypass_ht_init();
 229
 230         xen_pmu_init(0);
 231
 232         if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
 233                 BUG();
 234
 235         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 236                 panic("could not allocate xen_cpu_initialized_map\n");
 237
 238         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 239
 240         /* Restrict the possible_map according to max_cpus. */
 241         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 242                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 243                         continue;
 244                 set_cpu_possible(cpu, false);
 245         }
 246
 247         for_each_possible_cpu(cpu)
 248                 set_cpu_present(cpu, true);
 249 }
 250
 251 static int
 252 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 253 {
 254         struct vcpu_guest_context *ctxt;
 255         struct desc_struct *gdt;
 256         unsigned long gdt_mfn;
 257
 258         /* used to tell cpu_init() that it can proceed with initialization */
 259         cpumask_set_cpu(cpu, cpu_callout_mask);
 260         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 261                 return 0;
 262
 263         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 264         if (ctxt == NULL) {
 265                 cpumask_clear_cpu(cpu, xen_cpu_initialized_map);
 266                 cpumask_clear_cpu(cpu, cpu_callout_mask);
 267                 return -ENOMEM;
 268         }
 269
 270         gdt = get_cpu_gdt_rw(cpu);
 271
 272         /*
 273          * Bring up the CPU in cpu_bringup_and_idle() with the stack
 274          * pointing just below where pt_regs would be if it were a normal
 275          * kernel entry.
 276          */
 277         ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
 278         ctxt->flags = VGCF_IN_KERNEL;
 279         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 280         ctxt->user_regs.ds = __USER_DS;
 281         ctxt->user_regs.es = __USER_DS;
 282         ctxt->user_regs.ss = __KERNEL_DS;
 283         ctxt->user_regs.cs = __KERNEL_CS;
 284         ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
 285
 286         xen_copy_trap_info(ctxt->trap_ctxt);
 287
 288         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 289
 290         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 291         make_lowmem_page_readonly(gdt);
 292         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 293
 294         ctxt->gdt_frames[0] = gdt_mfn;
 295         ctxt->gdt_ents      = GDT_ENTRIES;
 296
 297         /*
 298          * Set SS:SP that Xen will use when entering guest kernel mode
 299          * from guest user mode.  Subsequent calls to load_sp0() can
 300          * change this value.
 301          */
 302         ctxt->kernel_ss = __KERNEL_DS;
 303         ctxt->kernel_sp = task_top_of_stack(idle);
 304
 305         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 306         ctxt->event_callback_eip    =
 307                 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
 308         ctxt->failsafe_callback_eip =
 309                 (unsigned long)xen_failsafe_callback;
 310         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 311
 312         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 313         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 314                 BUG();
 315
 316         kfree(ctxt);
 317         return 0;
 318 }
 319
 320 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
 321 {
 322         int rc;
 323
 324         rc = common_cpu_up(cpu, idle);
 325         if (rc)
 326                 return rc;
 327
 328         xen_setup_runstate_info(cpu);
 329
 330         /*
 331          * PV VCPUs are always successfully taken down (see 'while' loop
 332          * in xen_cpu_die()), so -EBUSY is an error.
 333          */
 334         rc = cpu_check_up_prepare(cpu);
 335         if (rc)
 336                 return rc;
 337
 338         /* make sure interrupts start blocked */
 339         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 340
 341         rc = cpu_initialize_context(cpu, idle);
 342         if (rc)
 343                 return rc;
 344
 345         xen_pmu_init(cpu);
 346
 347         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 348         BUG_ON(rc);
 349
 350         while (cpu_report_state(cpu) != CPU_ONLINE)
 351                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 352
 353         return 0;
 354 }
 355
 356 #ifdef CONFIG_HOTPLUG_CPU
 357 static int xen_pv_cpu_disable(void)
 358 {
 359         unsigned int cpu = smp_processor_id();
 360         if (cpu == 0)
 361                 return -EBUSY;
 362
 363         cpu_disable_common();
 364
 365         load_cr3(swapper_pg_dir);
 366         return 0;
 367 }
 368
 369 static void xen_pv_cpu_die(unsigned int cpu)
 370 {
 371         while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 372                                   xen_vcpu_nr(cpu), NULL)) {
 373                 __set_current_state(TASK_UNINTERRUPTIBLE);
 374                 schedule_timeout(HZ/10);
 375         }
 376
 377         if (common_cpu_die(cpu) == 0) {
 378                 xen_smp_intr_free(cpu);
 379                 xen_uninit_lock_cpu(cpu);
 380                 xen_teardown_timer(cpu);
 381                 xen_pmu_finish(cpu);
 382         }
 383 }
 384
 385 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
 386 {
 387         play_dead_common();
 388         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 389         cpu_bringup();
 390         /*
 391          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 392          * clears certain data that the cpu_idle loop (which called us
 393          * and that we return from) expects. The only way to get that
 394          * data back is to call:
 395          */
 396         tick_nohz_idle_enter();
 397         tick_nohz_idle_stop_tick_protected();
 398
 399         cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
 400 }
 401
 402 #else /* !CONFIG_HOTPLUG_CPU */
 403 static int xen_pv_cpu_disable(void)
 404 {
 405         return -ENOSYS;
 406 }
 407
 408 static void xen_pv_cpu_die(unsigned int cpu)
 409 {
 410         BUG();
 411 }
 412
 413 static void xen_pv_play_dead(void)
 414 {
 415         BUG();
 416 }
 417
 418 #endif
 419 static void stop_self(void *v)
 420 {
 421         int cpu = smp_processor_id();
 422
 423         /* make sure we're not pinning something down */
 424         load_cr3(swapper_pg_dir);
 425         /* should set up a minimal gdt */
 426
 427         set_cpu_online(cpu, false);
 428
 429         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 430         BUG();
 431 }
 432
 433 static void xen_pv_stop_other_cpus(int wait)
 434 {
 435         smp_call_function(stop_self, NULL, wait);
 436 }
 437
 438 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 439 {
 440         irq_work_run();
 441         inc_irq_stat(apic_irq_work_irqs);
 442
 443         return IRQ_HANDLED;
 444 }
 445
 446 static const struct smp_ops xen_smp_ops __initconst = {
 447         .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 448         .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
 449         .smp_cpus_done = xen_smp_cpus_done,
 450
 451         .cpu_up = xen_pv_cpu_up,
 452         .cpu_die = xen_pv_cpu_die,
 453         .cpu_disable = xen_pv_cpu_disable,
 454         .play_dead = xen_pv_play_dead,
 455
 456         .stop_other_cpus = xen_pv_stop_other_cpus,
 457         .smp_send_reschedule = xen_smp_send_reschedule,
 458
 459         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 460         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 461 };
 462
 463 void __init xen_smp_init(void)
 464 {
 465         smp_ops = xen_smp_ops;
 466
 467         /* Avoid searching for BIOS MP tables */
 468         x86_init.mpparse.find_smp_config = x86_init_noop;
 469         x86_init.mpparse.get_smp_config = _get_smp_config;
 470 }