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
  27 #include <asm/paravirt.h>
  28 #include <asm/desc.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/cpu.h>
  31 #include <asm/fpu/internal.h>
  32
  33 #include <xen/interface/xen.h>
  34 #include <xen/interface/vcpu.h>
  35 #include <xen/interface/xenpmu.h>
  36
  37 #include <asm/spec-ctrl.h>
  38 #include <asm/xen/interface.h>
  39 #include <asm/xen/hypercall.h>
  40
  41 #include <xen/xen.h>
  42 #include <xen/page.h>
  43 #include <xen/events.h>
  44
  45 #include <xen/hvc-console.h>
  46 #include "xen-ops.h"
  47 #include "mmu.h"
  48 #include "smp.h"
  49 #include "pmu.h"
  50
  51 cpumask_var_t xen_cpu_initialized_map;
  52
  53 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
  54 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
  55
  56 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
  57 void asm_cpu_bringup_and_idle(void);
  58
  59 static void cpu_bringup(void)
  60 {
  61         int cpu;
  62
  63         cr4_init();
  64         cpu_init();
  65         fpu__init_cpu();
  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         boot_init_stack_canary();
  97         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
  98         prevent_tail_call_optimization();
  99 }
 100
 101 void xen_smp_intr_free_pv(unsigned int cpu)
 102 {
 103         kfree(per_cpu(xen_irq_work, cpu).name);
 104         per_cpu(xen_irq_work, cpu).name = NULL;
 105         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 106                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 107                 per_cpu(xen_irq_work, cpu).irq = -1;
 108         }
 109
 110         kfree(per_cpu(xen_pmu_irq, cpu).name);
 111         per_cpu(xen_pmu_irq, cpu).name = NULL;
 112         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 113                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 114                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 115         }
 116 }
 117
 118 int xen_smp_intr_init_pv(unsigned int cpu)
 119 {
 120         int rc;
 121         char *callfunc_name, *pmu_name;
 122
 123         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
 124         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 125         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
 126                                     cpu,
 127                                     xen_irq_work_interrupt,
 128                                     IRQF_PERCPU|IRQF_NOBALANCING,
 129                                     callfunc_name,
 130                                     NULL);
 131         if (rc < 0)
 132                 goto fail;
 133         per_cpu(xen_irq_work, cpu).irq = rc;
 134
 135         if (is_xen_pmu) {
 136                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 137                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 138                 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
 139                                              xen_pmu_irq_handler,
 140                                              IRQF_PERCPU|IRQF_NOBALANCING,
 141                                              pmu_name, NULL);
 142                 if (rc < 0)
 143                         goto fail;
 144                 per_cpu(xen_pmu_irq, cpu).irq = rc;
 145         }
 146
 147         return 0;
 148
 149  fail:
 150         xen_smp_intr_free_pv(cpu);
 151         return rc;
 152 }
 153
 154 static void __init xen_fill_possible_map(void)
 155 {
 156         int i, rc;
 157
 158         if (xen_initial_domain())
 159                 return;
 160
 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                 }
 167         }
 168 }
 169
 170 static void __init xen_filter_cpu_maps(void)
 171 {
 172         int i, rc;
 173         unsigned int subtract = 0;
 174
 175         if (!xen_initial_domain())
 176                 return;
 177
 178         num_processors = 0;
 179         disabled_cpus = 0;
 180         for (i = 0; i < nr_cpu_ids; i++) {
 181                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 182                 if (rc >= 0) {
 183                         num_processors++;
 184                         set_cpu_possible(i, true);
 185                 } else {
 186                         set_cpu_possible(i, false);
 187                         set_cpu_present(i, false);
 188                         subtract++;
 189                 }
 190         }
 191 #ifdef CONFIG_HOTPLUG_CPU
 192         /* This is akin to using 'nr_cpus' on the Linux command line.
 193          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 194          * have up to X, while nr_cpu_ids is greater than X. This
 195          * normally is not a problem, except when CPU hotplugging
 196          * is involved and then there might be more than X CPUs
 197          * in the guest - which will not work as there is no
 198          * hypercall to expand the max number of VCPUs an already
 199          * running guest has. So cap it up to X. */
 200         if (subtract)
 201                 nr_cpu_ids = nr_cpu_ids - subtract;
 202 #endif
 203
 204 }
 205
 206 static void __init xen_pv_smp_prepare_boot_cpu(void)
 207 {
 208         BUG_ON(smp_processor_id() != 0);
 209         native_smp_prepare_boot_cpu();
 210
 211         if (!xen_feature(XENFEAT_writable_page_tables))
 212                 /* We've switched to the "real" per-cpu gdt, so make
 213                  * sure the old memory can be recycled. */
 214                 make_lowmem_page_readwrite(xen_initial_gdt);
 215
 216 #ifdef CONFIG_X86_32
 217         /*
 218          * Xen starts us with XEN_FLAT_RING1_DS, but linux code
 219          * expects __USER_DS
 220          */
 221         loadsegment(ds, __USER_DS);
 222         loadsegment(es, __USER_DS);
 223 #endif
 224
 225         xen_filter_cpu_maps();
 226         xen_setup_vcpu_info_placement();
 227
 228         /*
 229          * The alternative logic (which patches the unlock/lock) runs before
 230          * the smp bootup up code is activated. Hence we need to set this up
 231          * the core kernel is being patched. Otherwise we will have only
 232          * modules patched but not core code.
 233          */
 234         xen_init_spinlocks();
 235 }
 236
 237 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
 238 {
 239         unsigned cpu;
 240         unsigned int i;
 241
 242         if (skip_ioapic_setup) {
 243                 char *m = (max_cpus == 0) ?
 244                         "The nosmp parameter is incompatible with Xen; " \
 245                         "use Xen dom0_max_vcpus=1 parameter" :
 246                         "The noapic parameter is incompatible with Xen";
 247
 248                 xen_raw_printk(m);
 249                 panic(m);
 250         }
 251         xen_init_lock_cpu(0);
 252
 253         smp_store_boot_cpu_info();
 254         cpu_data(0).x86_max_cores = 1;
 255
 256         for_each_possible_cpu(i) {
 257                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 258                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 259                 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
 260                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 261         }
 262         set_cpu_sibling_map(0);
 263
 264         speculative_store_bypass_ht_init();
 265
 266         xen_pmu_init(0);
 267
 268         if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
 269                 BUG();
 270
 271         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 272                 panic("could not allocate xen_cpu_initialized_map\n");
 273
 274         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 275
 276         /* Restrict the possible_map according to max_cpus. */
 277         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 278                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 279                         continue;
 280                 set_cpu_possible(cpu, false);
 281         }
 282
 283         for_each_possible_cpu(cpu)
 284                 set_cpu_present(cpu, true);
 285 }
 286
 287 static int
 288 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 289 {
 290         struct vcpu_guest_context *ctxt;
 291         struct desc_struct *gdt;
 292         unsigned long gdt_mfn;
 293
 294         /* used to tell cpu_init() that it can proceed with initialization */
 295         cpumask_set_cpu(cpu, cpu_callout_mask);
 296         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 297                 return 0;
 298
 299         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 300         if (ctxt == NULL)
 301                 return -ENOMEM;
 302
 303         gdt = get_cpu_gdt_rw(cpu);
 304
 305 #ifdef CONFIG_X86_32
 306         ctxt->user_regs.fs = __KERNEL_PERCPU;
 307         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 308 #endif
 309         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 310
 311         /*
 312          * Bring up the CPU in cpu_bringup_and_idle() with the stack
 313          * pointing just below where pt_regs would be if it were a normal
 314          * kernel entry.
 315          */
 316         ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
 317         ctxt->flags = VGCF_IN_KERNEL;
 318         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 319         ctxt->user_regs.ds = __USER_DS;
 320         ctxt->user_regs.es = __USER_DS;
 321         ctxt->user_regs.ss = __KERNEL_DS;
 322         ctxt->user_regs.cs = __KERNEL_CS;
 323         ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
 324
 325         xen_copy_trap_info(ctxt->trap_ctxt);
 326
 327         ctxt->ldt_ents = 0;
 328
 329         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 330
 331         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 332         make_lowmem_page_readonly(gdt);
 333         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 334
 335         ctxt->gdt_frames[0] = gdt_mfn;
 336         ctxt->gdt_ents      = GDT_ENTRIES;
 337
 338         /*
 339          * Set SS:SP that Xen will use when entering guest kernel mode
 340          * from guest user mode.  Subsequent calls to load_sp0() can
 341          * change this value.
 342          */
 343         ctxt->kernel_ss = __KERNEL_DS;
 344         ctxt->kernel_sp = task_top_of_stack(idle);
 345
 346 #ifdef CONFIG_X86_32
 347         ctxt->event_callback_cs     = __KERNEL_CS;
 348         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 349 #else
 350         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 351 #endif
 352         ctxt->event_callback_eip    =
 353                 (unsigned long)xen_hypervisor_callback;
 354         ctxt->failsafe_callback_eip =
 355                 (unsigned long)xen_failsafe_callback;
 356         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 357
 358         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 359         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 360                 BUG();
 361
 362         kfree(ctxt);
 363         return 0;
 364 }
 365
 366 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
 367 {
 368         int rc;
 369
 370         rc = common_cpu_up(cpu, idle);
 371         if (rc)
 372                 return rc;
 373
 374         xen_setup_runstate_info(cpu);
 375
 376         /*
 377          * PV VCPUs are always successfully taken down (see 'while' loop
 378          * in xen_cpu_die()), so -EBUSY is an error.
 379          */
 380         rc = cpu_check_up_prepare(cpu);
 381         if (rc)
 382                 return rc;
 383
 384         /* make sure interrupts start blocked */
 385         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 386
 387         rc = cpu_initialize_context(cpu, idle);
 388         if (rc)
 389                 return rc;
 390
 391         xen_pmu_init(cpu);
 392
 393         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 394         BUG_ON(rc);
 395
 396         while (cpu_report_state(cpu) != CPU_ONLINE)
 397                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 398
 399         return 0;
 400 }
 401
 402 #ifdef CONFIG_HOTPLUG_CPU
 403 static int xen_pv_cpu_disable(void)
 404 {
 405         unsigned int cpu = smp_processor_id();
 406         if (cpu == 0)
 407                 return -EBUSY;
 408
 409         cpu_disable_common();
 410
 411         load_cr3(swapper_pg_dir);
 412         return 0;
 413 }
 414
 415 static void xen_pv_cpu_die(unsigned int cpu)
 416 {
 417         while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 418                                   xen_vcpu_nr(cpu), NULL)) {
 419                 __set_current_state(TASK_UNINTERRUPTIBLE);
 420                 schedule_timeout(HZ/10);
 421         }
 422
 423         if (common_cpu_die(cpu) == 0) {
 424                 xen_smp_intr_free(cpu);
 425                 xen_uninit_lock_cpu(cpu);
 426                 xen_teardown_timer(cpu);
 427                 xen_pmu_finish(cpu);
 428         }
 429 }
 430
 431 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
 432 {
 433         play_dead_common();
 434         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 435         cpu_bringup();
 436         /*
 437          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 438          * clears certain data that the cpu_idle loop (which called us
 439          * and that we return from) expects. The only way to get that
 440          * data back is to call:
 441          */
 442         tick_nohz_idle_enter();
 443         tick_nohz_idle_stop_tick_protected();
 444
 445         cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
 446 }
 447
 448 #else /* !CONFIG_HOTPLUG_CPU */
 449 static int xen_pv_cpu_disable(void)
 450 {
 451         return -ENOSYS;
 452 }
 453
 454 static void xen_pv_cpu_die(unsigned int cpu)
 455 {
 456         BUG();
 457 }
 458
 459 static void xen_pv_play_dead(void)
 460 {
 461         BUG();
 462 }
 463
 464 #endif
 465 static void stop_self(void *v)
 466 {
 467         int cpu = smp_processor_id();
 468
 469         /* make sure we're not pinning something down */
 470         load_cr3(swapper_pg_dir);
 471         /* should set up a minimal gdt */
 472
 473         set_cpu_online(cpu, false);
 474
 475         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 476         BUG();
 477 }
 478
 479 static void xen_pv_stop_other_cpus(int wait)
 480 {
 481         smp_call_function(stop_self, NULL, wait);
 482 }
 483
 484 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 485 {
 486         irq_enter();
 487         irq_work_run();
 488         inc_irq_stat(apic_irq_work_irqs);
 489         irq_exit();
 490
 491         return IRQ_HANDLED;
 492 }
 493
 494 static const struct smp_ops xen_smp_ops __initconst = {
 495         .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 496         .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
 497         .smp_cpus_done = xen_smp_cpus_done,
 498
 499         .cpu_up = xen_pv_cpu_up,
 500         .cpu_die = xen_pv_cpu_die,
 501         .cpu_disable = xen_pv_cpu_disable,
 502         .play_dead = xen_pv_play_dead,
 503
 504         .stop_other_cpus = xen_pv_stop_other_cpus,
 505         .smp_send_reschedule = xen_smp_send_reschedule,
 506
 507         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 508         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 509 };
 510
 511 void __init xen_smp_init(void)
 512 {
 513         smp_ops = xen_smp_ops;
 514         xen_fill_possible_map();
 515 }