linux/arch/x86/xen/smp_pv.c
<<
>>
Prefs
   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
  52cpumask_var_t xen_cpu_initialized_map;
  53
  54static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
  55static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
  56
  57static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
  58void asm_cpu_bringup_and_idle(void);
  59
  60static 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
  92asmlinkage __visible void cpu_bringup_and_idle(void)
  93{
  94        cpu_bringup();
  95        cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
  96}
  97
  98void 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
 115int 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(cpu)) {
 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
 151static void __init xen_fill_possible_map(void)
 152{
 153        int i, rc;
 154
 155        if (xen_initial_domain())
 156                return;
 157
 158        for (i = 0; i < nr_cpu_ids; i++) {
 159                rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 160                if (rc >= 0) {
 161                        num_processors++;
 162                        set_cpu_possible(i, true);
 163                }
 164        }
 165}
 166
 167static void __init xen_filter_cpu_maps(void)
 168{
 169        int i, rc;
 170        unsigned int subtract = 0;
 171
 172        if (!xen_initial_domain())
 173                return;
 174
 175        num_processors = 0;
 176        disabled_cpus = 0;
 177        for (i = 0; i < nr_cpu_ids; i++) {
 178                rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 179                if (rc >= 0) {
 180                        num_processors++;
 181                        set_cpu_possible(i, true);
 182                } else {
 183                        set_cpu_possible(i, false);
 184                        set_cpu_present(i, false);
 185                        subtract++;
 186                }
 187        }
 188#ifdef CONFIG_HOTPLUG_CPU
 189        /* This is akin to using 'nr_cpus' on the Linux command line.
 190         * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 191         * have up to X, while nr_cpu_ids is greater than X. This
 192         * normally is not a problem, except when CPU hotplugging
 193         * is involved and then there might be more than X CPUs
 194         * in the guest - which will not work as there is no
 195         * hypercall to expand the max number of VCPUs an already
 196         * running guest has. So cap it up to X. */
 197        if (subtract)
 198                nr_cpu_ids = nr_cpu_ids - subtract;
 199#endif
 200
 201}
 202
 203static void __init xen_pv_smp_prepare_boot_cpu(void)
 204{
 205        BUG_ON(smp_processor_id() != 0);
 206        native_smp_prepare_boot_cpu();
 207
 208        if (!xen_feature(XENFEAT_writable_page_tables))
 209                /* We've switched to the "real" per-cpu gdt, so make
 210                 * sure the old memory can be recycled. */
 211                make_lowmem_page_readwrite(xen_initial_gdt);
 212
 213        xen_filter_cpu_maps();
 214        xen_setup_vcpu_info_placement();
 215
 216        /*
 217         * The alternative logic (which patches the unlock/lock) runs before
 218         * the smp bootup up code is activated. Hence we need to set this up
 219         * the core kernel is being patched. Otherwise we will have only
 220         * modules patched but not core code.
 221         */
 222        xen_init_spinlocks();
 223}
 224
 225static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
 226{
 227        unsigned cpu;
 228        unsigned int i;
 229
 230        if (skip_ioapic_setup) {
 231                char *m = (max_cpus == 0) ?
 232                        "The nosmp parameter is incompatible with Xen; " \
 233                        "use Xen dom0_max_vcpus=1 parameter" :
 234                        "The noapic parameter is incompatible with Xen";
 235
 236                xen_raw_printk(m);
 237                panic(m);
 238        }
 239        xen_init_lock_cpu(0);
 240
 241        smp_store_boot_cpu_info();
 242        cpu_data(0).x86_max_cores = 1;
 243
 244        for_each_possible_cpu(i) {
 245                zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 246                zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 247                zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
 248                zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 249        }
 250        set_cpu_sibling_map(0);
 251
 252        speculative_store_bypass_ht_init();
 253
 254        xen_pmu_init(0);
 255
 256        if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
 257                BUG();
 258
 259        if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 260                panic("could not allocate xen_cpu_initialized_map\n");
 261
 262        cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 263
 264        /* Restrict the possible_map according to max_cpus. */
 265        while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 266                for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 267                        continue;
 268                set_cpu_possible(cpu, false);
 269        }
 270
 271        for_each_possible_cpu(cpu)
 272                set_cpu_present(cpu, true);
 273}
 274
 275static int
 276cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 277{
 278        struct vcpu_guest_context *ctxt;
 279        struct desc_struct *gdt;
 280        unsigned long gdt_mfn;
 281
 282        /* used to tell cpu_init() that it can proceed with initialization */
 283        cpumask_set_cpu(cpu, cpu_callout_mask);
 284        if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 285                return 0;
 286
 287        ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 288        if (ctxt == NULL)
 289                return -ENOMEM;
 290
 291        gdt = get_cpu_gdt_rw(cpu);
 292
 293        /*
 294         * Bring up the CPU in cpu_bringup_and_idle() with the stack
 295         * pointing just below where pt_regs would be if it were a normal
 296         * kernel entry.
 297         */
 298        ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
 299        ctxt->flags = VGCF_IN_KERNEL;
 300        ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 301        ctxt->user_regs.ds = __USER_DS;
 302        ctxt->user_regs.es = __USER_DS;
 303        ctxt->user_regs.ss = __KERNEL_DS;
 304        ctxt->user_regs.cs = __KERNEL_CS;
 305        ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
 306
 307        xen_copy_trap_info(ctxt->trap_ctxt);
 308
 309        BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 310
 311        gdt_mfn = arbitrary_virt_to_mfn(gdt);
 312        make_lowmem_page_readonly(gdt);
 313        make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 314
 315        ctxt->gdt_frames[0] = gdt_mfn;
 316        ctxt->gdt_ents      = GDT_ENTRIES;
 317
 318        /*
 319         * Set SS:SP that Xen will use when entering guest kernel mode
 320         * from guest user mode.  Subsequent calls to load_sp0() can
 321         * change this value.
 322         */
 323        ctxt->kernel_ss = __KERNEL_DS;
 324        ctxt->kernel_sp = task_top_of_stack(idle);
 325
 326        ctxt->gs_base_kernel = per_cpu_offset(cpu);
 327        ctxt->event_callback_eip    =
 328                (unsigned long)xen_asm_exc_xen_hypervisor_callback;
 329        ctxt->failsafe_callback_eip =
 330                (unsigned long)xen_failsafe_callback;
 331        per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 332
 333        ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 334        if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 335                BUG();
 336
 337        kfree(ctxt);
 338        return 0;
 339}
 340
 341static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
 342{
 343        int rc;
 344
 345        rc = common_cpu_up(cpu, idle);
 346        if (rc)
 347                return rc;
 348
 349        xen_setup_runstate_info(cpu);
 350
 351        /*
 352         * PV VCPUs are always successfully taken down (see 'while' loop
 353         * in xen_cpu_die()), so -EBUSY is an error.
 354         */
 355        rc = cpu_check_up_prepare(cpu);
 356        if (rc)
 357                return rc;
 358
 359        /* make sure interrupts start blocked */
 360        per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 361
 362        rc = cpu_initialize_context(cpu, idle);
 363        if (rc)
 364                return rc;
 365
 366        xen_pmu_init(cpu);
 367
 368        rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 369        BUG_ON(rc);
 370
 371        while (cpu_report_state(cpu) != CPU_ONLINE)
 372                HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 373
 374        return 0;
 375}
 376
 377#ifdef CONFIG_HOTPLUG_CPU
 378static int xen_pv_cpu_disable(void)
 379{
 380        unsigned int cpu = smp_processor_id();
 381        if (cpu == 0)
 382                return -EBUSY;
 383
 384        cpu_disable_common();
 385
 386        load_cr3(swapper_pg_dir);
 387        return 0;
 388}
 389
 390static void xen_pv_cpu_die(unsigned int cpu)
 391{
 392        while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 393                                  xen_vcpu_nr(cpu), NULL)) {
 394                __set_current_state(TASK_UNINTERRUPTIBLE);
 395                schedule_timeout(HZ/10);
 396        }
 397
 398        if (common_cpu_die(cpu) == 0) {
 399                xen_smp_intr_free(cpu);
 400                xen_uninit_lock_cpu(cpu);
 401                xen_teardown_timer(cpu);
 402                xen_pmu_finish(cpu);
 403        }
 404}
 405
 406static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
 407{
 408        play_dead_common();
 409        HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 410        cpu_bringup();
 411        /*
 412         * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 413         * clears certain data that the cpu_idle loop (which called us
 414         * and that we return from) expects. The only way to get that
 415         * data back is to call:
 416         */
 417        tick_nohz_idle_enter();
 418        tick_nohz_idle_stop_tick_protected();
 419
 420        cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
 421}
 422
 423#else /* !CONFIG_HOTPLUG_CPU */
 424static int xen_pv_cpu_disable(void)
 425{
 426        return -ENOSYS;
 427}
 428
 429static void xen_pv_cpu_die(unsigned int cpu)
 430{
 431        BUG();
 432}
 433
 434static void xen_pv_play_dead(void)
 435{
 436        BUG();
 437}
 438
 439#endif
 440static void stop_self(void *v)
 441{
 442        int cpu = smp_processor_id();
 443
 444        /* make sure we're not pinning something down */
 445        load_cr3(swapper_pg_dir);
 446        /* should set up a minimal gdt */
 447
 448        set_cpu_online(cpu, false);
 449
 450        HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 451        BUG();
 452}
 453
 454static void xen_pv_stop_other_cpus(int wait)
 455{
 456        smp_call_function(stop_self, NULL, wait);
 457}
 458
 459static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 460{
 461        irq_enter();
 462        irq_work_run();
 463        inc_irq_stat(apic_irq_work_irqs);
 464        irq_exit();
 465
 466        return IRQ_HANDLED;
 467}
 468
 469static const struct smp_ops xen_smp_ops __initconst = {
 470        .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 471        .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
 472        .smp_cpus_done = xen_smp_cpus_done,
 473
 474        .cpu_up = xen_pv_cpu_up,
 475        .cpu_die = xen_pv_cpu_die,
 476        .cpu_disable = xen_pv_cpu_disable,
 477        .play_dead = xen_pv_play_dead,
 478
 479        .stop_other_cpus = xen_pv_stop_other_cpus,
 480        .smp_send_reschedule = xen_smp_send_reschedule,
 481
 482        .send_call_func_ipi = xen_smp_send_call_function_ipi,
 483        .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 484};
 485
 486void __init xen_smp_init(void)
 487{
 488        smp_ops = xen_smp_ops;
 489        xen_fill_possible_map();
 490}
 491