1 // SPDX-License-Identifier: GPL-2.0
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.
9 * IPIs are handled through the Xen event mechanism.
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
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>
27 #include <asm/paravirt.h>
29 #include <asm/pgtable.h>
31 #include <asm/fpu/internal.h>
33 #include <xen/interface/xen.h>
34 #include <xen/interface/vcpu.h>
35 #include <xen/interface/xenpmu.h>
37 #include <asm/spec-ctrl.h>
38 #include <asm/xen/interface.h>
39 #include <asm/xen/hypercall.h>
43 #include <xen/events.h>
45 #include <xen/hvc-console.h>
51 cpumask_var_t xen_cpu_initialized_map;
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 };
56 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
57 void asm_cpu_bringup_and_idle(void);
59 static void cpu_bringup(void)
66 touch_softlockup_watchdog();
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();
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);
79 speculative_store_bypass_ht_init();
81 xen_setup_cpu_clockevents();
83 notify_cpu_starting(cpu);
85 set_cpu_online(cpu, true);
87 cpu_set_state_online(cpu); /* Implies full memory barrier. */
89 /* We can take interrupts now: we're officially "up". */
93 asmlinkage __visible void cpu_bringup_and_idle(void)
96 boot_init_stack_canary();
97 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
98 prevent_tail_call_optimization();
101 void xen_smp_intr_free_pv(unsigned int cpu)
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;
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;
118 int xen_smp_intr_init_pv(unsigned int cpu)
121 char *callfunc_name, *pmu_name;
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,
127 xen_irq_work_interrupt,
128 IRQF_PERCPU|IRQF_NOBALANCING,
133 per_cpu(xen_irq_work, cpu).irq = rc;
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,
140 IRQF_PERCPU|IRQF_NOBALANCING,
144 per_cpu(xen_pmu_irq, cpu).irq = rc;
150 xen_smp_intr_free_pv(cpu);
154 static void __init xen_fill_possible_map(void)
158 if (xen_initial_domain())
161 for (i = 0; i < nr_cpu_ids; i++) {
162 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
165 set_cpu_possible(i, true);
170 static void __init xen_filter_cpu_maps(void)
173 unsigned int subtract = 0;
175 if (!xen_initial_domain())
180 for (i = 0; i < nr_cpu_ids; i++) {
181 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
184 set_cpu_possible(i, true);
186 set_cpu_possible(i, false);
187 set_cpu_present(i, false);
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. */
201 nr_cpu_ids = nr_cpu_ids - subtract;
206 static void __init xen_pv_smp_prepare_boot_cpu(void)
208 BUG_ON(smp_processor_id() != 0);
209 native_smp_prepare_boot_cpu();
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);
218 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
221 loadsegment(ds, __USER_DS);
222 loadsegment(es, __USER_DS);
225 xen_filter_cpu_maps();
226 xen_setup_vcpu_info_placement();
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.
234 xen_init_spinlocks();
237 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
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";
251 xen_init_lock_cpu(0);
253 smp_store_boot_cpu_info();
254 cpu_data(0).x86_max_cores = 1;
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);
262 set_cpu_sibling_map(0);
264 speculative_store_bypass_ht_init();
268 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
271 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
272 panic("could not allocate xen_cpu_initialized_map\n");
274 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
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--)
280 set_cpu_possible(cpu, false);
283 for_each_possible_cpu(cpu)
284 set_cpu_present(cpu, true);
288 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
290 struct vcpu_guest_context *ctxt;
291 struct desc_struct *gdt;
292 unsigned long gdt_mfn;
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))
299 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
303 gdt = get_cpu_gdt_rw(cpu);
306 ctxt->user_regs.fs = __KERNEL_PERCPU;
307 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
309 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
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
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);
325 xen_copy_trap_info(ctxt->trap_ctxt);
329 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
331 gdt_mfn = arbitrary_virt_to_mfn(gdt);
332 make_lowmem_page_readonly(gdt);
333 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
335 ctxt->gdt_frames[0] = gdt_mfn;
336 ctxt->gdt_ents = GDT_ENTRIES;
339 * Set SS:SP that Xen will use when entering guest kernel mode
340 * from guest user mode. Subsequent calls to load_sp0() can
343 ctxt->kernel_ss = __KERNEL_DS;
344 ctxt->kernel_sp = task_top_of_stack(idle);
347 ctxt->event_callback_cs = __KERNEL_CS;
348 ctxt->failsafe_callback_cs = __KERNEL_CS;
350 ctxt->gs_base_kernel = per_cpu_offset(cpu);
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);
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))
366 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
370 rc = common_cpu_up(cpu, idle);
374 xen_setup_runstate_info(cpu);
377 * PV VCPUs are always successfully taken down (see 'while' loop
378 * in xen_cpu_die()), so -EBUSY is an error.
380 rc = cpu_check_up_prepare(cpu);
384 /* make sure interrupts start blocked */
385 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
387 rc = cpu_initialize_context(cpu, idle);
393 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
396 while (cpu_report_state(cpu) != CPU_ONLINE)
397 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
402 #ifdef CONFIG_HOTPLUG_CPU
403 static int xen_pv_cpu_disable(void)
405 unsigned int cpu = smp_processor_id();
409 cpu_disable_common();
411 load_cr3(swapper_pg_dir);
415 static void xen_pv_cpu_die(unsigned int cpu)
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);
423 if (common_cpu_die(cpu) == 0) {
424 xen_smp_intr_free(cpu);
425 xen_uninit_lock_cpu(cpu);
426 xen_teardown_timer(cpu);
431 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
434 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
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:
442 tick_nohz_idle_enter();
443 tick_nohz_idle_stop_tick_protected();
445 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
448 #else /* !CONFIG_HOTPLUG_CPU */
449 static int xen_pv_cpu_disable(void)
454 static void xen_pv_cpu_die(unsigned int cpu)
459 static void xen_pv_play_dead(void)
465 static void stop_self(void *v)
467 int cpu = smp_processor_id();
469 /* make sure we're not pinning something down */
470 load_cr3(swapper_pg_dir);
471 /* should set up a minimal gdt */
473 set_cpu_online(cpu, false);
475 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
479 static void xen_pv_stop_other_cpus(int wait)
481 smp_call_function(stop_self, NULL, wait);
484 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
488 inc_irq_stat(apic_irq_work_irqs);
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,
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,
504 .stop_other_cpus = xen_pv_stop_other_cpus,
505 .smp_send_reschedule = xen_smp_send_reschedule,
507 .send_call_func_ipi = xen_smp_send_call_function_ipi,
508 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
511 void __init xen_smp_init(void)
513 smp_ops = xen_smp_ops;
514 xen_fill_possible_map();