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>
26 #include <linux/pgtable.h>
28 #include <asm/paravirt.h>
29 #include <asm/idtentry.h>
32 #include <asm/io_apic.h>
33 #include <asm/fpu/internal.h>
35 #include <xen/interface/xen.h>
36 #include <xen/interface/vcpu.h>
37 #include <xen/interface/xenpmu.h>
39 #include <asm/spec-ctrl.h>
40 #include <asm/xen/interface.h>
41 #include <asm/xen/hypercall.h>
45 #include <xen/events.h>
47 #include <xen/hvc-console.h>
53 cpumask_var_t xen_cpu_initialized_map;
55 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
56 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
58 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
59 void asm_cpu_bringup_and_idle(void);
61 static void cpu_bringup(void)
68 touch_softlockup_watchdog();
71 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
72 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
73 xen_enable_sysenter();
76 cpu = smp_processor_id();
77 smp_store_cpu_info(cpu);
78 cpu_data(cpu).x86_max_cores = 1;
79 set_cpu_sibling_map(cpu);
81 speculative_store_bypass_ht_init();
83 xen_setup_cpu_clockevents();
85 notify_cpu_starting(cpu);
87 set_cpu_online(cpu, true);
89 cpu_set_state_online(cpu); /* Implies full memory barrier. */
91 /* We can take interrupts now: we're officially "up". */
95 asmlinkage __visible void cpu_bringup_and_idle(void)
98 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
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 _get_smp_config(unsigned int early)
157 unsigned int subtract = 0;
164 for (i = 0; i < nr_cpu_ids; i++) {
165 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
168 set_cpu_possible(i, true);
170 set_cpu_possible(i, false);
171 set_cpu_present(i, false);
175 #ifdef CONFIG_HOTPLUG_CPU
176 /* This is akin to using 'nr_cpus' on the Linux command line.
177 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
178 * have up to X, while nr_cpu_ids is greater than X. This
179 * normally is not a problem, except when CPU hotplugging
180 * is involved and then there might be more than X CPUs
181 * in the guest - which will not work as there is no
182 * hypercall to expand the max number of VCPUs an already
183 * running guest has. So cap it up to X. */
185 nr_cpu_ids = nr_cpu_ids - subtract;
190 static void __init xen_pv_smp_prepare_boot_cpu(void)
192 BUG_ON(smp_processor_id() != 0);
193 native_smp_prepare_boot_cpu();
195 if (!xen_feature(XENFEAT_writable_page_tables))
196 /* We've switched to the "real" per-cpu gdt, so make
197 * sure the old memory can be recycled. */
198 make_lowmem_page_readwrite(xen_initial_gdt);
200 xen_setup_vcpu_info_placement();
203 * The alternative logic (which patches the unlock/lock) runs before
204 * the smp bootup up code is activated. Hence we need to set this up
205 * the core kernel is being patched. Otherwise we will have only
206 * modules patched but not core code.
208 xen_init_spinlocks();
211 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
216 if (skip_ioapic_setup) {
217 char *m = (max_cpus == 0) ?
218 "The nosmp parameter is incompatible with Xen; " \
219 "use Xen dom0_max_vcpus=1 parameter" :
220 "The noapic parameter is incompatible with Xen";
225 xen_init_lock_cpu(0);
227 smp_store_boot_cpu_info();
228 cpu_data(0).x86_max_cores = 1;
230 for_each_possible_cpu(i) {
231 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
232 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
233 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
234 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
236 set_cpu_sibling_map(0);
238 speculative_store_bypass_ht_init();
242 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
245 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
246 panic("could not allocate xen_cpu_initialized_map\n");
248 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
250 /* Restrict the possible_map according to max_cpus. */
251 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
252 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
254 set_cpu_possible(cpu, false);
257 for_each_possible_cpu(cpu)
258 set_cpu_present(cpu, true);
262 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
264 struct vcpu_guest_context *ctxt;
265 struct desc_struct *gdt;
266 unsigned long gdt_mfn;
268 /* used to tell cpu_init() that it can proceed with initialization */
269 cpumask_set_cpu(cpu, cpu_callout_mask);
270 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
273 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
277 gdt = get_cpu_gdt_rw(cpu);
279 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
282 * Bring up the CPU in cpu_bringup_and_idle() with the stack
283 * pointing just below where pt_regs would be if it were a normal
286 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
287 ctxt->flags = VGCF_IN_KERNEL;
288 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
289 ctxt->user_regs.ds = __USER_DS;
290 ctxt->user_regs.es = __USER_DS;
291 ctxt->user_regs.ss = __KERNEL_DS;
292 ctxt->user_regs.cs = __KERNEL_CS;
293 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
295 xen_copy_trap_info(ctxt->trap_ctxt);
299 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
301 gdt_mfn = arbitrary_virt_to_mfn(gdt);
302 make_lowmem_page_readonly(gdt);
303 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
305 ctxt->gdt_frames[0] = gdt_mfn;
306 ctxt->gdt_ents = GDT_ENTRIES;
309 * Set SS:SP that Xen will use when entering guest kernel mode
310 * from guest user mode. Subsequent calls to load_sp0() can
313 ctxt->kernel_ss = __KERNEL_DS;
314 ctxt->kernel_sp = task_top_of_stack(idle);
316 ctxt->gs_base_kernel = per_cpu_offset(cpu);
317 ctxt->event_callback_eip =
318 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
319 ctxt->failsafe_callback_eip =
320 (unsigned long)xen_failsafe_callback;
321 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
323 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
324 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
331 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
335 rc = common_cpu_up(cpu, idle);
339 xen_setup_runstate_info(cpu);
342 * PV VCPUs are always successfully taken down (see 'while' loop
343 * in xen_cpu_die()), so -EBUSY is an error.
345 rc = cpu_check_up_prepare(cpu);
349 /* make sure interrupts start blocked */
350 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
352 rc = cpu_initialize_context(cpu, idle);
358 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
361 while (cpu_report_state(cpu) != CPU_ONLINE)
362 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
367 #ifdef CONFIG_HOTPLUG_CPU
368 static int xen_pv_cpu_disable(void)
370 unsigned int cpu = smp_processor_id();
374 cpu_disable_common();
376 load_cr3(swapper_pg_dir);
380 static void xen_pv_cpu_die(unsigned int cpu)
382 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
383 xen_vcpu_nr(cpu), NULL)) {
384 __set_current_state(TASK_UNINTERRUPTIBLE);
385 schedule_timeout(HZ/10);
388 if (common_cpu_die(cpu) == 0) {
389 xen_smp_intr_free(cpu);
390 xen_uninit_lock_cpu(cpu);
391 xen_teardown_timer(cpu);
396 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
399 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
402 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
403 * clears certain data that the cpu_idle loop (which called us
404 * and that we return from) expects. The only way to get that
405 * data back is to call:
407 tick_nohz_idle_enter();
408 tick_nohz_idle_stop_tick_protected();
410 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
413 #else /* !CONFIG_HOTPLUG_CPU */
414 static int xen_pv_cpu_disable(void)
419 static void xen_pv_cpu_die(unsigned int cpu)
424 static void xen_pv_play_dead(void)
430 static void stop_self(void *v)
432 int cpu = smp_processor_id();
434 /* make sure we're not pinning something down */
435 load_cr3(swapper_pg_dir);
436 /* should set up a minimal gdt */
438 set_cpu_online(cpu, false);
440 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
444 static void xen_pv_stop_other_cpus(int wait)
446 smp_call_function(stop_self, NULL, wait);
449 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
453 inc_irq_stat(apic_irq_work_irqs);
459 static const struct smp_ops xen_smp_ops __initconst = {
460 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
461 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
462 .smp_cpus_done = xen_smp_cpus_done,
464 .cpu_up = xen_pv_cpu_up,
465 .cpu_die = xen_pv_cpu_die,
466 .cpu_disable = xen_pv_cpu_disable,
467 .play_dead = xen_pv_play_dead,
469 .stop_other_cpus = xen_pv_stop_other_cpus,
470 .smp_send_reschedule = xen_smp_send_reschedule,
472 .send_call_func_ipi = xen_smp_send_call_function_ipi,
473 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
476 void __init xen_smp_init(void)
478 smp_ops = xen_smp_ops;
480 /* Avoid searching for BIOS MP tables */
481 x86_init.mpparse.find_smp_config = x86_init_noop;
482 x86_init.mpparse.get_smp_config = _get_smp_config;