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/apic.h>
 33#include <asm/io_apic.h>
 34
 35#include <xen/interface/xen.h>
 36#include <xen/interface/vcpu.h>
 37#include <xen/interface/xenpmu.h>
 38
 39#include <asm/spec-ctrl.h>
 40#include <asm/xen/interface.h>
 41#include <asm/xen/hypercall.h>
 42
 43#include <xen/xen.h>
 44#include <xen/page.h>
 45#include <xen/events.h>
 46
 47#include <xen/hvc-console.h>
 48#include "xen-ops.h"
 49
 50cpumask_var_t xen_cpu_initialized_map;
 51
 52static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
 53static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
 54
 55static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
 56
 57static void cpu_bringup(void)
 58{
 59	int cpu;
 60
 61	cr4_init();
 62	cpuhp_ap_sync_alive();
 63	cpu_init();
 64	fpu__init_cpu();
 65	touch_softlockup_watchdog();
 66
 67	/* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
 68	if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
 69		xen_enable_sysenter();
 70		xen_enable_syscall();
 71	}
 72	cpu = smp_processor_id();
 73	identify_secondary_cpu(cpu);
 74	set_cpu_sibling_map(cpu);
 75
 76	speculative_store_bypass_ht_init();
 77
 78	xen_setup_cpu_clockevents();
 79
 80	notify_cpu_starting(cpu);
 81
 82	set_cpu_online(cpu, true);
 83
 84	smp_mb();
 85
 86	/* We can take interrupts now: we're officially "up". */
 87	local_irq_enable();
 88}
 89
 90asmlinkage __visible void cpu_bringup_and_idle(void)
 91{
 92	cpu_bringup();
 93	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 94}
 95
 96void xen_smp_intr_free_pv(unsigned int cpu)
 97{
 98	kfree(per_cpu(xen_irq_work, cpu).name);
 99	per_cpu(xen_irq_work, cpu).name = NULL;
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	}
104
105	kfree(per_cpu(xen_pmu_irq, cpu).name);
106	per_cpu(xen_pmu_irq, cpu).name = NULL;
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	}
111}
112
113int xen_smp_intr_init_pv(unsigned int cpu)
114{
115	int rc;
116	char *callfunc_name, *pmu_name;
117
118	callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
119	per_cpu(xen_irq_work, cpu).name = callfunc_name;
120	rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
121				    cpu,
122				    xen_irq_work_interrupt,
123				    IRQF_PERCPU|IRQF_NOBALANCING,
124				    callfunc_name,
125				    NULL);
126	if (rc < 0)
127		goto fail;
128	per_cpu(xen_irq_work, cpu).irq = rc;
129
130	if (is_xen_pmu) {
131		pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
132		per_cpu(xen_pmu_irq, cpu).name = pmu_name;
133		rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
134					     xen_pmu_irq_handler,
135					     IRQF_PERCPU|IRQF_NOBALANCING,
136					     pmu_name, NULL);
137		if (rc < 0)
138			goto fail;
139		per_cpu(xen_pmu_irq, cpu).irq = rc;
140	}
141
142	return 0;
143
144 fail:
145	xen_smp_intr_free_pv(cpu);
146	return rc;
147}
148
149static void __init xen_pv_smp_config(void)
150{
151	u32 apicid = 0;
152	int i;
153
154	topology_register_boot_apic(apicid);
155
156	for (i = 0; i < nr_cpu_ids; i++)
157		topology_register_apic(apicid++, CPU_ACPIID_INVALID, true);
158
159	/* Pretend to be a proper enumerated system */
160	smp_found_config = 1;
161}
162
163static void __init xen_pv_smp_prepare_boot_cpu(void)
164{
165	BUG_ON(smp_processor_id() != 0);
166	native_smp_prepare_boot_cpu();
167
168	if (!xen_feature(XENFEAT_writable_page_tables))
169		/* We've switched to the "real" per-cpu gdt, so make
170		 * sure the old memory can be recycled. */
171		make_lowmem_page_readwrite(xen_initial_gdt);
172
173	xen_setup_vcpu_info_placement();
174
175	/*
176	 * The alternative logic (which patches the unlock/lock) runs before
177	 * the smp bootup up code is activated. Hence we need to set this up
178	 * the core kernel is being patched. Otherwise we will have only
179	 * modules patched but not core code.
180	 */
181	xen_init_spinlocks();
182}
183
184static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
185{
186	unsigned cpu;
187
188	if (ioapic_is_disabled) {
189		char *m = (max_cpus == 0) ?
190			"The nosmp parameter is incompatible with Xen; " \
191			"use Xen dom0_max_vcpus=1 parameter" :
192			"The noapic parameter is incompatible with Xen";
193
194		xen_raw_printk(m);
195		panic(m);
196	}
197	xen_init_lock_cpu(0);
198
199	smp_prepare_cpus_common();
200
201	speculative_store_bypass_ht_init();
202
203	xen_pmu_init(0);
204
205	if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
206		BUG();
207
208	if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
209		panic("could not allocate xen_cpu_initialized_map\n");
210
211	cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
212
213	/* Restrict the possible_map according to max_cpus. */
214	while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
215		for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
216			continue;
217		set_cpu_possible(cpu, false);
218	}
219
220	for_each_possible_cpu(cpu)
221		set_cpu_present(cpu, true);
222}
223
224static int
225cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
226{
227	struct vcpu_guest_context *ctxt;
228	struct desc_struct *gdt;
229	unsigned long gdt_mfn;
230
231	if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
232		return 0;
233
234	ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
235	if (ctxt == NULL) {
236		cpumask_clear_cpu(cpu, xen_cpu_initialized_map);
237		return -ENOMEM;
238	}
239
240	gdt = get_cpu_gdt_rw(cpu);
241
242	/*
243	 * Bring up the CPU in cpu_bringup_and_idle() with the stack
244	 * pointing just below where pt_regs would be if it were a normal
245	 * kernel entry.
246	 */
247	ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
248	ctxt->flags = VGCF_IN_KERNEL;
249	ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
250	ctxt->user_regs.ds = __USER_DS;
251	ctxt->user_regs.es = __USER_DS;
252	ctxt->user_regs.ss = __KERNEL_DS;
253	ctxt->user_regs.cs = __KERNEL_CS;
254	ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
255
256	xen_copy_trap_info(ctxt->trap_ctxt);
257
258	BUG_ON((unsigned long)gdt & ~PAGE_MASK);
259
260	gdt_mfn = arbitrary_virt_to_mfn(gdt);
261	make_lowmem_page_readonly(gdt);
262	make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
263
264	ctxt->gdt_frames[0] = gdt_mfn;
265	ctxt->gdt_ents      = GDT_ENTRIES;
266
267	/*
268	 * Set SS:SP that Xen will use when entering guest kernel mode
269	 * from guest user mode.  Subsequent calls to load_sp0() can
270	 * change this value.
271	 */
272	ctxt->kernel_ss = __KERNEL_DS;
273	ctxt->kernel_sp = task_top_of_stack(idle);
274
275	ctxt->gs_base_kernel = per_cpu_offset(cpu);
276	ctxt->event_callback_eip    =
277		(unsigned long)xen_asm_exc_xen_hypervisor_callback;
278	ctxt->failsafe_callback_eip =
279		(unsigned long)xen_failsafe_callback;
280	per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
281
282	ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
283	if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
284		BUG();
285
286	kfree(ctxt);
287	return 0;
288}
289
290static int xen_pv_kick_ap(unsigned int cpu, struct task_struct *idle)
291{
292	int rc;
293
294	rc = common_cpu_up(cpu, idle);
295	if (rc)
296		return rc;
297
298	xen_setup_runstate_info(cpu);
299
300	/* make sure interrupts start blocked */
301	per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
302
303	rc = cpu_initialize_context(cpu, idle);
304	if (rc)
305		return rc;
306
307	xen_pmu_init(cpu);
308
309	/*
310	 * Why is this a BUG? If the hypercall fails then everything can be
311	 * rolled back, no?
312	 */
313	BUG_ON(HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL));
314
315	return 0;
316}
317
318static void xen_pv_poll_sync_state(void)
319{
320	HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
321}
322
323#ifdef CONFIG_HOTPLUG_CPU
324static int xen_pv_cpu_disable(void)
325{
326	unsigned int cpu = smp_processor_id();
327	if (cpu == 0)
328		return -EBUSY;
329
330	cpu_disable_common();
331
332	load_cr3(swapper_pg_dir);
333	return 0;
334}
335
336static void xen_pv_cpu_die(unsigned int cpu)
337{
338	while (HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu), NULL)) {
339		__set_current_state(TASK_UNINTERRUPTIBLE);
340		schedule_timeout(HZ/10);
341	}
342}
343
344static void xen_pv_cleanup_dead_cpu(unsigned int cpu)
345{
346	xen_smp_intr_free(cpu);
347	xen_uninit_lock_cpu(cpu);
348	xen_teardown_timer(cpu);
349	xen_pmu_finish(cpu);
350}
351
352static void __noreturn xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
353{
354	play_dead_common();
355	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
356	xen_cpu_bringup_again((unsigned long)task_pt_regs(current));
357	BUG();
358}
359
360#else /* !CONFIG_HOTPLUG_CPU */
361static int xen_pv_cpu_disable(void)
362{
363	return -ENOSYS;
364}
365
366static void xen_pv_cpu_die(unsigned int cpu)
367{
368	BUG();
369}
370
371static void xen_pv_cleanup_dead_cpu(unsigned int cpu)
372{
373	BUG();
374}
375
376static void __noreturn xen_pv_play_dead(void)
377{
378	BUG();
379}
380
381#endif
382static void stop_self(void *v)
383{
384	int cpu = smp_processor_id();
385
386	/* make sure we're not pinning something down */
387	load_cr3(swapper_pg_dir);
388	/* should set up a minimal gdt */
389
390	set_cpu_online(cpu, false);
391
392	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
393	BUG();
394}
395
396static void xen_pv_stop_other_cpus(int wait)
397{
398	smp_call_function(stop_self, NULL, wait);
399}
400
401static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
402{
403	irq_work_run();
404	inc_irq_stat(apic_irq_work_irqs);
405
406	return IRQ_HANDLED;
407}
408
409void __init xen_smp_count_cpus(void)
410{
411	unsigned int cpus;
412
413	for (cpus = 0; cpus < nr_cpu_ids; cpus++) {
414		if (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpus, NULL) < 0)
415			break;
416	}
417
418	pr_info("Xen PV: Detected %u vCPUS\n", cpus);
419	if (cpus < nr_cpu_ids)
420		set_nr_cpu_ids(cpus);
421}
422
423static const struct smp_ops xen_smp_ops __initconst = {
424	.smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
425	.smp_prepare_cpus = xen_pv_smp_prepare_cpus,
426	.smp_cpus_done = xen_smp_cpus_done,
427
428	.kick_ap_alive = xen_pv_kick_ap,
429	.cpu_die = xen_pv_cpu_die,
430	.cleanup_dead_cpu = xen_pv_cleanup_dead_cpu,
431	.poll_sync_state = xen_pv_poll_sync_state,
432	.cpu_disable = xen_pv_cpu_disable,
433	.play_dead = xen_pv_play_dead,
434
435	.stop_other_cpus = xen_pv_stop_other_cpus,
436	.smp_send_reschedule = xen_smp_send_reschedule,
437
438	.send_call_func_ipi = xen_smp_send_call_function_ipi,
439	.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
440};
441
442void __init xen_smp_init(void)
443{
444	smp_ops = xen_smp_ops;
445
446	/* Avoid searching for BIOS MP tables */
447	x86_init.mpparse.find_mptable		= x86_init_noop;
448	x86_init.mpparse.early_parse_smp_cfg	= x86_init_noop;
449
450	/* XEN/PV Dom0 has halfways sane topology information via CPUID/MADT */
451	if (xen_initial_domain())
452		x86_init.mpparse.parse_smp_cfg	= x86_init_noop;
453	else
454		x86_init.mpparse.parse_smp_cfg	= xen_pv_smp_config;
455}