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
 26#include <asm/paravirt.h>
 27#include <asm/desc.h>
 28#include <asm/pgtable.h>
 29#include <asm/cpu.h>
 30
 31#include <xen/interface/xen.h>
 32#include <xen/interface/vcpu.h>
 33#include <xen/interface/xenpmu.h>
 34
 35#include <asm/xen/interface.h>
 36#include <asm/xen/hypercall.h>
 37
 38#include <xen/xen.h>
 39#include <xen/page.h>
 40#include <xen/events.h>
 41
 42#include <xen/hvc-console.h>
 43#include "xen-ops.h"
 44#include "mmu.h"
 45#include "smp.h"
 46#include "pmu.h"
 47
 48cpumask_var_t xen_cpu_initialized_map;
 49
 50static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
 51static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
 52
 53static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
 54
 55static void cpu_bringup(void)
 56{
 57	int cpu;
 58
 59	cpu_init();
 60	touch_softlockup_watchdog();
 61	preempt_disable();
 62
 63	/* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
 64	if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
 65		xen_enable_sysenter();
 66		xen_enable_syscall();
 67	}
 68	cpu = smp_processor_id();
 69	smp_store_cpu_info(cpu);
 70	cpu_data(cpu).x86_max_cores = 1;
 71	set_cpu_sibling_map(cpu);
 72
 73	xen_setup_cpu_clockevents();
 74
 75	notify_cpu_starting(cpu);
 76
 77	set_cpu_online(cpu, true);
 78
 79	cpu_set_state_online(cpu);  /* Implies full memory barrier. */
 80
 81	/* We can take interrupts now: we're officially "up". */
 82	local_irq_enable();
 83}
 84
 85asmlinkage __visible void cpu_bringup_and_idle(void)
 86{
 87	cpu_bringup();
 88	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 89}
 90
 91void xen_smp_intr_free_pv(unsigned int cpu)
 92{
 93	if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 94		unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 95		per_cpu(xen_irq_work, cpu).irq = -1;
 96		kfree(per_cpu(xen_irq_work, cpu).name);
 97		per_cpu(xen_irq_work, cpu).name = NULL;
 98	}
 99
100	if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
101		unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
102		per_cpu(xen_pmu_irq, cpu).irq = -1;
103		kfree(per_cpu(xen_pmu_irq, cpu).name);
104		per_cpu(xen_pmu_irq, cpu).name = NULL;
105	}
106}
107
108int xen_smp_intr_init_pv(unsigned int cpu)
109{
110	int rc;
111	char *callfunc_name, *pmu_name;
112
113	callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
114	rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
115				    cpu,
116				    xen_irq_work_interrupt,
117				    IRQF_PERCPU|IRQF_NOBALANCING,
118				    callfunc_name,
119				    NULL);
120	if (rc < 0)
121		goto fail;
122	per_cpu(xen_irq_work, cpu).irq = rc;
123	per_cpu(xen_irq_work, cpu).name = callfunc_name;
124
125	if (is_xen_pmu(cpu)) {
126		pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
127		rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
128					     xen_pmu_irq_handler,
129					     IRQF_PERCPU|IRQF_NOBALANCING,
130					     pmu_name, NULL);
131		if (rc < 0)
132			goto fail;
133		per_cpu(xen_pmu_irq, cpu).irq = rc;
134		per_cpu(xen_pmu_irq, cpu).name = pmu_name;
135	}
136
137	return 0;
138
139 fail:
140	xen_smp_intr_free_pv(cpu);
141	return rc;
142}
143
144static void __init xen_fill_possible_map(void)
145{
146	int i, rc;
147
148	if (xen_initial_domain())
149		return;
150
151	for (i = 0; i < nr_cpu_ids; i++) {
152		rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
153		if (rc >= 0) {
154			num_processors++;
155			set_cpu_possible(i, true);
156		}
157	}
158}
159
160static void __init xen_filter_cpu_maps(void)
161{
162	int i, rc;
163	unsigned int subtract = 0;
164
165	if (!xen_initial_domain())
166		return;
167
168	num_processors = 0;
169	disabled_cpus = 0;
170	for (i = 0; i < nr_cpu_ids; i++) {
171		rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
172		if (rc >= 0) {
173			num_processors++;
174			set_cpu_possible(i, true);
175		} else {
176			set_cpu_possible(i, false);
177			set_cpu_present(i, false);
178			subtract++;
179		}
180	}
181#ifdef CONFIG_HOTPLUG_CPU
182	/* This is akin to using 'nr_cpus' on the Linux command line.
183	 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
184	 * have up to X, while nr_cpu_ids is greater than X. This
185	 * normally is not a problem, except when CPU hotplugging
186	 * is involved and then there might be more than X CPUs
187	 * in the guest - which will not work as there is no
188	 * hypercall to expand the max number of VCPUs an already
189	 * running guest has. So cap it up to X. */
190	if (subtract)
191		nr_cpu_ids = nr_cpu_ids - subtract;
192#endif
193
194}
195
196static void __init xen_pv_smp_prepare_boot_cpu(void)
197{
198	BUG_ON(smp_processor_id() != 0);
199	native_smp_prepare_boot_cpu();
200
201	if (!xen_feature(XENFEAT_writable_page_tables))
202		/* We've switched to the "real" per-cpu gdt, so make
203		 * sure the old memory can be recycled. */
204		make_lowmem_page_readwrite(xen_initial_gdt);
205
206#ifdef CONFIG_X86_32
207	/*
208	 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
209	 * expects __USER_DS
210	 */
211	loadsegment(ds, __USER_DS);
212	loadsegment(es, __USER_DS);
213#endif
214
215	xen_filter_cpu_maps();
216	xen_setup_vcpu_info_placement();
217
218	/*
219	 * The alternative logic (which patches the unlock/lock) runs before
220	 * the smp bootup up code is activated. Hence we need to set this up
221	 * the core kernel is being patched. Otherwise we will have only
222	 * modules patched but not core code.
223	 */
224	xen_init_spinlocks();
225}
226
227static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
228{
229	unsigned cpu;
230	unsigned int i;
231
232	if (skip_ioapic_setup) {
233		char *m = (max_cpus == 0) ?
234			"The nosmp parameter is incompatible with Xen; " \
235			"use Xen dom0_max_vcpus=1 parameter" :
236			"The noapic parameter is incompatible with Xen";
237
238		xen_raw_printk(m);
239		panic(m);
240	}
241	xen_init_lock_cpu(0);
242
243	smp_store_boot_cpu_info();
244	cpu_data(0).x86_max_cores = 1;
245
246	for_each_possible_cpu(i) {
247		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
248		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
249		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
250	}
251	set_cpu_sibling_map(0);
252
253	xen_pmu_init(0);
254
255	if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
256		BUG();
257
258	if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
259		panic("could not allocate xen_cpu_initialized_map\n");
260
261	cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
262
263	/* Restrict the possible_map according to max_cpus. */
264	while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
265		for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
266			continue;
267		set_cpu_possible(cpu, false);
268	}
269
270	for_each_possible_cpu(cpu)
271		set_cpu_present(cpu, true);
272}
273
274static int
275cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
276{
277	struct vcpu_guest_context *ctxt;
278	struct desc_struct *gdt;
279	unsigned long gdt_mfn;
280
281	/* used to tell cpu_init() that it can proceed with initialization */
282	cpumask_set_cpu(cpu, cpu_callout_mask);
283	if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
284		return 0;
285
286	ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
287	if (ctxt == NULL)
288		return -ENOMEM;
289
290	gdt = get_cpu_gdt_rw(cpu);
291
292#ifdef CONFIG_X86_32
293	ctxt->user_regs.fs = __KERNEL_PERCPU;
294	ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
295#endif
296	memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
297
298	/*
299	 * Bring up the CPU in cpu_bringup_and_idle() with the stack
300	 * pointing just below where pt_regs would be if it were a normal
301	 * kernel entry.
302	 */
303	ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
304	ctxt->flags = VGCF_IN_KERNEL;
305	ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
306	ctxt->user_regs.ds = __USER_DS;
307	ctxt->user_regs.es = __USER_DS;
308	ctxt->user_regs.ss = __KERNEL_DS;
309	ctxt->user_regs.cs = __KERNEL_CS;
310	ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
311
312	xen_copy_trap_info(ctxt->trap_ctxt);
313
314	ctxt->ldt_ents = 0;
315
316	BUG_ON((unsigned long)gdt & ~PAGE_MASK);
317
318	gdt_mfn = arbitrary_virt_to_mfn(gdt);
319	make_lowmem_page_readonly(gdt);
320	make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
321
322	ctxt->gdt_frames[0] = gdt_mfn;
323	ctxt->gdt_ents      = GDT_ENTRIES;
324
325	/*
326	 * Set SS:SP that Xen will use when entering guest kernel mode
327	 * from guest user mode.  Subsequent calls to load_sp0() can
328	 * change this value.
329	 */
330	ctxt->kernel_ss = __KERNEL_DS;
331	ctxt->kernel_sp = task_top_of_stack(idle);
332
333#ifdef CONFIG_X86_32
334	ctxt->event_callback_cs     = __KERNEL_CS;
335	ctxt->failsafe_callback_cs  = __KERNEL_CS;
336#else
337	ctxt->gs_base_kernel = per_cpu_offset(cpu);
338#endif
339	ctxt->event_callback_eip    =
340		(unsigned long)xen_hypervisor_callback;
341	ctxt->failsafe_callback_eip =
342		(unsigned long)xen_failsafe_callback;
343	per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
344
345	ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
346	if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
347		BUG();
348
349	kfree(ctxt);
350	return 0;
351}
352
353static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
354{
355	int rc;
356
357	common_cpu_up(cpu, idle);
358
359	xen_setup_runstate_info(cpu);
360
361	/*
362	 * PV VCPUs are always successfully taken down (see 'while' loop
363	 * in xen_cpu_die()), so -EBUSY is an error.
364	 */
365	rc = cpu_check_up_prepare(cpu);
366	if (rc)
367		return rc;
368
369	/* make sure interrupts start blocked */
370	per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
371
372	rc = cpu_initialize_context(cpu, idle);
373	if (rc)
374		return rc;
375
376	xen_pmu_init(cpu);
377
378	rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
379	BUG_ON(rc);
380
381	while (cpu_report_state(cpu) != CPU_ONLINE)
382		HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
383
384	return 0;
385}
386
387#ifdef CONFIG_HOTPLUG_CPU
388static int xen_pv_cpu_disable(void)
389{
390	unsigned int cpu = smp_processor_id();
391	if (cpu == 0)
392		return -EBUSY;
393
394	cpu_disable_common();
395
396	load_cr3(swapper_pg_dir);
397	return 0;
398}
399
400static void xen_pv_cpu_die(unsigned int cpu)
401{
402	while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
403				  xen_vcpu_nr(cpu), NULL)) {
404		__set_current_state(TASK_UNINTERRUPTIBLE);
405		schedule_timeout(HZ/10);
406	}
407
408	if (common_cpu_die(cpu) == 0) {
409		xen_smp_intr_free(cpu);
410		xen_uninit_lock_cpu(cpu);
411		xen_teardown_timer(cpu);
412		xen_pmu_finish(cpu);
413	}
414}
415
416static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
417{
418	play_dead_common();
419	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
420	cpu_bringup();
421	/*
422	 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
423	 * clears certain data that the cpu_idle loop (which called us
424	 * and that we return from) expects. The only way to get that
425	 * data back is to call:
426	 */
427	tick_nohz_idle_enter();
428	tick_nohz_idle_stop_tick_protected();
429
430	cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
431}
432
433#else /* !CONFIG_HOTPLUG_CPU */
434static int xen_pv_cpu_disable(void)
435{
436	return -ENOSYS;
437}
438
439static void xen_pv_cpu_die(unsigned int cpu)
440{
441	BUG();
442}
443
444static void xen_pv_play_dead(void)
445{
446	BUG();
447}
448
449#endif
450static void stop_self(void *v)
451{
452	int cpu = smp_processor_id();
453
454	/* make sure we're not pinning something down */
455	load_cr3(swapper_pg_dir);
456	/* should set up a minimal gdt */
457
458	set_cpu_online(cpu, false);
459
460	HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
461	BUG();
462}
463
464static void xen_pv_stop_other_cpus(int wait)
465{
466	smp_call_function(stop_self, NULL, wait);
467}
468
469static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
470{
471	irq_enter();
472	irq_work_run();
473	inc_irq_stat(apic_irq_work_irqs);
474	irq_exit();
475
476	return IRQ_HANDLED;
477}
478
479static const struct smp_ops xen_smp_ops __initconst = {
480	.smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
481	.smp_prepare_cpus = xen_pv_smp_prepare_cpus,
482	.smp_cpus_done = xen_smp_cpus_done,
483
484	.cpu_up = xen_pv_cpu_up,
485	.cpu_die = xen_pv_cpu_die,
486	.cpu_disable = xen_pv_cpu_disable,
487	.play_dead = xen_pv_play_dead,
488
489	.stop_other_cpus = xen_pv_stop_other_cpus,
490	.smp_send_reschedule = xen_smp_send_reschedule,
491
492	.send_call_func_ipi = xen_smp_send_call_function_ipi,
493	.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
494};
495
496void __init xen_smp_init(void)
497{
498	smp_ops = xen_smp_ops;
499	xen_fill_possible_map();
500}