1// SPDX-License-Identifier: GPL-2.0-only
2#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4#include <linux/kvm_host.h>
5
6#include <asm/irq_remapping.h>
7#include <asm/cpu.h>
8
9#include "lapic.h"
10#include "irq.h"
11#include "posted_intr.h"
12#include "trace.h"
13#include "vmx.h"
14
15/*
16 * Maintain a per-CPU list of vCPUs that need to be awakened by wakeup_handler()
17 * when a WAKEUP_VECTOR interrupted is posted.  vCPUs are added to the list when
18 * the vCPU is scheduled out and is blocking (e.g. in HLT) with IRQs enabled.
19 * The vCPUs posted interrupt descriptor is updated at the same time to set its
20 * notification vector to WAKEUP_VECTOR, so that posted interrupt from devices
21 * wake the target vCPUs.  vCPUs are removed from the list and the notification
22 * vector is reset when the vCPU is scheduled in.
23 */
24static DEFINE_PER_CPU(struct list_head, wakeup_vcpus_on_cpu);
25/*
26 * Protect the per-CPU list with a per-CPU spinlock to handle task migration.
27 * When a blocking vCPU is awakened _and_ migrated to a different pCPU, the
28 * ->sched_in() path will need to take the vCPU off the list of the _previous_
29 * CPU.  IRQs must be disabled when taking this lock, otherwise deadlock will
30 * occur if a wakeup IRQ arrives and attempts to acquire the lock.
31 */
32static DEFINE_PER_CPU(raw_spinlock_t, wakeup_vcpus_on_cpu_lock);
33
34static inline struct pi_desc *vcpu_to_pi_desc(struct kvm_vcpu *vcpu)
35{
36	return &(to_vmx(vcpu)->pi_desc);
37}
38
39static int pi_try_set_control(struct pi_desc *pi_desc, u64 *pold, u64 new)
40{
41	/*
42	 * PID.ON can be set at any time by a different vCPU or by hardware,
43	 * e.g. a device.  PID.control must be written atomically, and the
44	 * update must be retried with a fresh snapshot an ON change causes
45	 * the cmpxchg to fail.
46	 */
47	if (!try_cmpxchg64(&pi_desc->control, pold, new))
48		return -EBUSY;
49
50	return 0;
51}
52
53void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu)
54{
55	struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
56	struct vcpu_vmx *vmx = to_vmx(vcpu);
57	struct pi_desc old, new;
58	unsigned long flags;
59	unsigned int dest;
60
61	/*
62	 * To simplify hot-plug and dynamic toggling of APICv, keep PI.NDST and
63	 * PI.SN up-to-date even if there is no assigned device or if APICv is
64	 * deactivated due to a dynamic inhibit bit, e.g. for Hyper-V's SyncIC.
65	 */
66	if (!enable_apicv || !lapic_in_kernel(vcpu))
67		return;
68
69	/*
70	 * If the vCPU wasn't on the wakeup list and wasn't migrated, then the
71	 * full update can be skipped as neither the vector nor the destination
72	 * needs to be changed.
73	 */
74	if (pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR && vcpu->cpu == cpu) {
75		/*
76		 * Clear SN if it was set due to being preempted.  Again, do
77		 * this even if there is no assigned device for simplicity.
78		 */
79		if (pi_test_and_clear_sn(pi_desc))
80			goto after_clear_sn;
81		return;
82	}
83
84	local_irq_save(flags);
85
86	/*
87	 * If the vCPU was waiting for wakeup, remove the vCPU from the wakeup
88	 * list of the _previous_ pCPU, which will not be the same as the
89	 * current pCPU if the task was migrated.
90	 */
91	if (pi_desc->nv == POSTED_INTR_WAKEUP_VECTOR) {
92		raw_spin_lock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu));
93		list_del(&vmx->pi_wakeup_list);
94		raw_spin_unlock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu));
95	}
96
97	dest = cpu_physical_id(cpu);
98	if (!x2apic_mode)
99		dest = (dest << 8) & 0xFF00;
100
101	old.control = READ_ONCE(pi_desc->control);
102	do {
103		new.control = old.control;
104
105		/*
106		 * Clear SN (as above) and refresh the destination APIC ID to
107		 * handle task migration (@cpu != vcpu->cpu).
108		 */
109		new.ndst = dest;
110		new.sn = 0;
111
112		/*
113		 * Restore the notification vector; in the blocking case, the
114		 * descriptor was modified on "put" to use the wakeup vector.
115		 */
116		new.nv = POSTED_INTR_VECTOR;
117	} while (pi_try_set_control(pi_desc, &old.control, new.control));
118
119	local_irq_restore(flags);
120
121after_clear_sn:
122
123	/*
124	 * Clear SN before reading the bitmap.  The VT-d firmware
125	 * writes the bitmap and reads SN atomically (5.2.3 in the
126	 * spec), so it doesn't really have a memory barrier that
127	 * pairs with this, but we cannot do that and we need one.
128	 */
129	smp_mb__after_atomic();
130
131	if (!pi_is_pir_empty(pi_desc))
132		pi_set_on(pi_desc);
133}
134
135static bool vmx_can_use_vtd_pi(struct kvm *kvm)
136{
137	return irqchip_in_kernel(kvm) && enable_apicv &&
138		kvm_arch_has_assigned_device(kvm) &&
139		irq_remapping_cap(IRQ_POSTING_CAP);
140}
141
142/*
143 * Put the vCPU on this pCPU's list of vCPUs that needs to be awakened and set
144 * WAKEUP as the notification vector in the PI descriptor.
145 */
146static void pi_enable_wakeup_handler(struct kvm_vcpu *vcpu)
147{
148	struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
149	struct vcpu_vmx *vmx = to_vmx(vcpu);
150	struct pi_desc old, new;
151	unsigned long flags;
152
153	local_irq_save(flags);
154
155	raw_spin_lock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu));
156	list_add_tail(&vmx->pi_wakeup_list,
157		      &per_cpu(wakeup_vcpus_on_cpu, vcpu->cpu));
158	raw_spin_unlock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu));
159
160	WARN(pi_desc->sn, "PI descriptor SN field set before blocking");
161
162	old.control = READ_ONCE(pi_desc->control);
163	do {
164		/* set 'NV' to 'wakeup vector' */
165		new.control = old.control;
166		new.nv = POSTED_INTR_WAKEUP_VECTOR;
167	} while (pi_try_set_control(pi_desc, &old.control, new.control));
168
169	/*
170	 * Send a wakeup IPI to this CPU if an interrupt may have been posted
171	 * before the notification vector was updated, in which case the IRQ
172	 * will arrive on the non-wakeup vector.  An IPI is needed as calling
173	 * try_to_wake_up() from ->sched_out() isn't allowed (IRQs are not
174	 * enabled until it is safe to call try_to_wake_up() on the task being
175	 * scheduled out).
176	 */
177	if (pi_test_on(&new))
178		__apic_send_IPI_self(POSTED_INTR_WAKEUP_VECTOR);
179
180	local_irq_restore(flags);
181}
182
183static bool vmx_needs_pi_wakeup(struct kvm_vcpu *vcpu)
184{
185	/*
186	 * The default posted interrupt vector does nothing when
187	 * invoked outside guest mode.   Return whether a blocked vCPU
188	 * can be the target of posted interrupts, as is the case when
189	 * using either IPI virtualization or VT-d PI, so that the
190	 * notification vector is switched to the one that calls
191	 * back to the pi_wakeup_handler() function.
192	 */
193	return vmx_can_use_ipiv(vcpu) || vmx_can_use_vtd_pi(vcpu->kvm);
194}
195
196void vmx_vcpu_pi_put(struct kvm_vcpu *vcpu)
197{
198	struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
199
200	if (!vmx_needs_pi_wakeup(vcpu))
201		return;
202
203	if (kvm_vcpu_is_blocking(vcpu) && !vmx_interrupt_blocked(vcpu))
204		pi_enable_wakeup_handler(vcpu);
205
206	/*
207	 * Set SN when the vCPU is preempted.  Note, the vCPU can both be seen
208	 * as blocking and preempted, e.g. if it's preempted between setting
209	 * its wait state and manually scheduling out.
210	 */
211	if (vcpu->preempted)
212		pi_set_sn(pi_desc);
213}
214
215/*
216 * Handler for POSTED_INTERRUPT_WAKEUP_VECTOR.
217 */
218void pi_wakeup_handler(void)
219{
220	int cpu = smp_processor_id();
221	struct list_head *wakeup_list = &per_cpu(wakeup_vcpus_on_cpu, cpu);
222	raw_spinlock_t *spinlock = &per_cpu(wakeup_vcpus_on_cpu_lock, cpu);
223	struct vcpu_vmx *vmx;
224
225	raw_spin_lock(spinlock);
226	list_for_each_entry(vmx, wakeup_list, pi_wakeup_list) {
227
228		if (pi_test_on(&vmx->pi_desc))
229			kvm_vcpu_wake_up(&vmx->vcpu);
230	}
231	raw_spin_unlock(spinlock);
232}
233
234void __init pi_init_cpu(int cpu)
235{
236	INIT_LIST_HEAD(&per_cpu(wakeup_vcpus_on_cpu, cpu));
237	raw_spin_lock_init(&per_cpu(wakeup_vcpus_on_cpu_lock, cpu));
238}
239
240bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu)
241{
242	struct pi_desc *pi_desc = vcpu_to_pi_desc(vcpu);
243
244	return pi_test_on(pi_desc) ||
245		(pi_test_sn(pi_desc) && !pi_is_pir_empty(pi_desc));
246}
247
248
249/*
250 * Bail out of the block loop if the VM has an assigned
251 * device, but the blocking vCPU didn't reconfigure the
252 * PI.NV to the wakeup vector, i.e. the assigned device
253 * came along after the initial check in vmx_vcpu_pi_put().
254 */
255void vmx_pi_start_assignment(struct kvm *kvm)
256{
257	if (!irq_remapping_cap(IRQ_POSTING_CAP))
258		return;
259
260	kvm_make_all_cpus_request(kvm, KVM_REQ_UNBLOCK);
261}
262
263/*
264 * vmx_pi_update_irte - set IRTE for Posted-Interrupts
265 *
266 * @kvm: kvm
267 * @host_irq: host irq of the interrupt
268 * @guest_irq: gsi of the interrupt
269 * @set: set or unset PI
270 * returns 0 on success, < 0 on failure
271 */
272int vmx_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
273		       uint32_t guest_irq, bool set)
274{
275	struct kvm_kernel_irq_routing_entry *e;
276	struct kvm_irq_routing_table *irq_rt;
277	struct kvm_lapic_irq irq;
278	struct kvm_vcpu *vcpu;
279	struct vcpu_data vcpu_info;
280	int idx, ret = 0;
281
282	if (!vmx_can_use_vtd_pi(kvm))
283		return 0;
284
285	idx = srcu_read_lock(&kvm->irq_srcu);
286	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
287	if (guest_irq >= irq_rt->nr_rt_entries ||
288	    hlist_empty(&irq_rt->map[guest_irq])) {
289		pr_warn_once("no route for guest_irq %u/%u (broken user space?)\n",
290			     guest_irq, irq_rt->nr_rt_entries);
291		goto out;
292	}
293
294	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
295		if (e->type != KVM_IRQ_ROUTING_MSI)
296			continue;
297		/*
298		 * VT-d PI cannot support posting multicast/broadcast
299		 * interrupts to a vCPU, we still use interrupt remapping
300		 * for these kind of interrupts.
301		 *
302		 * For lowest-priority interrupts, we only support
303		 * those with single CPU as the destination, e.g. user
304		 * configures the interrupts via /proc/irq or uses
305		 * irqbalance to make the interrupts single-CPU.
306		 *
307		 * We will support full lowest-priority interrupt later.
308		 *
309		 * In addition, we can only inject generic interrupts using
310		 * the PI mechanism, refuse to route others through it.
311		 */
312
313		kvm_set_msi_irq(kvm, e, &irq);
314		if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
315		    !kvm_irq_is_postable(&irq)) {
316			/*
317			 * Make sure the IRTE is in remapped mode if
318			 * we don't handle it in posted mode.
319			 */
320			ret = irq_set_vcpu_affinity(host_irq, NULL);
321			if (ret < 0) {
322				printk(KERN_INFO
323				   "failed to back to remapped mode, irq: %u\n",
324				   host_irq);
325				goto out;
326			}
327
328			continue;
329		}
330
331		vcpu_info.pi_desc_addr = __pa(vcpu_to_pi_desc(vcpu));
332		vcpu_info.vector = irq.vector;
333
334		trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id, e->gsi,
335				vcpu_info.vector, vcpu_info.pi_desc_addr, set);
336
337		if (set)
338			ret = irq_set_vcpu_affinity(host_irq, &vcpu_info);
339		else
340			ret = irq_set_vcpu_affinity(host_irq, NULL);
341
342		if (ret < 0) {
343			printk(KERN_INFO "%s: failed to update PI IRTE\n",
344					__func__);
345			goto out;
346		}
347	}
348
349	ret = 0;
350out:
351	srcu_read_unlock(&kvm->irq_srcu, idx);
352	return ret;
353}
354