xref: /openbmc/linux/arch/x86/kvm/xen.c (revision 748008e1)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright © 2019 Oracle and/or its affiliates. All rights reserved.
4  * Copyright © 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
5  *
6  * KVM Xen emulation
7  */
8 
9 #include "x86.h"
10 #include "xen.h"
11 #include "hyperv.h"
12 #include "lapic.h"
13 
14 #include <linux/eventfd.h>
15 #include <linux/kvm_host.h>
16 #include <linux/sched/stat.h>
17 
18 #include <trace/events/kvm.h>
19 #include <xen/interface/xen.h>
20 #include <xen/interface/vcpu.h>
21 #include <xen/interface/version.h>
22 #include <xen/interface/event_channel.h>
23 #include <xen/interface/sched.h>
24 
25 #include "trace.h"
26 
27 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm);
28 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data);
29 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r);
30 
31 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
32 
33 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
34 {
35 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
36 	struct pvclock_wall_clock *wc;
37 	gpa_t gpa = gfn_to_gpa(gfn);
38 	u32 *wc_sec_hi;
39 	u32 wc_version;
40 	u64 wall_nsec;
41 	int ret = 0;
42 	int idx = srcu_read_lock(&kvm->srcu);
43 
44 	if (gfn == GPA_INVALID) {
45 		kvm_gfn_to_pfn_cache_destroy(kvm, gpc);
46 		goto out;
47 	}
48 
49 	do {
50 		ret = kvm_gfn_to_pfn_cache_init(kvm, gpc, NULL, KVM_HOST_USES_PFN,
51 						gpa, PAGE_SIZE);
52 		if (ret)
53 			goto out;
54 
55 		/*
56 		 * This code mirrors kvm_write_wall_clock() except that it writes
57 		 * directly through the pfn cache and doesn't mark the page dirty.
58 		 */
59 		wall_nsec = ktime_get_real_ns() - get_kvmclock_ns(kvm);
60 
61 		/* It could be invalid again already, so we need to check */
62 		read_lock_irq(&gpc->lock);
63 
64 		if (gpc->valid)
65 			break;
66 
67 		read_unlock_irq(&gpc->lock);
68 	} while (1);
69 
70 	/* Paranoia checks on the 32-bit struct layout */
71 	BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
72 	BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
73 	BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
74 
75 #ifdef CONFIG_X86_64
76 	/* Paranoia checks on the 64-bit struct layout */
77 	BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
78 	BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
79 
80 	if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
81 		struct shared_info *shinfo = gpc->khva;
82 
83 		wc_sec_hi = &shinfo->wc_sec_hi;
84 		wc = &shinfo->wc;
85 	} else
86 #endif
87 	{
88 		struct compat_shared_info *shinfo = gpc->khva;
89 
90 		wc_sec_hi = &shinfo->arch.wc_sec_hi;
91 		wc = &shinfo->wc;
92 	}
93 
94 	/* Increment and ensure an odd value */
95 	wc_version = wc->version = (wc->version + 1) | 1;
96 	smp_wmb();
97 
98 	wc->nsec = do_div(wall_nsec,  1000000000);
99 	wc->sec = (u32)wall_nsec;
100 	*wc_sec_hi = wall_nsec >> 32;
101 	smp_wmb();
102 
103 	wc->version = wc_version + 1;
104 	read_unlock_irq(&gpc->lock);
105 
106 	kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
107 
108 out:
109 	srcu_read_unlock(&kvm->srcu, idx);
110 	return ret;
111 }
112 
113 void kvm_xen_inject_timer_irqs(struct kvm_vcpu *vcpu)
114 {
115 	if (atomic_read(&vcpu->arch.xen.timer_pending) > 0) {
116 		struct kvm_xen_evtchn e;
117 
118 		e.vcpu_id = vcpu->vcpu_id;
119 		e.vcpu_idx = vcpu->vcpu_idx;
120 		e.port = vcpu->arch.xen.timer_virq;
121 		e.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
122 
123 		kvm_xen_set_evtchn(&e, vcpu->kvm);
124 
125 		vcpu->arch.xen.timer_expires = 0;
126 		atomic_set(&vcpu->arch.xen.timer_pending, 0);
127 	}
128 }
129 
130 static enum hrtimer_restart xen_timer_callback(struct hrtimer *timer)
131 {
132 	struct kvm_vcpu *vcpu = container_of(timer, struct kvm_vcpu,
133 					     arch.xen.timer);
134 	if (atomic_read(&vcpu->arch.xen.timer_pending))
135 		return HRTIMER_NORESTART;
136 
137 	atomic_inc(&vcpu->arch.xen.timer_pending);
138 	kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
139 	kvm_vcpu_kick(vcpu);
140 
141 	return HRTIMER_NORESTART;
142 }
143 
144 static void kvm_xen_start_timer(struct kvm_vcpu *vcpu, u64 guest_abs, s64 delta_ns)
145 {
146 	atomic_set(&vcpu->arch.xen.timer_pending, 0);
147 	vcpu->arch.xen.timer_expires = guest_abs;
148 
149 	if (delta_ns <= 0) {
150 		xen_timer_callback(&vcpu->arch.xen.timer);
151 	} else {
152 		ktime_t ktime_now = ktime_get();
153 		hrtimer_start(&vcpu->arch.xen.timer,
154 			      ktime_add_ns(ktime_now, delta_ns),
155 			      HRTIMER_MODE_ABS_HARD);
156 	}
157 }
158 
159 static void kvm_xen_stop_timer(struct kvm_vcpu *vcpu)
160 {
161 	hrtimer_cancel(&vcpu->arch.xen.timer);
162 	vcpu->arch.xen.timer_expires = 0;
163 	atomic_set(&vcpu->arch.xen.timer_pending, 0);
164 }
165 
166 static void kvm_xen_init_timer(struct kvm_vcpu *vcpu)
167 {
168 	hrtimer_init(&vcpu->arch.xen.timer, CLOCK_MONOTONIC,
169 		     HRTIMER_MODE_ABS_HARD);
170 	vcpu->arch.xen.timer.function = xen_timer_callback;
171 }
172 
173 static void kvm_xen_update_runstate(struct kvm_vcpu *v, int state)
174 {
175 	struct kvm_vcpu_xen *vx = &v->arch.xen;
176 	u64 now = get_kvmclock_ns(v->kvm);
177 	u64 delta_ns = now - vx->runstate_entry_time;
178 	u64 run_delay = current->sched_info.run_delay;
179 
180 	if (unlikely(!vx->runstate_entry_time))
181 		vx->current_runstate = RUNSTATE_offline;
182 
183 	/*
184 	 * Time waiting for the scheduler isn't "stolen" if the
185 	 * vCPU wasn't running anyway.
186 	 */
187 	if (vx->current_runstate == RUNSTATE_running) {
188 		u64 steal_ns = run_delay - vx->last_steal;
189 
190 		delta_ns -= steal_ns;
191 
192 		vx->runstate_times[RUNSTATE_runnable] += steal_ns;
193 	}
194 	vx->last_steal = run_delay;
195 
196 	vx->runstate_times[vx->current_runstate] += delta_ns;
197 	vx->current_runstate = state;
198 	vx->runstate_entry_time = now;
199 }
200 
201 void kvm_xen_update_runstate_guest(struct kvm_vcpu *v, int state)
202 {
203 	struct kvm_vcpu_xen *vx = &v->arch.xen;
204 	struct gfn_to_pfn_cache *gpc = &vx->runstate_cache;
205 	uint64_t *user_times;
206 	unsigned long flags;
207 	size_t user_len;
208 	int *user_state;
209 
210 	kvm_xen_update_runstate(v, state);
211 
212 	if (!vx->runstate_cache.active)
213 		return;
214 
215 	if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode)
216 		user_len = sizeof(struct vcpu_runstate_info);
217 	else
218 		user_len = sizeof(struct compat_vcpu_runstate_info);
219 
220 	read_lock_irqsave(&gpc->lock, flags);
221 	while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
222 					   user_len)) {
223 		read_unlock_irqrestore(&gpc->lock, flags);
224 
225 		/* When invoked from kvm_sched_out() we cannot sleep */
226 		if (state == RUNSTATE_runnable)
227 			return;
228 
229 		if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa, user_len))
230 			return;
231 
232 		read_lock_irqsave(&gpc->lock, flags);
233 	}
234 
235 	/*
236 	 * The only difference between 32-bit and 64-bit versions of the
237 	 * runstate struct us the alignment of uint64_t in 32-bit, which
238 	 * means that the 64-bit version has an additional 4 bytes of
239 	 * padding after the first field 'state'.
240 	 *
241 	 * So we use 'int __user *user_state' to point to the state field,
242 	 * and 'uint64_t __user *user_times' for runstate_entry_time. So
243 	 * the actual array of time[] in each state starts at user_times[1].
244 	 */
245 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) != 0);
246 	BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state) != 0);
247 	BUILD_BUG_ON(sizeof(struct compat_vcpu_runstate_info) != 0x2c);
248 #ifdef CONFIG_X86_64
249 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
250 		     offsetof(struct compat_vcpu_runstate_info, state_entry_time) + 4);
251 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, time) !=
252 		     offsetof(struct compat_vcpu_runstate_info, time) + 4);
253 #endif
254 
255 	user_state = gpc->khva;
256 
257 	if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode)
258 		user_times = gpc->khva + offsetof(struct vcpu_runstate_info,
259 						  state_entry_time);
260 	else
261 		user_times = gpc->khva + offsetof(struct compat_vcpu_runstate_info,
262 						  state_entry_time);
263 
264 	/*
265 	 * First write the updated state_entry_time at the appropriate
266 	 * location determined by 'offset'.
267 	 */
268 	BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state_entry_time) !=
269 		     sizeof(user_times[0]));
270 	BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state_entry_time) !=
271 		     sizeof(user_times[0]));
272 
273 	user_times[0] = vx->runstate_entry_time | XEN_RUNSTATE_UPDATE;
274 	smp_wmb();
275 
276 	/*
277 	 * Next, write the new runstate. This is in the *same* place
278 	 * for 32-bit and 64-bit guests, asserted here for paranoia.
279 	 */
280 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
281 		     offsetof(struct compat_vcpu_runstate_info, state));
282 	BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, state) !=
283 		     sizeof(vx->current_runstate));
284 	BUILD_BUG_ON(sizeof_field(struct compat_vcpu_runstate_info, state) !=
285 		     sizeof(vx->current_runstate));
286 
287 	*user_state = vx->current_runstate;
288 
289 	/*
290 	 * Write the actual runstate times immediately after the
291 	 * runstate_entry_time.
292 	 */
293 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state_entry_time) !=
294 		     offsetof(struct vcpu_runstate_info, time) - sizeof(u64));
295 	BUILD_BUG_ON(offsetof(struct compat_vcpu_runstate_info, state_entry_time) !=
296 		     offsetof(struct compat_vcpu_runstate_info, time) - sizeof(u64));
297 	BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
298 		     sizeof_field(struct compat_vcpu_runstate_info, time));
299 	BUILD_BUG_ON(sizeof_field(struct vcpu_runstate_info, time) !=
300 		     sizeof(vx->runstate_times));
301 
302 	memcpy(user_times + 1, vx->runstate_times, sizeof(vx->runstate_times));
303 	smp_wmb();
304 
305 	/*
306 	 * Finally, clear the XEN_RUNSTATE_UPDATE bit in the guest's
307 	 * runstate_entry_time field.
308 	 */
309 	user_times[0] &= ~XEN_RUNSTATE_UPDATE;
310 	smp_wmb();
311 
312 	read_unlock_irqrestore(&gpc->lock, flags);
313 
314 	mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
315 }
316 
317 static void kvm_xen_inject_vcpu_vector(struct kvm_vcpu *v)
318 {
319 	struct kvm_lapic_irq irq = { };
320 	int r;
321 
322 	irq.dest_id = v->vcpu_id;
323 	irq.vector = v->arch.xen.upcall_vector;
324 	irq.dest_mode = APIC_DEST_PHYSICAL;
325 	irq.shorthand = APIC_DEST_NOSHORT;
326 	irq.delivery_mode = APIC_DM_FIXED;
327 	irq.level = 1;
328 
329 	/* The fast version will always work for physical unicast */
330 	WARN_ON_ONCE(!kvm_irq_delivery_to_apic_fast(v->kvm, NULL, &irq, &r, NULL));
331 }
332 
333 /*
334  * On event channel delivery, the vcpu_info may not have been accessible.
335  * In that case, there are bits in vcpu->arch.xen.evtchn_pending_sel which
336  * need to be marked into the vcpu_info (and evtchn_upcall_pending set).
337  * Do so now that we can sleep in the context of the vCPU to bring the
338  * page in, and refresh the pfn cache for it.
339  */
340 void kvm_xen_inject_pending_events(struct kvm_vcpu *v)
341 {
342 	unsigned long evtchn_pending_sel = READ_ONCE(v->arch.xen.evtchn_pending_sel);
343 	struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
344 	unsigned long flags;
345 
346 	if (!evtchn_pending_sel)
347 		return;
348 
349 	/*
350 	 * Yes, this is an open-coded loop. But that's just what put_user()
351 	 * does anyway. Page it in and retry the instruction. We're just a
352 	 * little more honest about it.
353 	 */
354 	read_lock_irqsave(&gpc->lock, flags);
355 	while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
356 					   sizeof(struct vcpu_info))) {
357 		read_unlock_irqrestore(&gpc->lock, flags);
358 
359 		if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa,
360 						 sizeof(struct vcpu_info)))
361 			return;
362 
363 		read_lock_irqsave(&gpc->lock, flags);
364 	}
365 
366 	/* Now gpc->khva is a valid kernel address for the vcpu_info */
367 	if (IS_ENABLED(CONFIG_64BIT) && v->kvm->arch.xen.long_mode) {
368 		struct vcpu_info *vi = gpc->khva;
369 
370 		asm volatile(LOCK_PREFIX "orq %0, %1\n"
371 			     "notq %0\n"
372 			     LOCK_PREFIX "andq %0, %2\n"
373 			     : "=r" (evtchn_pending_sel),
374 			       "+m" (vi->evtchn_pending_sel),
375 			       "+m" (v->arch.xen.evtchn_pending_sel)
376 			     : "0" (evtchn_pending_sel));
377 		WRITE_ONCE(vi->evtchn_upcall_pending, 1);
378 	} else {
379 		u32 evtchn_pending_sel32 = evtchn_pending_sel;
380 		struct compat_vcpu_info *vi = gpc->khva;
381 
382 		asm volatile(LOCK_PREFIX "orl %0, %1\n"
383 			     "notl %0\n"
384 			     LOCK_PREFIX "andl %0, %2\n"
385 			     : "=r" (evtchn_pending_sel32),
386 			       "+m" (vi->evtchn_pending_sel),
387 			       "+m" (v->arch.xen.evtchn_pending_sel)
388 			     : "0" (evtchn_pending_sel32));
389 		WRITE_ONCE(vi->evtchn_upcall_pending, 1);
390 	}
391 	read_unlock_irqrestore(&gpc->lock, flags);
392 
393 	/* For the per-vCPU lapic vector, deliver it as MSI. */
394 	if (v->arch.xen.upcall_vector)
395 		kvm_xen_inject_vcpu_vector(v);
396 
397 	mark_page_dirty_in_slot(v->kvm, gpc->memslot, gpc->gpa >> PAGE_SHIFT);
398 }
399 
400 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
401 {
402 	struct gfn_to_pfn_cache *gpc = &v->arch.xen.vcpu_info_cache;
403 	unsigned long flags;
404 	u8 rc = 0;
405 
406 	/*
407 	 * If the global upcall vector (HVMIRQ_callback_vector) is set and
408 	 * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
409 	 */
410 
411 	/* No need for compat handling here */
412 	BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
413 		     offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
414 	BUILD_BUG_ON(sizeof(rc) !=
415 		     sizeof_field(struct vcpu_info, evtchn_upcall_pending));
416 	BUILD_BUG_ON(sizeof(rc) !=
417 		     sizeof_field(struct compat_vcpu_info, evtchn_upcall_pending));
418 
419 	read_lock_irqsave(&gpc->lock, flags);
420 	while (!kvm_gfn_to_pfn_cache_check(v->kvm, gpc, gpc->gpa,
421 					   sizeof(struct vcpu_info))) {
422 		read_unlock_irqrestore(&gpc->lock, flags);
423 
424 		/*
425 		 * This function gets called from kvm_vcpu_block() after setting the
426 		 * task to TASK_INTERRUPTIBLE, to see if it needs to wake immediately
427 		 * from a HLT. So we really mustn't sleep. If the page ended up absent
428 		 * at that point, just return 1 in order to trigger an immediate wake,
429 		 * and we'll end up getting called again from a context where we *can*
430 		 * fault in the page and wait for it.
431 		 */
432 		if (in_atomic() || !task_is_running(current))
433 			return 1;
434 
435 		if (kvm_gfn_to_pfn_cache_refresh(v->kvm, gpc, gpc->gpa,
436 						 sizeof(struct vcpu_info))) {
437 			/*
438 			 * If this failed, userspace has screwed up the
439 			 * vcpu_info mapping. No interrupts for you.
440 			 */
441 			return 0;
442 		}
443 		read_lock_irqsave(&gpc->lock, flags);
444 	}
445 
446 	rc = ((struct vcpu_info *)gpc->khva)->evtchn_upcall_pending;
447 	read_unlock_irqrestore(&gpc->lock, flags);
448 	return rc;
449 }
450 
451 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
452 {
453 	int r = -ENOENT;
454 
455 
456 	switch (data->type) {
457 	case KVM_XEN_ATTR_TYPE_LONG_MODE:
458 		if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
459 			r = -EINVAL;
460 		} else {
461 			mutex_lock(&kvm->lock);
462 			kvm->arch.xen.long_mode = !!data->u.long_mode;
463 			mutex_unlock(&kvm->lock);
464 			r = 0;
465 		}
466 		break;
467 
468 	case KVM_XEN_ATTR_TYPE_SHARED_INFO:
469 		mutex_lock(&kvm->lock);
470 		r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
471 		mutex_unlock(&kvm->lock);
472 		break;
473 
474 	case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
475 		if (data->u.vector && data->u.vector < 0x10)
476 			r = -EINVAL;
477 		else {
478 			mutex_lock(&kvm->lock);
479 			kvm->arch.xen.upcall_vector = data->u.vector;
480 			mutex_unlock(&kvm->lock);
481 			r = 0;
482 		}
483 		break;
484 
485 	case KVM_XEN_ATTR_TYPE_EVTCHN:
486 		r = kvm_xen_setattr_evtchn(kvm, data);
487 		break;
488 
489 	case KVM_XEN_ATTR_TYPE_XEN_VERSION:
490 		mutex_lock(&kvm->lock);
491 		kvm->arch.xen.xen_version = data->u.xen_version;
492 		mutex_unlock(&kvm->lock);
493 		r = 0;
494 		break;
495 
496 	default:
497 		break;
498 	}
499 
500 	return r;
501 }
502 
503 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
504 {
505 	int r = -ENOENT;
506 
507 	mutex_lock(&kvm->lock);
508 
509 	switch (data->type) {
510 	case KVM_XEN_ATTR_TYPE_LONG_MODE:
511 		data->u.long_mode = kvm->arch.xen.long_mode;
512 		r = 0;
513 		break;
514 
515 	case KVM_XEN_ATTR_TYPE_SHARED_INFO:
516 		if (kvm->arch.xen.shinfo_cache.active)
517 			data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
518 		else
519 			data->u.shared_info.gfn = GPA_INVALID;
520 		r = 0;
521 		break;
522 
523 	case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
524 		data->u.vector = kvm->arch.xen.upcall_vector;
525 		r = 0;
526 		break;
527 
528 	case KVM_XEN_ATTR_TYPE_XEN_VERSION:
529 		data->u.xen_version = kvm->arch.xen.xen_version;
530 		r = 0;
531 		break;
532 
533 	default:
534 		break;
535 	}
536 
537 	mutex_unlock(&kvm->lock);
538 	return r;
539 }
540 
541 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
542 {
543 	int idx, r = -ENOENT;
544 
545 	mutex_lock(&vcpu->kvm->lock);
546 	idx = srcu_read_lock(&vcpu->kvm->srcu);
547 
548 	switch (data->type) {
549 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
550 		/* No compat necessary here. */
551 		BUILD_BUG_ON(sizeof(struct vcpu_info) !=
552 			     sizeof(struct compat_vcpu_info));
553 		BUILD_BUG_ON(offsetof(struct vcpu_info, time) !=
554 			     offsetof(struct compat_vcpu_info, time));
555 
556 		if (data->u.gpa == GPA_INVALID) {
557 			kvm_gfn_to_pfn_cache_destroy(vcpu->kvm, &vcpu->arch.xen.vcpu_info_cache);
558 			r = 0;
559 			break;
560 		}
561 
562 		r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
563 					      &vcpu->arch.xen.vcpu_info_cache,
564 					      NULL, KVM_HOST_USES_PFN, data->u.gpa,
565 					      sizeof(struct vcpu_info));
566 		if (!r)
567 			kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
568 
569 		break;
570 
571 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
572 		if (data->u.gpa == GPA_INVALID) {
573 			kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
574 						     &vcpu->arch.xen.vcpu_time_info_cache);
575 			r = 0;
576 			break;
577 		}
578 
579 		r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
580 					      &vcpu->arch.xen.vcpu_time_info_cache,
581 					      NULL, KVM_HOST_USES_PFN, data->u.gpa,
582 					      sizeof(struct pvclock_vcpu_time_info));
583 		if (!r)
584 			kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
585 		break;
586 
587 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
588 		if (!sched_info_on()) {
589 			r = -EOPNOTSUPP;
590 			break;
591 		}
592 		if (data->u.gpa == GPA_INVALID) {
593 			kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
594 						     &vcpu->arch.xen.runstate_cache);
595 			r = 0;
596 			break;
597 		}
598 
599 		r = kvm_gfn_to_pfn_cache_init(vcpu->kvm,
600 					      &vcpu->arch.xen.runstate_cache,
601 					      NULL, KVM_HOST_USES_PFN, data->u.gpa,
602 					      sizeof(struct vcpu_runstate_info));
603 		break;
604 
605 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
606 		if (!sched_info_on()) {
607 			r = -EOPNOTSUPP;
608 			break;
609 		}
610 		if (data->u.runstate.state > RUNSTATE_offline) {
611 			r = -EINVAL;
612 			break;
613 		}
614 
615 		kvm_xen_update_runstate(vcpu, data->u.runstate.state);
616 		r = 0;
617 		break;
618 
619 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
620 		if (!sched_info_on()) {
621 			r = -EOPNOTSUPP;
622 			break;
623 		}
624 		if (data->u.runstate.state > RUNSTATE_offline) {
625 			r = -EINVAL;
626 			break;
627 		}
628 		if (data->u.runstate.state_entry_time !=
629 		    (data->u.runstate.time_running +
630 		     data->u.runstate.time_runnable +
631 		     data->u.runstate.time_blocked +
632 		     data->u.runstate.time_offline)) {
633 			r = -EINVAL;
634 			break;
635 		}
636 		if (get_kvmclock_ns(vcpu->kvm) <
637 		    data->u.runstate.state_entry_time) {
638 			r = -EINVAL;
639 			break;
640 		}
641 
642 		vcpu->arch.xen.current_runstate = data->u.runstate.state;
643 		vcpu->arch.xen.runstate_entry_time =
644 			data->u.runstate.state_entry_time;
645 		vcpu->arch.xen.runstate_times[RUNSTATE_running] =
646 			data->u.runstate.time_running;
647 		vcpu->arch.xen.runstate_times[RUNSTATE_runnable] =
648 			data->u.runstate.time_runnable;
649 		vcpu->arch.xen.runstate_times[RUNSTATE_blocked] =
650 			data->u.runstate.time_blocked;
651 		vcpu->arch.xen.runstate_times[RUNSTATE_offline] =
652 			data->u.runstate.time_offline;
653 		vcpu->arch.xen.last_steal = current->sched_info.run_delay;
654 		r = 0;
655 		break;
656 
657 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
658 		if (!sched_info_on()) {
659 			r = -EOPNOTSUPP;
660 			break;
661 		}
662 		if (data->u.runstate.state > RUNSTATE_offline &&
663 		    data->u.runstate.state != (u64)-1) {
664 			r = -EINVAL;
665 			break;
666 		}
667 		/* The adjustment must add up */
668 		if (data->u.runstate.state_entry_time !=
669 		    (data->u.runstate.time_running +
670 		     data->u.runstate.time_runnable +
671 		     data->u.runstate.time_blocked +
672 		     data->u.runstate.time_offline)) {
673 			r = -EINVAL;
674 			break;
675 		}
676 
677 		if (get_kvmclock_ns(vcpu->kvm) <
678 		    (vcpu->arch.xen.runstate_entry_time +
679 		     data->u.runstate.state_entry_time)) {
680 			r = -EINVAL;
681 			break;
682 		}
683 
684 		vcpu->arch.xen.runstate_entry_time +=
685 			data->u.runstate.state_entry_time;
686 		vcpu->arch.xen.runstate_times[RUNSTATE_running] +=
687 			data->u.runstate.time_running;
688 		vcpu->arch.xen.runstate_times[RUNSTATE_runnable] +=
689 			data->u.runstate.time_runnable;
690 		vcpu->arch.xen.runstate_times[RUNSTATE_blocked] +=
691 			data->u.runstate.time_blocked;
692 		vcpu->arch.xen.runstate_times[RUNSTATE_offline] +=
693 			data->u.runstate.time_offline;
694 
695 		if (data->u.runstate.state <= RUNSTATE_offline)
696 			kvm_xen_update_runstate(vcpu, data->u.runstate.state);
697 		r = 0;
698 		break;
699 
700 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
701 		if (data->u.vcpu_id >= KVM_MAX_VCPUS)
702 			r = -EINVAL;
703 		else {
704 			vcpu->arch.xen.vcpu_id = data->u.vcpu_id;
705 			r = 0;
706 		}
707 		break;
708 
709 	case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
710 		if (data->u.timer.port) {
711 			if (data->u.timer.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL) {
712 				r = -EINVAL;
713 				break;
714 			}
715 			vcpu->arch.xen.timer_virq = data->u.timer.port;
716 			kvm_xen_init_timer(vcpu);
717 
718 			/* Restart the timer if it's set */
719 			if (data->u.timer.expires_ns)
720 				kvm_xen_start_timer(vcpu, data->u.timer.expires_ns,
721 						    data->u.timer.expires_ns -
722 						    get_kvmclock_ns(vcpu->kvm));
723 		} else if (kvm_xen_timer_enabled(vcpu)) {
724 			kvm_xen_stop_timer(vcpu);
725 			vcpu->arch.xen.timer_virq = 0;
726 		}
727 
728 		r = 0;
729 		break;
730 
731 	case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
732 		if (data->u.vector && data->u.vector < 0x10)
733 			r = -EINVAL;
734 		else {
735 			vcpu->arch.xen.upcall_vector = data->u.vector;
736 			r = 0;
737 		}
738 		break;
739 
740 	default:
741 		break;
742 	}
743 
744 	srcu_read_unlock(&vcpu->kvm->srcu, idx);
745 	mutex_unlock(&vcpu->kvm->lock);
746 	return r;
747 }
748 
749 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
750 {
751 	int r = -ENOENT;
752 
753 	mutex_lock(&vcpu->kvm->lock);
754 
755 	switch (data->type) {
756 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
757 		if (vcpu->arch.xen.vcpu_info_cache.active)
758 			data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
759 		else
760 			data->u.gpa = GPA_INVALID;
761 		r = 0;
762 		break;
763 
764 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
765 		if (vcpu->arch.xen.vcpu_time_info_cache.active)
766 			data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
767 		else
768 			data->u.gpa = GPA_INVALID;
769 		r = 0;
770 		break;
771 
772 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADDR:
773 		if (!sched_info_on()) {
774 			r = -EOPNOTSUPP;
775 			break;
776 		}
777 		if (vcpu->arch.xen.runstate_cache.active) {
778 			data->u.gpa = vcpu->arch.xen.runstate_cache.gpa;
779 			r = 0;
780 		}
781 		break;
782 
783 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_CURRENT:
784 		if (!sched_info_on()) {
785 			r = -EOPNOTSUPP;
786 			break;
787 		}
788 		data->u.runstate.state = vcpu->arch.xen.current_runstate;
789 		r = 0;
790 		break;
791 
792 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_DATA:
793 		if (!sched_info_on()) {
794 			r = -EOPNOTSUPP;
795 			break;
796 		}
797 		data->u.runstate.state = vcpu->arch.xen.current_runstate;
798 		data->u.runstate.state_entry_time =
799 			vcpu->arch.xen.runstate_entry_time;
800 		data->u.runstate.time_running =
801 			vcpu->arch.xen.runstate_times[RUNSTATE_running];
802 		data->u.runstate.time_runnable =
803 			vcpu->arch.xen.runstate_times[RUNSTATE_runnable];
804 		data->u.runstate.time_blocked =
805 			vcpu->arch.xen.runstate_times[RUNSTATE_blocked];
806 		data->u.runstate.time_offline =
807 			vcpu->arch.xen.runstate_times[RUNSTATE_offline];
808 		r = 0;
809 		break;
810 
811 	case KVM_XEN_VCPU_ATTR_TYPE_RUNSTATE_ADJUST:
812 		r = -EINVAL;
813 		break;
814 
815 	case KVM_XEN_VCPU_ATTR_TYPE_VCPU_ID:
816 		data->u.vcpu_id = vcpu->arch.xen.vcpu_id;
817 		r = 0;
818 		break;
819 
820 	case KVM_XEN_VCPU_ATTR_TYPE_TIMER:
821 		data->u.timer.port = vcpu->arch.xen.timer_virq;
822 		data->u.timer.priority = KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL;
823 		data->u.timer.expires_ns = vcpu->arch.xen.timer_expires;
824 		r = 0;
825 		break;
826 
827 	case KVM_XEN_VCPU_ATTR_TYPE_UPCALL_VECTOR:
828 		data->u.vector = vcpu->arch.xen.upcall_vector;
829 		r = 0;
830 		break;
831 
832 	default:
833 		break;
834 	}
835 
836 	mutex_unlock(&vcpu->kvm->lock);
837 	return r;
838 }
839 
840 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
841 {
842 	struct kvm *kvm = vcpu->kvm;
843 	u32 page_num = data & ~PAGE_MASK;
844 	u64 page_addr = data & PAGE_MASK;
845 	bool lm = is_long_mode(vcpu);
846 
847 	/* Latch long_mode for shared_info pages etc. */
848 	vcpu->kvm->arch.xen.long_mode = lm;
849 
850 	/*
851 	 * If Xen hypercall intercept is enabled, fill the hypercall
852 	 * page with VMCALL/VMMCALL instructions since that's what
853 	 * we catch. Else the VMM has provided the hypercall pages
854 	 * with instructions of its own choosing, so use those.
855 	 */
856 	if (kvm_xen_hypercall_enabled(kvm)) {
857 		u8 instructions[32];
858 		int i;
859 
860 		if (page_num)
861 			return 1;
862 
863 		/* mov imm32, %eax */
864 		instructions[0] = 0xb8;
865 
866 		/* vmcall / vmmcall */
867 		static_call(kvm_x86_patch_hypercall)(vcpu, instructions + 5);
868 
869 		/* ret */
870 		instructions[8] = 0xc3;
871 
872 		/* int3 to pad */
873 		memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
874 
875 		for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
876 			*(u32 *)&instructions[1] = i;
877 			if (kvm_vcpu_write_guest(vcpu,
878 						 page_addr + (i * sizeof(instructions)),
879 						 instructions, sizeof(instructions)))
880 				return 1;
881 		}
882 	} else {
883 		/*
884 		 * Note, truncation is a non-issue as 'lm' is guaranteed to be
885 		 * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
886 		 */
887 		hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
888 				     : kvm->arch.xen_hvm_config.blob_addr_32;
889 		u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
890 				  : kvm->arch.xen_hvm_config.blob_size_32;
891 		u8 *page;
892 
893 		if (page_num >= blob_size)
894 			return 1;
895 
896 		blob_addr += page_num * PAGE_SIZE;
897 
898 		page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
899 		if (IS_ERR(page))
900 			return PTR_ERR(page);
901 
902 		if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
903 			kfree(page);
904 			return 1;
905 		}
906 	}
907 	return 0;
908 }
909 
910 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
911 {
912 	/* Only some feature flags need to be *enabled* by userspace */
913 	u32 permitted_flags = KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL |
914 		KVM_XEN_HVM_CONFIG_EVTCHN_SEND;
915 
916 	if (xhc->flags & ~permitted_flags)
917 		return -EINVAL;
918 
919 	/*
920 	 * With hypercall interception the kernel generates its own
921 	 * hypercall page so it must not be provided.
922 	 */
923 	if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
924 	    (xhc->blob_addr_32 || xhc->blob_addr_64 ||
925 	     xhc->blob_size_32 || xhc->blob_size_64))
926 		return -EINVAL;
927 
928 	mutex_lock(&kvm->lock);
929 
930 	if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
931 		static_branch_inc(&kvm_xen_enabled.key);
932 	else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
933 		static_branch_slow_dec_deferred(&kvm_xen_enabled);
934 
935 	memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
936 
937 	mutex_unlock(&kvm->lock);
938 	return 0;
939 }
940 
941 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
942 {
943 	kvm_rax_write(vcpu, result);
944 	return kvm_skip_emulated_instruction(vcpu);
945 }
946 
947 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
948 {
949 	struct kvm_run *run = vcpu->run;
950 
951 	if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
952 		return 1;
953 
954 	return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
955 }
956 
957 static bool wait_pending_event(struct kvm_vcpu *vcpu, int nr_ports,
958 			       evtchn_port_t *ports)
959 {
960 	struct kvm *kvm = vcpu->kvm;
961 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
962 	unsigned long *pending_bits;
963 	unsigned long flags;
964 	bool ret = true;
965 	int idx, i;
966 
967 	read_lock_irqsave(&gpc->lock, flags);
968 	idx = srcu_read_lock(&kvm->srcu);
969 	if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, PAGE_SIZE))
970 		goto out_rcu;
971 
972 	ret = false;
973 	if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
974 		struct shared_info *shinfo = gpc->khva;
975 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
976 	} else {
977 		struct compat_shared_info *shinfo = gpc->khva;
978 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
979 	}
980 
981 	for (i = 0; i < nr_ports; i++) {
982 		if (test_bit(ports[i], pending_bits)) {
983 			ret = true;
984 			break;
985 		}
986 	}
987 
988  out_rcu:
989 	srcu_read_unlock(&kvm->srcu, idx);
990 	read_unlock_irqrestore(&gpc->lock, flags);
991 
992 	return ret;
993 }
994 
995 static bool kvm_xen_schedop_poll(struct kvm_vcpu *vcpu, bool longmode,
996 				 u64 param, u64 *r)
997 {
998 	int idx, i;
999 	struct sched_poll sched_poll;
1000 	evtchn_port_t port, *ports;
1001 	gpa_t gpa;
1002 
1003 	if (!longmode || !lapic_in_kernel(vcpu) ||
1004 	    !(vcpu->kvm->arch.xen_hvm_config.flags & KVM_XEN_HVM_CONFIG_EVTCHN_SEND))
1005 		return false;
1006 
1007 	idx = srcu_read_lock(&vcpu->kvm->srcu);
1008 	gpa = kvm_mmu_gva_to_gpa_system(vcpu, param, NULL);
1009 	srcu_read_unlock(&vcpu->kvm->srcu, idx);
1010 
1011 	if (!gpa || kvm_vcpu_read_guest(vcpu, gpa, &sched_poll,
1012 					sizeof(sched_poll))) {
1013 		*r = -EFAULT;
1014 		return true;
1015 	}
1016 
1017 	if (unlikely(sched_poll.nr_ports > 1)) {
1018 		/* Xen (unofficially) limits number of pollers to 128 */
1019 		if (sched_poll.nr_ports > 128) {
1020 			*r = -EINVAL;
1021 			return true;
1022 		}
1023 
1024 		ports = kmalloc_array(sched_poll.nr_ports,
1025 				      sizeof(*ports), GFP_KERNEL);
1026 		if (!ports) {
1027 			*r = -ENOMEM;
1028 			return true;
1029 		}
1030 	} else
1031 		ports = &port;
1032 
1033 	for (i = 0; i < sched_poll.nr_ports; i++) {
1034 		idx = srcu_read_lock(&vcpu->kvm->srcu);
1035 		gpa = kvm_mmu_gva_to_gpa_system(vcpu,
1036 						(gva_t)(sched_poll.ports + i),
1037 						NULL);
1038 		srcu_read_unlock(&vcpu->kvm->srcu, idx);
1039 
1040 		if (!gpa || kvm_vcpu_read_guest(vcpu, gpa,
1041 						&ports[i], sizeof(port))) {
1042 			*r = -EFAULT;
1043 			goto out;
1044 		}
1045 	}
1046 
1047 	if (sched_poll.nr_ports == 1)
1048 		vcpu->arch.xen.poll_evtchn = port;
1049 	else
1050 		vcpu->arch.xen.poll_evtchn = -1;
1051 
1052 	set_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1053 
1054 	if (!wait_pending_event(vcpu, sched_poll.nr_ports, ports)) {
1055 		vcpu->arch.mp_state = KVM_MP_STATE_HALTED;
1056 
1057 		if (sched_poll.timeout)
1058 			mod_timer(&vcpu->arch.xen.poll_timer,
1059 				  jiffies + nsecs_to_jiffies(sched_poll.timeout));
1060 
1061 		kvm_vcpu_halt(vcpu);
1062 
1063 		if (sched_poll.timeout)
1064 			del_timer(&vcpu->arch.xen.poll_timer);
1065 
1066 		vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
1067 		kvm_clear_request(KVM_REQ_UNHALT, vcpu);
1068 	}
1069 
1070 	vcpu->arch.xen.poll_evtchn = 0;
1071 	*r = 0;
1072 out:
1073 	/* Really, this is only needed in case of timeout */
1074 	clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask);
1075 
1076 	if (unlikely(sched_poll.nr_ports > 1))
1077 		kfree(ports);
1078 	return true;
1079 }
1080 
1081 static void cancel_evtchn_poll(struct timer_list *t)
1082 {
1083 	struct kvm_vcpu *vcpu = from_timer(vcpu, t, arch.xen.poll_timer);
1084 
1085 	kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1086 	kvm_vcpu_kick(vcpu);
1087 }
1088 
1089 static bool kvm_xen_hcall_sched_op(struct kvm_vcpu *vcpu, bool longmode,
1090 				   int cmd, u64 param, u64 *r)
1091 {
1092 	switch (cmd) {
1093 	case SCHEDOP_poll:
1094 		if (kvm_xen_schedop_poll(vcpu, longmode, param, r))
1095 			return true;
1096 		fallthrough;
1097 	case SCHEDOP_yield:
1098 		kvm_vcpu_on_spin(vcpu, true);
1099 		*r = 0;
1100 		return true;
1101 	default:
1102 		break;
1103 	}
1104 
1105 	return false;
1106 }
1107 
1108 struct compat_vcpu_set_singleshot_timer {
1109     uint64_t timeout_abs_ns;
1110     uint32_t flags;
1111 } __attribute__((packed));
1112 
1113 static bool kvm_xen_hcall_vcpu_op(struct kvm_vcpu *vcpu, bool longmode, int cmd,
1114 				  int vcpu_id, u64 param, u64 *r)
1115 {
1116 	struct vcpu_set_singleshot_timer oneshot;
1117 	s64 delta;
1118 	gpa_t gpa;
1119 	int idx;
1120 
1121 	if (!kvm_xen_timer_enabled(vcpu))
1122 		return false;
1123 
1124 	switch (cmd) {
1125 	case VCPUOP_set_singleshot_timer:
1126 		if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1127 			*r = -EINVAL;
1128 			return true;
1129 		}
1130 		idx = srcu_read_lock(&vcpu->kvm->srcu);
1131 		gpa = kvm_mmu_gva_to_gpa_system(vcpu, param, NULL);
1132 		srcu_read_unlock(&vcpu->kvm->srcu, idx);
1133 
1134 		/*
1135 		 * The only difference for 32-bit compat is the 4 bytes of
1136 		 * padding after the interesting part of the structure. So
1137 		 * for a faithful emulation of Xen we have to *try* to copy
1138 		 * the padding and return -EFAULT if we can't. Otherwise we
1139 		 * might as well just have copied the 12-byte 32-bit struct.
1140 		 */
1141 		BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1142 			     offsetof(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1143 		BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, timeout_abs_ns) !=
1144 			     sizeof_field(struct vcpu_set_singleshot_timer, timeout_abs_ns));
1145 		BUILD_BUG_ON(offsetof(struct compat_vcpu_set_singleshot_timer, flags) !=
1146 			     offsetof(struct vcpu_set_singleshot_timer, flags));
1147 		BUILD_BUG_ON(sizeof_field(struct compat_vcpu_set_singleshot_timer, flags) !=
1148 			     sizeof_field(struct vcpu_set_singleshot_timer, flags));
1149 
1150 		if (!gpa ||
1151 		    kvm_vcpu_read_guest(vcpu, gpa, &oneshot, longmode ? sizeof(oneshot) :
1152 					sizeof(struct compat_vcpu_set_singleshot_timer))) {
1153 			*r = -EFAULT;
1154 			return true;
1155 		}
1156 
1157 		delta = oneshot.timeout_abs_ns - get_kvmclock_ns(vcpu->kvm);
1158 		if ((oneshot.flags & VCPU_SSHOTTMR_future) && delta < 0) {
1159 			*r = -ETIME;
1160 			return true;
1161 		}
1162 
1163 		kvm_xen_start_timer(vcpu, oneshot.timeout_abs_ns, delta);
1164 		*r = 0;
1165 		return true;
1166 
1167 	case VCPUOP_stop_singleshot_timer:
1168 		if (vcpu->arch.xen.vcpu_id != vcpu_id) {
1169 			*r = -EINVAL;
1170 			return true;
1171 		}
1172 		kvm_xen_stop_timer(vcpu);
1173 		*r = 0;
1174 		return true;
1175 	}
1176 
1177 	return false;
1178 }
1179 
1180 static bool kvm_xen_hcall_set_timer_op(struct kvm_vcpu *vcpu, uint64_t timeout,
1181 				       u64 *r)
1182 {
1183 	if (!kvm_xen_timer_enabled(vcpu))
1184 		return false;
1185 
1186 	if (timeout) {
1187 		uint64_t guest_now = get_kvmclock_ns(vcpu->kvm);
1188 		int64_t delta = timeout - guest_now;
1189 
1190 		/* Xen has a 'Linux workaround' in do_set_timer_op() which
1191 		 * checks for negative absolute timeout values (caused by
1192 		 * integer overflow), and for values about 13 days in the
1193 		 * future (2^50ns) which would be caused by jiffies
1194 		 * overflow. For those cases, it sets the timeout 100ms in
1195 		 * the future (not *too* soon, since if a guest really did
1196 		 * set a long timeout on purpose we don't want to keep
1197 		 * churning CPU time by waking it up).
1198 		 */
1199 		if (unlikely((int64_t)timeout < 0 ||
1200 			     (delta > 0 && (uint32_t) (delta >> 50) != 0))) {
1201 			delta = 100 * NSEC_PER_MSEC;
1202 			timeout = guest_now + delta;
1203 		}
1204 
1205 		kvm_xen_start_timer(vcpu, timeout, delta);
1206 	} else {
1207 		kvm_xen_stop_timer(vcpu);
1208 	}
1209 
1210 	*r = 0;
1211 	return true;
1212 }
1213 
1214 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
1215 {
1216 	bool longmode;
1217 	u64 input, params[6], r = -ENOSYS;
1218 	bool handled = false;
1219 
1220 	input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
1221 
1222 	/* Hyper-V hypercalls get bit 31 set in EAX */
1223 	if ((input & 0x80000000) &&
1224 	    kvm_hv_hypercall_enabled(vcpu))
1225 		return kvm_hv_hypercall(vcpu);
1226 
1227 	longmode = is_64_bit_hypercall(vcpu);
1228 	if (!longmode) {
1229 		params[0] = (u32)kvm_rbx_read(vcpu);
1230 		params[1] = (u32)kvm_rcx_read(vcpu);
1231 		params[2] = (u32)kvm_rdx_read(vcpu);
1232 		params[3] = (u32)kvm_rsi_read(vcpu);
1233 		params[4] = (u32)kvm_rdi_read(vcpu);
1234 		params[5] = (u32)kvm_rbp_read(vcpu);
1235 	}
1236 #ifdef CONFIG_X86_64
1237 	else {
1238 		params[0] = (u64)kvm_rdi_read(vcpu);
1239 		params[1] = (u64)kvm_rsi_read(vcpu);
1240 		params[2] = (u64)kvm_rdx_read(vcpu);
1241 		params[3] = (u64)kvm_r10_read(vcpu);
1242 		params[4] = (u64)kvm_r8_read(vcpu);
1243 		params[5] = (u64)kvm_r9_read(vcpu);
1244 	}
1245 #endif
1246 	trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
1247 				params[3], params[4], params[5]);
1248 
1249 	switch (input) {
1250 	case __HYPERVISOR_xen_version:
1251 		if (params[0] == XENVER_version && vcpu->kvm->arch.xen.xen_version) {
1252 			r = vcpu->kvm->arch.xen.xen_version;
1253 			handled = true;
1254 		}
1255 		break;
1256 	case __HYPERVISOR_event_channel_op:
1257 		if (params[0] == EVTCHNOP_send)
1258 			handled = kvm_xen_hcall_evtchn_send(vcpu, params[1], &r);
1259 		break;
1260 	case __HYPERVISOR_sched_op:
1261 		handled = kvm_xen_hcall_sched_op(vcpu, longmode, params[0],
1262 						 params[1], &r);
1263 		break;
1264 	case __HYPERVISOR_vcpu_op:
1265 		handled = kvm_xen_hcall_vcpu_op(vcpu, longmode, params[0], params[1],
1266 						params[2], &r);
1267 		break;
1268 	case __HYPERVISOR_set_timer_op: {
1269 		u64 timeout = params[0];
1270 		/* In 32-bit mode, the 64-bit timeout is in two 32-bit params. */
1271 		if (!longmode)
1272 			timeout |= params[1] << 32;
1273 		handled = kvm_xen_hcall_set_timer_op(vcpu, timeout, &r);
1274 		break;
1275 	}
1276 	default:
1277 		break;
1278 	}
1279 
1280 	if (handled)
1281 		return kvm_xen_hypercall_set_result(vcpu, r);
1282 
1283 	vcpu->run->exit_reason = KVM_EXIT_XEN;
1284 	vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
1285 	vcpu->run->xen.u.hcall.longmode = longmode;
1286 	vcpu->run->xen.u.hcall.cpl = static_call(kvm_x86_get_cpl)(vcpu);
1287 	vcpu->run->xen.u.hcall.input = input;
1288 	vcpu->run->xen.u.hcall.params[0] = params[0];
1289 	vcpu->run->xen.u.hcall.params[1] = params[1];
1290 	vcpu->run->xen.u.hcall.params[2] = params[2];
1291 	vcpu->run->xen.u.hcall.params[3] = params[3];
1292 	vcpu->run->xen.u.hcall.params[4] = params[4];
1293 	vcpu->run->xen.u.hcall.params[5] = params[5];
1294 	vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
1295 	vcpu->arch.complete_userspace_io =
1296 		kvm_xen_hypercall_complete_userspace;
1297 
1298 	return 0;
1299 }
1300 
1301 static inline int max_evtchn_port(struct kvm *kvm)
1302 {
1303 	if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
1304 		return EVTCHN_2L_NR_CHANNELS;
1305 	else
1306 		return COMPAT_EVTCHN_2L_NR_CHANNELS;
1307 }
1308 
1309 static void kvm_xen_check_poller(struct kvm_vcpu *vcpu, int port)
1310 {
1311 	int poll_evtchn = vcpu->arch.xen.poll_evtchn;
1312 
1313 	if ((poll_evtchn == port || poll_evtchn == -1) &&
1314 	    test_and_clear_bit(vcpu->vcpu_idx, vcpu->kvm->arch.xen.poll_mask)) {
1315 		kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1316 		kvm_vcpu_kick(vcpu);
1317 	}
1318 }
1319 
1320 /*
1321  * The return value from this function is propagated to kvm_set_irq() API,
1322  * so it returns:
1323  *  < 0   Interrupt was ignored (masked or not delivered for other reasons)
1324  *  = 0   Interrupt was coalesced (previous irq is still pending)
1325  *  > 0   Number of CPUs interrupt was delivered to
1326  *
1327  * It is also called directly from kvm_arch_set_irq_inatomic(), where the
1328  * only check on its return value is a comparison with -EWOULDBLOCK'.
1329  */
1330 int kvm_xen_set_evtchn_fast(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1331 {
1332 	struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1333 	struct kvm_vcpu *vcpu;
1334 	unsigned long *pending_bits, *mask_bits;
1335 	unsigned long flags;
1336 	int port_word_bit;
1337 	bool kick_vcpu = false;
1338 	int vcpu_idx, idx, rc;
1339 
1340 	vcpu_idx = READ_ONCE(xe->vcpu_idx);
1341 	if (vcpu_idx >= 0)
1342 		vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1343 	else {
1344 		vcpu = kvm_get_vcpu_by_id(kvm, xe->vcpu_id);
1345 		if (!vcpu)
1346 			return -EINVAL;
1347 		WRITE_ONCE(xe->vcpu_idx, vcpu->vcpu_idx);
1348 	}
1349 
1350 	if (!vcpu->arch.xen.vcpu_info_cache.active)
1351 		return -EINVAL;
1352 
1353 	if (xe->port >= max_evtchn_port(kvm))
1354 		return -EINVAL;
1355 
1356 	rc = -EWOULDBLOCK;
1357 
1358 	idx = srcu_read_lock(&kvm->srcu);
1359 
1360 	read_lock_irqsave(&gpc->lock, flags);
1361 	if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, PAGE_SIZE))
1362 		goto out_rcu;
1363 
1364 	if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1365 		struct shared_info *shinfo = gpc->khva;
1366 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1367 		mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1368 		port_word_bit = xe->port / 64;
1369 	} else {
1370 		struct compat_shared_info *shinfo = gpc->khva;
1371 		pending_bits = (unsigned long *)&shinfo->evtchn_pending;
1372 		mask_bits = (unsigned long *)&shinfo->evtchn_mask;
1373 		port_word_bit = xe->port / 32;
1374 	}
1375 
1376 	/*
1377 	 * If this port wasn't already set, and if it isn't masked, then
1378 	 * we try to set the corresponding bit in the in-kernel shadow of
1379 	 * evtchn_pending_sel for the target vCPU. And if *that* wasn't
1380 	 * already set, then we kick the vCPU in question to write to the
1381 	 * *real* evtchn_pending_sel in its own guest vcpu_info struct.
1382 	 */
1383 	if (test_and_set_bit(xe->port, pending_bits)) {
1384 		rc = 0; /* It was already raised */
1385 	} else if (test_bit(xe->port, mask_bits)) {
1386 		rc = -ENOTCONN; /* Masked */
1387 		kvm_xen_check_poller(vcpu, xe->port);
1388 	} else {
1389 		rc = 1; /* Delivered to the bitmap in shared_info. */
1390 		/* Now switch to the vCPU's vcpu_info to set the index and pending_sel */
1391 		read_unlock_irqrestore(&gpc->lock, flags);
1392 		gpc = &vcpu->arch.xen.vcpu_info_cache;
1393 
1394 		read_lock_irqsave(&gpc->lock, flags);
1395 		if (!kvm_gfn_to_pfn_cache_check(kvm, gpc, gpc->gpa, sizeof(struct vcpu_info))) {
1396 			/*
1397 			 * Could not access the vcpu_info. Set the bit in-kernel
1398 			 * and prod the vCPU to deliver it for itself.
1399 			 */
1400 			if (!test_and_set_bit(port_word_bit, &vcpu->arch.xen.evtchn_pending_sel))
1401 				kick_vcpu = true;
1402 			goto out_rcu;
1403 		}
1404 
1405 		if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode) {
1406 			struct vcpu_info *vcpu_info = gpc->khva;
1407 			if (!test_and_set_bit(port_word_bit, &vcpu_info->evtchn_pending_sel)) {
1408 				WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1409 				kick_vcpu = true;
1410 			}
1411 		} else {
1412 			struct compat_vcpu_info *vcpu_info = gpc->khva;
1413 			if (!test_and_set_bit(port_word_bit,
1414 					      (unsigned long *)&vcpu_info->evtchn_pending_sel)) {
1415 				WRITE_ONCE(vcpu_info->evtchn_upcall_pending, 1);
1416 				kick_vcpu = true;
1417 			}
1418 		}
1419 
1420 		/* For the per-vCPU lapic vector, deliver it as MSI. */
1421 		if (kick_vcpu && vcpu->arch.xen.upcall_vector) {
1422 			kvm_xen_inject_vcpu_vector(vcpu);
1423 			kick_vcpu = false;
1424 		}
1425 	}
1426 
1427  out_rcu:
1428 	read_unlock_irqrestore(&gpc->lock, flags);
1429 	srcu_read_unlock(&kvm->srcu, idx);
1430 
1431 	if (kick_vcpu) {
1432 		kvm_make_request(KVM_REQ_UNBLOCK, vcpu);
1433 		kvm_vcpu_kick(vcpu);
1434 	}
1435 
1436 	return rc;
1437 }
1438 
1439 static int kvm_xen_set_evtchn(struct kvm_xen_evtchn *xe, struct kvm *kvm)
1440 {
1441 	bool mm_borrowed = false;
1442 	int rc;
1443 
1444 	rc = kvm_xen_set_evtchn_fast(xe, kvm);
1445 	if (rc != -EWOULDBLOCK)
1446 		return rc;
1447 
1448 	if (current->mm != kvm->mm) {
1449 		/*
1450 		 * If not on a thread which already belongs to this KVM,
1451 		 * we'd better be in the irqfd workqueue.
1452 		 */
1453 		if (WARN_ON_ONCE(current->mm))
1454 			return -EINVAL;
1455 
1456 		kthread_use_mm(kvm->mm);
1457 		mm_borrowed = true;
1458 	}
1459 
1460 	/*
1461 	 * For the irqfd workqueue, using the main kvm->lock mutex is
1462 	 * fine since this function is invoked from kvm_set_irq() with
1463 	 * no other lock held, no srcu. In future if it will be called
1464 	 * directly from a vCPU thread (e.g. on hypercall for an IPI)
1465 	 * then it may need to switch to using a leaf-node mutex for
1466 	 * serializing the shared_info mapping.
1467 	 */
1468 	mutex_lock(&kvm->lock);
1469 
1470 	/*
1471 	 * It is theoretically possible for the page to be unmapped
1472 	 * and the MMU notifier to invalidate the shared_info before
1473 	 * we even get to use it. In that case, this looks like an
1474 	 * infinite loop. It was tempting to do it via the userspace
1475 	 * HVA instead... but that just *hides* the fact that it's
1476 	 * an infinite loop, because if a fault occurs and it waits
1477 	 * for the page to come back, it can *still* immediately
1478 	 * fault and have to wait again, repeatedly.
1479 	 *
1480 	 * Conversely, the page could also have been reinstated by
1481 	 * another thread before we even obtain the mutex above, so
1482 	 * check again *first* before remapping it.
1483 	 */
1484 	do {
1485 		struct gfn_to_pfn_cache *gpc = &kvm->arch.xen.shinfo_cache;
1486 		int idx;
1487 
1488 		rc = kvm_xen_set_evtchn_fast(xe, kvm);
1489 		if (rc != -EWOULDBLOCK)
1490 			break;
1491 
1492 		idx = srcu_read_lock(&kvm->srcu);
1493 		rc = kvm_gfn_to_pfn_cache_refresh(kvm, gpc, gpc->gpa, PAGE_SIZE);
1494 		srcu_read_unlock(&kvm->srcu, idx);
1495 	} while(!rc);
1496 
1497 	mutex_unlock(&kvm->lock);
1498 
1499 	if (mm_borrowed)
1500 		kthread_unuse_mm(kvm->mm);
1501 
1502 	return rc;
1503 }
1504 
1505 /* This is the version called from kvm_set_irq() as the .set function */
1506 static int evtchn_set_fn(struct kvm_kernel_irq_routing_entry *e, struct kvm *kvm,
1507 			 int irq_source_id, int level, bool line_status)
1508 {
1509 	if (!level)
1510 		return -EINVAL;
1511 
1512 	return kvm_xen_set_evtchn(&e->xen_evtchn, kvm);
1513 }
1514 
1515 /*
1516  * Set up an event channel interrupt from the KVM IRQ routing table.
1517  * Used for e.g. PIRQ from passed through physical devices.
1518  */
1519 int kvm_xen_setup_evtchn(struct kvm *kvm,
1520 			 struct kvm_kernel_irq_routing_entry *e,
1521 			 const struct kvm_irq_routing_entry *ue)
1522 
1523 {
1524 	struct kvm_vcpu *vcpu;
1525 
1526 	if (ue->u.xen_evtchn.port >= max_evtchn_port(kvm))
1527 		return -EINVAL;
1528 
1529 	/* We only support 2 level event channels for now */
1530 	if (ue->u.xen_evtchn.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1531 		return -EINVAL;
1532 
1533 	/*
1534 	 * Xen gives us interesting mappings from vCPU index to APIC ID,
1535 	 * which means kvm_get_vcpu_by_id() has to iterate over all vCPUs
1536 	 * to find it. Do that once at setup time, instead of every time.
1537 	 * But beware that on live update / live migration, the routing
1538 	 * table might be reinstated before the vCPU threads have finished
1539 	 * recreating their vCPUs.
1540 	 */
1541 	vcpu = kvm_get_vcpu_by_id(kvm, ue->u.xen_evtchn.vcpu);
1542 	if (vcpu)
1543 		e->xen_evtchn.vcpu_idx = vcpu->vcpu_idx;
1544 	else
1545 		e->xen_evtchn.vcpu_idx = -1;
1546 
1547 	e->xen_evtchn.port = ue->u.xen_evtchn.port;
1548 	e->xen_evtchn.vcpu_id = ue->u.xen_evtchn.vcpu;
1549 	e->xen_evtchn.priority = ue->u.xen_evtchn.priority;
1550 	e->set = evtchn_set_fn;
1551 
1552 	return 0;
1553 }
1554 
1555 /*
1556  * Explicit event sending from userspace with KVM_XEN_HVM_EVTCHN_SEND ioctl.
1557  */
1558 int kvm_xen_hvm_evtchn_send(struct kvm *kvm, struct kvm_irq_routing_xen_evtchn *uxe)
1559 {
1560 	struct kvm_xen_evtchn e;
1561 	int ret;
1562 
1563 	if (!uxe->port || uxe->port >= max_evtchn_port(kvm))
1564 		return -EINVAL;
1565 
1566 	/* We only support 2 level event channels for now */
1567 	if (uxe->priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1568 		return -EINVAL;
1569 
1570 	e.port = uxe->port;
1571 	e.vcpu_id = uxe->vcpu;
1572 	e.vcpu_idx = -1;
1573 	e.priority = uxe->priority;
1574 
1575 	ret = kvm_xen_set_evtchn(&e, kvm);
1576 
1577 	/*
1578 	 * None of that 'return 1 if it actually got delivered' nonsense.
1579 	 * We don't care if it was masked (-ENOTCONN) either.
1580 	 */
1581 	if (ret > 0 || ret == -ENOTCONN)
1582 		ret = 0;
1583 
1584 	return ret;
1585 }
1586 
1587 /*
1588  * Support for *outbound* event channel events via the EVTCHNOP_send hypercall.
1589  */
1590 struct evtchnfd {
1591 	u32 send_port;
1592 	u32 type;
1593 	union {
1594 		struct kvm_xen_evtchn port;
1595 		struct {
1596 			u32 port; /* zero */
1597 			struct eventfd_ctx *ctx;
1598 		} eventfd;
1599 	} deliver;
1600 };
1601 
1602 /*
1603  * Update target vCPU or priority for a registered sending channel.
1604  */
1605 static int kvm_xen_eventfd_update(struct kvm *kvm,
1606 				  struct kvm_xen_hvm_attr *data)
1607 {
1608 	u32 port = data->u.evtchn.send_port;
1609 	struct evtchnfd *evtchnfd;
1610 
1611 	if (!port || port >= max_evtchn_port(kvm))
1612 		return -EINVAL;
1613 
1614 	mutex_lock(&kvm->lock);
1615 	evtchnfd = idr_find(&kvm->arch.xen.evtchn_ports, port);
1616 	mutex_unlock(&kvm->lock);
1617 
1618 	if (!evtchnfd)
1619 		return -ENOENT;
1620 
1621 	/* For an UPDATE, nothing may change except the priority/vcpu */
1622 	if (evtchnfd->type != data->u.evtchn.type)
1623 		return -EINVAL;
1624 
1625 	/*
1626 	 * Port cannot change, and if it's zero that was an eventfd
1627 	 * which can't be changed either.
1628 	 */
1629 	if (!evtchnfd->deliver.port.port ||
1630 	    evtchnfd->deliver.port.port != data->u.evtchn.deliver.port.port)
1631 		return -EINVAL;
1632 
1633 	/* We only support 2 level event channels for now */
1634 	if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1635 		return -EINVAL;
1636 
1637 	mutex_lock(&kvm->lock);
1638 	evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1639 	if (evtchnfd->deliver.port.vcpu_id != data->u.evtchn.deliver.port.vcpu) {
1640 		evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1641 		evtchnfd->deliver.port.vcpu_idx = -1;
1642 	}
1643 	mutex_unlock(&kvm->lock);
1644 	return 0;
1645 }
1646 
1647 /*
1648  * Configure the target (eventfd or local port delivery) for sending on
1649  * a given event channel.
1650  */
1651 static int kvm_xen_eventfd_assign(struct kvm *kvm,
1652 				  struct kvm_xen_hvm_attr *data)
1653 {
1654 	u32 port = data->u.evtchn.send_port;
1655 	struct eventfd_ctx *eventfd = NULL;
1656 	struct evtchnfd *evtchnfd = NULL;
1657 	int ret = -EINVAL;
1658 
1659 	if (!port || port >= max_evtchn_port(kvm))
1660 		return -EINVAL;
1661 
1662 	evtchnfd = kzalloc(sizeof(struct evtchnfd), GFP_KERNEL);
1663 	if (!evtchnfd)
1664 		return -ENOMEM;
1665 
1666 	switch(data->u.evtchn.type) {
1667 	case EVTCHNSTAT_ipi:
1668 		/* IPI  must map back to the same port# */
1669 		if (data->u.evtchn.deliver.port.port != data->u.evtchn.send_port)
1670 			goto out; /* -EINVAL */
1671 		break;
1672 
1673 	case EVTCHNSTAT_interdomain:
1674 		if (data->u.evtchn.deliver.port.port) {
1675 			if (data->u.evtchn.deliver.port.port >= max_evtchn_port(kvm))
1676 				goto out; /* -EINVAL */
1677 		} else {
1678 			eventfd = eventfd_ctx_fdget(data->u.evtchn.deliver.eventfd.fd);
1679 			if (IS_ERR(eventfd)) {
1680 				ret = PTR_ERR(eventfd);
1681 				goto out;
1682 			}
1683 		}
1684 		break;
1685 
1686 	case EVTCHNSTAT_virq:
1687 	case EVTCHNSTAT_closed:
1688 	case EVTCHNSTAT_unbound:
1689 	case EVTCHNSTAT_pirq:
1690 	default: /* Unknown event channel type */
1691 		goto out; /* -EINVAL */
1692 	}
1693 
1694 	evtchnfd->send_port = data->u.evtchn.send_port;
1695 	evtchnfd->type = data->u.evtchn.type;
1696 	if (eventfd) {
1697 		evtchnfd->deliver.eventfd.ctx = eventfd;
1698 	} else {
1699 		/* We only support 2 level event channels for now */
1700 		if (data->u.evtchn.deliver.port.priority != KVM_IRQ_ROUTING_XEN_EVTCHN_PRIO_2LEVEL)
1701 			goto out; /* -EINVAL; */
1702 
1703 		evtchnfd->deliver.port.port = data->u.evtchn.deliver.port.port;
1704 		evtchnfd->deliver.port.vcpu_id = data->u.evtchn.deliver.port.vcpu;
1705 		evtchnfd->deliver.port.vcpu_idx = -1;
1706 		evtchnfd->deliver.port.priority = data->u.evtchn.deliver.port.priority;
1707 	}
1708 
1709 	mutex_lock(&kvm->lock);
1710 	ret = idr_alloc(&kvm->arch.xen.evtchn_ports, evtchnfd, port, port + 1,
1711 			GFP_KERNEL);
1712 	mutex_unlock(&kvm->lock);
1713 	if (ret >= 0)
1714 		return 0;
1715 
1716 	if (ret == -ENOSPC)
1717 		ret = -EEXIST;
1718 out:
1719 	if (eventfd)
1720 		eventfd_ctx_put(eventfd);
1721 	kfree(evtchnfd);
1722 	return ret;
1723 }
1724 
1725 static int kvm_xen_eventfd_deassign(struct kvm *kvm, u32 port)
1726 {
1727 	struct evtchnfd *evtchnfd;
1728 
1729 	mutex_lock(&kvm->lock);
1730 	evtchnfd = idr_remove(&kvm->arch.xen.evtchn_ports, port);
1731 	mutex_unlock(&kvm->lock);
1732 
1733 	if (!evtchnfd)
1734 		return -ENOENT;
1735 
1736 	if (kvm)
1737 		synchronize_srcu(&kvm->srcu);
1738 	if (!evtchnfd->deliver.port.port)
1739 		eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1740 	kfree(evtchnfd);
1741 	return 0;
1742 }
1743 
1744 static int kvm_xen_eventfd_reset(struct kvm *kvm)
1745 {
1746 	struct evtchnfd *evtchnfd;
1747 	int i;
1748 
1749 	mutex_lock(&kvm->lock);
1750 	idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
1751 		idr_remove(&kvm->arch.xen.evtchn_ports, evtchnfd->send_port);
1752 		synchronize_srcu(&kvm->srcu);
1753 		if (!evtchnfd->deliver.port.port)
1754 			eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1755 		kfree(evtchnfd);
1756 	}
1757 	mutex_unlock(&kvm->lock);
1758 
1759 	return 0;
1760 }
1761 
1762 static int kvm_xen_setattr_evtchn(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
1763 {
1764 	u32 port = data->u.evtchn.send_port;
1765 
1766 	if (data->u.evtchn.flags == KVM_XEN_EVTCHN_RESET)
1767 		return kvm_xen_eventfd_reset(kvm);
1768 
1769 	if (!port || port >= max_evtchn_port(kvm))
1770 		return -EINVAL;
1771 
1772 	if (data->u.evtchn.flags == KVM_XEN_EVTCHN_DEASSIGN)
1773 		return kvm_xen_eventfd_deassign(kvm, port);
1774 	if (data->u.evtchn.flags == KVM_XEN_EVTCHN_UPDATE)
1775 		return kvm_xen_eventfd_update(kvm, data);
1776 	if (data->u.evtchn.flags)
1777 		return -EINVAL;
1778 
1779 	return kvm_xen_eventfd_assign(kvm, data);
1780 }
1781 
1782 static bool kvm_xen_hcall_evtchn_send(struct kvm_vcpu *vcpu, u64 param, u64 *r)
1783 {
1784 	struct evtchnfd *evtchnfd;
1785 	struct evtchn_send send;
1786 	gpa_t gpa;
1787 	int idx;
1788 
1789 	idx = srcu_read_lock(&vcpu->kvm->srcu);
1790 	gpa = kvm_mmu_gva_to_gpa_system(vcpu, param, NULL);
1791 	srcu_read_unlock(&vcpu->kvm->srcu, idx);
1792 
1793 	if (!gpa || kvm_vcpu_read_guest(vcpu, gpa, &send, sizeof(send))) {
1794 		*r = -EFAULT;
1795 		return true;
1796 	}
1797 
1798 	/* The evtchn_ports idr is protected by vcpu->kvm->srcu */
1799 	evtchnfd = idr_find(&vcpu->kvm->arch.xen.evtchn_ports, send.port);
1800 	if (!evtchnfd)
1801 		return false;
1802 
1803 	if (evtchnfd->deliver.port.port) {
1804 		int ret = kvm_xen_set_evtchn(&evtchnfd->deliver.port, vcpu->kvm);
1805 		if (ret < 0 && ret != -ENOTCONN)
1806 			return false;
1807 	} else {
1808 		eventfd_signal(evtchnfd->deliver.eventfd.ctx, 1);
1809 	}
1810 
1811 	*r = 0;
1812 	return true;
1813 }
1814 
1815 void kvm_xen_init_vcpu(struct kvm_vcpu *vcpu)
1816 {
1817 	vcpu->arch.xen.vcpu_id = vcpu->vcpu_idx;
1818 	vcpu->arch.xen.poll_evtchn = 0;
1819 	timer_setup(&vcpu->arch.xen.poll_timer, cancel_evtchn_poll, 0);
1820 }
1821 
1822 void kvm_xen_destroy_vcpu(struct kvm_vcpu *vcpu)
1823 {
1824 	if (kvm_xen_timer_enabled(vcpu))
1825 		kvm_xen_stop_timer(vcpu);
1826 
1827 	kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1828 				     &vcpu->arch.xen.runstate_cache);
1829 	kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1830 				     &vcpu->arch.xen.vcpu_info_cache);
1831 	kvm_gfn_to_pfn_cache_destroy(vcpu->kvm,
1832 				     &vcpu->arch.xen.vcpu_time_info_cache);
1833 	del_timer_sync(&vcpu->arch.xen.poll_timer);
1834 }
1835 
1836 void kvm_xen_init_vm(struct kvm *kvm)
1837 {
1838 	idr_init(&kvm->arch.xen.evtchn_ports);
1839 }
1840 
1841 void kvm_xen_destroy_vm(struct kvm *kvm)
1842 {
1843 	struct evtchnfd *evtchnfd;
1844 	int i;
1845 
1846 	kvm_gfn_to_pfn_cache_destroy(kvm, &kvm->arch.xen.shinfo_cache);
1847 
1848 	idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
1849 		if (!evtchnfd->deliver.port.port)
1850 			eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
1851 		kfree(evtchnfd);
1852 	}
1853 	idr_destroy(&kvm->arch.xen.evtchn_ports);
1854 
1855 	if (kvm->arch.xen_hvm_config.msr)
1856 		static_branch_slow_dec_deferred(&kvm_xen_enabled);
1857 }
1858