xref: /openbmc/linux/arch/x86/kvm/svm/avic.c (revision 8e7ae8ba)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel-based Virtual Machine driver for Linux
4  *
5  * AMD SVM support
6  *
7  * Copyright (C) 2006 Qumranet, Inc.
8  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
9  *
10  * Authors:
11  *   Yaniv Kamay  <yaniv@qumranet.com>
12  *   Avi Kivity   <avi@qumranet.com>
13  */
14 
15 #define pr_fmt(fmt) "SVM: " fmt
16 
17 #include <linux/kvm_types.h>
18 #include <linux/hashtable.h>
19 #include <linux/amd-iommu.h>
20 #include <linux/kvm_host.h>
21 
22 #include <asm/irq_remapping.h>
23 
24 #include "trace.h"
25 #include "lapic.h"
26 #include "x86.h"
27 #include "irq.h"
28 #include "svm.h"
29 
30 /* AVIC GATAG is encoded using VM and VCPU IDs */
31 #define AVIC_VCPU_ID_BITS		8
32 #define AVIC_VCPU_ID_MASK		((1 << AVIC_VCPU_ID_BITS) - 1)
33 
34 #define AVIC_VM_ID_BITS			24
35 #define AVIC_VM_ID_NR			(1 << AVIC_VM_ID_BITS)
36 #define AVIC_VM_ID_MASK			((1 << AVIC_VM_ID_BITS) - 1)
37 
38 #define AVIC_GATAG(x, y)		(((x & AVIC_VM_ID_MASK) << AVIC_VCPU_ID_BITS) | \
39 						(y & AVIC_VCPU_ID_MASK))
40 #define AVIC_GATAG_TO_VMID(x)		((x >> AVIC_VCPU_ID_BITS) & AVIC_VM_ID_MASK)
41 #define AVIC_GATAG_TO_VCPUID(x)		(x & AVIC_VCPU_ID_MASK)
42 
43 /* Note:
44  * This hash table is used to map VM_ID to a struct kvm_svm,
45  * when handling AMD IOMMU GALOG notification to schedule in
46  * a particular vCPU.
47  */
48 #define SVM_VM_DATA_HASH_BITS	8
49 static DEFINE_HASHTABLE(svm_vm_data_hash, SVM_VM_DATA_HASH_BITS);
50 static u32 next_vm_id = 0;
51 static bool next_vm_id_wrapped = 0;
52 static DEFINE_SPINLOCK(svm_vm_data_hash_lock);
53 
54 /*
55  * This is a wrapper of struct amd_iommu_ir_data.
56  */
57 struct amd_svm_iommu_ir {
58 	struct list_head node;	/* Used by SVM for per-vcpu ir_list */
59 	void *data;		/* Storing pointer to struct amd_ir_data */
60 };
61 
62 
63 /* Note:
64  * This function is called from IOMMU driver to notify
65  * SVM to schedule in a particular vCPU of a particular VM.
66  */
67 int avic_ga_log_notifier(u32 ga_tag)
68 {
69 	unsigned long flags;
70 	struct kvm_svm *kvm_svm;
71 	struct kvm_vcpu *vcpu = NULL;
72 	u32 vm_id = AVIC_GATAG_TO_VMID(ga_tag);
73 	u32 vcpu_id = AVIC_GATAG_TO_VCPUID(ga_tag);
74 
75 	pr_debug("SVM: %s: vm_id=%#x, vcpu_id=%#x\n", __func__, vm_id, vcpu_id);
76 	trace_kvm_avic_ga_log(vm_id, vcpu_id);
77 
78 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
79 	hash_for_each_possible(svm_vm_data_hash, kvm_svm, hnode, vm_id) {
80 		if (kvm_svm->avic_vm_id != vm_id)
81 			continue;
82 		vcpu = kvm_get_vcpu_by_id(&kvm_svm->kvm, vcpu_id);
83 		break;
84 	}
85 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
86 
87 	/* Note:
88 	 * At this point, the IOMMU should have already set the pending
89 	 * bit in the vAPIC backing page. So, we just need to schedule
90 	 * in the vcpu.
91 	 */
92 	if (vcpu)
93 		kvm_vcpu_wake_up(vcpu);
94 
95 	return 0;
96 }
97 
98 void avic_vm_destroy(struct kvm *kvm)
99 {
100 	unsigned long flags;
101 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
102 
103 	if (!enable_apicv)
104 		return;
105 
106 	if (kvm_svm->avic_logical_id_table_page)
107 		__free_page(kvm_svm->avic_logical_id_table_page);
108 	if (kvm_svm->avic_physical_id_table_page)
109 		__free_page(kvm_svm->avic_physical_id_table_page);
110 
111 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
112 	hash_del(&kvm_svm->hnode);
113 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
114 }
115 
116 int avic_vm_init(struct kvm *kvm)
117 {
118 	unsigned long flags;
119 	int err = -ENOMEM;
120 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
121 	struct kvm_svm *k2;
122 	struct page *p_page;
123 	struct page *l_page;
124 	u32 vm_id;
125 
126 	if (!enable_apicv)
127 		return 0;
128 
129 	/* Allocating physical APIC ID table (4KB) */
130 	p_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
131 	if (!p_page)
132 		goto free_avic;
133 
134 	kvm_svm->avic_physical_id_table_page = p_page;
135 
136 	/* Allocating logical APIC ID table (4KB) */
137 	l_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
138 	if (!l_page)
139 		goto free_avic;
140 
141 	kvm_svm->avic_logical_id_table_page = l_page;
142 
143 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
144  again:
145 	vm_id = next_vm_id = (next_vm_id + 1) & AVIC_VM_ID_MASK;
146 	if (vm_id == 0) { /* id is 1-based, zero is not okay */
147 		next_vm_id_wrapped = 1;
148 		goto again;
149 	}
150 	/* Is it still in use? Only possible if wrapped at least once */
151 	if (next_vm_id_wrapped) {
152 		hash_for_each_possible(svm_vm_data_hash, k2, hnode, vm_id) {
153 			if (k2->avic_vm_id == vm_id)
154 				goto again;
155 		}
156 	}
157 	kvm_svm->avic_vm_id = vm_id;
158 	hash_add(svm_vm_data_hash, &kvm_svm->hnode, kvm_svm->avic_vm_id);
159 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
160 
161 	return 0;
162 
163 free_avic:
164 	avic_vm_destroy(kvm);
165 	return err;
166 }
167 
168 void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
169 {
170 	struct kvm_svm *kvm_svm = to_kvm_svm(svm->vcpu.kvm);
171 	phys_addr_t bpa = __sme_set(page_to_phys(svm->avic_backing_page));
172 	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
173 	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
174 
175 	vmcb->control.avic_backing_page = bpa & AVIC_HPA_MASK;
176 	vmcb->control.avic_logical_id = lpa & AVIC_HPA_MASK;
177 	vmcb->control.avic_physical_id = ppa & AVIC_HPA_MASK;
178 	vmcb->control.avic_physical_id |= AVIC_MAX_PHYSICAL_ID_COUNT;
179 	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE & VMCB_AVIC_APIC_BAR_MASK;
180 
181 	if (kvm_apicv_activated(svm->vcpu.kvm))
182 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
183 	else
184 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
185 }
186 
187 static u64 *avic_get_physical_id_entry(struct kvm_vcpu *vcpu,
188 				       unsigned int index)
189 {
190 	u64 *avic_physical_id_table;
191 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
192 
193 	if (index >= AVIC_MAX_PHYSICAL_ID_COUNT)
194 		return NULL;
195 
196 	avic_physical_id_table = page_address(kvm_svm->avic_physical_id_table_page);
197 
198 	return &avic_physical_id_table[index];
199 }
200 
201 /*
202  * Note:
203  * AVIC hardware walks the nested page table to check permissions,
204  * but does not use the SPA address specified in the leaf page
205  * table entry since it uses  address in the AVIC_BACKING_PAGE pointer
206  * field of the VMCB. Therefore, we set up the
207  * APIC_ACCESS_PAGE_PRIVATE_MEMSLOT (4KB) here.
208  */
209 static int avic_alloc_access_page(struct kvm *kvm)
210 {
211 	void __user *ret;
212 	int r = 0;
213 
214 	mutex_lock(&kvm->slots_lock);
215 
216 	if (kvm->arch.apic_access_memslot_enabled)
217 		goto out;
218 
219 	ret = __x86_set_memory_region(kvm,
220 				      APIC_ACCESS_PAGE_PRIVATE_MEMSLOT,
221 				      APIC_DEFAULT_PHYS_BASE,
222 				      PAGE_SIZE);
223 	if (IS_ERR(ret)) {
224 		r = PTR_ERR(ret);
225 		goto out;
226 	}
227 
228 	kvm->arch.apic_access_memslot_enabled = true;
229 out:
230 	mutex_unlock(&kvm->slots_lock);
231 	return r;
232 }
233 
234 static int avic_init_backing_page(struct kvm_vcpu *vcpu)
235 {
236 	u64 *entry, new_entry;
237 	int id = vcpu->vcpu_id;
238 	struct vcpu_svm *svm = to_svm(vcpu);
239 
240 	if (id >= AVIC_MAX_PHYSICAL_ID_COUNT)
241 		return -EINVAL;
242 
243 	if (!vcpu->arch.apic->regs)
244 		return -EINVAL;
245 
246 	if (kvm_apicv_activated(vcpu->kvm)) {
247 		int ret;
248 
249 		ret = avic_alloc_access_page(vcpu->kvm);
250 		if (ret)
251 			return ret;
252 	}
253 
254 	svm->avic_backing_page = virt_to_page(vcpu->arch.apic->regs);
255 
256 	/* Setting AVIC backing page address in the phy APIC ID table */
257 	entry = avic_get_physical_id_entry(vcpu, id);
258 	if (!entry)
259 		return -EINVAL;
260 
261 	new_entry = __sme_set((page_to_phys(svm->avic_backing_page) &
262 			      AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK) |
263 			      AVIC_PHYSICAL_ID_ENTRY_VALID_MASK);
264 	WRITE_ONCE(*entry, new_entry);
265 
266 	svm->avic_physical_id_cache = entry;
267 
268 	return 0;
269 }
270 
271 void avic_ring_doorbell(struct kvm_vcpu *vcpu)
272 {
273 	/*
274 	 * Note, the vCPU could get migrated to a different pCPU at any point,
275 	 * which could result in signalling the wrong/previous pCPU.  But if
276 	 * that happens the vCPU is guaranteed to do a VMRUN (after being
277 	 * migrated) and thus will process pending interrupts, i.e. a doorbell
278 	 * is not needed (and the spurious one is harmless).
279 	 */
280 	int cpu = READ_ONCE(vcpu->cpu);
281 
282 	if (cpu != get_cpu())
283 		wrmsrl(MSR_AMD64_SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpu));
284 	put_cpu();
285 }
286 
287 /*
288  * A fast-path version of avic_kick_target_vcpus(), which attempts to match
289  * destination APIC ID to vCPU without looping through all vCPUs.
290  */
291 static int avic_kick_target_vcpus_fast(struct kvm *kvm, struct kvm_lapic *source,
292 				       u32 icrl, u32 icrh, u32 index)
293 {
294 	u32 dest, apic_id;
295 	struct kvm_vcpu *vcpu;
296 	int dest_mode = icrl & APIC_DEST_MASK;
297 	int shorthand = icrl & APIC_SHORT_MASK;
298 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
299 	u32 *avic_logical_id_table = page_address(kvm_svm->avic_logical_id_table_page);
300 
301 	if (shorthand != APIC_DEST_NOSHORT)
302 		return -EINVAL;
303 
304 	/*
305 	 * The AVIC incomplete IPI #vmexit info provides index into
306 	 * the physical APIC ID table, which can be used to derive
307 	 * guest physical APIC ID.
308 	 */
309 	if (dest_mode == APIC_DEST_PHYSICAL) {
310 		apic_id = index;
311 	} else {
312 		if (!apic_x2apic_mode(source)) {
313 			/* For xAPIC logical mode, the index is for logical APIC table. */
314 			apic_id = avic_logical_id_table[index] & 0x1ff;
315 		} else {
316 			return -EINVAL;
317 		}
318 	}
319 
320 	/*
321 	 * Assuming vcpu ID is the same as physical apic ID,
322 	 * and use it to retrieve the target vCPU.
323 	 */
324 	vcpu = kvm_get_vcpu_by_id(kvm, apic_id);
325 	if (!vcpu)
326 		return -EINVAL;
327 
328 	if (apic_x2apic_mode(vcpu->arch.apic))
329 		dest = icrh;
330 	else
331 		dest = GET_APIC_DEST_FIELD(icrh);
332 
333 	/*
334 	 * Try matching the destination APIC ID with the vCPU.
335 	 */
336 	if (kvm_apic_match_dest(vcpu, source, shorthand, dest, dest_mode)) {
337 		vcpu->arch.apic->irr_pending = true;
338 		svm_complete_interrupt_delivery(vcpu,
339 						icrl & APIC_MODE_MASK,
340 						icrl & APIC_INT_LEVELTRIG,
341 						icrl & APIC_VECTOR_MASK);
342 		return 0;
343 	}
344 
345 	return -EINVAL;
346 }
347 
348 static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
349 				   u32 icrl, u32 icrh, u32 index)
350 {
351 	unsigned long i;
352 	struct kvm_vcpu *vcpu;
353 
354 	if (!avic_kick_target_vcpus_fast(kvm, source, icrl, icrh, index))
355 		return;
356 
357 	trace_kvm_avic_kick_vcpu_slowpath(icrh, icrl, index);
358 
359 	/*
360 	 * Wake any target vCPUs that are blocking, i.e. waiting for a wake
361 	 * event.  There's no need to signal doorbells, as hardware has handled
362 	 * vCPUs that were in guest at the time of the IPI, and vCPUs that have
363 	 * since entered the guest will have processed pending IRQs at VMRUN.
364 	 */
365 	kvm_for_each_vcpu(i, vcpu, kvm) {
366 		if (kvm_apic_match_dest(vcpu, source, icrl & APIC_SHORT_MASK,
367 					GET_APIC_DEST_FIELD(icrh),
368 					icrl & APIC_DEST_MASK)) {
369 			vcpu->arch.apic->irr_pending = true;
370 			svm_complete_interrupt_delivery(vcpu,
371 							icrl & APIC_MODE_MASK,
372 							icrl & APIC_INT_LEVELTRIG,
373 							icrl & APIC_VECTOR_MASK);
374 		}
375 	}
376 }
377 
378 int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
379 {
380 	struct vcpu_svm *svm = to_svm(vcpu);
381 	u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
382 	u32 icrl = svm->vmcb->control.exit_info_1;
383 	u32 id = svm->vmcb->control.exit_info_2 >> 32;
384 	u32 index = svm->vmcb->control.exit_info_2 & 0x1FF;
385 	struct kvm_lapic *apic = vcpu->arch.apic;
386 
387 	trace_kvm_avic_incomplete_ipi(vcpu->vcpu_id, icrh, icrl, id, index);
388 
389 	switch (id) {
390 	case AVIC_IPI_FAILURE_INVALID_INT_TYPE:
391 		/*
392 		 * Emulate IPIs that are not handled by AVIC hardware, which
393 		 * only virtualizes Fixed, Edge-Triggered INTRs.  The exit is
394 		 * a trap, e.g. ICR holds the correct value and RIP has been
395 		 * advanced, KVM is responsible only for emulating the IPI.
396 		 * Sadly, hardware may sometimes leave the BUSY flag set, in
397 		 * which case KVM needs to emulate the ICR write as well in
398 		 * order to clear the BUSY flag.
399 		 */
400 		if (icrl & APIC_ICR_BUSY)
401 			kvm_apic_write_nodecode(vcpu, APIC_ICR);
402 		else
403 			kvm_apic_send_ipi(apic, icrl, icrh);
404 		break;
405 	case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING:
406 		/*
407 		 * At this point, we expect that the AVIC HW has already
408 		 * set the appropriate IRR bits on the valid target
409 		 * vcpus. So, we just need to kick the appropriate vcpu.
410 		 */
411 		avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh, index);
412 		break;
413 	case AVIC_IPI_FAILURE_INVALID_TARGET:
414 		break;
415 	case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
416 		WARN_ONCE(1, "Invalid backing page\n");
417 		break;
418 	default:
419 		pr_err("Unknown IPI interception\n");
420 	}
421 
422 	return 1;
423 }
424 
425 unsigned long avic_vcpu_get_apicv_inhibit_reasons(struct kvm_vcpu *vcpu)
426 {
427 	if (is_guest_mode(vcpu))
428 		return APICV_INHIBIT_REASON_NESTED;
429 	return 0;
430 }
431 
432 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
433 {
434 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
435 	int index;
436 	u32 *logical_apic_id_table;
437 	int dlid = GET_APIC_LOGICAL_ID(ldr);
438 
439 	if (!dlid)
440 		return NULL;
441 
442 	if (flat) { /* flat */
443 		index = ffs(dlid) - 1;
444 		if (index > 7)
445 			return NULL;
446 	} else { /* cluster */
447 		int cluster = (dlid & 0xf0) >> 4;
448 		int apic = ffs(dlid & 0x0f) - 1;
449 
450 		if ((apic < 0) || (apic > 7) ||
451 		    (cluster >= 0xf))
452 			return NULL;
453 		index = (cluster << 2) + apic;
454 	}
455 
456 	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
457 
458 	return &logical_apic_id_table[index];
459 }
460 
461 static int avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
462 {
463 	bool flat;
464 	u32 *entry, new_entry;
465 
466 	flat = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR) == APIC_DFR_FLAT;
467 	entry = avic_get_logical_id_entry(vcpu, ldr, flat);
468 	if (!entry)
469 		return -EINVAL;
470 
471 	new_entry = READ_ONCE(*entry);
472 	new_entry &= ~AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK;
473 	new_entry |= (g_physical_id & AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK);
474 	new_entry |= AVIC_LOGICAL_ID_ENTRY_VALID_MASK;
475 	WRITE_ONCE(*entry, new_entry);
476 
477 	return 0;
478 }
479 
480 static void avic_invalidate_logical_id_entry(struct kvm_vcpu *vcpu)
481 {
482 	struct vcpu_svm *svm = to_svm(vcpu);
483 	bool flat = svm->dfr_reg == APIC_DFR_FLAT;
484 	u32 *entry = avic_get_logical_id_entry(vcpu, svm->ldr_reg, flat);
485 
486 	if (entry)
487 		clear_bit(AVIC_LOGICAL_ID_ENTRY_VALID_BIT, (unsigned long *)entry);
488 }
489 
490 static int avic_handle_ldr_update(struct kvm_vcpu *vcpu)
491 {
492 	int ret = 0;
493 	struct vcpu_svm *svm = to_svm(vcpu);
494 	u32 ldr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_LDR);
495 	u32 id = kvm_xapic_id(vcpu->arch.apic);
496 
497 	if (ldr == svm->ldr_reg)
498 		return 0;
499 
500 	avic_invalidate_logical_id_entry(vcpu);
501 
502 	if (ldr)
503 		ret = avic_ldr_write(vcpu, id, ldr);
504 
505 	if (!ret)
506 		svm->ldr_reg = ldr;
507 
508 	return ret;
509 }
510 
511 static int avic_handle_apic_id_update(struct kvm_vcpu *vcpu)
512 {
513 	u64 *old, *new;
514 	struct vcpu_svm *svm = to_svm(vcpu);
515 	u32 id = kvm_xapic_id(vcpu->arch.apic);
516 
517 	if (vcpu->vcpu_id == id)
518 		return 0;
519 
520 	old = avic_get_physical_id_entry(vcpu, vcpu->vcpu_id);
521 	new = avic_get_physical_id_entry(vcpu, id);
522 	if (!new || !old)
523 		return 1;
524 
525 	/* We need to move physical_id_entry to new offset */
526 	*new = *old;
527 	*old = 0ULL;
528 	to_svm(vcpu)->avic_physical_id_cache = new;
529 
530 	/*
531 	 * Also update the guest physical APIC ID in the logical
532 	 * APIC ID table entry if already setup the LDR.
533 	 */
534 	if (svm->ldr_reg)
535 		avic_handle_ldr_update(vcpu);
536 
537 	return 0;
538 }
539 
540 static void avic_handle_dfr_update(struct kvm_vcpu *vcpu)
541 {
542 	struct vcpu_svm *svm = to_svm(vcpu);
543 	u32 dfr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR);
544 
545 	if (svm->dfr_reg == dfr)
546 		return;
547 
548 	avic_invalidate_logical_id_entry(vcpu);
549 	svm->dfr_reg = dfr;
550 }
551 
552 static int avic_unaccel_trap_write(struct kvm_vcpu *vcpu)
553 {
554 	u32 offset = to_svm(vcpu)->vmcb->control.exit_info_1 &
555 				AVIC_UNACCEL_ACCESS_OFFSET_MASK;
556 
557 	switch (offset) {
558 	case APIC_ID:
559 		if (avic_handle_apic_id_update(vcpu))
560 			return 0;
561 		break;
562 	case APIC_LDR:
563 		if (avic_handle_ldr_update(vcpu))
564 			return 0;
565 		break;
566 	case APIC_DFR:
567 		avic_handle_dfr_update(vcpu);
568 		break;
569 	default:
570 		break;
571 	}
572 
573 	kvm_apic_write_nodecode(vcpu, offset);
574 	return 1;
575 }
576 
577 static bool is_avic_unaccelerated_access_trap(u32 offset)
578 {
579 	bool ret = false;
580 
581 	switch (offset) {
582 	case APIC_ID:
583 	case APIC_EOI:
584 	case APIC_RRR:
585 	case APIC_LDR:
586 	case APIC_DFR:
587 	case APIC_SPIV:
588 	case APIC_ESR:
589 	case APIC_ICR:
590 	case APIC_LVTT:
591 	case APIC_LVTTHMR:
592 	case APIC_LVTPC:
593 	case APIC_LVT0:
594 	case APIC_LVT1:
595 	case APIC_LVTERR:
596 	case APIC_TMICT:
597 	case APIC_TDCR:
598 		ret = true;
599 		break;
600 	default:
601 		break;
602 	}
603 	return ret;
604 }
605 
606 int avic_unaccelerated_access_interception(struct kvm_vcpu *vcpu)
607 {
608 	struct vcpu_svm *svm = to_svm(vcpu);
609 	int ret = 0;
610 	u32 offset = svm->vmcb->control.exit_info_1 &
611 		     AVIC_UNACCEL_ACCESS_OFFSET_MASK;
612 	u32 vector = svm->vmcb->control.exit_info_2 &
613 		     AVIC_UNACCEL_ACCESS_VECTOR_MASK;
614 	bool write = (svm->vmcb->control.exit_info_1 >> 32) &
615 		     AVIC_UNACCEL_ACCESS_WRITE_MASK;
616 	bool trap = is_avic_unaccelerated_access_trap(offset);
617 
618 	trace_kvm_avic_unaccelerated_access(vcpu->vcpu_id, offset,
619 					    trap, write, vector);
620 	if (trap) {
621 		/* Handling Trap */
622 		WARN_ONCE(!write, "svm: Handling trap read.\n");
623 		ret = avic_unaccel_trap_write(vcpu);
624 	} else {
625 		/* Handling Fault */
626 		ret = kvm_emulate_instruction(vcpu, 0);
627 	}
628 
629 	return ret;
630 }
631 
632 int avic_init_vcpu(struct vcpu_svm *svm)
633 {
634 	int ret;
635 	struct kvm_vcpu *vcpu = &svm->vcpu;
636 
637 	if (!enable_apicv || !irqchip_in_kernel(vcpu->kvm))
638 		return 0;
639 
640 	ret = avic_init_backing_page(vcpu);
641 	if (ret)
642 		return ret;
643 
644 	INIT_LIST_HEAD(&svm->ir_list);
645 	spin_lock_init(&svm->ir_list_lock);
646 	svm->dfr_reg = APIC_DFR_FLAT;
647 
648 	return ret;
649 }
650 
651 void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu)
652 {
653 	if (avic_handle_apic_id_update(vcpu) != 0)
654 		return;
655 	avic_handle_dfr_update(vcpu);
656 	avic_handle_ldr_update(vcpu);
657 }
658 
659 static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
660 {
661 	int ret = 0;
662 	unsigned long flags;
663 	struct amd_svm_iommu_ir *ir;
664 	struct vcpu_svm *svm = to_svm(vcpu);
665 
666 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
667 		return 0;
668 
669 	/*
670 	 * Here, we go through the per-vcpu ir_list to update all existing
671 	 * interrupt remapping table entry targeting this vcpu.
672 	 */
673 	spin_lock_irqsave(&svm->ir_list_lock, flags);
674 
675 	if (list_empty(&svm->ir_list))
676 		goto out;
677 
678 	list_for_each_entry(ir, &svm->ir_list, node) {
679 		if (activate)
680 			ret = amd_iommu_activate_guest_mode(ir->data);
681 		else
682 			ret = amd_iommu_deactivate_guest_mode(ir->data);
683 		if (ret)
684 			break;
685 	}
686 out:
687 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
688 	return ret;
689 }
690 
691 static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
692 {
693 	unsigned long flags;
694 	struct amd_svm_iommu_ir *cur;
695 
696 	spin_lock_irqsave(&svm->ir_list_lock, flags);
697 	list_for_each_entry(cur, &svm->ir_list, node) {
698 		if (cur->data != pi->ir_data)
699 			continue;
700 		list_del(&cur->node);
701 		kfree(cur);
702 		break;
703 	}
704 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
705 }
706 
707 static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
708 {
709 	int ret = 0;
710 	unsigned long flags;
711 	struct amd_svm_iommu_ir *ir;
712 
713 	/**
714 	 * In some cases, the existing irte is updated and re-set,
715 	 * so we need to check here if it's already been * added
716 	 * to the ir_list.
717 	 */
718 	if (pi->ir_data && (pi->prev_ga_tag != 0)) {
719 		struct kvm *kvm = svm->vcpu.kvm;
720 		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
721 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
722 		struct vcpu_svm *prev_svm;
723 
724 		if (!prev_vcpu) {
725 			ret = -EINVAL;
726 			goto out;
727 		}
728 
729 		prev_svm = to_svm(prev_vcpu);
730 		svm_ir_list_del(prev_svm, pi);
731 	}
732 
733 	/**
734 	 * Allocating new amd_iommu_pi_data, which will get
735 	 * add to the per-vcpu ir_list.
736 	 */
737 	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_KERNEL_ACCOUNT);
738 	if (!ir) {
739 		ret = -ENOMEM;
740 		goto out;
741 	}
742 	ir->data = pi->ir_data;
743 
744 	spin_lock_irqsave(&svm->ir_list_lock, flags);
745 	list_add(&ir->node, &svm->ir_list);
746 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
747 out:
748 	return ret;
749 }
750 
751 /*
752  * Note:
753  * The HW cannot support posting multicast/broadcast
754  * interrupts to a vCPU. So, we still use legacy interrupt
755  * remapping for these kind of interrupts.
756  *
757  * For lowest-priority interrupts, we only support
758  * those with single CPU as the destination, e.g. user
759  * configures the interrupts via /proc/irq or uses
760  * irqbalance to make the interrupts single-CPU.
761  */
762 static int
763 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
764 		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
765 {
766 	struct kvm_lapic_irq irq;
767 	struct kvm_vcpu *vcpu = NULL;
768 
769 	kvm_set_msi_irq(kvm, e, &irq);
770 
771 	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
772 	    !kvm_irq_is_postable(&irq)) {
773 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
774 			 __func__, irq.vector);
775 		return -1;
776 	}
777 
778 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
779 		 irq.vector);
780 	*svm = to_svm(vcpu);
781 	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
782 	vcpu_info->vector = irq.vector;
783 
784 	return 0;
785 }
786 
787 /*
788  * avic_pi_update_irte - set IRTE for Posted-Interrupts
789  *
790  * @kvm: kvm
791  * @host_irq: host irq of the interrupt
792  * @guest_irq: gsi of the interrupt
793  * @set: set or unset PI
794  * returns 0 on success, < 0 on failure
795  */
796 int avic_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
797 			uint32_t guest_irq, bool set)
798 {
799 	struct kvm_kernel_irq_routing_entry *e;
800 	struct kvm_irq_routing_table *irq_rt;
801 	int idx, ret = 0;
802 
803 	if (!kvm_arch_has_assigned_device(kvm) ||
804 	    !irq_remapping_cap(IRQ_POSTING_CAP))
805 		return 0;
806 
807 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
808 		 __func__, host_irq, guest_irq, set);
809 
810 	idx = srcu_read_lock(&kvm->irq_srcu);
811 	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
812 
813 	if (guest_irq >= irq_rt->nr_rt_entries ||
814 		hlist_empty(&irq_rt->map[guest_irq])) {
815 		pr_warn_once("no route for guest_irq %u/%u (broken user space?)\n",
816 			     guest_irq, irq_rt->nr_rt_entries);
817 		goto out;
818 	}
819 
820 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
821 		struct vcpu_data vcpu_info;
822 		struct vcpu_svm *svm = NULL;
823 
824 		if (e->type != KVM_IRQ_ROUTING_MSI)
825 			continue;
826 
827 		/**
828 		 * Here, we setup with legacy mode in the following cases:
829 		 * 1. When cannot target interrupt to a specific vcpu.
830 		 * 2. Unsetting posted interrupt.
831 		 * 3. APIC virtualization is disabled for the vcpu.
832 		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
833 		 */
834 		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
835 		    kvm_vcpu_apicv_active(&svm->vcpu)) {
836 			struct amd_iommu_pi_data pi;
837 
838 			/* Try to enable guest_mode in IRTE */
839 			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
840 					    AVIC_HPA_MASK);
841 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
842 						     svm->vcpu.vcpu_id);
843 			pi.is_guest_mode = true;
844 			pi.vcpu_data = &vcpu_info;
845 			ret = irq_set_vcpu_affinity(host_irq, &pi);
846 
847 			/**
848 			 * Here, we successfully setting up vcpu affinity in
849 			 * IOMMU guest mode. Now, we need to store the posted
850 			 * interrupt information in a per-vcpu ir_list so that
851 			 * we can reference to them directly when we update vcpu
852 			 * scheduling information in IOMMU irte.
853 			 */
854 			if (!ret && pi.is_guest_mode)
855 				svm_ir_list_add(svm, &pi);
856 		} else {
857 			/* Use legacy mode in IRTE */
858 			struct amd_iommu_pi_data pi;
859 
860 			/**
861 			 * Here, pi is used to:
862 			 * - Tell IOMMU to use legacy mode for this interrupt.
863 			 * - Retrieve ga_tag of prior interrupt remapping data.
864 			 */
865 			pi.prev_ga_tag = 0;
866 			pi.is_guest_mode = false;
867 			ret = irq_set_vcpu_affinity(host_irq, &pi);
868 
869 			/**
870 			 * Check if the posted interrupt was previously
871 			 * setup with the guest_mode by checking if the ga_tag
872 			 * was cached. If so, we need to clean up the per-vcpu
873 			 * ir_list.
874 			 */
875 			if (!ret && pi.prev_ga_tag) {
876 				int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
877 				struct kvm_vcpu *vcpu;
878 
879 				vcpu = kvm_get_vcpu_by_id(kvm, id);
880 				if (vcpu)
881 					svm_ir_list_del(to_svm(vcpu), &pi);
882 			}
883 		}
884 
885 		if (!ret && svm) {
886 			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
887 						 e->gsi, vcpu_info.vector,
888 						 vcpu_info.pi_desc_addr, set);
889 		}
890 
891 		if (ret < 0) {
892 			pr_err("%s: failed to update PI IRTE\n", __func__);
893 			goto out;
894 		}
895 	}
896 
897 	ret = 0;
898 out:
899 	srcu_read_unlock(&kvm->irq_srcu, idx);
900 	return ret;
901 }
902 
903 bool avic_check_apicv_inhibit_reasons(enum kvm_apicv_inhibit reason)
904 {
905 	ulong supported = BIT(APICV_INHIBIT_REASON_DISABLE) |
906 			  BIT(APICV_INHIBIT_REASON_ABSENT) |
907 			  BIT(APICV_INHIBIT_REASON_HYPERV) |
908 			  BIT(APICV_INHIBIT_REASON_NESTED) |
909 			  BIT(APICV_INHIBIT_REASON_IRQWIN) |
910 			  BIT(APICV_INHIBIT_REASON_PIT_REINJ) |
911 			  BIT(APICV_INHIBIT_REASON_X2APIC) |
912 			  BIT(APICV_INHIBIT_REASON_BLOCKIRQ) |
913 			  BIT(APICV_INHIBIT_REASON_SEV);
914 
915 	return supported & BIT(reason);
916 }
917 
918 
919 static inline int
920 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
921 {
922 	int ret = 0;
923 	unsigned long flags;
924 	struct amd_svm_iommu_ir *ir;
925 	struct vcpu_svm *svm = to_svm(vcpu);
926 
927 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
928 		return 0;
929 
930 	/*
931 	 * Here, we go through the per-vcpu ir_list to update all existing
932 	 * interrupt remapping table entry targeting this vcpu.
933 	 */
934 	spin_lock_irqsave(&svm->ir_list_lock, flags);
935 
936 	if (list_empty(&svm->ir_list))
937 		goto out;
938 
939 	list_for_each_entry(ir, &svm->ir_list, node) {
940 		ret = amd_iommu_update_ga(cpu, r, ir->data);
941 		if (ret)
942 			break;
943 	}
944 out:
945 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
946 	return ret;
947 }
948 
949 void __avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
950 {
951 	u64 entry;
952 	int h_physical_id = kvm_cpu_get_apicid(cpu);
953 	struct vcpu_svm *svm = to_svm(vcpu);
954 
955 	lockdep_assert_preemption_disabled();
956 
957 	if (WARN_ON(h_physical_id & ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
958 		return;
959 
960 	/*
961 	 * No need to update anything if the vCPU is blocking, i.e. if the vCPU
962 	 * is being scheduled in after being preempted.  The CPU entries in the
963 	 * Physical APIC table and IRTE are consumed iff IsRun{ning} is '1'.
964 	 * If the vCPU was migrated, its new CPU value will be stuffed when the
965 	 * vCPU unblocks.
966 	 */
967 	if (kvm_vcpu_is_blocking(vcpu))
968 		return;
969 
970 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
971 	WARN_ON(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
972 
973 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
974 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
975 	entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
976 
977 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
978 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, true);
979 }
980 
981 void __avic_vcpu_put(struct kvm_vcpu *vcpu)
982 {
983 	u64 entry;
984 	struct vcpu_svm *svm = to_svm(vcpu);
985 
986 	lockdep_assert_preemption_disabled();
987 
988 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
989 
990 	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
991 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
992 		return;
993 
994 	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
995 
996 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
997 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
998 }
999 
1000 static void avic_vcpu_load(struct kvm_vcpu *vcpu)
1001 {
1002 	int cpu = get_cpu();
1003 
1004 	WARN_ON(cpu != vcpu->cpu);
1005 
1006 	__avic_vcpu_load(vcpu, cpu);
1007 
1008 	put_cpu();
1009 }
1010 
1011 static void avic_vcpu_put(struct kvm_vcpu *vcpu)
1012 {
1013 	preempt_disable();
1014 
1015 	__avic_vcpu_put(vcpu);
1016 
1017 	preempt_enable();
1018 }
1019 
1020 void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
1021 {
1022 	struct vcpu_svm *svm = to_svm(vcpu);
1023 	struct vmcb *vmcb = svm->vmcb01.ptr;
1024 	bool activated = kvm_vcpu_apicv_active(vcpu);
1025 
1026 	if (!enable_apicv)
1027 		return;
1028 
1029 	if (activated) {
1030 		/**
1031 		 * During AVIC temporary deactivation, guest could update
1032 		 * APIC ID, DFR and LDR registers, which would not be trapped
1033 		 * by avic_unaccelerated_access_interception(). In this case,
1034 		 * we need to check and update the AVIC logical APIC ID table
1035 		 * accordingly before re-activating.
1036 		 */
1037 		avic_apicv_post_state_restore(vcpu);
1038 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
1039 	} else {
1040 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
1041 	}
1042 	vmcb_mark_dirty(vmcb, VMCB_AVIC);
1043 
1044 	if (activated)
1045 		avic_vcpu_load(vcpu);
1046 	else
1047 		avic_vcpu_put(vcpu);
1048 
1049 	avic_set_pi_irte_mode(vcpu, activated);
1050 }
1051 
1052 void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
1053 {
1054 	if (!kvm_vcpu_apicv_active(vcpu))
1055 		return;
1056 
1057        /*
1058         * Unload the AVIC when the vCPU is about to block, _before_
1059         * the vCPU actually blocks.
1060         *
1061         * Any IRQs that arrive before IsRunning=0 will not cause an
1062         * incomplete IPI vmexit on the source, therefore vIRR will also
1063         * be checked by kvm_vcpu_check_block() before blocking.  The
1064         * memory barrier implicit in set_current_state orders writing
1065         * IsRunning=0 before reading the vIRR.  The processor needs a
1066         * matching memory barrier on interrupt delivery between writing
1067         * IRR and reading IsRunning; the lack of this barrier might be
1068         * the cause of errata #1235).
1069         */
1070 	avic_vcpu_put(vcpu);
1071 }
1072 
1073 void avic_vcpu_unblocking(struct kvm_vcpu *vcpu)
1074 {
1075 	if (!kvm_vcpu_apicv_active(vcpu))
1076 		return;
1077 
1078 	avic_vcpu_load(vcpu);
1079 }
1080