xref: /openbmc/linux/arch/x86/kvm/svm/avic.c (revision 0ad53fe3)
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 #define SVM_AVIC_DOORBELL	0xc001011b
31 
32 #define AVIC_HPA_MASK	~((0xFFFULL << 52) | 0xFFF)
33 
34 /*
35  * 0xff is broadcast, so the max index allowed for physical APIC ID
36  * table is 0xfe.  APIC IDs above 0xff are reserved.
37  */
38 #define AVIC_MAX_PHYSICAL_ID_COUNT	255
39 
40 #define AVIC_UNACCEL_ACCESS_WRITE_MASK		1
41 #define AVIC_UNACCEL_ACCESS_OFFSET_MASK		0xFF0
42 #define AVIC_UNACCEL_ACCESS_VECTOR_MASK		0xFFFFFFFF
43 
44 /* AVIC GATAG is encoded using VM and VCPU IDs */
45 #define AVIC_VCPU_ID_BITS		8
46 #define AVIC_VCPU_ID_MASK		((1 << AVIC_VCPU_ID_BITS) - 1)
47 
48 #define AVIC_VM_ID_BITS			24
49 #define AVIC_VM_ID_NR			(1 << AVIC_VM_ID_BITS)
50 #define AVIC_VM_ID_MASK			((1 << AVIC_VM_ID_BITS) - 1)
51 
52 #define AVIC_GATAG(x, y)		(((x & AVIC_VM_ID_MASK) << AVIC_VCPU_ID_BITS) | \
53 						(y & AVIC_VCPU_ID_MASK))
54 #define AVIC_GATAG_TO_VMID(x)		((x >> AVIC_VCPU_ID_BITS) & AVIC_VM_ID_MASK)
55 #define AVIC_GATAG_TO_VCPUID(x)		(x & AVIC_VCPU_ID_MASK)
56 
57 /* Note:
58  * This hash table is used to map VM_ID to a struct kvm_svm,
59  * when handling AMD IOMMU GALOG notification to schedule in
60  * a particular vCPU.
61  */
62 #define SVM_VM_DATA_HASH_BITS	8
63 static DEFINE_HASHTABLE(svm_vm_data_hash, SVM_VM_DATA_HASH_BITS);
64 static u32 next_vm_id = 0;
65 static bool next_vm_id_wrapped = 0;
66 static DEFINE_SPINLOCK(svm_vm_data_hash_lock);
67 
68 /*
69  * This is a wrapper of struct amd_iommu_ir_data.
70  */
71 struct amd_svm_iommu_ir {
72 	struct list_head node;	/* Used by SVM for per-vcpu ir_list */
73 	void *data;		/* Storing pointer to struct amd_ir_data */
74 };
75 
76 enum avic_ipi_failure_cause {
77 	AVIC_IPI_FAILURE_INVALID_INT_TYPE,
78 	AVIC_IPI_FAILURE_TARGET_NOT_RUNNING,
79 	AVIC_IPI_FAILURE_INVALID_TARGET,
80 	AVIC_IPI_FAILURE_INVALID_BACKING_PAGE,
81 };
82 
83 /* Note:
84  * This function is called from IOMMU driver to notify
85  * SVM to schedule in a particular vCPU of a particular VM.
86  */
87 int avic_ga_log_notifier(u32 ga_tag)
88 {
89 	unsigned long flags;
90 	struct kvm_svm *kvm_svm;
91 	struct kvm_vcpu *vcpu = NULL;
92 	u32 vm_id = AVIC_GATAG_TO_VMID(ga_tag);
93 	u32 vcpu_id = AVIC_GATAG_TO_VCPUID(ga_tag);
94 
95 	pr_debug("SVM: %s: vm_id=%#x, vcpu_id=%#x\n", __func__, vm_id, vcpu_id);
96 	trace_kvm_avic_ga_log(vm_id, vcpu_id);
97 
98 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
99 	hash_for_each_possible(svm_vm_data_hash, kvm_svm, hnode, vm_id) {
100 		if (kvm_svm->avic_vm_id != vm_id)
101 			continue;
102 		vcpu = kvm_get_vcpu_by_id(&kvm_svm->kvm, vcpu_id);
103 		break;
104 	}
105 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
106 
107 	/* Note:
108 	 * At this point, the IOMMU should have already set the pending
109 	 * bit in the vAPIC backing page. So, we just need to schedule
110 	 * in the vcpu.
111 	 */
112 	if (vcpu)
113 		kvm_vcpu_wake_up(vcpu);
114 
115 	return 0;
116 }
117 
118 void avic_vm_destroy(struct kvm *kvm)
119 {
120 	unsigned long flags;
121 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
122 
123 	if (!enable_apicv)
124 		return;
125 
126 	if (kvm_svm->avic_logical_id_table_page)
127 		__free_page(kvm_svm->avic_logical_id_table_page);
128 	if (kvm_svm->avic_physical_id_table_page)
129 		__free_page(kvm_svm->avic_physical_id_table_page);
130 
131 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
132 	hash_del(&kvm_svm->hnode);
133 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
134 }
135 
136 int avic_vm_init(struct kvm *kvm)
137 {
138 	unsigned long flags;
139 	int err = -ENOMEM;
140 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
141 	struct kvm_svm *k2;
142 	struct page *p_page;
143 	struct page *l_page;
144 	u32 vm_id;
145 
146 	if (!enable_apicv)
147 		return 0;
148 
149 	/* Allocating physical APIC ID table (4KB) */
150 	p_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
151 	if (!p_page)
152 		goto free_avic;
153 
154 	kvm_svm->avic_physical_id_table_page = p_page;
155 
156 	/* Allocating logical APIC ID table (4KB) */
157 	l_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
158 	if (!l_page)
159 		goto free_avic;
160 
161 	kvm_svm->avic_logical_id_table_page = l_page;
162 
163 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
164  again:
165 	vm_id = next_vm_id = (next_vm_id + 1) & AVIC_VM_ID_MASK;
166 	if (vm_id == 0) { /* id is 1-based, zero is not okay */
167 		next_vm_id_wrapped = 1;
168 		goto again;
169 	}
170 	/* Is it still in use? Only possible if wrapped at least once */
171 	if (next_vm_id_wrapped) {
172 		hash_for_each_possible(svm_vm_data_hash, k2, hnode, vm_id) {
173 			if (k2->avic_vm_id == vm_id)
174 				goto again;
175 		}
176 	}
177 	kvm_svm->avic_vm_id = vm_id;
178 	hash_add(svm_vm_data_hash, &kvm_svm->hnode, kvm_svm->avic_vm_id);
179 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
180 
181 	return 0;
182 
183 free_avic:
184 	avic_vm_destroy(kvm);
185 	return err;
186 }
187 
188 void avic_init_vmcb(struct vcpu_svm *svm)
189 {
190 	struct vmcb *vmcb = svm->vmcb;
191 	struct kvm_svm *kvm_svm = to_kvm_svm(svm->vcpu.kvm);
192 	phys_addr_t bpa = __sme_set(page_to_phys(svm->avic_backing_page));
193 	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
194 	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
195 
196 	vmcb->control.avic_backing_page = bpa & AVIC_HPA_MASK;
197 	vmcb->control.avic_logical_id = lpa & AVIC_HPA_MASK;
198 	vmcb->control.avic_physical_id = ppa & AVIC_HPA_MASK;
199 	vmcb->control.avic_physical_id |= AVIC_MAX_PHYSICAL_ID_COUNT;
200 	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE & VMCB_AVIC_APIC_BAR_MASK;
201 
202 	if (kvm_apicv_activated(svm->vcpu.kvm))
203 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
204 	else
205 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
206 }
207 
208 static u64 *avic_get_physical_id_entry(struct kvm_vcpu *vcpu,
209 				       unsigned int index)
210 {
211 	u64 *avic_physical_id_table;
212 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
213 
214 	if (index >= AVIC_MAX_PHYSICAL_ID_COUNT)
215 		return NULL;
216 
217 	avic_physical_id_table = page_address(kvm_svm->avic_physical_id_table_page);
218 
219 	return &avic_physical_id_table[index];
220 }
221 
222 /*
223  * Note:
224  * AVIC hardware walks the nested page table to check permissions,
225  * but does not use the SPA address specified in the leaf page
226  * table entry since it uses  address in the AVIC_BACKING_PAGE pointer
227  * field of the VMCB. Therefore, we set up the
228  * APIC_ACCESS_PAGE_PRIVATE_MEMSLOT (4KB) here.
229  */
230 static int avic_alloc_access_page(struct kvm *kvm)
231 {
232 	void __user *ret;
233 	int r = 0;
234 
235 	mutex_lock(&kvm->slots_lock);
236 
237 	if (kvm->arch.apic_access_memslot_enabled)
238 		goto out;
239 
240 	ret = __x86_set_memory_region(kvm,
241 				      APIC_ACCESS_PAGE_PRIVATE_MEMSLOT,
242 				      APIC_DEFAULT_PHYS_BASE,
243 				      PAGE_SIZE);
244 	if (IS_ERR(ret)) {
245 		r = PTR_ERR(ret);
246 		goto out;
247 	}
248 
249 	kvm->arch.apic_access_memslot_enabled = true;
250 out:
251 	mutex_unlock(&kvm->slots_lock);
252 	return r;
253 }
254 
255 static int avic_init_backing_page(struct kvm_vcpu *vcpu)
256 {
257 	u64 *entry, new_entry;
258 	int id = vcpu->vcpu_id;
259 	struct vcpu_svm *svm = to_svm(vcpu);
260 
261 	if (id >= AVIC_MAX_PHYSICAL_ID_COUNT)
262 		return -EINVAL;
263 
264 	if (!vcpu->arch.apic->regs)
265 		return -EINVAL;
266 
267 	if (kvm_apicv_activated(vcpu->kvm)) {
268 		int ret;
269 
270 		ret = avic_alloc_access_page(vcpu->kvm);
271 		if (ret)
272 			return ret;
273 	}
274 
275 	svm->avic_backing_page = virt_to_page(vcpu->arch.apic->regs);
276 
277 	/* Setting AVIC backing page address in the phy APIC ID table */
278 	entry = avic_get_physical_id_entry(vcpu, id);
279 	if (!entry)
280 		return -EINVAL;
281 
282 	new_entry = __sme_set((page_to_phys(svm->avic_backing_page) &
283 			      AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK) |
284 			      AVIC_PHYSICAL_ID_ENTRY_VALID_MASK);
285 	WRITE_ONCE(*entry, new_entry);
286 
287 	svm->avic_physical_id_cache = entry;
288 
289 	return 0;
290 }
291 
292 static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
293 				   u32 icrl, u32 icrh)
294 {
295 	struct kvm_vcpu *vcpu;
296 	int i;
297 
298 	kvm_for_each_vcpu(i, vcpu, kvm) {
299 		bool m = kvm_apic_match_dest(vcpu, source,
300 					     icrl & APIC_SHORT_MASK,
301 					     GET_APIC_DEST_FIELD(icrh),
302 					     icrl & APIC_DEST_MASK);
303 
304 		if (m && !avic_vcpu_is_running(vcpu))
305 			kvm_vcpu_wake_up(vcpu);
306 	}
307 }
308 
309 int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
310 {
311 	struct vcpu_svm *svm = to_svm(vcpu);
312 	u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
313 	u32 icrl = svm->vmcb->control.exit_info_1;
314 	u32 id = svm->vmcb->control.exit_info_2 >> 32;
315 	u32 index = svm->vmcb->control.exit_info_2 & 0xFF;
316 	struct kvm_lapic *apic = vcpu->arch.apic;
317 
318 	trace_kvm_avic_incomplete_ipi(vcpu->vcpu_id, icrh, icrl, id, index);
319 
320 	switch (id) {
321 	case AVIC_IPI_FAILURE_INVALID_INT_TYPE:
322 		/*
323 		 * AVIC hardware handles the generation of
324 		 * IPIs when the specified Message Type is Fixed
325 		 * (also known as fixed delivery mode) and
326 		 * the Trigger Mode is edge-triggered. The hardware
327 		 * also supports self and broadcast delivery modes
328 		 * specified via the Destination Shorthand(DSH)
329 		 * field of the ICRL. Logical and physical APIC ID
330 		 * formats are supported. All other IPI types cause
331 		 * a #VMEXIT, which needs to emulated.
332 		 */
333 		kvm_lapic_reg_write(apic, APIC_ICR2, icrh);
334 		kvm_lapic_reg_write(apic, APIC_ICR, icrl);
335 		break;
336 	case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING:
337 		/*
338 		 * At this point, we expect that the AVIC HW has already
339 		 * set the appropriate IRR bits on the valid target
340 		 * vcpus. So, we just need to kick the appropriate vcpu.
341 		 */
342 		avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh);
343 		break;
344 	case AVIC_IPI_FAILURE_INVALID_TARGET:
345 		WARN_ONCE(1, "Invalid IPI target: index=%u, vcpu=%d, icr=%#0x:%#0x\n",
346 			  index, vcpu->vcpu_id, icrh, icrl);
347 		break;
348 	case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
349 		WARN_ONCE(1, "Invalid backing page\n");
350 		break;
351 	default:
352 		pr_err("Unknown IPI interception\n");
353 	}
354 
355 	return 1;
356 }
357 
358 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
359 {
360 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
361 	int index;
362 	u32 *logical_apic_id_table;
363 	int dlid = GET_APIC_LOGICAL_ID(ldr);
364 
365 	if (!dlid)
366 		return NULL;
367 
368 	if (flat) { /* flat */
369 		index = ffs(dlid) - 1;
370 		if (index > 7)
371 			return NULL;
372 	} else { /* cluster */
373 		int cluster = (dlid & 0xf0) >> 4;
374 		int apic = ffs(dlid & 0x0f) - 1;
375 
376 		if ((apic < 0) || (apic > 7) ||
377 		    (cluster >= 0xf))
378 			return NULL;
379 		index = (cluster << 2) + apic;
380 	}
381 
382 	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
383 
384 	return &logical_apic_id_table[index];
385 }
386 
387 static int avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
388 {
389 	bool flat;
390 	u32 *entry, new_entry;
391 
392 	flat = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR) == APIC_DFR_FLAT;
393 	entry = avic_get_logical_id_entry(vcpu, ldr, flat);
394 	if (!entry)
395 		return -EINVAL;
396 
397 	new_entry = READ_ONCE(*entry);
398 	new_entry &= ~AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK;
399 	new_entry |= (g_physical_id & AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK);
400 	new_entry |= AVIC_LOGICAL_ID_ENTRY_VALID_MASK;
401 	WRITE_ONCE(*entry, new_entry);
402 
403 	return 0;
404 }
405 
406 static void avic_invalidate_logical_id_entry(struct kvm_vcpu *vcpu)
407 {
408 	struct vcpu_svm *svm = to_svm(vcpu);
409 	bool flat = svm->dfr_reg == APIC_DFR_FLAT;
410 	u32 *entry = avic_get_logical_id_entry(vcpu, svm->ldr_reg, flat);
411 
412 	if (entry)
413 		clear_bit(AVIC_LOGICAL_ID_ENTRY_VALID_BIT, (unsigned long *)entry);
414 }
415 
416 static int avic_handle_ldr_update(struct kvm_vcpu *vcpu)
417 {
418 	int ret = 0;
419 	struct vcpu_svm *svm = to_svm(vcpu);
420 	u32 ldr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_LDR);
421 	u32 id = kvm_xapic_id(vcpu->arch.apic);
422 
423 	if (ldr == svm->ldr_reg)
424 		return 0;
425 
426 	avic_invalidate_logical_id_entry(vcpu);
427 
428 	if (ldr)
429 		ret = avic_ldr_write(vcpu, id, ldr);
430 
431 	if (!ret)
432 		svm->ldr_reg = ldr;
433 
434 	return ret;
435 }
436 
437 static int avic_handle_apic_id_update(struct kvm_vcpu *vcpu)
438 {
439 	u64 *old, *new;
440 	struct vcpu_svm *svm = to_svm(vcpu);
441 	u32 id = kvm_xapic_id(vcpu->arch.apic);
442 
443 	if (vcpu->vcpu_id == id)
444 		return 0;
445 
446 	old = avic_get_physical_id_entry(vcpu, vcpu->vcpu_id);
447 	new = avic_get_physical_id_entry(vcpu, id);
448 	if (!new || !old)
449 		return 1;
450 
451 	/* We need to move physical_id_entry to new offset */
452 	*new = *old;
453 	*old = 0ULL;
454 	to_svm(vcpu)->avic_physical_id_cache = new;
455 
456 	/*
457 	 * Also update the guest physical APIC ID in the logical
458 	 * APIC ID table entry if already setup the LDR.
459 	 */
460 	if (svm->ldr_reg)
461 		avic_handle_ldr_update(vcpu);
462 
463 	return 0;
464 }
465 
466 static void avic_handle_dfr_update(struct kvm_vcpu *vcpu)
467 {
468 	struct vcpu_svm *svm = to_svm(vcpu);
469 	u32 dfr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR);
470 
471 	if (svm->dfr_reg == dfr)
472 		return;
473 
474 	avic_invalidate_logical_id_entry(vcpu);
475 	svm->dfr_reg = dfr;
476 }
477 
478 static int avic_unaccel_trap_write(struct vcpu_svm *svm)
479 {
480 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
481 	u32 offset = svm->vmcb->control.exit_info_1 &
482 				AVIC_UNACCEL_ACCESS_OFFSET_MASK;
483 
484 	switch (offset) {
485 	case APIC_ID:
486 		if (avic_handle_apic_id_update(&svm->vcpu))
487 			return 0;
488 		break;
489 	case APIC_LDR:
490 		if (avic_handle_ldr_update(&svm->vcpu))
491 			return 0;
492 		break;
493 	case APIC_DFR:
494 		avic_handle_dfr_update(&svm->vcpu);
495 		break;
496 	default:
497 		break;
498 	}
499 
500 	kvm_lapic_reg_write(apic, offset, kvm_lapic_get_reg(apic, offset));
501 
502 	return 1;
503 }
504 
505 static bool is_avic_unaccelerated_access_trap(u32 offset)
506 {
507 	bool ret = false;
508 
509 	switch (offset) {
510 	case APIC_ID:
511 	case APIC_EOI:
512 	case APIC_RRR:
513 	case APIC_LDR:
514 	case APIC_DFR:
515 	case APIC_SPIV:
516 	case APIC_ESR:
517 	case APIC_ICR:
518 	case APIC_LVTT:
519 	case APIC_LVTTHMR:
520 	case APIC_LVTPC:
521 	case APIC_LVT0:
522 	case APIC_LVT1:
523 	case APIC_LVTERR:
524 	case APIC_TMICT:
525 	case APIC_TDCR:
526 		ret = true;
527 		break;
528 	default:
529 		break;
530 	}
531 	return ret;
532 }
533 
534 int avic_unaccelerated_access_interception(struct kvm_vcpu *vcpu)
535 {
536 	struct vcpu_svm *svm = to_svm(vcpu);
537 	int ret = 0;
538 	u32 offset = svm->vmcb->control.exit_info_1 &
539 		     AVIC_UNACCEL_ACCESS_OFFSET_MASK;
540 	u32 vector = svm->vmcb->control.exit_info_2 &
541 		     AVIC_UNACCEL_ACCESS_VECTOR_MASK;
542 	bool write = (svm->vmcb->control.exit_info_1 >> 32) &
543 		     AVIC_UNACCEL_ACCESS_WRITE_MASK;
544 	bool trap = is_avic_unaccelerated_access_trap(offset);
545 
546 	trace_kvm_avic_unaccelerated_access(vcpu->vcpu_id, offset,
547 					    trap, write, vector);
548 	if (trap) {
549 		/* Handling Trap */
550 		WARN_ONCE(!write, "svm: Handling trap read.\n");
551 		ret = avic_unaccel_trap_write(svm);
552 	} else {
553 		/* Handling Fault */
554 		ret = kvm_emulate_instruction(vcpu, 0);
555 	}
556 
557 	return ret;
558 }
559 
560 int avic_init_vcpu(struct vcpu_svm *svm)
561 {
562 	int ret;
563 	struct kvm_vcpu *vcpu = &svm->vcpu;
564 
565 	if (!enable_apicv || !irqchip_in_kernel(vcpu->kvm))
566 		return 0;
567 
568 	ret = avic_init_backing_page(vcpu);
569 	if (ret)
570 		return ret;
571 
572 	INIT_LIST_HEAD(&svm->ir_list);
573 	spin_lock_init(&svm->ir_list_lock);
574 	svm->dfr_reg = APIC_DFR_FLAT;
575 
576 	return ret;
577 }
578 
579 void avic_post_state_restore(struct kvm_vcpu *vcpu)
580 {
581 	if (avic_handle_apic_id_update(vcpu) != 0)
582 		return;
583 	avic_handle_dfr_update(vcpu);
584 	avic_handle_ldr_update(vcpu);
585 }
586 
587 void svm_set_virtual_apic_mode(struct kvm_vcpu *vcpu)
588 {
589 	return;
590 }
591 
592 void svm_hwapic_irr_update(struct kvm_vcpu *vcpu, int max_irr)
593 {
594 }
595 
596 void svm_hwapic_isr_update(struct kvm_vcpu *vcpu, int max_isr)
597 {
598 }
599 
600 static int svm_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
601 {
602 	int ret = 0;
603 	unsigned long flags;
604 	struct amd_svm_iommu_ir *ir;
605 	struct vcpu_svm *svm = to_svm(vcpu);
606 
607 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
608 		return 0;
609 
610 	/*
611 	 * Here, we go through the per-vcpu ir_list to update all existing
612 	 * interrupt remapping table entry targeting this vcpu.
613 	 */
614 	spin_lock_irqsave(&svm->ir_list_lock, flags);
615 
616 	if (list_empty(&svm->ir_list))
617 		goto out;
618 
619 	list_for_each_entry(ir, &svm->ir_list, node) {
620 		if (activate)
621 			ret = amd_iommu_activate_guest_mode(ir->data);
622 		else
623 			ret = amd_iommu_deactivate_guest_mode(ir->data);
624 		if (ret)
625 			break;
626 	}
627 out:
628 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
629 	return ret;
630 }
631 
632 void svm_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
633 {
634 	struct vcpu_svm *svm = to_svm(vcpu);
635 	struct vmcb *vmcb = svm->vmcb01.ptr;
636 	bool activated = kvm_vcpu_apicv_active(vcpu);
637 
638 	if (!enable_apicv)
639 		return;
640 
641 	if (activated) {
642 		/**
643 		 * During AVIC temporary deactivation, guest could update
644 		 * APIC ID, DFR and LDR registers, which would not be trapped
645 		 * by avic_unaccelerated_access_interception(). In this case,
646 		 * we need to check and update the AVIC logical APIC ID table
647 		 * accordingly before re-activating.
648 		 */
649 		avic_post_state_restore(vcpu);
650 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
651 	} else {
652 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
653 	}
654 	vmcb_mark_dirty(vmcb, VMCB_AVIC);
655 
656 	if (activated)
657 		avic_vcpu_load(vcpu, vcpu->cpu);
658 	else
659 		avic_vcpu_put(vcpu);
660 
661 	svm_set_pi_irte_mode(vcpu, activated);
662 }
663 
664 void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)
665 {
666 	return;
667 }
668 
669 int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
670 {
671 	if (!vcpu->arch.apicv_active)
672 		return -1;
673 
674 	kvm_lapic_set_irr(vec, vcpu->arch.apic);
675 	smp_mb__after_atomic();
676 
677 	if (avic_vcpu_is_running(vcpu)) {
678 		int cpuid = vcpu->cpu;
679 
680 		if (cpuid != get_cpu())
681 			wrmsrl(SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpuid));
682 		put_cpu();
683 	} else
684 		kvm_vcpu_wake_up(vcpu);
685 
686 	return 0;
687 }
688 
689 bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
690 {
691 	return false;
692 }
693 
694 static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
695 {
696 	unsigned long flags;
697 	struct amd_svm_iommu_ir *cur;
698 
699 	spin_lock_irqsave(&svm->ir_list_lock, flags);
700 	list_for_each_entry(cur, &svm->ir_list, node) {
701 		if (cur->data != pi->ir_data)
702 			continue;
703 		list_del(&cur->node);
704 		kfree(cur);
705 		break;
706 	}
707 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
708 }
709 
710 static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
711 {
712 	int ret = 0;
713 	unsigned long flags;
714 	struct amd_svm_iommu_ir *ir;
715 
716 	/**
717 	 * In some cases, the existing irte is updated and re-set,
718 	 * so we need to check here if it's already been * added
719 	 * to the ir_list.
720 	 */
721 	if (pi->ir_data && (pi->prev_ga_tag != 0)) {
722 		struct kvm *kvm = svm->vcpu.kvm;
723 		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
724 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
725 		struct vcpu_svm *prev_svm;
726 
727 		if (!prev_vcpu) {
728 			ret = -EINVAL;
729 			goto out;
730 		}
731 
732 		prev_svm = to_svm(prev_vcpu);
733 		svm_ir_list_del(prev_svm, pi);
734 	}
735 
736 	/**
737 	 * Allocating new amd_iommu_pi_data, which will get
738 	 * add to the per-vcpu ir_list.
739 	 */
740 	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_KERNEL_ACCOUNT);
741 	if (!ir) {
742 		ret = -ENOMEM;
743 		goto out;
744 	}
745 	ir->data = pi->ir_data;
746 
747 	spin_lock_irqsave(&svm->ir_list_lock, flags);
748 	list_add(&ir->node, &svm->ir_list);
749 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
750 out:
751 	return ret;
752 }
753 
754 /*
755  * Note:
756  * The HW cannot support posting multicast/broadcast
757  * interrupts to a vCPU. So, we still use legacy interrupt
758  * remapping for these kind of interrupts.
759  *
760  * For lowest-priority interrupts, we only support
761  * those with single CPU as the destination, e.g. user
762  * configures the interrupts via /proc/irq or uses
763  * irqbalance to make the interrupts single-CPU.
764  */
765 static int
766 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
767 		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
768 {
769 	struct kvm_lapic_irq irq;
770 	struct kvm_vcpu *vcpu = NULL;
771 
772 	kvm_set_msi_irq(kvm, e, &irq);
773 
774 	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
775 	    !kvm_irq_is_postable(&irq)) {
776 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
777 			 __func__, irq.vector);
778 		return -1;
779 	}
780 
781 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
782 		 irq.vector);
783 	*svm = to_svm(vcpu);
784 	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
785 	vcpu_info->vector = irq.vector;
786 
787 	return 0;
788 }
789 
790 /*
791  * svm_update_pi_irte - set IRTE for Posted-Interrupts
792  *
793  * @kvm: kvm
794  * @host_irq: host irq of the interrupt
795  * @guest_irq: gsi of the interrupt
796  * @set: set or unset PI
797  * returns 0 on success, < 0 on failure
798  */
799 int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
800 		       uint32_t guest_irq, bool set)
801 {
802 	struct kvm_kernel_irq_routing_entry *e;
803 	struct kvm_irq_routing_table *irq_rt;
804 	int idx, ret = -EINVAL;
805 
806 	if (!kvm_arch_has_assigned_device(kvm) ||
807 	    !irq_remapping_cap(IRQ_POSTING_CAP))
808 		return 0;
809 
810 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
811 		 __func__, host_irq, guest_irq, set);
812 
813 	idx = srcu_read_lock(&kvm->irq_srcu);
814 	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
815 	WARN_ON(guest_irq >= irq_rt->nr_rt_entries);
816 
817 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
818 		struct vcpu_data vcpu_info;
819 		struct vcpu_svm *svm = NULL;
820 
821 		if (e->type != KVM_IRQ_ROUTING_MSI)
822 			continue;
823 
824 		/**
825 		 * Here, we setup with legacy mode in the following cases:
826 		 * 1. When cannot target interrupt to a specific vcpu.
827 		 * 2. Unsetting posted interrupt.
828 		 * 3. APIC virtualization is disabled for the vcpu.
829 		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
830 		 */
831 		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
832 		    kvm_vcpu_apicv_active(&svm->vcpu)) {
833 			struct amd_iommu_pi_data pi;
834 
835 			/* Try to enable guest_mode in IRTE */
836 			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
837 					    AVIC_HPA_MASK);
838 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
839 						     svm->vcpu.vcpu_id);
840 			pi.is_guest_mode = true;
841 			pi.vcpu_data = &vcpu_info;
842 			ret = irq_set_vcpu_affinity(host_irq, &pi);
843 
844 			/**
845 			 * Here, we successfully setting up vcpu affinity in
846 			 * IOMMU guest mode. Now, we need to store the posted
847 			 * interrupt information in a per-vcpu ir_list so that
848 			 * we can reference to them directly when we update vcpu
849 			 * scheduling information in IOMMU irte.
850 			 */
851 			if (!ret && pi.is_guest_mode)
852 				svm_ir_list_add(svm, &pi);
853 		} else {
854 			/* Use legacy mode in IRTE */
855 			struct amd_iommu_pi_data pi;
856 
857 			/**
858 			 * Here, pi is used to:
859 			 * - Tell IOMMU to use legacy mode for this interrupt.
860 			 * - Retrieve ga_tag of prior interrupt remapping data.
861 			 */
862 			pi.prev_ga_tag = 0;
863 			pi.is_guest_mode = false;
864 			ret = irq_set_vcpu_affinity(host_irq, &pi);
865 
866 			/**
867 			 * Check if the posted interrupt was previously
868 			 * setup with the guest_mode by checking if the ga_tag
869 			 * was cached. If so, we need to clean up the per-vcpu
870 			 * ir_list.
871 			 */
872 			if (!ret && pi.prev_ga_tag) {
873 				int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
874 				struct kvm_vcpu *vcpu;
875 
876 				vcpu = kvm_get_vcpu_by_id(kvm, id);
877 				if (vcpu)
878 					svm_ir_list_del(to_svm(vcpu), &pi);
879 			}
880 		}
881 
882 		if (!ret && svm) {
883 			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
884 						 e->gsi, vcpu_info.vector,
885 						 vcpu_info.pi_desc_addr, set);
886 		}
887 
888 		if (ret < 0) {
889 			pr_err("%s: failed to update PI IRTE\n", __func__);
890 			goto out;
891 		}
892 	}
893 
894 	ret = 0;
895 out:
896 	srcu_read_unlock(&kvm->irq_srcu, idx);
897 	return ret;
898 }
899 
900 bool svm_check_apicv_inhibit_reasons(ulong bit)
901 {
902 	ulong supported = BIT(APICV_INHIBIT_REASON_DISABLE) |
903 			  BIT(APICV_INHIBIT_REASON_HYPERV) |
904 			  BIT(APICV_INHIBIT_REASON_NESTED) |
905 			  BIT(APICV_INHIBIT_REASON_IRQWIN) |
906 			  BIT(APICV_INHIBIT_REASON_PIT_REINJ) |
907 			  BIT(APICV_INHIBIT_REASON_X2APIC);
908 
909 	return supported & BIT(bit);
910 }
911 
912 
913 static inline int
914 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
915 {
916 	int ret = 0;
917 	unsigned long flags;
918 	struct amd_svm_iommu_ir *ir;
919 	struct vcpu_svm *svm = to_svm(vcpu);
920 
921 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
922 		return 0;
923 
924 	/*
925 	 * Here, we go through the per-vcpu ir_list to update all existing
926 	 * interrupt remapping table entry targeting this vcpu.
927 	 */
928 	spin_lock_irqsave(&svm->ir_list_lock, flags);
929 
930 	if (list_empty(&svm->ir_list))
931 		goto out;
932 
933 	list_for_each_entry(ir, &svm->ir_list, node) {
934 		ret = amd_iommu_update_ga(cpu, r, ir->data);
935 		if (ret)
936 			break;
937 	}
938 out:
939 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
940 	return ret;
941 }
942 
943 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
944 {
945 	u64 entry;
946 	/* ID = 0xff (broadcast), ID > 0xff (reserved) */
947 	int h_physical_id = kvm_cpu_get_apicid(cpu);
948 	struct vcpu_svm *svm = to_svm(vcpu);
949 
950 	/*
951 	 * Since the host physical APIC id is 8 bits,
952 	 * we can support host APIC ID upto 255.
953 	 */
954 	if (WARN_ON(h_physical_id > AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
955 		return;
956 
957 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
958 	WARN_ON(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
959 
960 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
961 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
962 
963 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
964 	if (svm->avic_is_running)
965 		entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
966 
967 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
968 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id,
969 					svm->avic_is_running);
970 }
971 
972 void avic_vcpu_put(struct kvm_vcpu *vcpu)
973 {
974 	u64 entry;
975 	struct vcpu_svm *svm = to_svm(vcpu);
976 
977 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
978 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
979 		avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
980 
981 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
982 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
983 }
984 
985 /*
986  * This function is called during VCPU halt/unhalt.
987  */
988 static void avic_set_running(struct kvm_vcpu *vcpu, bool is_run)
989 {
990 	struct vcpu_svm *svm = to_svm(vcpu);
991 
992 	svm->avic_is_running = is_run;
993 
994 	if (!kvm_vcpu_apicv_active(vcpu))
995 		return;
996 
997 	if (is_run)
998 		avic_vcpu_load(vcpu, vcpu->cpu);
999 	else
1000 		avic_vcpu_put(vcpu);
1001 }
1002 
1003 void svm_vcpu_blocking(struct kvm_vcpu *vcpu)
1004 {
1005 	avic_set_running(vcpu, false);
1006 }
1007 
1008 void svm_vcpu_unblocking(struct kvm_vcpu *vcpu)
1009 {
1010 	if (kvm_check_request(KVM_REQ_APICV_UPDATE, vcpu))
1011 		kvm_vcpu_update_apicv(vcpu);
1012 	avic_set_running(vcpu, true);
1013 }
1014