xref: /openbmc/linux/arch/x86/kvm/svm/avic.c (revision 877013bc)
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 /* enable / disable AVIC */
31 int avic;
32 #ifdef CONFIG_X86_LOCAL_APIC
33 module_param(avic, int, S_IRUGO);
34 #endif
35 
36 #define SVM_AVIC_DOORBELL	0xc001011b
37 
38 #define AVIC_HPA_MASK	~((0xFFFULL << 52) | 0xFFF)
39 
40 /*
41  * 0xff is broadcast, so the max index allowed for physical APIC ID
42  * table is 0xfe.  APIC IDs above 0xff are reserved.
43  */
44 #define AVIC_MAX_PHYSICAL_ID_COUNT	255
45 
46 #define AVIC_UNACCEL_ACCESS_WRITE_MASK		1
47 #define AVIC_UNACCEL_ACCESS_OFFSET_MASK		0xFF0
48 #define AVIC_UNACCEL_ACCESS_VECTOR_MASK		0xFFFFFFFF
49 
50 /* AVIC GATAG is encoded using VM and VCPU IDs */
51 #define AVIC_VCPU_ID_BITS		8
52 #define AVIC_VCPU_ID_MASK		((1 << AVIC_VCPU_ID_BITS) - 1)
53 
54 #define AVIC_VM_ID_BITS			24
55 #define AVIC_VM_ID_NR			(1 << AVIC_VM_ID_BITS)
56 #define AVIC_VM_ID_MASK			((1 << AVIC_VM_ID_BITS) - 1)
57 
58 #define AVIC_GATAG(x, y)		(((x & AVIC_VM_ID_MASK) << AVIC_VCPU_ID_BITS) | \
59 						(y & AVIC_VCPU_ID_MASK))
60 #define AVIC_GATAG_TO_VMID(x)		((x >> AVIC_VCPU_ID_BITS) & AVIC_VM_ID_MASK)
61 #define AVIC_GATAG_TO_VCPUID(x)		(x & AVIC_VCPU_ID_MASK)
62 
63 /* Note:
64  * This hash table is used to map VM_ID to a struct kvm_svm,
65  * when handling AMD IOMMU GALOG notification to schedule in
66  * a particular vCPU.
67  */
68 #define SVM_VM_DATA_HASH_BITS	8
69 static DEFINE_HASHTABLE(svm_vm_data_hash, SVM_VM_DATA_HASH_BITS);
70 static u32 next_vm_id = 0;
71 static bool next_vm_id_wrapped = 0;
72 static DEFINE_SPINLOCK(svm_vm_data_hash_lock);
73 
74 /*
75  * This is a wrapper of struct amd_iommu_ir_data.
76  */
77 struct amd_svm_iommu_ir {
78 	struct list_head node;	/* Used by SVM for per-vcpu ir_list */
79 	void *data;		/* Storing pointer to struct amd_ir_data */
80 };
81 
82 enum avic_ipi_failure_cause {
83 	AVIC_IPI_FAILURE_INVALID_INT_TYPE,
84 	AVIC_IPI_FAILURE_TARGET_NOT_RUNNING,
85 	AVIC_IPI_FAILURE_INVALID_TARGET,
86 	AVIC_IPI_FAILURE_INVALID_BACKING_PAGE,
87 };
88 
89 /* Note:
90  * This function is called from IOMMU driver to notify
91  * SVM to schedule in a particular vCPU of a particular VM.
92  */
93 int avic_ga_log_notifier(u32 ga_tag)
94 {
95 	unsigned long flags;
96 	struct kvm_svm *kvm_svm;
97 	struct kvm_vcpu *vcpu = NULL;
98 	u32 vm_id = AVIC_GATAG_TO_VMID(ga_tag);
99 	u32 vcpu_id = AVIC_GATAG_TO_VCPUID(ga_tag);
100 
101 	pr_debug("SVM: %s: vm_id=%#x, vcpu_id=%#x\n", __func__, vm_id, vcpu_id);
102 	trace_kvm_avic_ga_log(vm_id, vcpu_id);
103 
104 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
105 	hash_for_each_possible(svm_vm_data_hash, kvm_svm, hnode, vm_id) {
106 		if (kvm_svm->avic_vm_id != vm_id)
107 			continue;
108 		vcpu = kvm_get_vcpu_by_id(&kvm_svm->kvm, vcpu_id);
109 		break;
110 	}
111 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
112 
113 	/* Note:
114 	 * At this point, the IOMMU should have already set the pending
115 	 * bit in the vAPIC backing page. So, we just need to schedule
116 	 * in the vcpu.
117 	 */
118 	if (vcpu)
119 		kvm_vcpu_wake_up(vcpu);
120 
121 	return 0;
122 }
123 
124 void avic_vm_destroy(struct kvm *kvm)
125 {
126 	unsigned long flags;
127 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
128 
129 	if (!avic)
130 		return;
131 
132 	if (kvm_svm->avic_logical_id_table_page)
133 		__free_page(kvm_svm->avic_logical_id_table_page);
134 	if (kvm_svm->avic_physical_id_table_page)
135 		__free_page(kvm_svm->avic_physical_id_table_page);
136 
137 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
138 	hash_del(&kvm_svm->hnode);
139 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
140 }
141 
142 int avic_vm_init(struct kvm *kvm)
143 {
144 	unsigned long flags;
145 	int err = -ENOMEM;
146 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
147 	struct kvm_svm *k2;
148 	struct page *p_page;
149 	struct page *l_page;
150 	u32 vm_id;
151 
152 	if (!avic)
153 		return 0;
154 
155 	/* Allocating physical APIC ID table (4KB) */
156 	p_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
157 	if (!p_page)
158 		goto free_avic;
159 
160 	kvm_svm->avic_physical_id_table_page = p_page;
161 
162 	/* Allocating logical APIC ID table (4KB) */
163 	l_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
164 	if (!l_page)
165 		goto free_avic;
166 
167 	kvm_svm->avic_logical_id_table_page = l_page;
168 
169 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
170  again:
171 	vm_id = next_vm_id = (next_vm_id + 1) & AVIC_VM_ID_MASK;
172 	if (vm_id == 0) { /* id is 1-based, zero is not okay */
173 		next_vm_id_wrapped = 1;
174 		goto again;
175 	}
176 	/* Is it still in use? Only possible if wrapped at least once */
177 	if (next_vm_id_wrapped) {
178 		hash_for_each_possible(svm_vm_data_hash, k2, hnode, vm_id) {
179 			if (k2->avic_vm_id == vm_id)
180 				goto again;
181 		}
182 	}
183 	kvm_svm->avic_vm_id = vm_id;
184 	hash_add(svm_vm_data_hash, &kvm_svm->hnode, kvm_svm->avic_vm_id);
185 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
186 
187 	return 0;
188 
189 free_avic:
190 	avic_vm_destroy(kvm);
191 	return err;
192 }
193 
194 void avic_init_vmcb(struct vcpu_svm *svm)
195 {
196 	struct vmcb *vmcb = svm->vmcb;
197 	struct kvm_svm *kvm_svm = to_kvm_svm(svm->vcpu.kvm);
198 	phys_addr_t bpa = __sme_set(page_to_phys(svm->avic_backing_page));
199 	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
200 	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
201 
202 	vmcb->control.avic_backing_page = bpa & AVIC_HPA_MASK;
203 	vmcb->control.avic_logical_id = lpa & AVIC_HPA_MASK;
204 	vmcb->control.avic_physical_id = ppa & AVIC_HPA_MASK;
205 	vmcb->control.avic_physical_id |= AVIC_MAX_PHYSICAL_ID_COUNT;
206 	if (kvm_apicv_activated(svm->vcpu.kvm))
207 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
208 	else
209 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
210 }
211 
212 static u64 *avic_get_physical_id_entry(struct kvm_vcpu *vcpu,
213 				       unsigned int index)
214 {
215 	u64 *avic_physical_id_table;
216 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
217 
218 	if (index >= AVIC_MAX_PHYSICAL_ID_COUNT)
219 		return NULL;
220 
221 	avic_physical_id_table = page_address(kvm_svm->avic_physical_id_table_page);
222 
223 	return &avic_physical_id_table[index];
224 }
225 
226 /**
227  * Note:
228  * AVIC hardware walks the nested page table to check permissions,
229  * but does not use the SPA address specified in the leaf page
230  * table entry since it uses  address in the AVIC_BACKING_PAGE pointer
231  * field of the VMCB. Therefore, we set up the
232  * APIC_ACCESS_PAGE_PRIVATE_MEMSLOT (4KB) here.
233  */
234 static int avic_update_access_page(struct kvm *kvm, bool activate)
235 {
236 	void __user *ret;
237 	int r = 0;
238 
239 	mutex_lock(&kvm->slots_lock);
240 	/*
241 	 * During kvm_destroy_vm(), kvm_pit_set_reinject() could trigger
242 	 * APICv mode change, which update APIC_ACCESS_PAGE_PRIVATE_MEMSLOT
243 	 * memory region. So, we need to ensure that kvm->mm == current->mm.
244 	 */
245 	if ((kvm->arch.apic_access_page_done == activate) ||
246 	    (kvm->mm != current->mm))
247 		goto out;
248 
249 	ret = __x86_set_memory_region(kvm,
250 				      APIC_ACCESS_PAGE_PRIVATE_MEMSLOT,
251 				      APIC_DEFAULT_PHYS_BASE,
252 				      activate ? PAGE_SIZE : 0);
253 	if (IS_ERR(ret)) {
254 		r = PTR_ERR(ret);
255 		goto out;
256 	}
257 
258 	kvm->arch.apic_access_page_done = activate;
259 out:
260 	mutex_unlock(&kvm->slots_lock);
261 	return r;
262 }
263 
264 static int avic_init_backing_page(struct kvm_vcpu *vcpu)
265 {
266 	u64 *entry, new_entry;
267 	int id = vcpu->vcpu_id;
268 	struct vcpu_svm *svm = to_svm(vcpu);
269 
270 	if (id >= AVIC_MAX_PHYSICAL_ID_COUNT)
271 		return -EINVAL;
272 
273 	if (!vcpu->arch.apic->regs)
274 		return -EINVAL;
275 
276 	if (kvm_apicv_activated(vcpu->kvm)) {
277 		int ret;
278 
279 		ret = avic_update_access_page(vcpu->kvm, true);
280 		if (ret)
281 			return ret;
282 	}
283 
284 	svm->avic_backing_page = virt_to_page(vcpu->arch.apic->regs);
285 
286 	/* Setting AVIC backing page address in the phy APIC ID table */
287 	entry = avic_get_physical_id_entry(vcpu, id);
288 	if (!entry)
289 		return -EINVAL;
290 
291 	new_entry = __sme_set((page_to_phys(svm->avic_backing_page) &
292 			      AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK) |
293 			      AVIC_PHYSICAL_ID_ENTRY_VALID_MASK);
294 	WRITE_ONCE(*entry, new_entry);
295 
296 	svm->avic_physical_id_cache = entry;
297 
298 	return 0;
299 }
300 
301 static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
302 				   u32 icrl, u32 icrh)
303 {
304 	struct kvm_vcpu *vcpu;
305 	int i;
306 
307 	kvm_for_each_vcpu(i, vcpu, kvm) {
308 		bool m = kvm_apic_match_dest(vcpu, source,
309 					     icrl & APIC_SHORT_MASK,
310 					     GET_APIC_DEST_FIELD(icrh),
311 					     icrl & APIC_DEST_MASK);
312 
313 		if (m && !avic_vcpu_is_running(vcpu))
314 			kvm_vcpu_wake_up(vcpu);
315 	}
316 }
317 
318 int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
319 {
320 	struct vcpu_svm *svm = to_svm(vcpu);
321 	u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
322 	u32 icrl = svm->vmcb->control.exit_info_1;
323 	u32 id = svm->vmcb->control.exit_info_2 >> 32;
324 	u32 index = svm->vmcb->control.exit_info_2 & 0xFF;
325 	struct kvm_lapic *apic = vcpu->arch.apic;
326 
327 	trace_kvm_avic_incomplete_ipi(vcpu->vcpu_id, icrh, icrl, id, index);
328 
329 	switch (id) {
330 	case AVIC_IPI_FAILURE_INVALID_INT_TYPE:
331 		/*
332 		 * AVIC hardware handles the generation of
333 		 * IPIs when the specified Message Type is Fixed
334 		 * (also known as fixed delivery mode) and
335 		 * the Trigger Mode is edge-triggered. The hardware
336 		 * also supports self and broadcast delivery modes
337 		 * specified via the Destination Shorthand(DSH)
338 		 * field of the ICRL. Logical and physical APIC ID
339 		 * formats are supported. All other IPI types cause
340 		 * a #VMEXIT, which needs to emulated.
341 		 */
342 		kvm_lapic_reg_write(apic, APIC_ICR2, icrh);
343 		kvm_lapic_reg_write(apic, APIC_ICR, icrl);
344 		break;
345 	case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING:
346 		/*
347 		 * At this point, we expect that the AVIC HW has already
348 		 * set the appropriate IRR bits on the valid target
349 		 * vcpus. So, we just need to kick the appropriate vcpu.
350 		 */
351 		avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh);
352 		break;
353 	case AVIC_IPI_FAILURE_INVALID_TARGET:
354 		WARN_ONCE(1, "Invalid IPI target: index=%u, vcpu=%d, icr=%#0x:%#0x\n",
355 			  index, vcpu->vcpu_id, icrh, icrl);
356 		break;
357 	case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
358 		WARN_ONCE(1, "Invalid backing page\n");
359 		break;
360 	default:
361 		pr_err("Unknown IPI interception\n");
362 	}
363 
364 	return 1;
365 }
366 
367 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
368 {
369 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
370 	int index;
371 	u32 *logical_apic_id_table;
372 	int dlid = GET_APIC_LOGICAL_ID(ldr);
373 
374 	if (!dlid)
375 		return NULL;
376 
377 	if (flat) { /* flat */
378 		index = ffs(dlid) - 1;
379 		if (index > 7)
380 			return NULL;
381 	} else { /* cluster */
382 		int cluster = (dlid & 0xf0) >> 4;
383 		int apic = ffs(dlid & 0x0f) - 1;
384 
385 		if ((apic < 0) || (apic > 7) ||
386 		    (cluster >= 0xf))
387 			return NULL;
388 		index = (cluster << 2) + apic;
389 	}
390 
391 	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
392 
393 	return &logical_apic_id_table[index];
394 }
395 
396 static int avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
397 {
398 	bool flat;
399 	u32 *entry, new_entry;
400 
401 	flat = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR) == APIC_DFR_FLAT;
402 	entry = avic_get_logical_id_entry(vcpu, ldr, flat);
403 	if (!entry)
404 		return -EINVAL;
405 
406 	new_entry = READ_ONCE(*entry);
407 	new_entry &= ~AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK;
408 	new_entry |= (g_physical_id & AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK);
409 	new_entry |= AVIC_LOGICAL_ID_ENTRY_VALID_MASK;
410 	WRITE_ONCE(*entry, new_entry);
411 
412 	return 0;
413 }
414 
415 static void avic_invalidate_logical_id_entry(struct kvm_vcpu *vcpu)
416 {
417 	struct vcpu_svm *svm = to_svm(vcpu);
418 	bool flat = svm->dfr_reg == APIC_DFR_FLAT;
419 	u32 *entry = avic_get_logical_id_entry(vcpu, svm->ldr_reg, flat);
420 
421 	if (entry)
422 		clear_bit(AVIC_LOGICAL_ID_ENTRY_VALID_BIT, (unsigned long *)entry);
423 }
424 
425 static int avic_handle_ldr_update(struct kvm_vcpu *vcpu)
426 {
427 	int ret = 0;
428 	struct vcpu_svm *svm = to_svm(vcpu);
429 	u32 ldr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_LDR);
430 	u32 id = kvm_xapic_id(vcpu->arch.apic);
431 
432 	if (ldr == svm->ldr_reg)
433 		return 0;
434 
435 	avic_invalidate_logical_id_entry(vcpu);
436 
437 	if (ldr)
438 		ret = avic_ldr_write(vcpu, id, ldr);
439 
440 	if (!ret)
441 		svm->ldr_reg = ldr;
442 
443 	return ret;
444 }
445 
446 static int avic_handle_apic_id_update(struct kvm_vcpu *vcpu)
447 {
448 	u64 *old, *new;
449 	struct vcpu_svm *svm = to_svm(vcpu);
450 	u32 id = kvm_xapic_id(vcpu->arch.apic);
451 
452 	if (vcpu->vcpu_id == id)
453 		return 0;
454 
455 	old = avic_get_physical_id_entry(vcpu, vcpu->vcpu_id);
456 	new = avic_get_physical_id_entry(vcpu, id);
457 	if (!new || !old)
458 		return 1;
459 
460 	/* We need to move physical_id_entry to new offset */
461 	*new = *old;
462 	*old = 0ULL;
463 	to_svm(vcpu)->avic_physical_id_cache = new;
464 
465 	/*
466 	 * Also update the guest physical APIC ID in the logical
467 	 * APIC ID table entry if already setup the LDR.
468 	 */
469 	if (svm->ldr_reg)
470 		avic_handle_ldr_update(vcpu);
471 
472 	return 0;
473 }
474 
475 static void avic_handle_dfr_update(struct kvm_vcpu *vcpu)
476 {
477 	struct vcpu_svm *svm = to_svm(vcpu);
478 	u32 dfr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR);
479 
480 	if (svm->dfr_reg == dfr)
481 		return;
482 
483 	avic_invalidate_logical_id_entry(vcpu);
484 	svm->dfr_reg = dfr;
485 }
486 
487 static int avic_unaccel_trap_write(struct vcpu_svm *svm)
488 {
489 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
490 	u32 offset = svm->vmcb->control.exit_info_1 &
491 				AVIC_UNACCEL_ACCESS_OFFSET_MASK;
492 
493 	switch (offset) {
494 	case APIC_ID:
495 		if (avic_handle_apic_id_update(&svm->vcpu))
496 			return 0;
497 		break;
498 	case APIC_LDR:
499 		if (avic_handle_ldr_update(&svm->vcpu))
500 			return 0;
501 		break;
502 	case APIC_DFR:
503 		avic_handle_dfr_update(&svm->vcpu);
504 		break;
505 	default:
506 		break;
507 	}
508 
509 	kvm_lapic_reg_write(apic, offset, kvm_lapic_get_reg(apic, offset));
510 
511 	return 1;
512 }
513 
514 static bool is_avic_unaccelerated_access_trap(u32 offset)
515 {
516 	bool ret = false;
517 
518 	switch (offset) {
519 	case APIC_ID:
520 	case APIC_EOI:
521 	case APIC_RRR:
522 	case APIC_LDR:
523 	case APIC_DFR:
524 	case APIC_SPIV:
525 	case APIC_ESR:
526 	case APIC_ICR:
527 	case APIC_LVTT:
528 	case APIC_LVTTHMR:
529 	case APIC_LVTPC:
530 	case APIC_LVT0:
531 	case APIC_LVT1:
532 	case APIC_LVTERR:
533 	case APIC_TMICT:
534 	case APIC_TDCR:
535 		ret = true;
536 		break;
537 	default:
538 		break;
539 	}
540 	return ret;
541 }
542 
543 int avic_unaccelerated_access_interception(struct kvm_vcpu *vcpu)
544 {
545 	struct vcpu_svm *svm = to_svm(vcpu);
546 	int ret = 0;
547 	u32 offset = svm->vmcb->control.exit_info_1 &
548 		     AVIC_UNACCEL_ACCESS_OFFSET_MASK;
549 	u32 vector = svm->vmcb->control.exit_info_2 &
550 		     AVIC_UNACCEL_ACCESS_VECTOR_MASK;
551 	bool write = (svm->vmcb->control.exit_info_1 >> 32) &
552 		     AVIC_UNACCEL_ACCESS_WRITE_MASK;
553 	bool trap = is_avic_unaccelerated_access_trap(offset);
554 
555 	trace_kvm_avic_unaccelerated_access(vcpu->vcpu_id, offset,
556 					    trap, write, vector);
557 	if (trap) {
558 		/* Handling Trap */
559 		WARN_ONCE(!write, "svm: Handling trap read.\n");
560 		ret = avic_unaccel_trap_write(svm);
561 	} else {
562 		/* Handling Fault */
563 		ret = kvm_emulate_instruction(vcpu, 0);
564 	}
565 
566 	return ret;
567 }
568 
569 int avic_init_vcpu(struct vcpu_svm *svm)
570 {
571 	int ret;
572 	struct kvm_vcpu *vcpu = &svm->vcpu;
573 
574 	if (!avic || !irqchip_in_kernel(vcpu->kvm))
575 		return 0;
576 
577 	ret = avic_init_backing_page(vcpu);
578 	if (ret)
579 		return ret;
580 
581 	INIT_LIST_HEAD(&svm->ir_list);
582 	spin_lock_init(&svm->ir_list_lock);
583 	svm->dfr_reg = APIC_DFR_FLAT;
584 
585 	return ret;
586 }
587 
588 void avic_post_state_restore(struct kvm_vcpu *vcpu)
589 {
590 	if (avic_handle_apic_id_update(vcpu) != 0)
591 		return;
592 	avic_handle_dfr_update(vcpu);
593 	avic_handle_ldr_update(vcpu);
594 }
595 
596 void svm_toggle_avic_for_irq_window(struct kvm_vcpu *vcpu, bool activate)
597 {
598 	if (!avic || !lapic_in_kernel(vcpu))
599 		return;
600 
601 	srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
602 	kvm_request_apicv_update(vcpu->kvm, activate,
603 				 APICV_INHIBIT_REASON_IRQWIN);
604 	vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
605 }
606 
607 void svm_set_virtual_apic_mode(struct kvm_vcpu *vcpu)
608 {
609 	return;
610 }
611 
612 void svm_hwapic_irr_update(struct kvm_vcpu *vcpu, int max_irr)
613 {
614 }
615 
616 void svm_hwapic_isr_update(struct kvm_vcpu *vcpu, int max_isr)
617 {
618 }
619 
620 static int svm_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
621 {
622 	int ret = 0;
623 	unsigned long flags;
624 	struct amd_svm_iommu_ir *ir;
625 	struct vcpu_svm *svm = to_svm(vcpu);
626 
627 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
628 		return 0;
629 
630 	/*
631 	 * Here, we go through the per-vcpu ir_list to update all existing
632 	 * interrupt remapping table entry targeting this vcpu.
633 	 */
634 	spin_lock_irqsave(&svm->ir_list_lock, flags);
635 
636 	if (list_empty(&svm->ir_list))
637 		goto out;
638 
639 	list_for_each_entry(ir, &svm->ir_list, node) {
640 		if (activate)
641 			ret = amd_iommu_activate_guest_mode(ir->data);
642 		else
643 			ret = amd_iommu_deactivate_guest_mode(ir->data);
644 		if (ret)
645 			break;
646 	}
647 out:
648 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
649 	return ret;
650 }
651 
652 void svm_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
653 {
654 	struct vcpu_svm *svm = to_svm(vcpu);
655 	struct vmcb *vmcb = svm->vmcb;
656 	bool activated = kvm_vcpu_apicv_active(vcpu);
657 
658 	if (!avic)
659 		return;
660 
661 	if (activated) {
662 		/**
663 		 * During AVIC temporary deactivation, guest could update
664 		 * APIC ID, DFR and LDR registers, which would not be trapped
665 		 * by avic_unaccelerated_access_interception(). In this case,
666 		 * we need to check and update the AVIC logical APIC ID table
667 		 * accordingly before re-activating.
668 		 */
669 		avic_post_state_restore(vcpu);
670 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
671 	} else {
672 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
673 	}
674 	vmcb_mark_dirty(vmcb, VMCB_AVIC);
675 
676 	svm_set_pi_irte_mode(vcpu, activated);
677 }
678 
679 void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)
680 {
681 	return;
682 }
683 
684 int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
685 {
686 	if (!vcpu->arch.apicv_active)
687 		return -1;
688 
689 	kvm_lapic_set_irr(vec, vcpu->arch.apic);
690 	smp_mb__after_atomic();
691 
692 	if (avic_vcpu_is_running(vcpu)) {
693 		int cpuid = vcpu->cpu;
694 
695 		if (cpuid != get_cpu())
696 			wrmsrl(SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpuid));
697 		put_cpu();
698 	} else
699 		kvm_vcpu_wake_up(vcpu);
700 
701 	return 0;
702 }
703 
704 bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
705 {
706 	return false;
707 }
708 
709 static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
710 {
711 	unsigned long flags;
712 	struct amd_svm_iommu_ir *cur;
713 
714 	spin_lock_irqsave(&svm->ir_list_lock, flags);
715 	list_for_each_entry(cur, &svm->ir_list, node) {
716 		if (cur->data != pi->ir_data)
717 			continue;
718 		list_del(&cur->node);
719 		kfree(cur);
720 		break;
721 	}
722 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
723 }
724 
725 static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
726 {
727 	int ret = 0;
728 	unsigned long flags;
729 	struct amd_svm_iommu_ir *ir;
730 
731 	/**
732 	 * In some cases, the existing irte is updated and re-set,
733 	 * so we need to check here if it's already been * added
734 	 * to the ir_list.
735 	 */
736 	if (pi->ir_data && (pi->prev_ga_tag != 0)) {
737 		struct kvm *kvm = svm->vcpu.kvm;
738 		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
739 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
740 		struct vcpu_svm *prev_svm;
741 
742 		if (!prev_vcpu) {
743 			ret = -EINVAL;
744 			goto out;
745 		}
746 
747 		prev_svm = to_svm(prev_vcpu);
748 		svm_ir_list_del(prev_svm, pi);
749 	}
750 
751 	/**
752 	 * Allocating new amd_iommu_pi_data, which will get
753 	 * add to the per-vcpu ir_list.
754 	 */
755 	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_KERNEL_ACCOUNT);
756 	if (!ir) {
757 		ret = -ENOMEM;
758 		goto out;
759 	}
760 	ir->data = pi->ir_data;
761 
762 	spin_lock_irqsave(&svm->ir_list_lock, flags);
763 	list_add(&ir->node, &svm->ir_list);
764 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
765 out:
766 	return ret;
767 }
768 
769 /**
770  * Note:
771  * The HW cannot support posting multicast/broadcast
772  * interrupts to a vCPU. So, we still use legacy interrupt
773  * remapping for these kind of interrupts.
774  *
775  * For lowest-priority interrupts, we only support
776  * those with single CPU as the destination, e.g. user
777  * configures the interrupts via /proc/irq or uses
778  * irqbalance to make the interrupts single-CPU.
779  */
780 static int
781 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
782 		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
783 {
784 	struct kvm_lapic_irq irq;
785 	struct kvm_vcpu *vcpu = NULL;
786 
787 	kvm_set_msi_irq(kvm, e, &irq);
788 
789 	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
790 	    !kvm_irq_is_postable(&irq)) {
791 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
792 			 __func__, irq.vector);
793 		return -1;
794 	}
795 
796 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
797 		 irq.vector);
798 	*svm = to_svm(vcpu);
799 	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
800 	vcpu_info->vector = irq.vector;
801 
802 	return 0;
803 }
804 
805 /*
806  * svm_update_pi_irte - set IRTE for Posted-Interrupts
807  *
808  * @kvm: kvm
809  * @host_irq: host irq of the interrupt
810  * @guest_irq: gsi of the interrupt
811  * @set: set or unset PI
812  * returns 0 on success, < 0 on failure
813  */
814 int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
815 		       uint32_t guest_irq, bool set)
816 {
817 	struct kvm_kernel_irq_routing_entry *e;
818 	struct kvm_irq_routing_table *irq_rt;
819 	int idx, ret = -EINVAL;
820 
821 	if (!kvm_arch_has_assigned_device(kvm) ||
822 	    !irq_remapping_cap(IRQ_POSTING_CAP))
823 		return 0;
824 
825 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
826 		 __func__, host_irq, guest_irq, set);
827 
828 	idx = srcu_read_lock(&kvm->irq_srcu);
829 	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
830 	WARN_ON(guest_irq >= irq_rt->nr_rt_entries);
831 
832 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
833 		struct vcpu_data vcpu_info;
834 		struct vcpu_svm *svm = NULL;
835 
836 		if (e->type != KVM_IRQ_ROUTING_MSI)
837 			continue;
838 
839 		/**
840 		 * Here, we setup with legacy mode in the following cases:
841 		 * 1. When cannot target interrupt to a specific vcpu.
842 		 * 2. Unsetting posted interrupt.
843 		 * 3. APIC virtualization is disabled for the vcpu.
844 		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
845 		 */
846 		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
847 		    kvm_vcpu_apicv_active(&svm->vcpu)) {
848 			struct amd_iommu_pi_data pi;
849 
850 			/* Try to enable guest_mode in IRTE */
851 			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
852 					    AVIC_HPA_MASK);
853 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
854 						     svm->vcpu.vcpu_id);
855 			pi.is_guest_mode = true;
856 			pi.vcpu_data = &vcpu_info;
857 			ret = irq_set_vcpu_affinity(host_irq, &pi);
858 
859 			/**
860 			 * Here, we successfully setting up vcpu affinity in
861 			 * IOMMU guest mode. Now, we need to store the posted
862 			 * interrupt information in a per-vcpu ir_list so that
863 			 * we can reference to them directly when we update vcpu
864 			 * scheduling information in IOMMU irte.
865 			 */
866 			if (!ret && pi.is_guest_mode)
867 				svm_ir_list_add(svm, &pi);
868 		} else {
869 			/* Use legacy mode in IRTE */
870 			struct amd_iommu_pi_data pi;
871 
872 			/**
873 			 * Here, pi is used to:
874 			 * - Tell IOMMU to use legacy mode for this interrupt.
875 			 * - Retrieve ga_tag of prior interrupt remapping data.
876 			 */
877 			pi.prev_ga_tag = 0;
878 			pi.is_guest_mode = false;
879 			ret = irq_set_vcpu_affinity(host_irq, &pi);
880 
881 			/**
882 			 * Check if the posted interrupt was previously
883 			 * setup with the guest_mode by checking if the ga_tag
884 			 * was cached. If so, we need to clean up the per-vcpu
885 			 * ir_list.
886 			 */
887 			if (!ret && pi.prev_ga_tag) {
888 				int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
889 				struct kvm_vcpu *vcpu;
890 
891 				vcpu = kvm_get_vcpu_by_id(kvm, id);
892 				if (vcpu)
893 					svm_ir_list_del(to_svm(vcpu), &pi);
894 			}
895 		}
896 
897 		if (!ret && svm) {
898 			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
899 						 e->gsi, vcpu_info.vector,
900 						 vcpu_info.pi_desc_addr, set);
901 		}
902 
903 		if (ret < 0) {
904 			pr_err("%s: failed to update PI IRTE\n", __func__);
905 			goto out;
906 		}
907 	}
908 
909 	ret = 0;
910 out:
911 	srcu_read_unlock(&kvm->irq_srcu, idx);
912 	return ret;
913 }
914 
915 bool svm_check_apicv_inhibit_reasons(ulong bit)
916 {
917 	ulong supported = BIT(APICV_INHIBIT_REASON_DISABLE) |
918 			  BIT(APICV_INHIBIT_REASON_HYPERV) |
919 			  BIT(APICV_INHIBIT_REASON_NESTED) |
920 			  BIT(APICV_INHIBIT_REASON_IRQWIN) |
921 			  BIT(APICV_INHIBIT_REASON_PIT_REINJ) |
922 			  BIT(APICV_INHIBIT_REASON_X2APIC);
923 
924 	return supported & BIT(bit);
925 }
926 
927 void svm_pre_update_apicv_exec_ctrl(struct kvm *kvm, bool activate)
928 {
929 	avic_update_access_page(kvm, activate);
930 }
931 
932 static inline int
933 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
934 {
935 	int ret = 0;
936 	unsigned long flags;
937 	struct amd_svm_iommu_ir *ir;
938 	struct vcpu_svm *svm = to_svm(vcpu);
939 
940 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
941 		return 0;
942 
943 	/*
944 	 * Here, we go through the per-vcpu ir_list to update all existing
945 	 * interrupt remapping table entry targeting this vcpu.
946 	 */
947 	spin_lock_irqsave(&svm->ir_list_lock, flags);
948 
949 	if (list_empty(&svm->ir_list))
950 		goto out;
951 
952 	list_for_each_entry(ir, &svm->ir_list, node) {
953 		ret = amd_iommu_update_ga(cpu, r, ir->data);
954 		if (ret)
955 			break;
956 	}
957 out:
958 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
959 	return ret;
960 }
961 
962 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
963 {
964 	u64 entry;
965 	/* ID = 0xff (broadcast), ID > 0xff (reserved) */
966 	int h_physical_id = kvm_cpu_get_apicid(cpu);
967 	struct vcpu_svm *svm = to_svm(vcpu);
968 
969 	if (!kvm_vcpu_apicv_active(vcpu))
970 		return;
971 
972 	/*
973 	 * Since the host physical APIC id is 8 bits,
974 	 * we can support host APIC ID upto 255.
975 	 */
976 	if (WARN_ON(h_physical_id > AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
977 		return;
978 
979 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
980 	WARN_ON(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
981 
982 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
983 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
984 
985 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
986 	if (svm->avic_is_running)
987 		entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
988 
989 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
990 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id,
991 					svm->avic_is_running);
992 }
993 
994 void avic_vcpu_put(struct kvm_vcpu *vcpu)
995 {
996 	u64 entry;
997 	struct vcpu_svm *svm = to_svm(vcpu);
998 
999 	if (!kvm_vcpu_apicv_active(vcpu))
1000 		return;
1001 
1002 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
1003 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
1004 		avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
1005 
1006 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
1007 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
1008 }
1009 
1010 /**
1011  * This function is called during VCPU halt/unhalt.
1012  */
1013 static void avic_set_running(struct kvm_vcpu *vcpu, bool is_run)
1014 {
1015 	struct vcpu_svm *svm = to_svm(vcpu);
1016 
1017 	svm->avic_is_running = is_run;
1018 	if (is_run)
1019 		avic_vcpu_load(vcpu, vcpu->cpu);
1020 	else
1021 		avic_vcpu_put(vcpu);
1022 }
1023 
1024 void svm_vcpu_blocking(struct kvm_vcpu *vcpu)
1025 {
1026 	avic_set_running(vcpu, false);
1027 }
1028 
1029 void svm_vcpu_unblocking(struct kvm_vcpu *vcpu)
1030 {
1031 	if (kvm_check_request(KVM_REQ_APICV_UPDATE, vcpu))
1032 		kvm_vcpu_update_apicv(vcpu);
1033 	avic_set_running(vcpu, true);
1034 }
1035