xref: /openbmc/linux/arch/x86/kvm/svm/avic.c (revision 8938c48f)
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 (!svm->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(svm->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 vcpu_svm *svm)
319 {
320 	u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
321 	u32 icrl = svm->vmcb->control.exit_info_1;
322 	u32 id = svm->vmcb->control.exit_info_2 >> 32;
323 	u32 index = svm->vmcb->control.exit_info_2 & 0xFF;
324 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
325 
326 	trace_kvm_avic_incomplete_ipi(svm->vcpu.vcpu_id, icrh, icrl, id, index);
327 
328 	switch (id) {
329 	case AVIC_IPI_FAILURE_INVALID_INT_TYPE:
330 		/*
331 		 * AVIC hardware handles the generation of
332 		 * IPIs when the specified Message Type is Fixed
333 		 * (also known as fixed delivery mode) and
334 		 * the Trigger Mode is edge-triggered. The hardware
335 		 * also supports self and broadcast delivery modes
336 		 * specified via the Destination Shorthand(DSH)
337 		 * field of the ICRL. Logical and physical APIC ID
338 		 * formats are supported. All other IPI types cause
339 		 * a #VMEXIT, which needs to emulated.
340 		 */
341 		kvm_lapic_reg_write(apic, APIC_ICR2, icrh);
342 		kvm_lapic_reg_write(apic, APIC_ICR, icrl);
343 		break;
344 	case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING:
345 		/*
346 		 * At this point, we expect that the AVIC HW has already
347 		 * set the appropriate IRR bits on the valid target
348 		 * vcpus. So, we just need to kick the appropriate vcpu.
349 		 */
350 		avic_kick_target_vcpus(svm->vcpu.kvm, apic, icrl, icrh);
351 		break;
352 	case AVIC_IPI_FAILURE_INVALID_TARGET:
353 		WARN_ONCE(1, "Invalid IPI target: index=%u, vcpu=%d, icr=%#0x:%#0x\n",
354 			  index, svm->vcpu.vcpu_id, icrh, icrl);
355 		break;
356 	case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
357 		WARN_ONCE(1, "Invalid backing page\n");
358 		break;
359 	default:
360 		pr_err("Unknown IPI interception\n");
361 	}
362 
363 	return 1;
364 }
365 
366 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
367 {
368 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
369 	int index;
370 	u32 *logical_apic_id_table;
371 	int dlid = GET_APIC_LOGICAL_ID(ldr);
372 
373 	if (!dlid)
374 		return NULL;
375 
376 	if (flat) { /* flat */
377 		index = ffs(dlid) - 1;
378 		if (index > 7)
379 			return NULL;
380 	} else { /* cluster */
381 		int cluster = (dlid & 0xf0) >> 4;
382 		int apic = ffs(dlid & 0x0f) - 1;
383 
384 		if ((apic < 0) || (apic > 7) ||
385 		    (cluster >= 0xf))
386 			return NULL;
387 		index = (cluster << 2) + apic;
388 	}
389 
390 	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
391 
392 	return &logical_apic_id_table[index];
393 }
394 
395 static int avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
396 {
397 	bool flat;
398 	u32 *entry, new_entry;
399 
400 	flat = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR) == APIC_DFR_FLAT;
401 	entry = avic_get_logical_id_entry(vcpu, ldr, flat);
402 	if (!entry)
403 		return -EINVAL;
404 
405 	new_entry = READ_ONCE(*entry);
406 	new_entry &= ~AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK;
407 	new_entry |= (g_physical_id & AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK);
408 	new_entry |= AVIC_LOGICAL_ID_ENTRY_VALID_MASK;
409 	WRITE_ONCE(*entry, new_entry);
410 
411 	return 0;
412 }
413 
414 static void avic_invalidate_logical_id_entry(struct kvm_vcpu *vcpu)
415 {
416 	struct vcpu_svm *svm = to_svm(vcpu);
417 	bool flat = svm->dfr_reg == APIC_DFR_FLAT;
418 	u32 *entry = avic_get_logical_id_entry(vcpu, svm->ldr_reg, flat);
419 
420 	if (entry)
421 		clear_bit(AVIC_LOGICAL_ID_ENTRY_VALID_BIT, (unsigned long *)entry);
422 }
423 
424 static int avic_handle_ldr_update(struct kvm_vcpu *vcpu)
425 {
426 	int ret = 0;
427 	struct vcpu_svm *svm = to_svm(vcpu);
428 	u32 ldr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_LDR);
429 	u32 id = kvm_xapic_id(vcpu->arch.apic);
430 
431 	if (ldr == svm->ldr_reg)
432 		return 0;
433 
434 	avic_invalidate_logical_id_entry(vcpu);
435 
436 	if (ldr)
437 		ret = avic_ldr_write(vcpu, id, ldr);
438 
439 	if (!ret)
440 		svm->ldr_reg = ldr;
441 
442 	return ret;
443 }
444 
445 static int avic_handle_apic_id_update(struct kvm_vcpu *vcpu)
446 {
447 	u64 *old, *new;
448 	struct vcpu_svm *svm = to_svm(vcpu);
449 	u32 id = kvm_xapic_id(vcpu->arch.apic);
450 
451 	if (vcpu->vcpu_id == id)
452 		return 0;
453 
454 	old = avic_get_physical_id_entry(vcpu, vcpu->vcpu_id);
455 	new = avic_get_physical_id_entry(vcpu, id);
456 	if (!new || !old)
457 		return 1;
458 
459 	/* We need to move physical_id_entry to new offset */
460 	*new = *old;
461 	*old = 0ULL;
462 	to_svm(vcpu)->avic_physical_id_cache = new;
463 
464 	/*
465 	 * Also update the guest physical APIC ID in the logical
466 	 * APIC ID table entry if already setup the LDR.
467 	 */
468 	if (svm->ldr_reg)
469 		avic_handle_ldr_update(vcpu);
470 
471 	return 0;
472 }
473 
474 static void avic_handle_dfr_update(struct kvm_vcpu *vcpu)
475 {
476 	struct vcpu_svm *svm = to_svm(vcpu);
477 	u32 dfr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR);
478 
479 	if (svm->dfr_reg == dfr)
480 		return;
481 
482 	avic_invalidate_logical_id_entry(vcpu);
483 	svm->dfr_reg = dfr;
484 }
485 
486 static int avic_unaccel_trap_write(struct vcpu_svm *svm)
487 {
488 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
489 	u32 offset = svm->vmcb->control.exit_info_1 &
490 				AVIC_UNACCEL_ACCESS_OFFSET_MASK;
491 
492 	switch (offset) {
493 	case APIC_ID:
494 		if (avic_handle_apic_id_update(&svm->vcpu))
495 			return 0;
496 		break;
497 	case APIC_LDR:
498 		if (avic_handle_ldr_update(&svm->vcpu))
499 			return 0;
500 		break;
501 	case APIC_DFR:
502 		avic_handle_dfr_update(&svm->vcpu);
503 		break;
504 	default:
505 		break;
506 	}
507 
508 	kvm_lapic_reg_write(apic, offset, kvm_lapic_get_reg(apic, offset));
509 
510 	return 1;
511 }
512 
513 static bool is_avic_unaccelerated_access_trap(u32 offset)
514 {
515 	bool ret = false;
516 
517 	switch (offset) {
518 	case APIC_ID:
519 	case APIC_EOI:
520 	case APIC_RRR:
521 	case APIC_LDR:
522 	case APIC_DFR:
523 	case APIC_SPIV:
524 	case APIC_ESR:
525 	case APIC_ICR:
526 	case APIC_LVTT:
527 	case APIC_LVTTHMR:
528 	case APIC_LVTPC:
529 	case APIC_LVT0:
530 	case APIC_LVT1:
531 	case APIC_LVTERR:
532 	case APIC_TMICT:
533 	case APIC_TDCR:
534 		ret = true;
535 		break;
536 	default:
537 		break;
538 	}
539 	return ret;
540 }
541 
542 int avic_unaccelerated_access_interception(struct vcpu_svm *svm)
543 {
544 	int ret = 0;
545 	u32 offset = svm->vmcb->control.exit_info_1 &
546 		     AVIC_UNACCEL_ACCESS_OFFSET_MASK;
547 	u32 vector = svm->vmcb->control.exit_info_2 &
548 		     AVIC_UNACCEL_ACCESS_VECTOR_MASK;
549 	bool write = (svm->vmcb->control.exit_info_1 >> 32) &
550 		     AVIC_UNACCEL_ACCESS_WRITE_MASK;
551 	bool trap = is_avic_unaccelerated_access_trap(offset);
552 
553 	trace_kvm_avic_unaccelerated_access(svm->vcpu.vcpu_id, offset,
554 					    trap, write, vector);
555 	if (trap) {
556 		/* Handling Trap */
557 		WARN_ONCE(!write, "svm: Handling trap read.\n");
558 		ret = avic_unaccel_trap_write(svm);
559 	} else {
560 		/* Handling Fault */
561 		ret = kvm_emulate_instruction(&svm->vcpu, 0);
562 	}
563 
564 	return ret;
565 }
566 
567 int avic_init_vcpu(struct vcpu_svm *svm)
568 {
569 	int ret;
570 	struct kvm_vcpu *vcpu = &svm->vcpu;
571 
572 	if (!avic || !irqchip_in_kernel(vcpu->kvm))
573 		return 0;
574 
575 	ret = avic_init_backing_page(&svm->vcpu);
576 	if (ret)
577 		return ret;
578 
579 	INIT_LIST_HEAD(&svm->ir_list);
580 	spin_lock_init(&svm->ir_list_lock);
581 	svm->dfr_reg = APIC_DFR_FLAT;
582 
583 	return ret;
584 }
585 
586 void avic_post_state_restore(struct kvm_vcpu *vcpu)
587 {
588 	if (avic_handle_apic_id_update(vcpu) != 0)
589 		return;
590 	avic_handle_dfr_update(vcpu);
591 	avic_handle_ldr_update(vcpu);
592 }
593 
594 void svm_toggle_avic_for_irq_window(struct kvm_vcpu *vcpu, bool activate)
595 {
596 	if (!avic || !lapic_in_kernel(vcpu))
597 		return;
598 
599 	srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
600 	kvm_request_apicv_update(vcpu->kvm, activate,
601 				 APICV_INHIBIT_REASON_IRQWIN);
602 	vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
603 }
604 
605 void svm_set_virtual_apic_mode(struct kvm_vcpu *vcpu)
606 {
607 	return;
608 }
609 
610 void svm_hwapic_irr_update(struct kvm_vcpu *vcpu, int max_irr)
611 {
612 }
613 
614 void svm_hwapic_isr_update(struct kvm_vcpu *vcpu, int max_isr)
615 {
616 }
617 
618 static int svm_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
619 {
620 	int ret = 0;
621 	unsigned long flags;
622 	struct amd_svm_iommu_ir *ir;
623 	struct vcpu_svm *svm = to_svm(vcpu);
624 
625 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
626 		return 0;
627 
628 	/*
629 	 * Here, we go through the per-vcpu ir_list to update all existing
630 	 * interrupt remapping table entry targeting this vcpu.
631 	 */
632 	spin_lock_irqsave(&svm->ir_list_lock, flags);
633 
634 	if (list_empty(&svm->ir_list))
635 		goto out;
636 
637 	list_for_each_entry(ir, &svm->ir_list, node) {
638 		if (activate)
639 			ret = amd_iommu_activate_guest_mode(ir->data);
640 		else
641 			ret = amd_iommu_deactivate_guest_mode(ir->data);
642 		if (ret)
643 			break;
644 	}
645 out:
646 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
647 	return ret;
648 }
649 
650 void svm_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
651 {
652 	struct vcpu_svm *svm = to_svm(vcpu);
653 	struct vmcb *vmcb = svm->vmcb;
654 	bool activated = kvm_vcpu_apicv_active(vcpu);
655 
656 	if (!avic)
657 		return;
658 
659 	if (activated) {
660 		/**
661 		 * During AVIC temporary deactivation, guest could update
662 		 * APIC ID, DFR and LDR registers, which would not be trapped
663 		 * by avic_unaccelerated_access_interception(). In this case,
664 		 * we need to check and update the AVIC logical APIC ID table
665 		 * accordingly before re-activating.
666 		 */
667 		avic_post_state_restore(vcpu);
668 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
669 	} else {
670 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
671 	}
672 	vmcb_mark_dirty(vmcb, VMCB_AVIC);
673 
674 	svm_set_pi_irte_mode(vcpu, activated);
675 }
676 
677 void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)
678 {
679 	return;
680 }
681 
682 int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
683 {
684 	if (!vcpu->arch.apicv_active)
685 		return -1;
686 
687 	kvm_lapic_set_irr(vec, vcpu->arch.apic);
688 	smp_mb__after_atomic();
689 
690 	if (avic_vcpu_is_running(vcpu)) {
691 		int cpuid = vcpu->cpu;
692 
693 		if (cpuid != get_cpu())
694 			wrmsrl(SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpuid));
695 		put_cpu();
696 	} else
697 		kvm_vcpu_wake_up(vcpu);
698 
699 	return 0;
700 }
701 
702 bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
703 {
704 	return false;
705 }
706 
707 static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
708 {
709 	unsigned long flags;
710 	struct amd_svm_iommu_ir *cur;
711 
712 	spin_lock_irqsave(&svm->ir_list_lock, flags);
713 	list_for_each_entry(cur, &svm->ir_list, node) {
714 		if (cur->data != pi->ir_data)
715 			continue;
716 		list_del(&cur->node);
717 		kfree(cur);
718 		break;
719 	}
720 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
721 }
722 
723 static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
724 {
725 	int ret = 0;
726 	unsigned long flags;
727 	struct amd_svm_iommu_ir *ir;
728 
729 	/**
730 	 * In some cases, the existing irte is updaed and re-set,
731 	 * so we need to check here if it's already been * added
732 	 * to the ir_list.
733 	 */
734 	if (pi->ir_data && (pi->prev_ga_tag != 0)) {
735 		struct kvm *kvm = svm->vcpu.kvm;
736 		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
737 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
738 		struct vcpu_svm *prev_svm;
739 
740 		if (!prev_vcpu) {
741 			ret = -EINVAL;
742 			goto out;
743 		}
744 
745 		prev_svm = to_svm(prev_vcpu);
746 		svm_ir_list_del(prev_svm, pi);
747 	}
748 
749 	/**
750 	 * Allocating new amd_iommu_pi_data, which will get
751 	 * add to the per-vcpu ir_list.
752 	 */
753 	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_KERNEL_ACCOUNT);
754 	if (!ir) {
755 		ret = -ENOMEM;
756 		goto out;
757 	}
758 	ir->data = pi->ir_data;
759 
760 	spin_lock_irqsave(&svm->ir_list_lock, flags);
761 	list_add(&ir->node, &svm->ir_list);
762 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
763 out:
764 	return ret;
765 }
766 
767 /**
768  * Note:
769  * The HW cannot support posting multicast/broadcast
770  * interrupts to a vCPU. So, we still use legacy interrupt
771  * remapping for these kind of interrupts.
772  *
773  * For lowest-priority interrupts, we only support
774  * those with single CPU as the destination, e.g. user
775  * configures the interrupts via /proc/irq or uses
776  * irqbalance to make the interrupts single-CPU.
777  */
778 static int
779 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
780 		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
781 {
782 	struct kvm_lapic_irq irq;
783 	struct kvm_vcpu *vcpu = NULL;
784 
785 	kvm_set_msi_irq(kvm, e, &irq);
786 
787 	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
788 	    !kvm_irq_is_postable(&irq)) {
789 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
790 			 __func__, irq.vector);
791 		return -1;
792 	}
793 
794 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
795 		 irq.vector);
796 	*svm = to_svm(vcpu);
797 	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
798 	vcpu_info->vector = irq.vector;
799 
800 	return 0;
801 }
802 
803 /*
804  * svm_update_pi_irte - set IRTE for Posted-Interrupts
805  *
806  * @kvm: kvm
807  * @host_irq: host irq of the interrupt
808  * @guest_irq: gsi of the interrupt
809  * @set: set or unset PI
810  * returns 0 on success, < 0 on failure
811  */
812 int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
813 		       uint32_t guest_irq, bool set)
814 {
815 	struct kvm_kernel_irq_routing_entry *e;
816 	struct kvm_irq_routing_table *irq_rt;
817 	int idx, ret = -EINVAL;
818 
819 	if (!kvm_arch_has_assigned_device(kvm) ||
820 	    !irq_remapping_cap(IRQ_POSTING_CAP))
821 		return 0;
822 
823 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
824 		 __func__, host_irq, guest_irq, set);
825 
826 	idx = srcu_read_lock(&kvm->irq_srcu);
827 	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
828 	WARN_ON(guest_irq >= irq_rt->nr_rt_entries);
829 
830 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
831 		struct vcpu_data vcpu_info;
832 		struct vcpu_svm *svm = NULL;
833 
834 		if (e->type != KVM_IRQ_ROUTING_MSI)
835 			continue;
836 
837 		/**
838 		 * Here, we setup with legacy mode in the following cases:
839 		 * 1. When cannot target interrupt to a specific vcpu.
840 		 * 2. Unsetting posted interrupt.
841 		 * 3. APIC virtialization is disabled for the vcpu.
842 		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
843 		 */
844 		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
845 		    kvm_vcpu_apicv_active(&svm->vcpu)) {
846 			struct amd_iommu_pi_data pi;
847 
848 			/* Try to enable guest_mode in IRTE */
849 			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
850 					    AVIC_HPA_MASK);
851 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
852 						     svm->vcpu.vcpu_id);
853 			pi.is_guest_mode = true;
854 			pi.vcpu_data = &vcpu_info;
855 			ret = irq_set_vcpu_affinity(host_irq, &pi);
856 
857 			/**
858 			 * Here, we successfully setting up vcpu affinity in
859 			 * IOMMU guest mode. Now, we need to store the posted
860 			 * interrupt information in a per-vcpu ir_list so that
861 			 * we can reference to them directly when we update vcpu
862 			 * scheduling information in IOMMU irte.
863 			 */
864 			if (!ret && pi.is_guest_mode)
865 				svm_ir_list_add(svm, &pi);
866 		} else {
867 			/* Use legacy mode in IRTE */
868 			struct amd_iommu_pi_data pi;
869 
870 			/**
871 			 * Here, pi is used to:
872 			 * - Tell IOMMU to use legacy mode for this interrupt.
873 			 * - Retrieve ga_tag of prior interrupt remapping data.
874 			 */
875 			pi.prev_ga_tag = 0;
876 			pi.is_guest_mode = false;
877 			ret = irq_set_vcpu_affinity(host_irq, &pi);
878 
879 			/**
880 			 * Check if the posted interrupt was previously
881 			 * setup with the guest_mode by checking if the ga_tag
882 			 * was cached. If so, we need to clean up the per-vcpu
883 			 * ir_list.
884 			 */
885 			if (!ret && pi.prev_ga_tag) {
886 				int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
887 				struct kvm_vcpu *vcpu;
888 
889 				vcpu = kvm_get_vcpu_by_id(kvm, id);
890 				if (vcpu)
891 					svm_ir_list_del(to_svm(vcpu), &pi);
892 			}
893 		}
894 
895 		if (!ret && svm) {
896 			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
897 						 e->gsi, vcpu_info.vector,
898 						 vcpu_info.pi_desc_addr, set);
899 		}
900 
901 		if (ret < 0) {
902 			pr_err("%s: failed to update PI IRTE\n", __func__);
903 			goto out;
904 		}
905 	}
906 
907 	ret = 0;
908 out:
909 	srcu_read_unlock(&kvm->irq_srcu, idx);
910 	return ret;
911 }
912 
913 bool svm_check_apicv_inhibit_reasons(ulong bit)
914 {
915 	ulong supported = BIT(APICV_INHIBIT_REASON_DISABLE) |
916 			  BIT(APICV_INHIBIT_REASON_HYPERV) |
917 			  BIT(APICV_INHIBIT_REASON_NESTED) |
918 			  BIT(APICV_INHIBIT_REASON_IRQWIN) |
919 			  BIT(APICV_INHIBIT_REASON_PIT_REINJ) |
920 			  BIT(APICV_INHIBIT_REASON_X2APIC);
921 
922 	return supported & BIT(bit);
923 }
924 
925 void svm_pre_update_apicv_exec_ctrl(struct kvm *kvm, bool activate)
926 {
927 	avic_update_access_page(kvm, activate);
928 }
929 
930 static inline int
931 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
932 {
933 	int ret = 0;
934 	unsigned long flags;
935 	struct amd_svm_iommu_ir *ir;
936 	struct vcpu_svm *svm = to_svm(vcpu);
937 
938 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
939 		return 0;
940 
941 	/*
942 	 * Here, we go through the per-vcpu ir_list to update all existing
943 	 * interrupt remapping table entry targeting this vcpu.
944 	 */
945 	spin_lock_irqsave(&svm->ir_list_lock, flags);
946 
947 	if (list_empty(&svm->ir_list))
948 		goto out;
949 
950 	list_for_each_entry(ir, &svm->ir_list, node) {
951 		ret = amd_iommu_update_ga(cpu, r, ir->data);
952 		if (ret)
953 			break;
954 	}
955 out:
956 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
957 	return ret;
958 }
959 
960 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
961 {
962 	u64 entry;
963 	/* ID = 0xff (broadcast), ID > 0xff (reserved) */
964 	int h_physical_id = kvm_cpu_get_apicid(cpu);
965 	struct vcpu_svm *svm = to_svm(vcpu);
966 
967 	if (!kvm_vcpu_apicv_active(vcpu))
968 		return;
969 
970 	/*
971 	 * Since the host physical APIC id is 8 bits,
972 	 * we can support host APIC ID upto 255.
973 	 */
974 	if (WARN_ON(h_physical_id > AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
975 		return;
976 
977 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
978 	WARN_ON(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
979 
980 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
981 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
982 
983 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
984 	if (svm->avic_is_running)
985 		entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
986 
987 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
988 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id,
989 					svm->avic_is_running);
990 }
991 
992 void avic_vcpu_put(struct kvm_vcpu *vcpu)
993 {
994 	u64 entry;
995 	struct vcpu_svm *svm = to_svm(vcpu);
996 
997 	if (!kvm_vcpu_apicv_active(vcpu))
998 		return;
999 
1000 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
1001 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
1002 		avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
1003 
1004 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
1005 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
1006 }
1007 
1008 /**
1009  * This function is called during VCPU halt/unhalt.
1010  */
1011 static void avic_set_running(struct kvm_vcpu *vcpu, bool is_run)
1012 {
1013 	struct vcpu_svm *svm = to_svm(vcpu);
1014 
1015 	svm->avic_is_running = is_run;
1016 	if (is_run)
1017 		avic_vcpu_load(vcpu, vcpu->cpu);
1018 	else
1019 		avic_vcpu_put(vcpu);
1020 }
1021 
1022 void svm_vcpu_blocking(struct kvm_vcpu *vcpu)
1023 {
1024 	avic_set_running(vcpu, false);
1025 }
1026 
1027 void svm_vcpu_unblocking(struct kvm_vcpu *vcpu)
1028 {
1029 	if (kvm_check_request(KVM_REQ_APICV_UPDATE, vcpu))
1030 		kvm_vcpu_update_apicv(vcpu);
1031 	avic_set_running(vcpu, true);
1032 }
1033