xref: /openbmc/linux/arch/x86/kvm/svm/avic.c (revision 0bc92e7f)
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 int avic_incomplete_ipi_interception(struct vcpu_svm *svm)
302 {
303 	u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
304 	u32 icrl = svm->vmcb->control.exit_info_1;
305 	u32 id = svm->vmcb->control.exit_info_2 >> 32;
306 	u32 index = svm->vmcb->control.exit_info_2 & 0xFF;
307 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
308 
309 	trace_kvm_avic_incomplete_ipi(svm->vcpu.vcpu_id, icrh, icrl, id, index);
310 
311 	switch (id) {
312 	case AVIC_IPI_FAILURE_INVALID_INT_TYPE:
313 		/*
314 		 * AVIC hardware handles the generation of
315 		 * IPIs when the specified Message Type is Fixed
316 		 * (also known as fixed delivery mode) and
317 		 * the Trigger Mode is edge-triggered. The hardware
318 		 * also supports self and broadcast delivery modes
319 		 * specified via the Destination Shorthand(DSH)
320 		 * field of the ICRL. Logical and physical APIC ID
321 		 * formats are supported. All other IPI types cause
322 		 * a #VMEXIT, which needs to emulated.
323 		 */
324 		kvm_lapic_reg_write(apic, APIC_ICR2, icrh);
325 		kvm_lapic_reg_write(apic, APIC_ICR, icrl);
326 		break;
327 	case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING: {
328 		int i;
329 		struct kvm_vcpu *vcpu;
330 		struct kvm *kvm = svm->vcpu.kvm;
331 		struct kvm_lapic *apic = svm->vcpu.arch.apic;
332 
333 		/*
334 		 * At this point, we expect that the AVIC HW has already
335 		 * set the appropriate IRR bits on the valid target
336 		 * vcpus. So, we just need to kick the appropriate vcpu.
337 		 */
338 		kvm_for_each_vcpu(i, vcpu, kvm) {
339 			bool m = kvm_apic_match_dest(vcpu, apic,
340 						     icrl & APIC_SHORT_MASK,
341 						     GET_APIC_DEST_FIELD(icrh),
342 						     icrl & APIC_DEST_MASK);
343 
344 			if (m && !avic_vcpu_is_running(vcpu))
345 				kvm_vcpu_wake_up(vcpu);
346 		}
347 		break;
348 	}
349 	case AVIC_IPI_FAILURE_INVALID_TARGET:
350 		WARN_ONCE(1, "Invalid IPI target: index=%u, vcpu=%d, icr=%#0x:%#0x\n",
351 			  index, svm->vcpu.vcpu_id, icrh, icrl);
352 		break;
353 	case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
354 		WARN_ONCE(1, "Invalid backing page\n");
355 		break;
356 	default:
357 		pr_err("Unknown IPI interception\n");
358 	}
359 
360 	return 1;
361 }
362 
363 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
364 {
365 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
366 	int index;
367 	u32 *logical_apic_id_table;
368 	int dlid = GET_APIC_LOGICAL_ID(ldr);
369 
370 	if (!dlid)
371 		return NULL;
372 
373 	if (flat) { /* flat */
374 		index = ffs(dlid) - 1;
375 		if (index > 7)
376 			return NULL;
377 	} else { /* cluster */
378 		int cluster = (dlid & 0xf0) >> 4;
379 		int apic = ffs(dlid & 0x0f) - 1;
380 
381 		if ((apic < 0) || (apic > 7) ||
382 		    (cluster >= 0xf))
383 			return NULL;
384 		index = (cluster << 2) + apic;
385 	}
386 
387 	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
388 
389 	return &logical_apic_id_table[index];
390 }
391 
392 static int avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
393 {
394 	bool flat;
395 	u32 *entry, new_entry;
396 
397 	flat = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR) == APIC_DFR_FLAT;
398 	entry = avic_get_logical_id_entry(vcpu, ldr, flat);
399 	if (!entry)
400 		return -EINVAL;
401 
402 	new_entry = READ_ONCE(*entry);
403 	new_entry &= ~AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK;
404 	new_entry |= (g_physical_id & AVIC_LOGICAL_ID_ENTRY_GUEST_PHYSICAL_ID_MASK);
405 	new_entry |= AVIC_LOGICAL_ID_ENTRY_VALID_MASK;
406 	WRITE_ONCE(*entry, new_entry);
407 
408 	return 0;
409 }
410 
411 static void avic_invalidate_logical_id_entry(struct kvm_vcpu *vcpu)
412 {
413 	struct vcpu_svm *svm = to_svm(vcpu);
414 	bool flat = svm->dfr_reg == APIC_DFR_FLAT;
415 	u32 *entry = avic_get_logical_id_entry(vcpu, svm->ldr_reg, flat);
416 
417 	if (entry)
418 		clear_bit(AVIC_LOGICAL_ID_ENTRY_VALID_BIT, (unsigned long *)entry);
419 }
420 
421 static int avic_handle_ldr_update(struct kvm_vcpu *vcpu)
422 {
423 	int ret = 0;
424 	struct vcpu_svm *svm = to_svm(vcpu);
425 	u32 ldr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_LDR);
426 	u32 id = kvm_xapic_id(vcpu->arch.apic);
427 
428 	if (ldr == svm->ldr_reg)
429 		return 0;
430 
431 	avic_invalidate_logical_id_entry(vcpu);
432 
433 	if (ldr)
434 		ret = avic_ldr_write(vcpu, id, ldr);
435 
436 	if (!ret)
437 		svm->ldr_reg = ldr;
438 
439 	return ret;
440 }
441 
442 static int avic_handle_apic_id_update(struct kvm_vcpu *vcpu)
443 {
444 	u64 *old, *new;
445 	struct vcpu_svm *svm = to_svm(vcpu);
446 	u32 id = kvm_xapic_id(vcpu->arch.apic);
447 
448 	if (vcpu->vcpu_id == id)
449 		return 0;
450 
451 	old = avic_get_physical_id_entry(vcpu, vcpu->vcpu_id);
452 	new = avic_get_physical_id_entry(vcpu, id);
453 	if (!new || !old)
454 		return 1;
455 
456 	/* We need to move physical_id_entry to new offset */
457 	*new = *old;
458 	*old = 0ULL;
459 	to_svm(vcpu)->avic_physical_id_cache = new;
460 
461 	/*
462 	 * Also update the guest physical APIC ID in the logical
463 	 * APIC ID table entry if already setup the LDR.
464 	 */
465 	if (svm->ldr_reg)
466 		avic_handle_ldr_update(vcpu);
467 
468 	return 0;
469 }
470 
471 static void avic_handle_dfr_update(struct kvm_vcpu *vcpu)
472 {
473 	struct vcpu_svm *svm = to_svm(vcpu);
474 	u32 dfr = kvm_lapic_get_reg(vcpu->arch.apic, APIC_DFR);
475 
476 	if (svm->dfr_reg == dfr)
477 		return;
478 
479 	avic_invalidate_logical_id_entry(vcpu);
480 	svm->dfr_reg = dfr;
481 }
482 
483 static int avic_unaccel_trap_write(struct vcpu_svm *svm)
484 {
485 	struct kvm_lapic *apic = svm->vcpu.arch.apic;
486 	u32 offset = svm->vmcb->control.exit_info_1 &
487 				AVIC_UNACCEL_ACCESS_OFFSET_MASK;
488 
489 	switch (offset) {
490 	case APIC_ID:
491 		if (avic_handle_apic_id_update(&svm->vcpu))
492 			return 0;
493 		break;
494 	case APIC_LDR:
495 		if (avic_handle_ldr_update(&svm->vcpu))
496 			return 0;
497 		break;
498 	case APIC_DFR:
499 		avic_handle_dfr_update(&svm->vcpu);
500 		break;
501 	default:
502 		break;
503 	}
504 
505 	kvm_lapic_reg_write(apic, offset, kvm_lapic_get_reg(apic, offset));
506 
507 	return 1;
508 }
509 
510 static bool is_avic_unaccelerated_access_trap(u32 offset)
511 {
512 	bool ret = false;
513 
514 	switch (offset) {
515 	case APIC_ID:
516 	case APIC_EOI:
517 	case APIC_RRR:
518 	case APIC_LDR:
519 	case APIC_DFR:
520 	case APIC_SPIV:
521 	case APIC_ESR:
522 	case APIC_ICR:
523 	case APIC_LVTT:
524 	case APIC_LVTTHMR:
525 	case APIC_LVTPC:
526 	case APIC_LVT0:
527 	case APIC_LVT1:
528 	case APIC_LVTERR:
529 	case APIC_TMICT:
530 	case APIC_TDCR:
531 		ret = true;
532 		break;
533 	default:
534 		break;
535 	}
536 	return ret;
537 }
538 
539 int avic_unaccelerated_access_interception(struct vcpu_svm *svm)
540 {
541 	int ret = 0;
542 	u32 offset = svm->vmcb->control.exit_info_1 &
543 		     AVIC_UNACCEL_ACCESS_OFFSET_MASK;
544 	u32 vector = svm->vmcb->control.exit_info_2 &
545 		     AVIC_UNACCEL_ACCESS_VECTOR_MASK;
546 	bool write = (svm->vmcb->control.exit_info_1 >> 32) &
547 		     AVIC_UNACCEL_ACCESS_WRITE_MASK;
548 	bool trap = is_avic_unaccelerated_access_trap(offset);
549 
550 	trace_kvm_avic_unaccelerated_access(svm->vcpu.vcpu_id, offset,
551 					    trap, write, vector);
552 	if (trap) {
553 		/* Handling Trap */
554 		WARN_ONCE(!write, "svm: Handling trap read.\n");
555 		ret = avic_unaccel_trap_write(svm);
556 	} else {
557 		/* Handling Fault */
558 		ret = kvm_emulate_instruction(&svm->vcpu, 0);
559 	}
560 
561 	return ret;
562 }
563 
564 int avic_init_vcpu(struct vcpu_svm *svm)
565 {
566 	int ret;
567 	struct kvm_vcpu *vcpu = &svm->vcpu;
568 
569 	if (!avic || !irqchip_in_kernel(vcpu->kvm))
570 		return 0;
571 
572 	ret = avic_init_backing_page(&svm->vcpu);
573 	if (ret)
574 		return ret;
575 
576 	INIT_LIST_HEAD(&svm->ir_list);
577 	spin_lock_init(&svm->ir_list_lock);
578 	svm->dfr_reg = APIC_DFR_FLAT;
579 
580 	return ret;
581 }
582 
583 void avic_post_state_restore(struct kvm_vcpu *vcpu)
584 {
585 	if (avic_handle_apic_id_update(vcpu) != 0)
586 		return;
587 	avic_handle_dfr_update(vcpu);
588 	avic_handle_ldr_update(vcpu);
589 }
590 
591 void svm_toggle_avic_for_irq_window(struct kvm_vcpu *vcpu, bool activate)
592 {
593 	if (!avic || !lapic_in_kernel(vcpu))
594 		return;
595 
596 	srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
597 	kvm_request_apicv_update(vcpu->kvm, activate,
598 				 APICV_INHIBIT_REASON_IRQWIN);
599 	vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
600 }
601 
602 void svm_set_virtual_apic_mode(struct kvm_vcpu *vcpu)
603 {
604 	return;
605 }
606 
607 void svm_hwapic_irr_update(struct kvm_vcpu *vcpu, int max_irr)
608 {
609 }
610 
611 void svm_hwapic_isr_update(struct kvm_vcpu *vcpu, int max_isr)
612 {
613 }
614 
615 static int svm_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
616 {
617 	int ret = 0;
618 	unsigned long flags;
619 	struct amd_svm_iommu_ir *ir;
620 	struct vcpu_svm *svm = to_svm(vcpu);
621 
622 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
623 		return 0;
624 
625 	/*
626 	 * Here, we go through the per-vcpu ir_list to update all existing
627 	 * interrupt remapping table entry targeting this vcpu.
628 	 */
629 	spin_lock_irqsave(&svm->ir_list_lock, flags);
630 
631 	if (list_empty(&svm->ir_list))
632 		goto out;
633 
634 	list_for_each_entry(ir, &svm->ir_list, node) {
635 		if (activate)
636 			ret = amd_iommu_activate_guest_mode(ir->data);
637 		else
638 			ret = amd_iommu_deactivate_guest_mode(ir->data);
639 		if (ret)
640 			break;
641 	}
642 out:
643 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
644 	return ret;
645 }
646 
647 void svm_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
648 {
649 	struct vcpu_svm *svm = to_svm(vcpu);
650 	struct vmcb *vmcb = svm->vmcb;
651 	bool activated = kvm_vcpu_apicv_active(vcpu);
652 
653 	if (!avic)
654 		return;
655 
656 	if (activated) {
657 		/**
658 		 * During AVIC temporary deactivation, guest could update
659 		 * APIC ID, DFR and LDR registers, which would not be trapped
660 		 * by avic_unaccelerated_access_interception(). In this case,
661 		 * we need to check and update the AVIC logical APIC ID table
662 		 * accordingly before re-activating.
663 		 */
664 		avic_post_state_restore(vcpu);
665 		vmcb->control.int_ctl |= AVIC_ENABLE_MASK;
666 	} else {
667 		vmcb->control.int_ctl &= ~AVIC_ENABLE_MASK;
668 	}
669 	vmcb_mark_dirty(vmcb, VMCB_AVIC);
670 
671 	svm_set_pi_irte_mode(vcpu, activated);
672 }
673 
674 void svm_load_eoi_exitmap(struct kvm_vcpu *vcpu, u64 *eoi_exit_bitmap)
675 {
676 	return;
677 }
678 
679 int svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
680 {
681 	if (!vcpu->arch.apicv_active)
682 		return -1;
683 
684 	kvm_lapic_set_irr(vec, vcpu->arch.apic);
685 	smp_mb__after_atomic();
686 
687 	if (avic_vcpu_is_running(vcpu)) {
688 		int cpuid = vcpu->cpu;
689 
690 		if (cpuid != get_cpu())
691 			wrmsrl(SVM_AVIC_DOORBELL, kvm_cpu_get_apicid(cpuid));
692 		put_cpu();
693 	} else
694 		kvm_vcpu_wake_up(vcpu);
695 
696 	return 0;
697 }
698 
699 bool svm_dy_apicv_has_pending_interrupt(struct kvm_vcpu *vcpu)
700 {
701 	return false;
702 }
703 
704 static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
705 {
706 	unsigned long flags;
707 	struct amd_svm_iommu_ir *cur;
708 
709 	spin_lock_irqsave(&svm->ir_list_lock, flags);
710 	list_for_each_entry(cur, &svm->ir_list, node) {
711 		if (cur->data != pi->ir_data)
712 			continue;
713 		list_del(&cur->node);
714 		kfree(cur);
715 		break;
716 	}
717 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
718 }
719 
720 static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
721 {
722 	int ret = 0;
723 	unsigned long flags;
724 	struct amd_svm_iommu_ir *ir;
725 
726 	/**
727 	 * In some cases, the existing irte is updaed and re-set,
728 	 * so we need to check here if it's already been * added
729 	 * to the ir_list.
730 	 */
731 	if (pi->ir_data && (pi->prev_ga_tag != 0)) {
732 		struct kvm *kvm = svm->vcpu.kvm;
733 		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
734 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
735 		struct vcpu_svm *prev_svm;
736 
737 		if (!prev_vcpu) {
738 			ret = -EINVAL;
739 			goto out;
740 		}
741 
742 		prev_svm = to_svm(prev_vcpu);
743 		svm_ir_list_del(prev_svm, pi);
744 	}
745 
746 	/**
747 	 * Allocating new amd_iommu_pi_data, which will get
748 	 * add to the per-vcpu ir_list.
749 	 */
750 	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_KERNEL_ACCOUNT);
751 	if (!ir) {
752 		ret = -ENOMEM;
753 		goto out;
754 	}
755 	ir->data = pi->ir_data;
756 
757 	spin_lock_irqsave(&svm->ir_list_lock, flags);
758 	list_add(&ir->node, &svm->ir_list);
759 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
760 out:
761 	return ret;
762 }
763 
764 /**
765  * Note:
766  * The HW cannot support posting multicast/broadcast
767  * interrupts to a vCPU. So, we still use legacy interrupt
768  * remapping for these kind of interrupts.
769  *
770  * For lowest-priority interrupts, we only support
771  * those with single CPU as the destination, e.g. user
772  * configures the interrupts via /proc/irq or uses
773  * irqbalance to make the interrupts single-CPU.
774  */
775 static int
776 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
777 		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
778 {
779 	struct kvm_lapic_irq irq;
780 	struct kvm_vcpu *vcpu = NULL;
781 
782 	kvm_set_msi_irq(kvm, e, &irq);
783 
784 	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
785 	    !kvm_irq_is_postable(&irq)) {
786 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
787 			 __func__, irq.vector);
788 		return -1;
789 	}
790 
791 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
792 		 irq.vector);
793 	*svm = to_svm(vcpu);
794 	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
795 	vcpu_info->vector = irq.vector;
796 
797 	return 0;
798 }
799 
800 /*
801  * svm_update_pi_irte - set IRTE for Posted-Interrupts
802  *
803  * @kvm: kvm
804  * @host_irq: host irq of the interrupt
805  * @guest_irq: gsi of the interrupt
806  * @set: set or unset PI
807  * returns 0 on success, < 0 on failure
808  */
809 int svm_update_pi_irte(struct kvm *kvm, unsigned int host_irq,
810 		       uint32_t guest_irq, bool set)
811 {
812 	struct kvm_kernel_irq_routing_entry *e;
813 	struct kvm_irq_routing_table *irq_rt;
814 	int idx, ret = -EINVAL;
815 
816 	if (!kvm_arch_has_assigned_device(kvm) ||
817 	    !irq_remapping_cap(IRQ_POSTING_CAP))
818 		return 0;
819 
820 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
821 		 __func__, host_irq, guest_irq, set);
822 
823 	idx = srcu_read_lock(&kvm->irq_srcu);
824 	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
825 	WARN_ON(guest_irq >= irq_rt->nr_rt_entries);
826 
827 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
828 		struct vcpu_data vcpu_info;
829 		struct vcpu_svm *svm = NULL;
830 
831 		if (e->type != KVM_IRQ_ROUTING_MSI)
832 			continue;
833 
834 		/**
835 		 * Here, we setup with legacy mode in the following cases:
836 		 * 1. When cannot target interrupt to a specific vcpu.
837 		 * 2. Unsetting posted interrupt.
838 		 * 3. APIC virtialization is disabled for the vcpu.
839 		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
840 		 */
841 		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
842 		    kvm_vcpu_apicv_active(&svm->vcpu)) {
843 			struct amd_iommu_pi_data pi;
844 
845 			/* Try to enable guest_mode in IRTE */
846 			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
847 					    AVIC_HPA_MASK);
848 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
849 						     svm->vcpu.vcpu_id);
850 			pi.is_guest_mode = true;
851 			pi.vcpu_data = &vcpu_info;
852 			ret = irq_set_vcpu_affinity(host_irq, &pi);
853 
854 			/**
855 			 * Here, we successfully setting up vcpu affinity in
856 			 * IOMMU guest mode. Now, we need to store the posted
857 			 * interrupt information in a per-vcpu ir_list so that
858 			 * we can reference to them directly when we update vcpu
859 			 * scheduling information in IOMMU irte.
860 			 */
861 			if (!ret && pi.is_guest_mode)
862 				svm_ir_list_add(svm, &pi);
863 		} else {
864 			/* Use legacy mode in IRTE */
865 			struct amd_iommu_pi_data pi;
866 
867 			/**
868 			 * Here, pi is used to:
869 			 * - Tell IOMMU to use legacy mode for this interrupt.
870 			 * - Retrieve ga_tag of prior interrupt remapping data.
871 			 */
872 			pi.prev_ga_tag = 0;
873 			pi.is_guest_mode = false;
874 			ret = irq_set_vcpu_affinity(host_irq, &pi);
875 
876 			/**
877 			 * Check if the posted interrupt was previously
878 			 * setup with the guest_mode by checking if the ga_tag
879 			 * was cached. If so, we need to clean up the per-vcpu
880 			 * ir_list.
881 			 */
882 			if (!ret && pi.prev_ga_tag) {
883 				int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
884 				struct kvm_vcpu *vcpu;
885 
886 				vcpu = kvm_get_vcpu_by_id(kvm, id);
887 				if (vcpu)
888 					svm_ir_list_del(to_svm(vcpu), &pi);
889 			}
890 		}
891 
892 		if (!ret && svm) {
893 			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
894 						 e->gsi, vcpu_info.vector,
895 						 vcpu_info.pi_desc_addr, set);
896 		}
897 
898 		if (ret < 0) {
899 			pr_err("%s: failed to update PI IRTE\n", __func__);
900 			goto out;
901 		}
902 	}
903 
904 	ret = 0;
905 out:
906 	srcu_read_unlock(&kvm->irq_srcu, idx);
907 	return ret;
908 }
909 
910 bool svm_check_apicv_inhibit_reasons(ulong bit)
911 {
912 	ulong supported = BIT(APICV_INHIBIT_REASON_DISABLE) |
913 			  BIT(APICV_INHIBIT_REASON_HYPERV) |
914 			  BIT(APICV_INHIBIT_REASON_NESTED) |
915 			  BIT(APICV_INHIBIT_REASON_IRQWIN) |
916 			  BIT(APICV_INHIBIT_REASON_PIT_REINJ) |
917 			  BIT(APICV_INHIBIT_REASON_X2APIC);
918 
919 	return supported & BIT(bit);
920 }
921 
922 void svm_pre_update_apicv_exec_ctrl(struct kvm *kvm, bool activate)
923 {
924 	avic_update_access_page(kvm, activate);
925 }
926 
927 static inline int
928 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
929 {
930 	int ret = 0;
931 	unsigned long flags;
932 	struct amd_svm_iommu_ir *ir;
933 	struct vcpu_svm *svm = to_svm(vcpu);
934 
935 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
936 		return 0;
937 
938 	/*
939 	 * Here, we go through the per-vcpu ir_list to update all existing
940 	 * interrupt remapping table entry targeting this vcpu.
941 	 */
942 	spin_lock_irqsave(&svm->ir_list_lock, flags);
943 
944 	if (list_empty(&svm->ir_list))
945 		goto out;
946 
947 	list_for_each_entry(ir, &svm->ir_list, node) {
948 		ret = amd_iommu_update_ga(cpu, r, ir->data);
949 		if (ret)
950 			break;
951 	}
952 out:
953 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
954 	return ret;
955 }
956 
957 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
958 {
959 	u64 entry;
960 	/* ID = 0xff (broadcast), ID > 0xff (reserved) */
961 	int h_physical_id = kvm_cpu_get_apicid(cpu);
962 	struct vcpu_svm *svm = to_svm(vcpu);
963 
964 	if (!kvm_vcpu_apicv_active(vcpu))
965 		return;
966 
967 	/*
968 	 * Since the host physical APIC id is 8 bits,
969 	 * we can support host APIC ID upto 255.
970 	 */
971 	if (WARN_ON(h_physical_id > AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
972 		return;
973 
974 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
975 	WARN_ON(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
976 
977 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
978 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
979 
980 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
981 	if (svm->avic_is_running)
982 		entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
983 
984 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
985 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id,
986 					svm->avic_is_running);
987 }
988 
989 void avic_vcpu_put(struct kvm_vcpu *vcpu)
990 {
991 	u64 entry;
992 	struct vcpu_svm *svm = to_svm(vcpu);
993 
994 	if (!kvm_vcpu_apicv_active(vcpu))
995 		return;
996 
997 	entry = READ_ONCE(*(svm->avic_physical_id_cache));
998 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
999 		avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
1000 
1001 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
1002 	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
1003 }
1004 
1005 /**
1006  * This function is called during VCPU halt/unhalt.
1007  */
1008 static void avic_set_running(struct kvm_vcpu *vcpu, bool is_run)
1009 {
1010 	struct vcpu_svm *svm = to_svm(vcpu);
1011 
1012 	svm->avic_is_running = is_run;
1013 	if (is_run)
1014 		avic_vcpu_load(vcpu, vcpu->cpu);
1015 	else
1016 		avic_vcpu_put(vcpu);
1017 }
1018 
1019 void svm_vcpu_blocking(struct kvm_vcpu *vcpu)
1020 {
1021 	avic_set_running(vcpu, false);
1022 }
1023 
1024 void svm_vcpu_unblocking(struct kvm_vcpu *vcpu)
1025 {
1026 	if (kvm_check_request(KVM_REQ_APICV_UPDATE, vcpu))
1027 		kvm_vcpu_update_apicv(vcpu);
1028 	avic_set_running(vcpu, true);
1029 }
1030