xref: /linux/drivers/iommu/hyperv-iommu.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /*
4  * Hyper-V stub IOMMU driver.
5  *
6  * Copyright (C) 2019, Microsoft, Inc.
7  *
8  * Author : Lan Tianyu <Tianyu.Lan@microsoft.com>
9  */
10 
11 #include <linux/types.h>
12 #include <linux/interrupt.h>
13 #include <linux/irq.h>
14 #include <linux/iommu.h>
15 #include <linux/module.h>
16 
17 #include <asm/apic.h>
18 #include <asm/cpu.h>
19 #include <asm/hw_irq.h>
20 #include <asm/io_apic.h>
21 #include <asm/irq_remapping.h>
22 #include <asm/hypervisor.h>
23 #include <asm/mshyperv.h>
24 
25 #include "irq_remapping.h"
26 
27 #ifdef CONFIG_IRQ_REMAP
28 
29 /*
30  * According 82093AA IO-APIC spec , IO APIC has a 24-entry Interrupt
31  * Redirection Table. Hyper-V exposes one single IO-APIC and so define
32  * 24 IO APIC remmapping entries.
33  */
34 #define IOAPIC_REMAPPING_ENTRY 24
35 
36 static cpumask_t ioapic_max_cpumask = { CPU_BITS_NONE };
37 static struct irq_domain *ioapic_ir_domain;
38 
39 static int hyperv_ir_set_affinity(struct irq_data *data,
40 		const struct cpumask *mask, bool force)
41 {
42 	struct irq_data *parent = data->parent_data;
43 	struct irq_cfg *cfg = irqd_cfg(data);
44 	int ret;
45 
46 	/* Return error If new irq affinity is out of ioapic_max_cpumask. */
47 	if (!cpumask_subset(mask, &ioapic_max_cpumask))
48 		return -EINVAL;
49 
50 	ret = parent->chip->irq_set_affinity(parent, mask, force);
51 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
52 		return ret;
53 
54 	vector_schedule_cleanup(cfg);
55 
56 	return 0;
57 }
58 
59 static struct irq_chip hyperv_ir_chip = {
60 	.name			= "HYPERV-IR",
61 	.irq_ack		= apic_ack_irq,
62 	.irq_set_affinity	= hyperv_ir_set_affinity,
63 };
64 
65 static int hyperv_irq_remapping_alloc(struct irq_domain *domain,
66 				     unsigned int virq, unsigned int nr_irqs,
67 				     void *arg)
68 {
69 	struct irq_alloc_info *info = arg;
70 	struct irq_data *irq_data;
71 	int ret = 0;
72 
73 	if (!info || info->type != X86_IRQ_ALLOC_TYPE_IOAPIC || nr_irqs > 1)
74 		return -EINVAL;
75 
76 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
77 	if (ret < 0)
78 		return ret;
79 
80 	irq_data = irq_domain_get_irq_data(domain, virq);
81 	if (!irq_data) {
82 		irq_domain_free_irqs_common(domain, virq, nr_irqs);
83 		return -EINVAL;
84 	}
85 
86 	irq_data->chip = &hyperv_ir_chip;
87 
88 	/*
89 	 * Hypver-V IO APIC irq affinity should be in the scope of
90 	 * ioapic_max_cpumask because no irq remapping support.
91 	 */
92 	irq_data_update_affinity(irq_data, &ioapic_max_cpumask);
93 
94 	return 0;
95 }
96 
97 static void hyperv_irq_remapping_free(struct irq_domain *domain,
98 				 unsigned int virq, unsigned int nr_irqs)
99 {
100 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
101 }
102 
103 static int hyperv_irq_remapping_select(struct irq_domain *d,
104 				       struct irq_fwspec *fwspec,
105 				       enum irq_domain_bus_token bus_token)
106 {
107 	/* Claim the only I/O APIC emulated by Hyper-V */
108 	return x86_fwspec_is_ioapic(fwspec);
109 }
110 
111 static const struct irq_domain_ops hyperv_ir_domain_ops = {
112 	.select = hyperv_irq_remapping_select,
113 	.alloc = hyperv_irq_remapping_alloc,
114 	.free = hyperv_irq_remapping_free,
115 };
116 
117 static const struct irq_domain_ops hyperv_root_ir_domain_ops;
118 static int __init hyperv_prepare_irq_remapping(void)
119 {
120 	struct fwnode_handle *fn;
121 	int i;
122 	const char *name;
123 	const struct irq_domain_ops *ops;
124 
125 	/*
126 	 * For a Hyper-V root partition, ms_hyperv_msi_ext_dest_id()
127 	 * will always return false.
128 	 */
129 	if (!hypervisor_is_type(X86_HYPER_MS_HYPERV) ||
130 	    x86_init.hyper.msi_ext_dest_id())
131 		return -ENODEV;
132 
133 	if (hv_root_partition) {
134 		name = "HYPERV-ROOT-IR";
135 		ops = &hyperv_root_ir_domain_ops;
136 	} else {
137 		name = "HYPERV-IR";
138 		ops = &hyperv_ir_domain_ops;
139 	}
140 
141 	fn = irq_domain_alloc_named_id_fwnode(name, 0);
142 	if (!fn)
143 		return -ENOMEM;
144 
145 	ioapic_ir_domain =
146 		irq_domain_create_hierarchy(arch_get_ir_parent_domain(),
147 				0, IOAPIC_REMAPPING_ENTRY, fn, ops, NULL);
148 
149 	if (!ioapic_ir_domain) {
150 		irq_domain_free_fwnode(fn);
151 		return -ENOMEM;
152 	}
153 
154 	if (hv_root_partition)
155 		return 0; /* The rest is only relevant to guests */
156 
157 	/*
158 	 * Hyper-V doesn't provide irq remapping function for
159 	 * IO-APIC and so IO-APIC only accepts 8-bit APIC ID.
160 	 * Cpu's APIC ID is read from ACPI MADT table and APIC IDs
161 	 * in the MADT table on Hyper-v are sorted monotonic increasingly.
162 	 * APIC ID reflects cpu topology. There maybe some APIC ID
163 	 * gaps when cpu number in a socket is not power of two. Prepare
164 	 * max cpu affinity for IOAPIC irqs. Scan cpu 0-255 and set cpu
165 	 * into ioapic_max_cpumask if its APIC ID is less than 256.
166 	 */
167 	for (i = min_t(unsigned int, num_possible_cpus() - 1, 255); i >= 0; i--)
168 		if (cpu_physical_id(i) < 256)
169 			cpumask_set_cpu(i, &ioapic_max_cpumask);
170 
171 	return 0;
172 }
173 
174 static int __init hyperv_enable_irq_remapping(void)
175 {
176 	if (x2apic_supported())
177 		return IRQ_REMAP_X2APIC_MODE;
178 	return IRQ_REMAP_XAPIC_MODE;
179 }
180 
181 struct irq_remap_ops hyperv_irq_remap_ops = {
182 	.prepare		= hyperv_prepare_irq_remapping,
183 	.enable			= hyperv_enable_irq_remapping,
184 };
185 
186 /* IRQ remapping domain when Linux runs as the root partition */
187 struct hyperv_root_ir_data {
188 	u8 ioapic_id;
189 	bool is_level;
190 	struct hv_interrupt_entry entry;
191 };
192 
193 static void
194 hyperv_root_ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
195 {
196 	u64 status;
197 	u32 vector;
198 	struct irq_cfg *cfg;
199 	int ioapic_id;
200 	const struct cpumask *affinity;
201 	int cpu;
202 	struct hv_interrupt_entry entry;
203 	struct hyperv_root_ir_data *data = irq_data->chip_data;
204 	struct IO_APIC_route_entry e;
205 
206 	cfg = irqd_cfg(irq_data);
207 	affinity = irq_data_get_effective_affinity_mask(irq_data);
208 	cpu = cpumask_first_and(affinity, cpu_online_mask);
209 
210 	vector = cfg->vector;
211 	ioapic_id = data->ioapic_id;
212 
213 	if (data->entry.source == HV_DEVICE_TYPE_IOAPIC
214 	    && data->entry.ioapic_rte.as_uint64) {
215 		entry = data->entry;
216 
217 		status = hv_unmap_ioapic_interrupt(ioapic_id, &entry);
218 
219 		if (status != HV_STATUS_SUCCESS)
220 			pr_debug("%s: unexpected unmap status %lld\n", __func__, status);
221 
222 		data->entry.ioapic_rte.as_uint64 = 0;
223 		data->entry.source = 0; /* Invalid source */
224 	}
225 
226 
227 	status = hv_map_ioapic_interrupt(ioapic_id, data->is_level, cpu,
228 					vector, &entry);
229 
230 	if (status != HV_STATUS_SUCCESS) {
231 		pr_err("%s: map hypercall failed, status %lld\n", __func__, status);
232 		return;
233 	}
234 
235 	data->entry = entry;
236 
237 	/* Turn it into an IO_APIC_route_entry, and generate MSI MSG. */
238 	e.w1 = entry.ioapic_rte.low_uint32;
239 	e.w2 = entry.ioapic_rte.high_uint32;
240 
241 	memset(msg, 0, sizeof(*msg));
242 	msg->arch_data.vector = e.vector;
243 	msg->arch_data.delivery_mode = e.delivery_mode;
244 	msg->arch_addr_lo.dest_mode_logical = e.dest_mode_logical;
245 	msg->arch_addr_lo.dmar_format = e.ir_format;
246 	msg->arch_addr_lo.dmar_index_0_14 = e.ir_index_0_14;
247 }
248 
249 static int hyperv_root_ir_set_affinity(struct irq_data *data,
250 		const struct cpumask *mask, bool force)
251 {
252 	struct irq_data *parent = data->parent_data;
253 	struct irq_cfg *cfg = irqd_cfg(data);
254 	int ret;
255 
256 	ret = parent->chip->irq_set_affinity(parent, mask, force);
257 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
258 		return ret;
259 
260 	vector_schedule_cleanup(cfg);
261 
262 	return 0;
263 }
264 
265 static struct irq_chip hyperv_root_ir_chip = {
266 	.name			= "HYPERV-ROOT-IR",
267 	.irq_ack		= apic_ack_irq,
268 	.irq_set_affinity	= hyperv_root_ir_set_affinity,
269 	.irq_compose_msi_msg	= hyperv_root_ir_compose_msi_msg,
270 };
271 
272 static int hyperv_root_irq_remapping_alloc(struct irq_domain *domain,
273 				     unsigned int virq, unsigned int nr_irqs,
274 				     void *arg)
275 {
276 	struct irq_alloc_info *info = arg;
277 	struct irq_data *irq_data;
278 	struct hyperv_root_ir_data *data;
279 	int ret = 0;
280 
281 	if (!info || info->type != X86_IRQ_ALLOC_TYPE_IOAPIC || nr_irqs > 1)
282 		return -EINVAL;
283 
284 	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
285 	if (ret < 0)
286 		return ret;
287 
288 	data = kzalloc(sizeof(*data), GFP_KERNEL);
289 	if (!data) {
290 		irq_domain_free_irqs_common(domain, virq, nr_irqs);
291 		return -ENOMEM;
292 	}
293 
294 	irq_data = irq_domain_get_irq_data(domain, virq);
295 	if (!irq_data) {
296 		kfree(data);
297 		irq_domain_free_irqs_common(domain, virq, nr_irqs);
298 		return -EINVAL;
299 	}
300 
301 	data->ioapic_id = info->devid;
302 	data->is_level = info->ioapic.is_level;
303 
304 	irq_data->chip = &hyperv_root_ir_chip;
305 	irq_data->chip_data = data;
306 
307 	return 0;
308 }
309 
310 static void hyperv_root_irq_remapping_free(struct irq_domain *domain,
311 				 unsigned int virq, unsigned int nr_irqs)
312 {
313 	struct irq_data *irq_data;
314 	struct hyperv_root_ir_data *data;
315 	struct hv_interrupt_entry *e;
316 	int i;
317 
318 	for (i = 0; i < nr_irqs; i++) {
319 		irq_data = irq_domain_get_irq_data(domain, virq + i);
320 
321 		if (irq_data && irq_data->chip_data) {
322 			data = irq_data->chip_data;
323 			e = &data->entry;
324 
325 			if (e->source == HV_DEVICE_TYPE_IOAPIC
326 			      && e->ioapic_rte.as_uint64)
327 				hv_unmap_ioapic_interrupt(data->ioapic_id,
328 							&data->entry);
329 
330 			kfree(data);
331 		}
332 	}
333 
334 	irq_domain_free_irqs_common(domain, virq, nr_irqs);
335 }
336 
337 static const struct irq_domain_ops hyperv_root_ir_domain_ops = {
338 	.select = hyperv_irq_remapping_select,
339 	.alloc = hyperv_root_irq_remapping_alloc,
340 	.free = hyperv_root_irq_remapping_free,
341 };
342 
343 #endif
344