1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Support of MSI, HPET and DMAR interrupts.
4  *
5  * Copyright (C) 1997, 1998, 1999, 2000, 2009 Ingo Molnar, Hajnalka Szabo
6  *	Moved from arch/x86/kernel/apic/io_apic.c.
7  * Jiang Liu <jiang.liu@linux.intel.com>
8  *	Convert to hierarchical irqdomain
9  */
10 #include <linux/mm.h>
11 #include <linux/interrupt.h>
12 #include <linux/irq.h>
13 #include <linux/pci.h>
14 #include <linux/dmar.h>
15 #include <linux/hpet.h>
16 #include <linux/msi.h>
17 #include <asm/irqdomain.h>
18 #include <asm/hpet.h>
19 #include <asm/hw_irq.h>
20 #include <asm/apic.h>
21 #include <asm/irq_remapping.h>
22 #include <asm/xen/hypervisor.h>
23 
24 struct irq_domain *x86_pci_msi_default_domain __ro_after_init;
25 
irq_msi_update_msg(struct irq_data * irqd,struct irq_cfg * cfg)26 static void irq_msi_update_msg(struct irq_data *irqd, struct irq_cfg *cfg)
27 {
28 	struct msi_msg msg[2] = { [1] = { }, };
29 
30 	__irq_msi_compose_msg(cfg, msg, false);
31 	irq_data_get_irq_chip(irqd)->irq_write_msi_msg(irqd, msg);
32 }
33 
34 static int
msi_set_affinity(struct irq_data * irqd,const struct cpumask * mask,bool force)35 msi_set_affinity(struct irq_data *irqd, const struct cpumask *mask, bool force)
36 {
37 	struct irq_cfg old_cfg, *cfg = irqd_cfg(irqd);
38 	struct irq_data *parent = irqd->parent_data;
39 	unsigned int cpu;
40 	int ret;
41 
42 	/* Save the current configuration */
43 	cpu = cpumask_first(irq_data_get_effective_affinity_mask(irqd));
44 	old_cfg = *cfg;
45 
46 	/* Allocate a new target vector */
47 	ret = parent->chip->irq_set_affinity(parent, mask, force);
48 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
49 		return ret;
50 
51 	/*
52 	 * For non-maskable and non-remapped MSI interrupts the migration
53 	 * to a different destination CPU and a different vector has to be
54 	 * done careful to handle the possible stray interrupt which can be
55 	 * caused by the non-atomic update of the address/data pair.
56 	 *
57 	 * Direct update is possible when:
58 	 * - The MSI is maskable (remapped MSI does not use this code path)).
59 	 *   The quirk bit is not set in this case.
60 	 * - The new vector is the same as the old vector
61 	 * - The old vector is MANAGED_IRQ_SHUTDOWN_VECTOR (interrupt starts up)
62 	 * - The interrupt is not yet started up
63 	 * - The new destination CPU is the same as the old destination CPU
64 	 */
65 	if (!irqd_msi_nomask_quirk(irqd) ||
66 	    cfg->vector == old_cfg.vector ||
67 	    old_cfg.vector == MANAGED_IRQ_SHUTDOWN_VECTOR ||
68 	    !irqd_is_started(irqd) ||
69 	    cfg->dest_apicid == old_cfg.dest_apicid) {
70 		irq_msi_update_msg(irqd, cfg);
71 		return ret;
72 	}
73 
74 	/*
75 	 * Paranoia: Validate that the interrupt target is the local
76 	 * CPU.
77 	 */
78 	if (WARN_ON_ONCE(cpu != smp_processor_id())) {
79 		irq_msi_update_msg(irqd, cfg);
80 		return ret;
81 	}
82 
83 	/*
84 	 * Redirect the interrupt to the new vector on the current CPU
85 	 * first. This might cause a spurious interrupt on this vector if
86 	 * the device raises an interrupt right between this update and the
87 	 * update to the final destination CPU.
88 	 *
89 	 * If the vector is in use then the installed device handler will
90 	 * denote it as spurious which is no harm as this is a rare event
91 	 * and interrupt handlers have to cope with spurious interrupts
92 	 * anyway. If the vector is unused, then it is marked so it won't
93 	 * trigger the 'No irq handler for vector' warning in
94 	 * common_interrupt().
95 	 *
96 	 * This requires to hold vector lock to prevent concurrent updates to
97 	 * the affected vector.
98 	 */
99 	lock_vector_lock();
100 
101 	/*
102 	 * Mark the new target vector on the local CPU if it is currently
103 	 * unused. Reuse the VECTOR_RETRIGGERED state which is also used in
104 	 * the CPU hotplug path for a similar purpose. This cannot be
105 	 * undone here as the current CPU has interrupts disabled and
106 	 * cannot handle the interrupt before the whole set_affinity()
107 	 * section is done. In the CPU unplug case, the current CPU is
108 	 * about to vanish and will not handle any interrupts anymore. The
109 	 * vector is cleaned up when the CPU comes online again.
110 	 */
111 	if (IS_ERR_OR_NULL(this_cpu_read(vector_irq[cfg->vector])))
112 		this_cpu_write(vector_irq[cfg->vector], VECTOR_RETRIGGERED);
113 
114 	/* Redirect it to the new vector on the local CPU temporarily */
115 	old_cfg.vector = cfg->vector;
116 	irq_msi_update_msg(irqd, &old_cfg);
117 
118 	/* Now transition it to the target CPU */
119 	irq_msi_update_msg(irqd, cfg);
120 
121 	/*
122 	 * All interrupts after this point are now targeted at the new
123 	 * vector/CPU.
124 	 *
125 	 * Drop vector lock before testing whether the temporary assignment
126 	 * to the local CPU was hit by an interrupt raised in the device,
127 	 * because the retrigger function acquires vector lock again.
128 	 */
129 	unlock_vector_lock();
130 
131 	/*
132 	 * Check whether the transition raced with a device interrupt and
133 	 * is pending in the local APICs IRR. It is safe to do this outside
134 	 * of vector lock as the irq_desc::lock of this interrupt is still
135 	 * held and interrupts are disabled: The check is not accessing the
136 	 * underlying vector store. It's just checking the local APIC's
137 	 * IRR.
138 	 */
139 	if (lapic_vector_set_in_irr(cfg->vector))
140 		irq_data_get_irq_chip(irqd)->irq_retrigger(irqd);
141 
142 	return ret;
143 }
144 
145 /*
146  * IRQ Chip for MSI PCI/PCI-X/PCI-Express Devices,
147  * which implement the MSI or MSI-X Capability Structure.
148  */
149 static struct irq_chip pci_msi_controller = {
150 	.name			= "PCI-MSI",
151 	.irq_unmask		= pci_msi_unmask_irq,
152 	.irq_mask		= pci_msi_mask_irq,
153 	.irq_ack		= irq_chip_ack_parent,
154 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
155 	.irq_set_affinity	= msi_set_affinity,
156 	.flags			= IRQCHIP_SKIP_SET_WAKE |
157 				  IRQCHIP_AFFINITY_PRE_STARTUP,
158 };
159 
pci_msi_prepare(struct irq_domain * domain,struct device * dev,int nvec,msi_alloc_info_t * arg)160 int pci_msi_prepare(struct irq_domain *domain, struct device *dev, int nvec,
161 		    msi_alloc_info_t *arg)
162 {
163 	init_irq_alloc_info(arg, NULL);
164 	if (to_pci_dev(dev)->msix_enabled) {
165 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSIX;
166 	} else {
167 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSI;
168 		arg->flags |= X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
169 	}
170 
171 	return 0;
172 }
173 EXPORT_SYMBOL_GPL(pci_msi_prepare);
174 
175 static struct msi_domain_ops pci_msi_domain_ops = {
176 	.msi_prepare	= pci_msi_prepare,
177 };
178 
179 static struct msi_domain_info pci_msi_domain_info = {
180 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
181 			  MSI_FLAG_PCI_MSIX,
182 	.ops		= &pci_msi_domain_ops,
183 	.chip		= &pci_msi_controller,
184 	.handler	= handle_edge_irq,
185 	.handler_name	= "edge",
186 };
187 
native_create_pci_msi_domain(void)188 struct irq_domain * __init native_create_pci_msi_domain(void)
189 {
190 	struct fwnode_handle *fn;
191 	struct irq_domain *d;
192 
193 	if (disable_apic)
194 		return NULL;
195 
196 	fn = irq_domain_alloc_named_fwnode("PCI-MSI");
197 	if (!fn)
198 		return NULL;
199 
200 	d = pci_msi_create_irq_domain(fn, &pci_msi_domain_info,
201 				      x86_vector_domain);
202 	if (!d) {
203 		irq_domain_free_fwnode(fn);
204 		pr_warn("Failed to initialize PCI-MSI irqdomain.\n");
205 	} else {
206 		d->flags |= IRQ_DOMAIN_MSI_NOMASK_QUIRK;
207 	}
208 	return d;
209 }
210 
x86_create_pci_msi_domain(void)211 void __init x86_create_pci_msi_domain(void)
212 {
213 	x86_pci_msi_default_domain = x86_init.irqs.create_pci_msi_domain();
214 }
215 
216 #ifdef CONFIG_IRQ_REMAP
217 static struct irq_chip pci_msi_ir_controller = {
218 	.name			= "IR-PCI-MSI",
219 	.irq_unmask		= pci_msi_unmask_irq,
220 	.irq_mask		= pci_msi_mask_irq,
221 	.irq_ack		= irq_chip_ack_parent,
222 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
223 	.flags			= IRQCHIP_SKIP_SET_WAKE |
224 				  IRQCHIP_AFFINITY_PRE_STARTUP,
225 };
226 
227 static struct msi_domain_info pci_msi_ir_domain_info = {
228 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
229 			  MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX,
230 	.ops		= &pci_msi_domain_ops,
231 	.chip		= &pci_msi_ir_controller,
232 	.handler	= handle_edge_irq,
233 	.handler_name	= "edge",
234 };
235 
arch_create_remap_msi_irq_domain(struct irq_domain * parent,const char * name,int id)236 struct irq_domain *arch_create_remap_msi_irq_domain(struct irq_domain *parent,
237 						    const char *name, int id)
238 {
239 	struct fwnode_handle *fn;
240 	struct irq_domain *d;
241 
242 	fn = irq_domain_alloc_named_id_fwnode(name, id);
243 	if (!fn)
244 		return NULL;
245 	d = pci_msi_create_irq_domain(fn, &pci_msi_ir_domain_info, parent);
246 	if (!d)
247 		irq_domain_free_fwnode(fn);
248 	return d;
249 }
250 #endif
251 
252 #ifdef CONFIG_DMAR_TABLE
253 /*
254  * The Intel IOMMU (ab)uses the high bits of the MSI address to contain the
255  * high bits of the destination APIC ID. This can't be done in the general
256  * case for MSIs as it would be targeting real memory above 4GiB not the
257  * APIC.
258  */
dmar_msi_compose_msg(struct irq_data * data,struct msi_msg * msg)259 static void dmar_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
260 {
261 	__irq_msi_compose_msg(irqd_cfg(data), msg, true);
262 }
263 
dmar_msi_write_msg(struct irq_data * data,struct msi_msg * msg)264 static void dmar_msi_write_msg(struct irq_data *data, struct msi_msg *msg)
265 {
266 	dmar_msi_write(data->irq, msg);
267 }
268 
269 static struct irq_chip dmar_msi_controller = {
270 	.name			= "DMAR-MSI",
271 	.irq_unmask		= dmar_msi_unmask,
272 	.irq_mask		= dmar_msi_mask,
273 	.irq_ack		= irq_chip_ack_parent,
274 	.irq_set_affinity	= msi_domain_set_affinity,
275 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
276 	.irq_compose_msi_msg	= dmar_msi_compose_msg,
277 	.irq_write_msi_msg	= dmar_msi_write_msg,
278 	.flags			= IRQCHIP_SKIP_SET_WAKE |
279 				  IRQCHIP_AFFINITY_PRE_STARTUP,
280 };
281 
dmar_msi_init(struct irq_domain * domain,struct msi_domain_info * info,unsigned int virq,irq_hw_number_t hwirq,msi_alloc_info_t * arg)282 static int dmar_msi_init(struct irq_domain *domain,
283 			 struct msi_domain_info *info, unsigned int virq,
284 			 irq_hw_number_t hwirq, msi_alloc_info_t *arg)
285 {
286 	irq_domain_set_info(domain, virq, arg->devid, info->chip, NULL,
287 			    handle_edge_irq, arg->data, "edge");
288 
289 	return 0;
290 }
291 
292 static struct msi_domain_ops dmar_msi_domain_ops = {
293 	.msi_init	= dmar_msi_init,
294 };
295 
296 static struct msi_domain_info dmar_msi_domain_info = {
297 	.ops		= &dmar_msi_domain_ops,
298 	.chip		= &dmar_msi_controller,
299 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS,
300 };
301 
dmar_get_irq_domain(void)302 static struct irq_domain *dmar_get_irq_domain(void)
303 {
304 	static struct irq_domain *dmar_domain;
305 	static DEFINE_MUTEX(dmar_lock);
306 	struct fwnode_handle *fn;
307 
308 	mutex_lock(&dmar_lock);
309 	if (dmar_domain)
310 		goto out;
311 
312 	fn = irq_domain_alloc_named_fwnode("DMAR-MSI");
313 	if (fn) {
314 		dmar_domain = msi_create_irq_domain(fn, &dmar_msi_domain_info,
315 						    x86_vector_domain);
316 		if (!dmar_domain)
317 			irq_domain_free_fwnode(fn);
318 	}
319 out:
320 	mutex_unlock(&dmar_lock);
321 	return dmar_domain;
322 }
323 
dmar_alloc_hwirq(int id,int node,void * arg)324 int dmar_alloc_hwirq(int id, int node, void *arg)
325 {
326 	struct irq_domain *domain = dmar_get_irq_domain();
327 	struct irq_alloc_info info;
328 
329 	if (!domain)
330 		return -1;
331 
332 	init_irq_alloc_info(&info, NULL);
333 	info.type = X86_IRQ_ALLOC_TYPE_DMAR;
334 	info.devid = id;
335 	info.hwirq = id;
336 	info.data = arg;
337 
338 	return irq_domain_alloc_irqs(domain, 1, node, &info);
339 }
340 
dmar_free_hwirq(int irq)341 void dmar_free_hwirq(int irq)
342 {
343 	irq_domain_free_irqs(irq, 1);
344 }
345 #endif
346 
arch_restore_msi_irqs(struct pci_dev * dev)347 bool arch_restore_msi_irqs(struct pci_dev *dev)
348 {
349 	return xen_initdom_restore_msi(dev);
350 }
351