xref: /openbmc/linux/arch/x86/kernel/apic/msi.c (revision 887069f4)
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 
23 struct irq_domain *x86_pci_msi_default_domain __ro_after_init;
24 
25 static void irq_msi_update_msg(struct irq_data *irqd, struct irq_cfg *cfg)
26 {
27 	struct msi_msg msg[2] = { [1] = { }, };
28 
29 	__irq_msi_compose_msg(cfg, msg, false);
30 	irq_data_get_irq_chip(irqd)->irq_write_msi_msg(irqd, msg);
31 }
32 
33 static int
34 msi_set_affinity(struct irq_data *irqd, const struct cpumask *mask, bool force)
35 {
36 	struct irq_cfg old_cfg, *cfg = irqd_cfg(irqd);
37 	struct irq_data *parent = irqd->parent_data;
38 	unsigned int cpu;
39 	int ret;
40 
41 	/* Save the current configuration */
42 	cpu = cpumask_first(irq_data_get_effective_affinity_mask(irqd));
43 	old_cfg = *cfg;
44 
45 	/* Allocate a new target vector */
46 	ret = parent->chip->irq_set_affinity(parent, mask, force);
47 	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
48 		return ret;
49 
50 	/*
51 	 * For non-maskable and non-remapped MSI interrupts the migration
52 	 * to a different destination CPU and a different vector has to be
53 	 * done careful to handle the possible stray interrupt which can be
54 	 * caused by the non-atomic update of the address/data pair.
55 	 *
56 	 * Direct update is possible when:
57 	 * - The MSI is maskable (remapped MSI does not use this code path)).
58 	 *   The quirk bit is not set in this case.
59 	 * - The new vector is the same as the old vector
60 	 * - The old vector is MANAGED_IRQ_SHUTDOWN_VECTOR (interrupt starts up)
61 	 * - The interrupt is not yet started up
62 	 * - The new destination CPU is the same as the old destination CPU
63 	 */
64 	if (!irqd_msi_nomask_quirk(irqd) ||
65 	    cfg->vector == old_cfg.vector ||
66 	    old_cfg.vector == MANAGED_IRQ_SHUTDOWN_VECTOR ||
67 	    !irqd_is_started(irqd) ||
68 	    cfg->dest_apicid == old_cfg.dest_apicid) {
69 		irq_msi_update_msg(irqd, cfg);
70 		return ret;
71 	}
72 
73 	/*
74 	 * Paranoia: Validate that the interrupt target is the local
75 	 * CPU.
76 	 */
77 	if (WARN_ON_ONCE(cpu != smp_processor_id())) {
78 		irq_msi_update_msg(irqd, cfg);
79 		return ret;
80 	}
81 
82 	/*
83 	 * Redirect the interrupt to the new vector on the current CPU
84 	 * first. This might cause a spurious interrupt on this vector if
85 	 * the device raises an interrupt right between this update and the
86 	 * update to the final destination CPU.
87 	 *
88 	 * If the vector is in use then the installed device handler will
89 	 * denote it as spurious which is no harm as this is a rare event
90 	 * and interrupt handlers have to cope with spurious interrupts
91 	 * anyway. If the vector is unused, then it is marked so it won't
92 	 * trigger the 'No irq handler for vector' warning in
93 	 * common_interrupt().
94 	 *
95 	 * This requires to hold vector lock to prevent concurrent updates to
96 	 * the affected vector.
97 	 */
98 	lock_vector_lock();
99 
100 	/*
101 	 * Mark the new target vector on the local CPU if it is currently
102 	 * unused. Reuse the VECTOR_RETRIGGERED state which is also used in
103 	 * the CPU hotplug path for a similar purpose. This cannot be
104 	 * undone here as the current CPU has interrupts disabled and
105 	 * cannot handle the interrupt before the whole set_affinity()
106 	 * section is done. In the CPU unplug case, the current CPU is
107 	 * about to vanish and will not handle any interrupts anymore. The
108 	 * vector is cleaned up when the CPU comes online again.
109 	 */
110 	if (IS_ERR_OR_NULL(this_cpu_read(vector_irq[cfg->vector])))
111 		this_cpu_write(vector_irq[cfg->vector], VECTOR_RETRIGGERED);
112 
113 	/* Redirect it to the new vector on the local CPU temporarily */
114 	old_cfg.vector = cfg->vector;
115 	irq_msi_update_msg(irqd, &old_cfg);
116 
117 	/* Now transition it to the target CPU */
118 	irq_msi_update_msg(irqd, cfg);
119 
120 	/*
121 	 * All interrupts after this point are now targeted at the new
122 	 * vector/CPU.
123 	 *
124 	 * Drop vector lock before testing whether the temporary assignment
125 	 * to the local CPU was hit by an interrupt raised in the device,
126 	 * because the retrigger function acquires vector lock again.
127 	 */
128 	unlock_vector_lock();
129 
130 	/*
131 	 * Check whether the transition raced with a device interrupt and
132 	 * is pending in the local APICs IRR. It is safe to do this outside
133 	 * of vector lock as the irq_desc::lock of this interrupt is still
134 	 * held and interrupts are disabled: The check is not accessing the
135 	 * underlying vector store. It's just checking the local APIC's
136 	 * IRR.
137 	 */
138 	if (lapic_vector_set_in_irr(cfg->vector))
139 		irq_data_get_irq_chip(irqd)->irq_retrigger(irqd);
140 
141 	return ret;
142 }
143 
144 /*
145  * IRQ Chip for MSI PCI/PCI-X/PCI-Express Devices,
146  * which implement the MSI or MSI-X Capability Structure.
147  */
148 static struct irq_chip pci_msi_controller = {
149 	.name			= "PCI-MSI",
150 	.irq_unmask		= pci_msi_unmask_irq,
151 	.irq_mask		= pci_msi_mask_irq,
152 	.irq_ack		= irq_chip_ack_parent,
153 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
154 	.irq_set_affinity	= msi_set_affinity,
155 	.flags			= IRQCHIP_SKIP_SET_WAKE |
156 				  IRQCHIP_AFFINITY_PRE_STARTUP,
157 };
158 
159 int pci_msi_prepare(struct irq_domain *domain, struct device *dev, int nvec,
160 		    msi_alloc_info_t *arg)
161 {
162 	struct pci_dev *pdev = to_pci_dev(dev);
163 	struct msi_desc *desc = first_pci_msi_entry(pdev);
164 
165 	init_irq_alloc_info(arg, NULL);
166 	if (desc->msi_attrib.is_msix) {
167 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSIX;
168 	} else {
169 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSI;
170 		arg->flags |= X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
171 	}
172 
173 	return 0;
174 }
175 EXPORT_SYMBOL_GPL(pci_msi_prepare);
176 
177 static struct msi_domain_ops pci_msi_domain_ops = {
178 	.msi_prepare	= pci_msi_prepare,
179 };
180 
181 static struct msi_domain_info pci_msi_domain_info = {
182 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
183 			  MSI_FLAG_PCI_MSIX,
184 	.ops		= &pci_msi_domain_ops,
185 	.chip		= &pci_msi_controller,
186 	.handler	= handle_edge_irq,
187 	.handler_name	= "edge",
188 };
189 
190 struct irq_domain * __init native_create_pci_msi_domain(void)
191 {
192 	struct fwnode_handle *fn;
193 	struct irq_domain *d;
194 
195 	if (disable_apic)
196 		return NULL;
197 
198 	fn = irq_domain_alloc_named_fwnode("PCI-MSI");
199 	if (!fn)
200 		return NULL;
201 
202 	d = pci_msi_create_irq_domain(fn, &pci_msi_domain_info,
203 				      x86_vector_domain);
204 	if (!d) {
205 		irq_domain_free_fwnode(fn);
206 		pr_warn("Failed to initialize PCI-MSI irqdomain.\n");
207 	} else {
208 		d->flags |= IRQ_DOMAIN_MSI_NOMASK_QUIRK;
209 	}
210 	return d;
211 }
212 
213 void __init x86_create_pci_msi_domain(void)
214 {
215 	x86_pci_msi_default_domain = x86_init.irqs.create_pci_msi_domain();
216 }
217 
218 #ifdef CONFIG_IRQ_REMAP
219 static struct irq_chip pci_msi_ir_controller = {
220 	.name			= "IR-PCI-MSI",
221 	.irq_unmask		= pci_msi_unmask_irq,
222 	.irq_mask		= pci_msi_mask_irq,
223 	.irq_ack		= irq_chip_ack_parent,
224 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
225 	.flags			= IRQCHIP_SKIP_SET_WAKE |
226 				  IRQCHIP_AFFINITY_PRE_STARTUP,
227 };
228 
229 static struct msi_domain_info pci_msi_ir_domain_info = {
230 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
231 			  MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX,
232 	.ops		= &pci_msi_domain_ops,
233 	.chip		= &pci_msi_ir_controller,
234 	.handler	= handle_edge_irq,
235 	.handler_name	= "edge",
236 };
237 
238 struct irq_domain *arch_create_remap_msi_irq_domain(struct irq_domain *parent,
239 						    const char *name, int id)
240 {
241 	struct fwnode_handle *fn;
242 	struct irq_domain *d;
243 
244 	fn = irq_domain_alloc_named_id_fwnode(name, id);
245 	if (!fn)
246 		return NULL;
247 	d = pci_msi_create_irq_domain(fn, &pci_msi_ir_domain_info, parent);
248 	if (!d)
249 		irq_domain_free_fwnode(fn);
250 	return d;
251 }
252 #endif
253 
254 #ifdef CONFIG_DMAR_TABLE
255 /*
256  * The Intel IOMMU (ab)uses the high bits of the MSI address to contain the
257  * high bits of the destination APIC ID. This can't be done in the general
258  * case for MSIs as it would be targeting real memory above 4GiB not the
259  * APIC.
260  */
261 static void dmar_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
262 {
263 	__irq_msi_compose_msg(irqd_cfg(data), msg, true);
264 }
265 
266 static void dmar_msi_write_msg(struct irq_data *data, struct msi_msg *msg)
267 {
268 	dmar_msi_write(data->irq, msg);
269 }
270 
271 static struct irq_chip dmar_msi_controller = {
272 	.name			= "DMAR-MSI",
273 	.irq_unmask		= dmar_msi_unmask,
274 	.irq_mask		= dmar_msi_mask,
275 	.irq_ack		= irq_chip_ack_parent,
276 	.irq_set_affinity	= msi_domain_set_affinity,
277 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
278 	.irq_compose_msi_msg	= dmar_msi_compose_msg,
279 	.irq_write_msi_msg	= dmar_msi_write_msg,
280 	.flags			= IRQCHIP_SKIP_SET_WAKE |
281 				  IRQCHIP_AFFINITY_PRE_STARTUP,
282 };
283 
284 static int dmar_msi_init(struct irq_domain *domain,
285 			 struct msi_domain_info *info, unsigned int virq,
286 			 irq_hw_number_t hwirq, msi_alloc_info_t *arg)
287 {
288 	irq_domain_set_info(domain, virq, arg->devid, info->chip, NULL,
289 			    handle_edge_irq, arg->data, "edge");
290 
291 	return 0;
292 }
293 
294 static struct msi_domain_ops dmar_msi_domain_ops = {
295 	.msi_init	= dmar_msi_init,
296 };
297 
298 static struct msi_domain_info dmar_msi_domain_info = {
299 	.ops		= &dmar_msi_domain_ops,
300 	.chip		= &dmar_msi_controller,
301 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS,
302 };
303 
304 static struct irq_domain *dmar_get_irq_domain(void)
305 {
306 	static struct irq_domain *dmar_domain;
307 	static DEFINE_MUTEX(dmar_lock);
308 	struct fwnode_handle *fn;
309 
310 	mutex_lock(&dmar_lock);
311 	if (dmar_domain)
312 		goto out;
313 
314 	fn = irq_domain_alloc_named_fwnode("DMAR-MSI");
315 	if (fn) {
316 		dmar_domain = msi_create_irq_domain(fn, &dmar_msi_domain_info,
317 						    x86_vector_domain);
318 		if (!dmar_domain)
319 			irq_domain_free_fwnode(fn);
320 	}
321 out:
322 	mutex_unlock(&dmar_lock);
323 	return dmar_domain;
324 }
325 
326 int dmar_alloc_hwirq(int id, int node, void *arg)
327 {
328 	struct irq_domain *domain = dmar_get_irq_domain();
329 	struct irq_alloc_info info;
330 
331 	if (!domain)
332 		return -1;
333 
334 	init_irq_alloc_info(&info, NULL);
335 	info.type = X86_IRQ_ALLOC_TYPE_DMAR;
336 	info.devid = id;
337 	info.hwirq = id;
338 	info.data = arg;
339 
340 	return irq_domain_alloc_irqs(domain, 1, node, &info);
341 }
342 
343 void dmar_free_hwirq(int irq)
344 {
345 	irq_domain_free_irqs(irq, 1);
346 }
347 #endif
348