xref: /openbmc/linux/arch/x86/kernel/apic/msi.c (revision b737eecd)
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 new destination CPU is the same as the old destination CPU
62 	 */
63 	if (!irqd_msi_nomask_quirk(irqd) ||
64 	    cfg->vector == old_cfg.vector ||
65 	    old_cfg.vector == MANAGED_IRQ_SHUTDOWN_VECTOR ||
66 	    cfg->dest_apicid == old_cfg.dest_apicid) {
67 		irq_msi_update_msg(irqd, cfg);
68 		return ret;
69 	}
70 
71 	/*
72 	 * Paranoia: Validate that the interrupt target is the local
73 	 * CPU.
74 	 */
75 	if (WARN_ON_ONCE(cpu != smp_processor_id())) {
76 		irq_msi_update_msg(irqd, cfg);
77 		return ret;
78 	}
79 
80 	/*
81 	 * Redirect the interrupt to the new vector on the current CPU
82 	 * first. This might cause a spurious interrupt on this vector if
83 	 * the device raises an interrupt right between this update and the
84 	 * update to the final destination CPU.
85 	 *
86 	 * If the vector is in use then the installed device handler will
87 	 * denote it as spurious which is no harm as this is a rare event
88 	 * and interrupt handlers have to cope with spurious interrupts
89 	 * anyway. If the vector is unused, then it is marked so it won't
90 	 * trigger the 'No irq handler for vector' warning in
91 	 * common_interrupt().
92 	 *
93 	 * This requires to hold vector lock to prevent concurrent updates to
94 	 * the affected vector.
95 	 */
96 	lock_vector_lock();
97 
98 	/*
99 	 * Mark the new target vector on the local CPU if it is currently
100 	 * unused. Reuse the VECTOR_RETRIGGERED state which is also used in
101 	 * the CPU hotplug path for a similar purpose. This cannot be
102 	 * undone here as the current CPU has interrupts disabled and
103 	 * cannot handle the interrupt before the whole set_affinity()
104 	 * section is done. In the CPU unplug case, the current CPU is
105 	 * about to vanish and will not handle any interrupts anymore. The
106 	 * vector is cleaned up when the CPU comes online again.
107 	 */
108 	if (IS_ERR_OR_NULL(this_cpu_read(vector_irq[cfg->vector])))
109 		this_cpu_write(vector_irq[cfg->vector], VECTOR_RETRIGGERED);
110 
111 	/* Redirect it to the new vector on the local CPU temporarily */
112 	old_cfg.vector = cfg->vector;
113 	irq_msi_update_msg(irqd, &old_cfg);
114 
115 	/* Now transition it to the target CPU */
116 	irq_msi_update_msg(irqd, cfg);
117 
118 	/*
119 	 * All interrupts after this point are now targeted at the new
120 	 * vector/CPU.
121 	 *
122 	 * Drop vector lock before testing whether the temporary assignment
123 	 * to the local CPU was hit by an interrupt raised in the device,
124 	 * because the retrigger function acquires vector lock again.
125 	 */
126 	unlock_vector_lock();
127 
128 	/*
129 	 * Check whether the transition raced with a device interrupt and
130 	 * is pending in the local APICs IRR. It is safe to do this outside
131 	 * of vector lock as the irq_desc::lock of this interrupt is still
132 	 * held and interrupts are disabled: The check is not accessing the
133 	 * underlying vector store. It's just checking the local APIC's
134 	 * IRR.
135 	 */
136 	if (lapic_vector_set_in_irr(cfg->vector))
137 		irq_data_get_irq_chip(irqd)->irq_retrigger(irqd);
138 
139 	return ret;
140 }
141 
142 /*
143  * IRQ Chip for MSI PCI/PCI-X/PCI-Express Devices,
144  * which implement the MSI or MSI-X Capability Structure.
145  */
146 static struct irq_chip pci_msi_controller = {
147 	.name			= "PCI-MSI",
148 	.irq_unmask		= pci_msi_unmask_irq,
149 	.irq_mask		= pci_msi_mask_irq,
150 	.irq_ack		= irq_chip_ack_parent,
151 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
152 	.irq_set_affinity	= msi_set_affinity,
153 	.flags			= IRQCHIP_SKIP_SET_WAKE,
154 };
155 
156 int pci_msi_prepare(struct irq_domain *domain, struct device *dev, int nvec,
157 		    msi_alloc_info_t *arg)
158 {
159 	struct pci_dev *pdev = to_pci_dev(dev);
160 	struct msi_desc *desc = first_pci_msi_entry(pdev);
161 
162 	init_irq_alloc_info(arg, NULL);
163 	if (desc->msi_attrib.is_msix) {
164 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSIX;
165 	} else {
166 		arg->type = X86_IRQ_ALLOC_TYPE_PCI_MSI;
167 		arg->flags |= X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
168 	}
169 
170 	return 0;
171 }
172 EXPORT_SYMBOL_GPL(pci_msi_prepare);
173 
174 static struct msi_domain_ops pci_msi_domain_ops = {
175 	.msi_prepare	= pci_msi_prepare,
176 };
177 
178 static struct msi_domain_info pci_msi_domain_info = {
179 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
180 			  MSI_FLAG_PCI_MSIX,
181 	.ops		= &pci_msi_domain_ops,
182 	.chip		= &pci_msi_controller,
183 	.handler	= handle_edge_irq,
184 	.handler_name	= "edge",
185 };
186 
187 struct irq_domain * __init native_create_pci_msi_domain(void)
188 {
189 	struct fwnode_handle *fn;
190 	struct irq_domain *d;
191 
192 	if (disable_apic)
193 		return NULL;
194 
195 	fn = irq_domain_alloc_named_fwnode("PCI-MSI");
196 	if (!fn)
197 		return NULL;
198 
199 	d = pci_msi_create_irq_domain(fn, &pci_msi_domain_info,
200 				      x86_vector_domain);
201 	if (!d) {
202 		irq_domain_free_fwnode(fn);
203 		pr_warn("Failed to initialize PCI-MSI irqdomain.\n");
204 	} else {
205 		d->flags |= IRQ_DOMAIN_MSI_NOMASK_QUIRK;
206 	}
207 	return d;
208 }
209 
210 void __init x86_create_pci_msi_domain(void)
211 {
212 	x86_pci_msi_default_domain = x86_init.irqs.create_pci_msi_domain();
213 }
214 
215 #ifdef CONFIG_IRQ_REMAP
216 static struct irq_chip pci_msi_ir_controller = {
217 	.name			= "IR-PCI-MSI",
218 	.irq_unmask		= pci_msi_unmask_irq,
219 	.irq_mask		= pci_msi_mask_irq,
220 	.irq_ack		= irq_chip_ack_parent,
221 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
222 	.flags			= IRQCHIP_SKIP_SET_WAKE,
223 };
224 
225 static struct msi_domain_info pci_msi_ir_domain_info = {
226 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
227 			  MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX,
228 	.ops		= &pci_msi_domain_ops,
229 	.chip		= &pci_msi_ir_controller,
230 	.handler	= handle_edge_irq,
231 	.handler_name	= "edge",
232 };
233 
234 struct irq_domain *arch_create_remap_msi_irq_domain(struct irq_domain *parent,
235 						    const char *name, int id)
236 {
237 	struct fwnode_handle *fn;
238 	struct irq_domain *d;
239 
240 	fn = irq_domain_alloc_named_id_fwnode(name, id);
241 	if (!fn)
242 		return NULL;
243 	d = pci_msi_create_irq_domain(fn, &pci_msi_ir_domain_info, parent);
244 	if (!d)
245 		irq_domain_free_fwnode(fn);
246 	return d;
247 }
248 #endif
249 
250 #ifdef CONFIG_DMAR_TABLE
251 /*
252  * The Intel IOMMU (ab)uses the high bits of the MSI address to contain the
253  * high bits of the destination APIC ID. This can't be done in the general
254  * case for MSIs as it would be targeting real memory above 4GiB not the
255  * APIC.
256  */
257 static void dmar_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
258 {
259 	__irq_msi_compose_msg(irqd_cfg(data), msg, true);
260 }
261 
262 static void dmar_msi_write_msg(struct irq_data *data, struct msi_msg *msg)
263 {
264 	dmar_msi_write(data->irq, msg);
265 }
266 
267 static struct irq_chip dmar_msi_controller = {
268 	.name			= "DMAR-MSI",
269 	.irq_unmask		= dmar_msi_unmask,
270 	.irq_mask		= dmar_msi_mask,
271 	.irq_ack		= irq_chip_ack_parent,
272 	.irq_set_affinity	= msi_domain_set_affinity,
273 	.irq_retrigger		= irq_chip_retrigger_hierarchy,
274 	.irq_compose_msi_msg	= dmar_msi_compose_msg,
275 	.irq_write_msi_msg	= dmar_msi_write_msg,
276 	.flags			= IRQCHIP_SKIP_SET_WAKE,
277 };
278 
279 static int dmar_msi_init(struct irq_domain *domain,
280 			 struct msi_domain_info *info, unsigned int virq,
281 			 irq_hw_number_t hwirq, msi_alloc_info_t *arg)
282 {
283 	irq_domain_set_info(domain, virq, arg->devid, info->chip, NULL,
284 			    handle_edge_irq, arg->data, "edge");
285 
286 	return 0;
287 }
288 
289 static struct msi_domain_ops dmar_msi_domain_ops = {
290 	.msi_init	= dmar_msi_init,
291 };
292 
293 static struct msi_domain_info dmar_msi_domain_info = {
294 	.ops		= &dmar_msi_domain_ops,
295 	.chip		= &dmar_msi_controller,
296 	.flags		= MSI_FLAG_USE_DEF_DOM_OPS,
297 };
298 
299 static struct irq_domain *dmar_get_irq_domain(void)
300 {
301 	static struct irq_domain *dmar_domain;
302 	static DEFINE_MUTEX(dmar_lock);
303 	struct fwnode_handle *fn;
304 
305 	mutex_lock(&dmar_lock);
306 	if (dmar_domain)
307 		goto out;
308 
309 	fn = irq_domain_alloc_named_fwnode("DMAR-MSI");
310 	if (fn) {
311 		dmar_domain = msi_create_irq_domain(fn, &dmar_msi_domain_info,
312 						    x86_vector_domain);
313 		if (!dmar_domain)
314 			irq_domain_free_fwnode(fn);
315 	}
316 out:
317 	mutex_unlock(&dmar_lock);
318 	return dmar_domain;
319 }
320 
321 int dmar_alloc_hwirq(int id, int node, void *arg)
322 {
323 	struct irq_domain *domain = dmar_get_irq_domain();
324 	struct irq_alloc_info info;
325 
326 	if (!domain)
327 		return -1;
328 
329 	init_irq_alloc_info(&info, NULL);
330 	info.type = X86_IRQ_ALLOC_TYPE_DMAR;
331 	info.devid = id;
332 	info.hwirq = id;
333 	info.data = arg;
334 
335 	return irq_domain_alloc_irqs(domain, 1, node, &info);
336 }
337 
338 void dmar_free_hwirq(int irq)
339 {
340 	irq_domain_free_irqs(irq, 1);
341 }
342 #endif
343