xref: /linux/arch/x86/pci/xen.c (revision a82ebb093fc7bdd88f8df17b2fa303d7535fa43b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Xen PCI - handle PCI (INTx) and MSI infrastructure calls for PV, HVM and
4  * initial domain support. We also handle the DSDT _PRT callbacks for GSI's
5  * used in HVM and initial domain mode (PV does not parse ACPI, so it has no
6  * concept of GSIs). Under PV we hook under the pnbbios API for IRQs and
7  * 0xcf8 PCI configuration read/write.
8  *
9  *   Author: Ryan Wilson <hap9@epoch.ncsc.mil>
10  *           Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
11  *           Stefano Stabellini <stefano.stabellini@eu.citrix.com>
12  */
13 #include <linux/export.h>
14 #include <linux/init.h>
15 #include <linux/pci.h>
16 #include <linux/acpi.h>
17 
18 #include <linux/io.h>
19 #include <asm/io_apic.h>
20 #include <asm/pci_x86.h>
21 
22 #include <asm/xen/hypervisor.h>
23 
24 #include <xen/features.h>
25 #include <xen/events.h>
26 #include <xen/pci.h>
27 #include <asm/xen/pci.h>
28 #include <asm/xen/cpuid.h>
29 #include <asm/apic.h>
30 #include <asm/acpi.h>
31 #include <asm/i8259.h>
32 
33 static int xen_pcifront_enable_irq(struct pci_dev *dev)
34 {
35 	int rc;
36 	int share = 1;
37 	int pirq;
38 	u8 gsi;
39 
40 	rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi);
41 	if (rc < 0) {
42 		dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n",
43 			 rc);
44 		return rc;
45 	}
46 	/* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/
47 	pirq = gsi;
48 
49 	if (gsi < nr_legacy_irqs())
50 		share = 0;
51 
52 	rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront");
53 	if (rc < 0) {
54 		dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n",
55 			 gsi, pirq, rc);
56 		return rc;
57 	}
58 
59 	dev->irq = rc;
60 	dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq);
61 	return 0;
62 }
63 
64 #ifdef CONFIG_ACPI
65 static int xen_register_pirq(u32 gsi, int triggering, bool set_pirq)
66 {
67 	int rc, pirq = -1, irq;
68 	struct physdev_map_pirq map_irq;
69 	int shareable = 0;
70 	char *name;
71 
72 	irq = xen_irq_from_gsi(gsi);
73 	if (irq > 0)
74 		return irq;
75 
76 	if (set_pirq)
77 		pirq = gsi;
78 
79 	map_irq.domid = DOMID_SELF;
80 	map_irq.type = MAP_PIRQ_TYPE_GSI;
81 	map_irq.index = gsi;
82 	map_irq.pirq = pirq;
83 
84 	rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
85 	if (rc) {
86 		printk(KERN_WARNING "xen map irq failed %d\n", rc);
87 		return -1;
88 	}
89 
90 	if (triggering == ACPI_EDGE_SENSITIVE) {
91 		shareable = 0;
92 		name = "ioapic-edge";
93 	} else {
94 		shareable = 1;
95 		name = "ioapic-level";
96 	}
97 
98 	irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name);
99 	if (irq < 0)
100 		goto out;
101 
102 	printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi);
103 out:
104 	return irq;
105 }
106 
107 static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi,
108 				     int trigger, int polarity)
109 {
110 	if (!xen_hvm_domain())
111 		return -1;
112 
113 	return xen_register_pirq(gsi, trigger,
114 				 false /* no mapping of GSI to PIRQ */);
115 }
116 
117 #ifdef CONFIG_XEN_PV_DOM0
118 static int xen_register_gsi(u32 gsi, int triggering, int polarity)
119 {
120 	int rc, irq;
121 	struct physdev_setup_gsi setup_gsi;
122 
123 	if (!xen_pv_domain())
124 		return -1;
125 
126 	printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n",
127 			gsi, triggering, polarity);
128 
129 	irq = xen_register_pirq(gsi, triggering, true);
130 
131 	setup_gsi.gsi = gsi;
132 	setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1);
133 	setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
134 
135 	rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
136 	if (rc == -EEXIST)
137 		printk(KERN_INFO "Already setup the GSI :%d\n", gsi);
138 	else if (rc) {
139 		printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n",
140 				gsi, rc);
141 	}
142 
143 	return irq;
144 }
145 
146 static int acpi_register_gsi_xen(struct device *dev, u32 gsi,
147 				 int trigger, int polarity)
148 {
149 	return xen_register_gsi(gsi, trigger, polarity);
150 }
151 #endif
152 #endif
153 
154 #if defined(CONFIG_PCI_MSI)
155 #include <linux/msi.h>
156 
157 struct xen_pci_frontend_ops *xen_pci_frontend;
158 EXPORT_SYMBOL_GPL(xen_pci_frontend);
159 
160 struct xen_msi_ops {
161 	int (*setup_msi_irqs)(struct pci_dev *dev, int nvec, int type);
162 	void (*teardown_msi_irqs)(struct pci_dev *dev);
163 };
164 
165 static struct xen_msi_ops xen_msi_ops __ro_after_init;
166 
167 static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
168 {
169 	int irq, ret, i;
170 	struct msi_desc *msidesc;
171 	int *v;
172 
173 	if (type == PCI_CAP_ID_MSI && nvec > 1)
174 		return 1;
175 
176 	v = kcalloc(max(1, nvec), sizeof(int), GFP_KERNEL);
177 	if (!v)
178 		return -ENOMEM;
179 
180 	if (type == PCI_CAP_ID_MSIX)
181 		ret = xen_pci_frontend_enable_msix(dev, v, nvec);
182 	else
183 		ret = xen_pci_frontend_enable_msi(dev, v);
184 	if (ret)
185 		goto error;
186 	i = 0;
187 	msi_for_each_desc(msidesc, &dev->dev, MSI_DESC_NOTASSOCIATED) {
188 		irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i],
189 					       (type == PCI_CAP_ID_MSI) ? nvec : 1,
190 					       (type == PCI_CAP_ID_MSIX) ?
191 					       "pcifront-msi-x" :
192 					       "pcifront-msi",
193 						DOMID_SELF);
194 		if (irq < 0) {
195 			ret = irq;
196 			goto free;
197 		}
198 		i++;
199 	}
200 	kfree(v);
201 	return 0;
202 
203 error:
204 	if (ret == -ENOSYS)
205 		dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n");
206 	else if (ret)
207 		dev_err(&dev->dev, "Xen PCI frontend error: %d!\n", ret);
208 free:
209 	kfree(v);
210 	return ret;
211 }
212 
213 static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq,
214 		struct msi_msg *msg)
215 {
216 	/*
217 	 * We set vector == 0 to tell the hypervisor we don't care about
218 	 * it, but we want a pirq setup instead.  We use the dest_id fields
219 	 * to pass the pirq that we want.
220 	 */
221 	memset(msg, 0, sizeof(*msg));
222 	msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
223 	msg->arch_addr_hi.destid_8_31 = pirq >> 8;
224 	msg->arch_addr_lo.destid_0_7 = pirq & 0xFF;
225 	msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
226 	msg->arch_data.delivery_mode = APIC_DELIVERY_MODE_EXTINT;
227 }
228 
229 static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
230 {
231 	int irq, pirq;
232 	struct msi_desc *msidesc;
233 	struct msi_msg msg;
234 
235 	if (type == PCI_CAP_ID_MSI && nvec > 1)
236 		return 1;
237 
238 	msi_for_each_desc(msidesc, &dev->dev, MSI_DESC_NOTASSOCIATED) {
239 		pirq = xen_allocate_pirq_msi(dev, msidesc);
240 		if (pirq < 0) {
241 			irq = -ENODEV;
242 			goto error;
243 		}
244 		xen_msi_compose_msg(dev, pirq, &msg);
245 		__pci_write_msi_msg(msidesc, &msg);
246 		dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq);
247 		irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq,
248 					       (type == PCI_CAP_ID_MSI) ? nvec : 1,
249 					       (type == PCI_CAP_ID_MSIX) ?
250 					       "msi-x" : "msi",
251 					       DOMID_SELF);
252 		if (irq < 0)
253 			goto error;
254 		dev_dbg(&dev->dev,
255 			"xen: msi --> pirq=%d --> irq=%d\n", pirq, irq);
256 	}
257 	return 0;
258 
259 error:
260 	dev_err(&dev->dev, "Failed to create MSI%s! ret=%d!\n",
261 		type == PCI_CAP_ID_MSI ? "" : "-X", irq);
262 	return irq;
263 }
264 
265 #ifdef CONFIG_XEN_PV_DOM0
266 static bool __read_mostly pci_seg_supported = true;
267 
268 static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
269 {
270 	int ret = 0;
271 	struct msi_desc *msidesc;
272 
273 	msi_for_each_desc(msidesc, &dev->dev, MSI_DESC_NOTASSOCIATED) {
274 		struct physdev_map_pirq map_irq;
275 		domid_t domid;
276 
277 		domid = ret = xen_find_device_domain_owner(dev);
278 		/* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED,
279 		 * hence check ret value for < 0. */
280 		if (ret < 0)
281 			domid = DOMID_SELF;
282 
283 		memset(&map_irq, 0, sizeof(map_irq));
284 		map_irq.domid = domid;
285 		map_irq.type = MAP_PIRQ_TYPE_MSI_SEG;
286 		map_irq.index = -1;
287 		map_irq.pirq = -1;
288 		map_irq.bus = dev->bus->number |
289 			      (pci_domain_nr(dev->bus) << 16);
290 		map_irq.devfn = dev->devfn;
291 
292 		if (type == PCI_CAP_ID_MSI && nvec > 1) {
293 			map_irq.type = MAP_PIRQ_TYPE_MULTI_MSI;
294 			map_irq.entry_nr = nvec;
295 		} else if (type == PCI_CAP_ID_MSIX) {
296 			int pos;
297 			unsigned long flags;
298 			u32 table_offset, bir;
299 
300 			pos = dev->msix_cap;
301 			pci_read_config_dword(dev, pos + PCI_MSIX_TABLE,
302 					      &table_offset);
303 			bir = (u8)(table_offset & PCI_MSIX_TABLE_BIR);
304 			flags = pci_resource_flags(dev, bir);
305 			if (!flags || (flags & IORESOURCE_UNSET))
306 				return -EINVAL;
307 
308 			map_irq.table_base = pci_resource_start(dev, bir);
309 			map_irq.entry_nr = msidesc->msi_index;
310 		}
311 
312 		ret = -EINVAL;
313 		if (pci_seg_supported)
314 			ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
315 						    &map_irq);
316 		if (type == PCI_CAP_ID_MSI && nvec > 1 && ret) {
317 			/*
318 			 * If MAP_PIRQ_TYPE_MULTI_MSI is not available
319 			 * there's nothing else we can do in this case.
320 			 * Just set ret > 0 so driver can retry with
321 			 * single MSI.
322 			 */
323 			ret = 1;
324 			goto out;
325 		}
326 		if (ret == -EINVAL && !pci_domain_nr(dev->bus)) {
327 			map_irq.type = MAP_PIRQ_TYPE_MSI;
328 			map_irq.index = -1;
329 			map_irq.pirq = -1;
330 			map_irq.bus = dev->bus->number;
331 			ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
332 						    &map_irq);
333 			if (ret != -EINVAL)
334 				pci_seg_supported = false;
335 		}
336 		if (ret) {
337 			dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n",
338 				 ret, domid);
339 			goto out;
340 		}
341 
342 		ret = xen_bind_pirq_msi_to_irq(dev, msidesc, map_irq.pirq,
343 		                               (type == PCI_CAP_ID_MSI) ? nvec : 1,
344 		                               (type == PCI_CAP_ID_MSIX) ? "msi-x" : "msi",
345 		                               domid);
346 		if (ret < 0)
347 			goto out;
348 	}
349 	ret = 0;
350 out:
351 	return ret;
352 }
353 
354 bool xen_initdom_restore_msi(struct pci_dev *dev)
355 {
356 	int ret = 0;
357 
358 	if (!xen_initial_domain())
359 		return true;
360 
361 	if (pci_seg_supported) {
362 		struct physdev_pci_device restore_ext;
363 
364 		restore_ext.seg = pci_domain_nr(dev->bus);
365 		restore_ext.bus = dev->bus->number;
366 		restore_ext.devfn = dev->devfn;
367 		ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi_ext,
368 					&restore_ext);
369 		if (ret == -ENOSYS)
370 			pci_seg_supported = false;
371 		WARN(ret && ret != -ENOSYS, "restore_msi_ext -> %d\n", ret);
372 	}
373 	if (!pci_seg_supported) {
374 		struct physdev_restore_msi restore;
375 
376 		restore.bus = dev->bus->number;
377 		restore.devfn = dev->devfn;
378 		ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi, &restore);
379 		WARN(ret && ret != -ENOSYS, "restore_msi -> %d\n", ret);
380 	}
381 	return false;
382 }
383 #else /* CONFIG_XEN_PV_DOM0 */
384 #define xen_initdom_setup_msi_irqs	NULL
385 #endif /* !CONFIG_XEN_PV_DOM0 */
386 
387 static void xen_teardown_msi_irqs(struct pci_dev *dev)
388 {
389 	struct msi_desc *msidesc;
390 	int i;
391 
392 	msi_for_each_desc(msidesc, &dev->dev, MSI_DESC_ASSOCIATED) {
393 		for (i = 0; i < msidesc->nvec_used; i++)
394 			xen_destroy_irq(msidesc->irq + i);
395 	}
396 }
397 
398 static void xen_pv_teardown_msi_irqs(struct pci_dev *dev)
399 {
400 	if (dev->msix_enabled)
401 		xen_pci_frontend_disable_msix(dev);
402 	else
403 		xen_pci_frontend_disable_msi(dev);
404 
405 	xen_teardown_msi_irqs(dev);
406 }
407 
408 static int xen_msi_domain_alloc_irqs(struct irq_domain *domain,
409 				     struct device *dev,  int nvec)
410 {
411 	int type;
412 
413 	if (WARN_ON_ONCE(!dev_is_pci(dev)))
414 		return -EINVAL;
415 
416 	type = to_pci_dev(dev)->msix_enabled ? PCI_CAP_ID_MSIX : PCI_CAP_ID_MSI;
417 
418 	return xen_msi_ops.setup_msi_irqs(to_pci_dev(dev), nvec, type);
419 }
420 
421 static void xen_msi_domain_free_irqs(struct irq_domain *domain,
422 				     struct device *dev)
423 {
424 	if (WARN_ON_ONCE(!dev_is_pci(dev)))
425 		return;
426 
427 	xen_msi_ops.teardown_msi_irqs(to_pci_dev(dev));
428 }
429 
430 static struct msi_domain_ops xen_pci_msi_domain_ops = {
431 	.domain_alloc_irqs	= xen_msi_domain_alloc_irqs,
432 	.domain_free_irqs	= xen_msi_domain_free_irqs,
433 };
434 
435 static struct msi_domain_info xen_pci_msi_domain_info = {
436 	.ops			= &xen_pci_msi_domain_ops,
437 };
438 
439 /*
440  * This irq domain is a blatant violation of the irq domain design, but
441  * distangling XEN into real irq domains is not a job for mere mortals with
442  * limited XENology. But it's the least dangerous way for a mere mortal to
443  * get rid of the arch_*_msi_irqs() hackery in order to store the irq
444  * domain pointer in struct device. This irq domain wrappery allows to do
445  * that without breaking XEN terminally.
446  */
447 static __init struct irq_domain *xen_create_pci_msi_domain(void)
448 {
449 	struct irq_domain *d = NULL;
450 	struct fwnode_handle *fn;
451 
452 	fn = irq_domain_alloc_named_fwnode("XEN-MSI");
453 	if (fn)
454 		d = msi_create_irq_domain(fn, &xen_pci_msi_domain_info, NULL);
455 
456 	/* FIXME: No idea how to survive if this fails */
457 	BUG_ON(!d);
458 
459 	return d;
460 }
461 
462 static __init void xen_setup_pci_msi(void)
463 {
464 	if (xen_pv_domain()) {
465 		if (xen_initial_domain())
466 			xen_msi_ops.setup_msi_irqs = xen_initdom_setup_msi_irqs;
467 		else
468 			xen_msi_ops.setup_msi_irqs = xen_setup_msi_irqs;
469 		xen_msi_ops.teardown_msi_irqs = xen_pv_teardown_msi_irqs;
470 		pci_msi_ignore_mask = 1;
471 	} else if (xen_hvm_domain()) {
472 		xen_msi_ops.setup_msi_irqs = xen_hvm_setup_msi_irqs;
473 		xen_msi_ops.teardown_msi_irqs = xen_teardown_msi_irqs;
474 	} else {
475 		WARN_ON_ONCE(1);
476 		return;
477 	}
478 
479 	/*
480 	 * Override the PCI/MSI irq domain init function. No point
481 	 * in allocating the native domain and never use it.
482 	 */
483 	x86_init.irqs.create_pci_msi_domain = xen_create_pci_msi_domain;
484 }
485 
486 #else /* CONFIG_PCI_MSI */
487 static inline void xen_setup_pci_msi(void) { }
488 #endif /* CONFIG_PCI_MSI */
489 
490 int __init pci_xen_init(void)
491 {
492 	if (!xen_pv_domain() || xen_initial_domain())
493 		return -ENODEV;
494 
495 	printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
496 
497 	pcibios_set_cache_line_size();
498 
499 	pcibios_enable_irq = xen_pcifront_enable_irq;
500 	pcibios_disable_irq = NULL;
501 
502 	/* Keep ACPI out of the picture */
503 	acpi_noirq_set();
504 
505 	xen_setup_pci_msi();
506 	return 0;
507 }
508 
509 #ifdef CONFIG_PCI_MSI
510 static void __init xen_hvm_msi_init(void)
511 {
512 	if (!disable_apic) {
513 		/*
514 		 * If hardware supports (x2)APIC virtualization (as indicated
515 		 * by hypervisor's leaf 4) then we don't need to use pirqs/
516 		 * event channels for MSI handling and instead use regular
517 		 * APIC processing
518 		 */
519 		uint32_t eax = cpuid_eax(xen_cpuid_base() + 4);
520 
521 		if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) ||
522 		    ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC)))
523 			return;
524 	}
525 	xen_setup_pci_msi();
526 }
527 #endif
528 
529 int __init pci_xen_hvm_init(void)
530 {
531 	if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs))
532 		return 0;
533 
534 #ifdef CONFIG_ACPI
535 	/*
536 	 * We don't want to change the actual ACPI delivery model,
537 	 * just how GSIs get registered.
538 	 */
539 	__acpi_register_gsi = acpi_register_gsi_xen_hvm;
540 	__acpi_unregister_gsi = NULL;
541 #endif
542 
543 #ifdef CONFIG_PCI_MSI
544 	/*
545 	 * We need to wait until after x2apic is initialized
546 	 * before we can set MSI IRQ ops.
547 	 */
548 	x86_platform.apic_post_init = xen_hvm_msi_init;
549 #endif
550 	return 0;
551 }
552 
553 #ifdef CONFIG_XEN_PV_DOM0
554 int __init pci_xen_initial_domain(void)
555 {
556 	int irq;
557 
558 	xen_setup_pci_msi();
559 	__acpi_register_gsi = acpi_register_gsi_xen;
560 	__acpi_unregister_gsi = NULL;
561 	/*
562 	 * Pre-allocate the legacy IRQs.  Use NR_LEGACY_IRQS here
563 	 * because we don't have a PIC and thus nr_legacy_irqs() is zero.
564 	 */
565 	for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
566 		int trigger, polarity;
567 
568 		if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
569 			continue;
570 
571 		xen_register_pirq(irq,
572 			trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
573 			true /* Map GSI to PIRQ */);
574 	}
575 	if (0 == nr_ioapics) {
576 		for (irq = 0; irq < nr_legacy_irqs(); irq++)
577 			xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
578 	}
579 	return 0;
580 }
581 #endif
582 
583