xref: /linux/arch/x86/pci/xen.c (revision 2bc46b3ad3c15165f91459b07ff8682478683194)
1 /*
2  * Xen PCI - handle PCI (INTx) and MSI infrastructure calls for PV, HVM and
3  * initial domain support. We also handle the DSDT _PRT callbacks for GSI's
4  * used in HVM and initial domain mode (PV does not parse ACPI, so it has no
5  * concept of GSIs). Under PV we hook under the pnbbios API for IRQs and
6  * 0xcf8 PCI configuration read/write.
7  *
8  *   Author: Ryan Wilson <hap9@epoch.ncsc.mil>
9  *           Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
10  *           Stefano Stabellini <stefano.stabellini@eu.citrix.com>
11  */
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/pci.h>
15 #include <linux/acpi.h>
16 
17 #include <linux/io.h>
18 #include <asm/io_apic.h>
19 #include <asm/pci_x86.h>
20 
21 #include <asm/xen/hypervisor.h>
22 
23 #include <xen/features.h>
24 #include <xen/events.h>
25 #include <asm/xen/pci.h>
26 #include <asm/xen/cpuid.h>
27 #include <asm/apic.h>
28 #include <asm/i8259.h>
29 
30 static int xen_pcifront_enable_irq(struct pci_dev *dev)
31 {
32 	int rc;
33 	int share = 1;
34 	int pirq;
35 	u8 gsi;
36 
37 	rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi);
38 	if (rc < 0) {
39 		dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n",
40 			 rc);
41 		return rc;
42 	}
43 	/* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/
44 	pirq = gsi;
45 
46 	if (gsi < nr_legacy_irqs())
47 		share = 0;
48 
49 	rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront");
50 	if (rc < 0) {
51 		dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n",
52 			 gsi, pirq, rc);
53 		return rc;
54 	}
55 
56 	dev->irq = rc;
57 	dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq);
58 	return 0;
59 }
60 
61 #ifdef CONFIG_ACPI
62 static int xen_register_pirq(u32 gsi, int gsi_override, int triggering,
63 			     bool set_pirq)
64 {
65 	int rc, pirq = -1, irq = -1;
66 	struct physdev_map_pirq map_irq;
67 	int shareable = 0;
68 	char *name;
69 
70 	irq = xen_irq_from_gsi(gsi);
71 	if (irq > 0)
72 		return irq;
73 
74 	if (set_pirq)
75 		pirq = gsi;
76 
77 	map_irq.domid = DOMID_SELF;
78 	map_irq.type = MAP_PIRQ_TYPE_GSI;
79 	map_irq.index = gsi;
80 	map_irq.pirq = pirq;
81 
82 	rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
83 	if (rc) {
84 		printk(KERN_WARNING "xen map irq failed %d\n", rc);
85 		return -1;
86 	}
87 
88 	if (triggering == ACPI_EDGE_SENSITIVE) {
89 		shareable = 0;
90 		name = "ioapic-edge";
91 	} else {
92 		shareable = 1;
93 		name = "ioapic-level";
94 	}
95 
96 	if (gsi_override >= 0)
97 		gsi = gsi_override;
98 
99 	irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name);
100 	if (irq < 0)
101 		goto out;
102 
103 	printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi);
104 out:
105 	return irq;
106 }
107 
108 static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi,
109 				     int trigger, int polarity)
110 {
111 	if (!xen_hvm_domain())
112 		return -1;
113 
114 	return xen_register_pirq(gsi, -1 /* no GSI override */, trigger,
115 				 false /* no mapping of GSI to PIRQ */);
116 }
117 
118 #ifdef CONFIG_XEN_DOM0
119 static int xen_register_gsi(u32 gsi, int gsi_override, int triggering, int polarity)
120 {
121 	int rc, irq;
122 	struct physdev_setup_gsi setup_gsi;
123 
124 	if (!xen_pv_domain())
125 		return -1;
126 
127 	printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n",
128 			gsi, triggering, polarity);
129 
130 	irq = xen_register_pirq(gsi, gsi_override, triggering, true);
131 
132 	setup_gsi.gsi = gsi;
133 	setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1);
134 	setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
135 
136 	rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
137 	if (rc == -EEXIST)
138 		printk(KERN_INFO "Already setup the GSI :%d\n", gsi);
139 	else if (rc) {
140 		printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n",
141 				gsi, rc);
142 	}
143 
144 	return irq;
145 }
146 
147 static int acpi_register_gsi_xen(struct device *dev, u32 gsi,
148 				 int trigger, int polarity)
149 {
150 	return xen_register_gsi(gsi, -1 /* no GSI override */, trigger, polarity);
151 }
152 #endif
153 #endif
154 
155 #if defined(CONFIG_PCI_MSI)
156 #include <linux/msi.h>
157 #include <asm/msidef.h>
158 
159 struct xen_pci_frontend_ops *xen_pci_frontend;
160 EXPORT_SYMBOL_GPL(xen_pci_frontend);
161 
162 static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
163 {
164 	int irq, ret, i;
165 	struct msi_desc *msidesc;
166 	int *v;
167 
168 	if (type == PCI_CAP_ID_MSI && nvec > 1)
169 		return 1;
170 
171 	v = kzalloc(sizeof(int) * max(1, nvec), GFP_KERNEL);
172 	if (!v)
173 		return -ENOMEM;
174 
175 	if (type == PCI_CAP_ID_MSIX)
176 		ret = xen_pci_frontend_enable_msix(dev, v, nvec);
177 	else
178 		ret = xen_pci_frontend_enable_msi(dev, v);
179 	if (ret)
180 		goto error;
181 	i = 0;
182 	for_each_pci_msi_entry(msidesc, dev) {
183 		irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i],
184 					       (type == PCI_CAP_ID_MSI) ? nvec : 1,
185 					       (type == PCI_CAP_ID_MSIX) ?
186 					       "pcifront-msi-x" :
187 					       "pcifront-msi",
188 						DOMID_SELF);
189 		if (irq < 0) {
190 			ret = irq;
191 			goto free;
192 		}
193 		i++;
194 	}
195 	kfree(v);
196 	return 0;
197 
198 error:
199 	if (ret == -ENOSYS)
200 		dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n");
201 	else if (ret)
202 		dev_err(&dev->dev, "Xen PCI frontend error: %d!\n", ret);
203 free:
204 	kfree(v);
205 	return ret;
206 }
207 
208 #define XEN_PIRQ_MSI_DATA  (MSI_DATA_TRIGGER_EDGE | \
209 		MSI_DATA_LEVEL_ASSERT | (3 << 8) | MSI_DATA_VECTOR(0))
210 
211 static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq,
212 		struct msi_msg *msg)
213 {
214 	/* We set vector == 0 to tell the hypervisor we don't care about it,
215 	 * but we want a pirq setup instead.
216 	 * We use the dest_id field to pass the pirq that we want. */
217 	msg->address_hi = MSI_ADDR_BASE_HI | MSI_ADDR_EXT_DEST_ID(pirq);
218 	msg->address_lo =
219 		MSI_ADDR_BASE_LO |
220 		MSI_ADDR_DEST_MODE_PHYSICAL |
221 		MSI_ADDR_REDIRECTION_CPU |
222 		MSI_ADDR_DEST_ID(pirq);
223 
224 	msg->data = XEN_PIRQ_MSI_DATA;
225 }
226 
227 static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
228 {
229 	int irq, pirq;
230 	struct msi_desc *msidesc;
231 	struct msi_msg msg;
232 
233 	if (type == PCI_CAP_ID_MSI && nvec > 1)
234 		return 1;
235 
236 	for_each_pci_msi_entry(msidesc, dev) {
237 		__pci_read_msi_msg(msidesc, &msg);
238 		pirq = MSI_ADDR_EXT_DEST_ID(msg.address_hi) |
239 			((msg.address_lo >> MSI_ADDR_DEST_ID_SHIFT) & 0xff);
240 		if (msg.data != XEN_PIRQ_MSI_DATA ||
241 		    xen_irq_from_pirq(pirq) < 0) {
242 			pirq = xen_allocate_pirq_msi(dev, msidesc);
243 			if (pirq < 0) {
244 				irq = -ENODEV;
245 				goto error;
246 			}
247 			xen_msi_compose_msg(dev, pirq, &msg);
248 			__pci_write_msi_msg(msidesc, &msg);
249 			dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq);
250 		} else {
251 			dev_dbg(&dev->dev,
252 				"xen: msi already bound to pirq=%d\n", pirq);
253 		}
254 		irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq,
255 					       (type == PCI_CAP_ID_MSI) ? nvec : 1,
256 					       (type == PCI_CAP_ID_MSIX) ?
257 					       "msi-x" : "msi",
258 					       DOMID_SELF);
259 		if (irq < 0)
260 			goto error;
261 		dev_dbg(&dev->dev,
262 			"xen: msi --> pirq=%d --> irq=%d\n", pirq, irq);
263 	}
264 	return 0;
265 
266 error:
267 	dev_err(&dev->dev,
268 		"Xen PCI frontend has not registered MSI/MSI-X support!\n");
269 	return irq;
270 }
271 
272 #ifdef CONFIG_XEN_DOM0
273 static bool __read_mostly pci_seg_supported = true;
274 
275 static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
276 {
277 	int ret = 0;
278 	struct msi_desc *msidesc;
279 
280 	for_each_pci_msi_entry(msidesc, dev) {
281 		struct physdev_map_pirq map_irq;
282 		domid_t domid;
283 
284 		domid = ret = xen_find_device_domain_owner(dev);
285 		/* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED,
286 		 * hence check ret value for < 0. */
287 		if (ret < 0)
288 			domid = DOMID_SELF;
289 
290 		memset(&map_irq, 0, sizeof(map_irq));
291 		map_irq.domid = domid;
292 		map_irq.type = MAP_PIRQ_TYPE_MSI_SEG;
293 		map_irq.index = -1;
294 		map_irq.pirq = -1;
295 		map_irq.bus = dev->bus->number |
296 			      (pci_domain_nr(dev->bus) << 16);
297 		map_irq.devfn = dev->devfn;
298 
299 		if (type == PCI_CAP_ID_MSI && nvec > 1) {
300 			map_irq.type = MAP_PIRQ_TYPE_MULTI_MSI;
301 			map_irq.entry_nr = nvec;
302 		} else if (type == PCI_CAP_ID_MSIX) {
303 			int pos;
304 			unsigned long flags;
305 			u32 table_offset, bir;
306 
307 			pos = dev->msix_cap;
308 			pci_read_config_dword(dev, pos + PCI_MSIX_TABLE,
309 					      &table_offset);
310 			bir = (u8)(table_offset & PCI_MSIX_TABLE_BIR);
311 			flags = pci_resource_flags(dev, bir);
312 			if (!flags || (flags & IORESOURCE_UNSET))
313 				return -EINVAL;
314 
315 			map_irq.table_base = pci_resource_start(dev, bir);
316 			map_irq.entry_nr = msidesc->msi_attrib.entry_nr;
317 		}
318 
319 		ret = -EINVAL;
320 		if (pci_seg_supported)
321 			ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
322 						    &map_irq);
323 		if (type == PCI_CAP_ID_MSI && nvec > 1 && ret) {
324 			/*
325 			 * If MAP_PIRQ_TYPE_MULTI_MSI is not available
326 			 * there's nothing else we can do in this case.
327 			 * Just set ret > 0 so driver can retry with
328 			 * single MSI.
329 			 */
330 			ret = 1;
331 			goto out;
332 		}
333 		if (ret == -EINVAL && !pci_domain_nr(dev->bus)) {
334 			map_irq.type = MAP_PIRQ_TYPE_MSI;
335 			map_irq.index = -1;
336 			map_irq.pirq = -1;
337 			map_irq.bus = dev->bus->number;
338 			ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq,
339 						    &map_irq);
340 			if (ret != -EINVAL)
341 				pci_seg_supported = false;
342 		}
343 		if (ret) {
344 			dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n",
345 				 ret, domid);
346 			goto out;
347 		}
348 
349 		ret = xen_bind_pirq_msi_to_irq(dev, msidesc, map_irq.pirq,
350 		                               (type == PCI_CAP_ID_MSI) ? nvec : 1,
351 		                               (type == PCI_CAP_ID_MSIX) ? "msi-x" : "msi",
352 		                               domid);
353 		if (ret < 0)
354 			goto out;
355 	}
356 	ret = 0;
357 out:
358 	return ret;
359 }
360 
361 static void xen_initdom_restore_msi_irqs(struct pci_dev *dev)
362 {
363 	int ret = 0;
364 
365 	if (pci_seg_supported) {
366 		struct physdev_pci_device restore_ext;
367 
368 		restore_ext.seg = pci_domain_nr(dev->bus);
369 		restore_ext.bus = dev->bus->number;
370 		restore_ext.devfn = dev->devfn;
371 		ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi_ext,
372 					&restore_ext);
373 		if (ret == -ENOSYS)
374 			pci_seg_supported = false;
375 		WARN(ret && ret != -ENOSYS, "restore_msi_ext -> %d\n", ret);
376 	}
377 	if (!pci_seg_supported) {
378 		struct physdev_restore_msi restore;
379 
380 		restore.bus = dev->bus->number;
381 		restore.devfn = dev->devfn;
382 		ret = HYPERVISOR_physdev_op(PHYSDEVOP_restore_msi, &restore);
383 		WARN(ret && ret != -ENOSYS, "restore_msi -> %d\n", ret);
384 	}
385 }
386 #endif
387 
388 static void xen_teardown_msi_irqs(struct pci_dev *dev)
389 {
390 	struct msi_desc *msidesc;
391 
392 	msidesc = first_pci_msi_entry(dev);
393 	if (msidesc->msi_attrib.is_msix)
394 		xen_pci_frontend_disable_msix(dev);
395 	else
396 		xen_pci_frontend_disable_msi(dev);
397 
398 	/* Free the IRQ's and the msidesc using the generic code. */
399 	default_teardown_msi_irqs(dev);
400 }
401 
402 static void xen_teardown_msi_irq(unsigned int irq)
403 {
404 	xen_destroy_irq(irq);
405 }
406 
407 #endif
408 
409 int __init pci_xen_init(void)
410 {
411 	if (!xen_pv_domain() || xen_initial_domain())
412 		return -ENODEV;
413 
414 	printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
415 
416 	pcibios_set_cache_line_size();
417 
418 	pcibios_enable_irq = xen_pcifront_enable_irq;
419 	pcibios_disable_irq = NULL;
420 
421 #ifdef CONFIG_ACPI
422 	/* Keep ACPI out of the picture */
423 	acpi_noirq = 1;
424 #endif
425 
426 #ifdef CONFIG_PCI_MSI
427 	x86_msi.setup_msi_irqs = xen_setup_msi_irqs;
428 	x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
429 	x86_msi.teardown_msi_irqs = xen_teardown_msi_irqs;
430 	pci_msi_ignore_mask = 1;
431 #endif
432 	return 0;
433 }
434 
435 #ifdef CONFIG_PCI_MSI
436 void __init xen_msi_init(void)
437 {
438 	if (!disable_apic) {
439 		/*
440 		 * If hardware supports (x2)APIC virtualization (as indicated
441 		 * by hypervisor's leaf 4) then we don't need to use pirqs/
442 		 * event channels for MSI handling and instead use regular
443 		 * APIC processing
444 		 */
445 		uint32_t eax = cpuid_eax(xen_cpuid_base() + 4);
446 
447 		if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) ||
448 		    ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC)))
449 			return;
450 	}
451 
452 	x86_msi.setup_msi_irqs = xen_hvm_setup_msi_irqs;
453 	x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
454 }
455 #endif
456 
457 int __init pci_xen_hvm_init(void)
458 {
459 	if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs))
460 		return 0;
461 
462 #ifdef CONFIG_ACPI
463 	/*
464 	 * We don't want to change the actual ACPI delivery model,
465 	 * just how GSIs get registered.
466 	 */
467 	__acpi_register_gsi = acpi_register_gsi_xen_hvm;
468 	__acpi_unregister_gsi = NULL;
469 #endif
470 
471 #ifdef CONFIG_PCI_MSI
472 	/*
473 	 * We need to wait until after x2apic is initialized
474 	 * before we can set MSI IRQ ops.
475 	 */
476 	x86_platform.apic_post_init = xen_msi_init;
477 #endif
478 	return 0;
479 }
480 
481 #ifdef CONFIG_XEN_DOM0
482 int __init pci_xen_initial_domain(void)
483 {
484 	int irq;
485 
486 #ifdef CONFIG_PCI_MSI
487 	x86_msi.setup_msi_irqs = xen_initdom_setup_msi_irqs;
488 	x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
489 	x86_msi.restore_msi_irqs = xen_initdom_restore_msi_irqs;
490 	pci_msi_ignore_mask = 1;
491 #endif
492 	__acpi_register_gsi = acpi_register_gsi_xen;
493 	__acpi_unregister_gsi = NULL;
494 	/*
495 	 * Pre-allocate the legacy IRQs.  Use NR_LEGACY_IRQS here
496 	 * because we don't have a PIC and thus nr_legacy_irqs() is zero.
497 	 */
498 	for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
499 		int trigger, polarity;
500 
501 		if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
502 			continue;
503 
504 		xen_register_pirq(irq, -1 /* no GSI override */,
505 			trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
506 			true /* Map GSI to PIRQ */);
507 	}
508 	if (0 == nr_ioapics) {
509 		for (irq = 0; irq < nr_legacy_irqs(); irq++)
510 			xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
511 	}
512 	return 0;
513 }
514 
515 struct xen_device_domain_owner {
516 	domid_t domain;
517 	struct pci_dev *dev;
518 	struct list_head list;
519 };
520 
521 static DEFINE_SPINLOCK(dev_domain_list_spinlock);
522 static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
523 
524 static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
525 {
526 	struct xen_device_domain_owner *owner;
527 
528 	list_for_each_entry(owner, &dev_domain_list, list) {
529 		if (owner->dev == dev)
530 			return owner;
531 	}
532 	return NULL;
533 }
534 
535 int xen_find_device_domain_owner(struct pci_dev *dev)
536 {
537 	struct xen_device_domain_owner *owner;
538 	int domain = -ENODEV;
539 
540 	spin_lock(&dev_domain_list_spinlock);
541 	owner = find_device(dev);
542 	if (owner)
543 		domain = owner->domain;
544 	spin_unlock(&dev_domain_list_spinlock);
545 	return domain;
546 }
547 EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
548 
549 int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
550 {
551 	struct xen_device_domain_owner *owner;
552 
553 	owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
554 	if (!owner)
555 		return -ENODEV;
556 
557 	spin_lock(&dev_domain_list_spinlock);
558 	if (find_device(dev)) {
559 		spin_unlock(&dev_domain_list_spinlock);
560 		kfree(owner);
561 		return -EEXIST;
562 	}
563 	owner->domain = domain;
564 	owner->dev = dev;
565 	list_add_tail(&owner->list, &dev_domain_list);
566 	spin_unlock(&dev_domain_list_spinlock);
567 	return 0;
568 }
569 EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
570 
571 int xen_unregister_device_domain_owner(struct pci_dev *dev)
572 {
573 	struct xen_device_domain_owner *owner;
574 
575 	spin_lock(&dev_domain_list_spinlock);
576 	owner = find_device(dev);
577 	if (!owner) {
578 		spin_unlock(&dev_domain_list_spinlock);
579 		return -ENODEV;
580 	}
581 	list_del(&owner->list);
582 	spin_unlock(&dev_domain_list_spinlock);
583 	kfree(owner);
584 	return 0;
585 }
586 EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
587 #endif
588