Loading...
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
27static int xen_pcifront_enable_irq(struct pci_dev *dev)
28{
29 int rc;
30 int share = 1;
31 int pirq;
32 u8 gsi;
33
34 rc = pci_read_config_byte(dev, PCI_INTERRUPT_LINE, &gsi);
35 if (rc < 0) {
36 dev_warn(&dev->dev, "Xen PCI: failed to read interrupt line: %d\n",
37 rc);
38 return rc;
39 }
40 /* In PV DomU the Xen PCI backend puts the PIRQ in the interrupt line.*/
41 pirq = gsi;
42
43 if (gsi < NR_IRQS_LEGACY)
44 share = 0;
45
46 rc = xen_bind_pirq_gsi_to_irq(gsi, pirq, share, "pcifront");
47 if (rc < 0) {
48 dev_warn(&dev->dev, "Xen PCI: failed to bind GSI%d (PIRQ%d) to IRQ: %d\n",
49 gsi, pirq, rc);
50 return rc;
51 }
52
53 dev->irq = rc;
54 dev_info(&dev->dev, "Xen PCI mapped GSI%d to IRQ%d\n", gsi, dev->irq);
55 return 0;
56}
57
58#ifdef CONFIG_ACPI
59static int xen_register_pirq(u32 gsi, int gsi_override, int triggering,
60 bool set_pirq)
61{
62 int rc, pirq = -1, irq = -1;
63 struct physdev_map_pirq map_irq;
64 int shareable = 0;
65 char *name;
66
67 if (set_pirq)
68 pirq = gsi;
69
70 map_irq.domid = DOMID_SELF;
71 map_irq.type = MAP_PIRQ_TYPE_GSI;
72 map_irq.index = gsi;
73 map_irq.pirq = pirq;
74
75 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
76 if (rc) {
77 printk(KERN_WARNING "xen map irq failed %d\n", rc);
78 return -1;
79 }
80
81 if (triggering == ACPI_EDGE_SENSITIVE) {
82 shareable = 0;
83 name = "ioapic-edge";
84 } else {
85 shareable = 1;
86 name = "ioapic-level";
87 }
88
89 if (gsi_override >= 0)
90 gsi = gsi_override;
91
92 irq = xen_bind_pirq_gsi_to_irq(gsi, map_irq.pirq, shareable, name);
93 if (irq < 0)
94 goto out;
95
96 printk(KERN_DEBUG "xen: --> pirq=%d -> irq=%d (gsi=%d)\n", map_irq.pirq, irq, gsi);
97out:
98 return irq;
99}
100
101static int acpi_register_gsi_xen_hvm(struct device *dev, u32 gsi,
102 int trigger, int polarity)
103{
104 if (!xen_hvm_domain())
105 return -1;
106
107 return xen_register_pirq(gsi, -1 /* no GSI override */, trigger,
108 false /* no mapping of GSI to PIRQ */);
109}
110
111#ifdef CONFIG_XEN_DOM0
112static int xen_register_gsi(u32 gsi, int gsi_override, int triggering, int polarity)
113{
114 int rc, irq;
115 struct physdev_setup_gsi setup_gsi;
116
117 if (!xen_pv_domain())
118 return -1;
119
120 printk(KERN_DEBUG "xen: registering gsi %u triggering %d polarity %d\n",
121 gsi, triggering, polarity);
122
123 irq = xen_register_pirq(gsi, gsi_override, triggering, true);
124
125 setup_gsi.gsi = gsi;
126 setup_gsi.triggering = (triggering == ACPI_EDGE_SENSITIVE ? 0 : 1);
127 setup_gsi.polarity = (polarity == ACPI_ACTIVE_HIGH ? 0 : 1);
128
129 rc = HYPERVISOR_physdev_op(PHYSDEVOP_setup_gsi, &setup_gsi);
130 if (rc == -EEXIST)
131 printk(KERN_INFO "Already setup the GSI :%d\n", gsi);
132 else if (rc) {
133 printk(KERN_ERR "Failed to setup GSI :%d, err_code:%d\n",
134 gsi, rc);
135 }
136
137 return irq;
138}
139
140static int acpi_register_gsi_xen(struct device *dev, u32 gsi,
141 int trigger, int polarity)
142{
143 return xen_register_gsi(gsi, -1 /* no GSI override */, trigger, polarity);
144}
145#endif
146#endif
147
148#if defined(CONFIG_PCI_MSI)
149#include <linux/msi.h>
150#include <asm/msidef.h>
151
152struct xen_pci_frontend_ops *xen_pci_frontend;
153EXPORT_SYMBOL_GPL(xen_pci_frontend);
154
155static int xen_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
156{
157 int irq, ret, i;
158 struct msi_desc *msidesc;
159 int *v;
160
161 v = kzalloc(sizeof(int) * max(1, nvec), GFP_KERNEL);
162 if (!v)
163 return -ENOMEM;
164
165 if (type == PCI_CAP_ID_MSIX)
166 ret = xen_pci_frontend_enable_msix(dev, v, nvec);
167 else
168 ret = xen_pci_frontend_enable_msi(dev, v);
169 if (ret)
170 goto error;
171 i = 0;
172 list_for_each_entry(msidesc, &dev->msi_list, list) {
173 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, v[i], 0,
174 (type == PCI_CAP_ID_MSIX) ?
175 "pcifront-msi-x" :
176 "pcifront-msi",
177 DOMID_SELF);
178 if (irq < 0)
179 goto free;
180 i++;
181 }
182 kfree(v);
183 return 0;
184
185error:
186 dev_err(&dev->dev, "Xen PCI frontend has not registered MSI/MSI-X support!\n");
187free:
188 kfree(v);
189 return ret;
190}
191
192#define XEN_PIRQ_MSI_DATA (MSI_DATA_TRIGGER_EDGE | \
193 MSI_DATA_LEVEL_ASSERT | (3 << 8) | MSI_DATA_VECTOR(0))
194
195static void xen_msi_compose_msg(struct pci_dev *pdev, unsigned int pirq,
196 struct msi_msg *msg)
197{
198 /* We set vector == 0 to tell the hypervisor we don't care about it,
199 * but we want a pirq setup instead.
200 * We use the dest_id field to pass the pirq that we want. */
201 msg->address_hi = MSI_ADDR_BASE_HI | MSI_ADDR_EXT_DEST_ID(pirq);
202 msg->address_lo =
203 MSI_ADDR_BASE_LO |
204 MSI_ADDR_DEST_MODE_PHYSICAL |
205 MSI_ADDR_REDIRECTION_CPU |
206 MSI_ADDR_DEST_ID(pirq);
207
208 msg->data = XEN_PIRQ_MSI_DATA;
209}
210
211static int xen_hvm_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
212{
213 int irq, pirq;
214 struct msi_desc *msidesc;
215 struct msi_msg msg;
216
217 list_for_each_entry(msidesc, &dev->msi_list, list) {
218 __read_msi_msg(msidesc, &msg);
219 pirq = MSI_ADDR_EXT_DEST_ID(msg.address_hi) |
220 ((msg.address_lo >> MSI_ADDR_DEST_ID_SHIFT) & 0xff);
221 if (msg.data != XEN_PIRQ_MSI_DATA ||
222 xen_irq_from_pirq(pirq) < 0) {
223 pirq = xen_allocate_pirq_msi(dev, msidesc);
224 if (pirq < 0)
225 goto error;
226 xen_msi_compose_msg(dev, pirq, &msg);
227 __write_msi_msg(msidesc, &msg);
228 dev_dbg(&dev->dev, "xen: msi bound to pirq=%d\n", pirq);
229 } else {
230 dev_dbg(&dev->dev,
231 "xen: msi already bound to pirq=%d\n", pirq);
232 }
233 irq = xen_bind_pirq_msi_to_irq(dev, msidesc, pirq, 0,
234 (type == PCI_CAP_ID_MSIX) ?
235 "msi-x" : "msi",
236 DOMID_SELF);
237 if (irq < 0)
238 goto error;
239 dev_dbg(&dev->dev,
240 "xen: msi --> pirq=%d --> irq=%d\n", pirq, irq);
241 }
242 return 0;
243
244error:
245 dev_err(&dev->dev,
246 "Xen PCI frontend has not registered MSI/MSI-X support!\n");
247 return -ENODEV;
248}
249
250#ifdef CONFIG_XEN_DOM0
251static int xen_initdom_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
252{
253 int ret = 0;
254 struct msi_desc *msidesc;
255
256 list_for_each_entry(msidesc, &dev->msi_list, list) {
257 struct physdev_map_pirq map_irq;
258 domid_t domid;
259
260 domid = ret = xen_find_device_domain_owner(dev);
261 /* N.B. Casting int's -ENODEV to uint16_t results in 0xFFED,
262 * hence check ret value for < 0. */
263 if (ret < 0)
264 domid = DOMID_SELF;
265
266 memset(&map_irq, 0, sizeof(map_irq));
267 map_irq.domid = domid;
268 map_irq.type = MAP_PIRQ_TYPE_MSI;
269 map_irq.index = -1;
270 map_irq.pirq = -1;
271 map_irq.bus = dev->bus->number;
272 map_irq.devfn = dev->devfn;
273
274 if (type == PCI_CAP_ID_MSIX) {
275 int pos;
276 u32 table_offset, bir;
277
278 pos = pci_find_capability(dev, PCI_CAP_ID_MSIX);
279
280 pci_read_config_dword(dev, pos + PCI_MSIX_TABLE,
281 &table_offset);
282 bir = (u8)(table_offset & PCI_MSIX_FLAGS_BIRMASK);
283
284 map_irq.table_base = pci_resource_start(dev, bir);
285 map_irq.entry_nr = msidesc->msi_attrib.entry_nr;
286 }
287
288 ret = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
289 if (ret) {
290 dev_warn(&dev->dev, "xen map irq failed %d for %d domain\n",
291 ret, domid);
292 goto out;
293 }
294
295 ret = xen_bind_pirq_msi_to_irq(dev, msidesc,
296 map_irq.pirq, map_irq.index,
297 (type == PCI_CAP_ID_MSIX) ?
298 "msi-x" : "msi",
299 domid);
300 if (ret < 0)
301 goto out;
302 }
303 ret = 0;
304out:
305 return ret;
306}
307#endif
308
309static void xen_teardown_msi_irqs(struct pci_dev *dev)
310{
311 struct msi_desc *msidesc;
312
313 msidesc = list_entry(dev->msi_list.next, struct msi_desc, list);
314 if (msidesc->msi_attrib.is_msix)
315 xen_pci_frontend_disable_msix(dev);
316 else
317 xen_pci_frontend_disable_msi(dev);
318
319 /* Free the IRQ's and the msidesc using the generic code. */
320 default_teardown_msi_irqs(dev);
321}
322
323static void xen_teardown_msi_irq(unsigned int irq)
324{
325 xen_destroy_irq(irq);
326}
327
328#endif
329
330int __init pci_xen_init(void)
331{
332 if (!xen_pv_domain() || xen_initial_domain())
333 return -ENODEV;
334
335 printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
336
337 pcibios_set_cache_line_size();
338
339 pcibios_enable_irq = xen_pcifront_enable_irq;
340 pcibios_disable_irq = NULL;
341
342#ifdef CONFIG_ACPI
343 /* Keep ACPI out of the picture */
344 acpi_noirq = 1;
345#endif
346
347#ifdef CONFIG_PCI_MSI
348 x86_msi.setup_msi_irqs = xen_setup_msi_irqs;
349 x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
350 x86_msi.teardown_msi_irqs = xen_teardown_msi_irqs;
351#endif
352 return 0;
353}
354
355int __init pci_xen_hvm_init(void)
356{
357 if (!xen_feature(XENFEAT_hvm_pirqs))
358 return 0;
359
360#ifdef CONFIG_ACPI
361 /*
362 * We don't want to change the actual ACPI delivery model,
363 * just how GSIs get registered.
364 */
365 __acpi_register_gsi = acpi_register_gsi_xen_hvm;
366#endif
367
368#ifdef CONFIG_PCI_MSI
369 x86_msi.setup_msi_irqs = xen_hvm_setup_msi_irqs;
370 x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
371#endif
372 return 0;
373}
374
375#ifdef CONFIG_XEN_DOM0
376static __init void xen_setup_acpi_sci(void)
377{
378 int rc;
379 int trigger, polarity;
380 int gsi = acpi_sci_override_gsi;
381 int irq = -1;
382 int gsi_override = -1;
383
384 if (!gsi)
385 return;
386
387 rc = acpi_get_override_irq(gsi, &trigger, &polarity);
388 if (rc) {
389 printk(KERN_WARNING "xen: acpi_get_override_irq failed for acpi"
390 " sci, rc=%d\n", rc);
391 return;
392 }
393 trigger = trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE;
394 polarity = polarity ? ACPI_ACTIVE_LOW : ACPI_ACTIVE_HIGH;
395
396 printk(KERN_INFO "xen: sci override: global_irq=%d trigger=%d "
397 "polarity=%d\n", gsi, trigger, polarity);
398
399 /* Before we bind the GSI to a Linux IRQ, check whether
400 * we need to override it with bus_irq (IRQ) value. Usually for
401 * IRQs below IRQ_LEGACY_IRQ this holds IRQ == GSI, as so:
402 * ACPI: INT_SRC_OVR (bus 0 bus_irq 9 global_irq 9 low level)
403 * but there are oddballs where the IRQ != GSI:
404 * ACPI: INT_SRC_OVR (bus 0 bus_irq 9 global_irq 20 low level)
405 * which ends up being: gsi_to_irq[9] == 20
406 * (which is what acpi_gsi_to_irq ends up calling when starting the
407 * the ACPI interpreter and keels over since IRQ 9 has not been
408 * setup as we had setup IRQ 20 for it).
409 */
410 if (acpi_gsi_to_irq(gsi, &irq) == 0) {
411 /* Use the provided value if it's valid. */
412 if (irq >= 0)
413 gsi_override = irq;
414 }
415
416 gsi = xen_register_gsi(gsi, gsi_override, trigger, polarity);
417 printk(KERN_INFO "xen: acpi sci %d\n", gsi);
418
419 return;
420}
421
422int __init pci_xen_initial_domain(void)
423{
424 int irq;
425
426#ifdef CONFIG_PCI_MSI
427 x86_msi.setup_msi_irqs = xen_initdom_setup_msi_irqs;
428 x86_msi.teardown_msi_irq = xen_teardown_msi_irq;
429#endif
430 xen_setup_acpi_sci();
431 __acpi_register_gsi = acpi_register_gsi_xen;
432 /* Pre-allocate legacy irqs */
433 for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
434 int trigger, polarity;
435
436 if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
437 continue;
438
439 xen_register_pirq(irq, -1 /* no GSI override */,
440 trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
441 true /* Map GSI to PIRQ */);
442 }
443 if (0 == nr_ioapics) {
444 for (irq = 0; irq < NR_IRQS_LEGACY; irq++)
445 xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
446 }
447 return 0;
448}
449
450struct xen_device_domain_owner {
451 domid_t domain;
452 struct pci_dev *dev;
453 struct list_head list;
454};
455
456static DEFINE_SPINLOCK(dev_domain_list_spinlock);
457static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
458
459static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
460{
461 struct xen_device_domain_owner *owner;
462
463 list_for_each_entry(owner, &dev_domain_list, list) {
464 if (owner->dev == dev)
465 return owner;
466 }
467 return NULL;
468}
469
470int xen_find_device_domain_owner(struct pci_dev *dev)
471{
472 struct xen_device_domain_owner *owner;
473 int domain = -ENODEV;
474
475 spin_lock(&dev_domain_list_spinlock);
476 owner = find_device(dev);
477 if (owner)
478 domain = owner->domain;
479 spin_unlock(&dev_domain_list_spinlock);
480 return domain;
481}
482EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
483
484int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
485{
486 struct xen_device_domain_owner *owner;
487
488 owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
489 if (!owner)
490 return -ENODEV;
491
492 spin_lock(&dev_domain_list_spinlock);
493 if (find_device(dev)) {
494 spin_unlock(&dev_domain_list_spinlock);
495 kfree(owner);
496 return -EEXIST;
497 }
498 owner->domain = domain;
499 owner->dev = dev;
500 list_add_tail(&owner->list, &dev_domain_list);
501 spin_unlock(&dev_domain_list_spinlock);
502 return 0;
503}
504EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
505
506int xen_unregister_device_domain_owner(struct pci_dev *dev)
507{
508 struct xen_device_domain_owner *owner;
509
510 spin_lock(&dev_domain_list_spinlock);
511 owner = find_device(dev);
512 if (!owner) {
513 spin_unlock(&dev_domain_list_spinlock);
514 return -ENODEV;
515 }
516 list_del(&owner->list);
517 spin_unlock(&dev_domain_list_spinlock);
518 kfree(owner);
519 return 0;
520}
521EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
522#endif
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
33static 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
65static 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);
103out:
104 return irq;
105}
106
107static 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
118static 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
146static 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
157struct xen_pci_frontend_ops *xen_pci_frontend;
158EXPORT_SYMBOL_GPL(xen_pci_frontend);
159
160struct 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
165static struct xen_msi_ops xen_msi_ops __ro_after_init;
166
167static 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
203error:
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);
208free:
209 kfree(v);
210 return ret;
211}
212
213static 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
229static 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
259error:
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
266static bool __read_mostly pci_seg_supported = true;
267
268static 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;
350out:
351 return ret;
352}
353
354bool 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
387static 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 msidesc->irq = 0;
396 }
397}
398
399static void xen_pv_teardown_msi_irqs(struct pci_dev *dev)
400{
401 if (dev->msix_enabled)
402 xen_pci_frontend_disable_msix(dev);
403 else
404 xen_pci_frontend_disable_msi(dev);
405
406 xen_teardown_msi_irqs(dev);
407}
408
409static int xen_msi_domain_alloc_irqs(struct irq_domain *domain,
410 struct device *dev, int nvec)
411{
412 int type;
413
414 if (WARN_ON_ONCE(!dev_is_pci(dev)))
415 return -EINVAL;
416
417 type = to_pci_dev(dev)->msix_enabled ? PCI_CAP_ID_MSIX : PCI_CAP_ID_MSI;
418
419 return xen_msi_ops.setup_msi_irqs(to_pci_dev(dev), nvec, type);
420}
421
422static void xen_msi_domain_free_irqs(struct irq_domain *domain,
423 struct device *dev)
424{
425 if (WARN_ON_ONCE(!dev_is_pci(dev)))
426 return;
427
428 xen_msi_ops.teardown_msi_irqs(to_pci_dev(dev));
429}
430
431static struct msi_domain_ops xen_pci_msi_domain_ops = {
432 .domain_alloc_irqs = xen_msi_domain_alloc_irqs,
433 .domain_free_irqs = xen_msi_domain_free_irqs,
434};
435
436static struct msi_domain_info xen_pci_msi_domain_info = {
437 .flags = MSI_FLAG_PCI_MSIX | MSI_FLAG_FREE_MSI_DESCS | MSI_FLAG_DEV_SYSFS,
438 .ops = &xen_pci_msi_domain_ops,
439};
440
441/*
442 * This irq domain is a blatant violation of the irq domain design, but
443 * distangling XEN into real irq domains is not a job for mere mortals with
444 * limited XENology. But it's the least dangerous way for a mere mortal to
445 * get rid of the arch_*_msi_irqs() hackery in order to store the irq
446 * domain pointer in struct device. This irq domain wrappery allows to do
447 * that without breaking XEN terminally.
448 */
449static __init struct irq_domain *xen_create_pci_msi_domain(void)
450{
451 struct irq_domain *d = NULL;
452 struct fwnode_handle *fn;
453
454 fn = irq_domain_alloc_named_fwnode("XEN-MSI");
455 if (fn)
456 d = msi_create_irq_domain(fn, &xen_pci_msi_domain_info, NULL);
457
458 /* FIXME: No idea how to survive if this fails */
459 BUG_ON(!d);
460
461 return d;
462}
463
464static __init void xen_setup_pci_msi(void)
465{
466 if (xen_pv_domain()) {
467 if (xen_initial_domain())
468 xen_msi_ops.setup_msi_irqs = xen_initdom_setup_msi_irqs;
469 else
470 xen_msi_ops.setup_msi_irqs = xen_setup_msi_irqs;
471 xen_msi_ops.teardown_msi_irqs = xen_pv_teardown_msi_irqs;
472 } else if (xen_hvm_domain()) {
473 xen_msi_ops.setup_msi_irqs = xen_hvm_setup_msi_irqs;
474 xen_msi_ops.teardown_msi_irqs = xen_teardown_msi_irqs;
475 } else {
476 WARN_ON_ONCE(1);
477 return;
478 }
479
480 /*
481 * Override the PCI/MSI irq domain init function. No point
482 * in allocating the native domain and never use it.
483 */
484 x86_init.irqs.create_pci_msi_domain = xen_create_pci_msi_domain;
485 /*
486 * With XEN PIRQ/Eventchannels in use PCI/MSI[-X] masking is solely
487 * controlled by the hypervisor.
488 */
489 pci_msi_ignore_mask = 1;
490}
491
492#else /* CONFIG_PCI_MSI */
493static inline void xen_setup_pci_msi(void) { }
494#endif /* CONFIG_PCI_MSI */
495
496int __init pci_xen_init(void)
497{
498 if (!xen_pv_domain() || xen_initial_domain())
499 return -ENODEV;
500
501 printk(KERN_INFO "PCI: setting up Xen PCI frontend stub\n");
502
503 pcibios_set_cache_line_size();
504
505 pcibios_enable_irq = xen_pcifront_enable_irq;
506 pcibios_disable_irq = NULL;
507
508 /* Keep ACPI out of the picture */
509 acpi_noirq_set();
510
511 xen_setup_pci_msi();
512 return 0;
513}
514
515#ifdef CONFIG_PCI_MSI
516static void __init xen_hvm_msi_init(void)
517{
518 if (!disable_apic) {
519 /*
520 * If hardware supports (x2)APIC virtualization (as indicated
521 * by hypervisor's leaf 4) then we don't need to use pirqs/
522 * event channels for MSI handling and instead use regular
523 * APIC processing
524 */
525 uint32_t eax = cpuid_eax(xen_cpuid_base() + 4);
526
527 if (((eax & XEN_HVM_CPUID_X2APIC_VIRT) && x2apic_mode) ||
528 ((eax & XEN_HVM_CPUID_APIC_ACCESS_VIRT) && boot_cpu_has(X86_FEATURE_APIC)))
529 return;
530 }
531 xen_setup_pci_msi();
532}
533#endif
534
535int __init pci_xen_hvm_init(void)
536{
537 if (!xen_have_vector_callback || !xen_feature(XENFEAT_hvm_pirqs))
538 return 0;
539
540#ifdef CONFIG_ACPI
541 /*
542 * We don't want to change the actual ACPI delivery model,
543 * just how GSIs get registered.
544 */
545 __acpi_register_gsi = acpi_register_gsi_xen_hvm;
546 __acpi_unregister_gsi = NULL;
547#endif
548
549#ifdef CONFIG_PCI_MSI
550 /*
551 * We need to wait until after x2apic is initialized
552 * before we can set MSI IRQ ops.
553 */
554 x86_platform.apic_post_init = xen_hvm_msi_init;
555#endif
556 return 0;
557}
558
559#ifdef CONFIG_XEN_PV_DOM0
560int __init pci_xen_initial_domain(void)
561{
562 int irq;
563
564 xen_setup_pci_msi();
565 __acpi_register_gsi = acpi_register_gsi_xen;
566 __acpi_unregister_gsi = NULL;
567 /*
568 * Pre-allocate the legacy IRQs. Use NR_LEGACY_IRQS here
569 * because we don't have a PIC and thus nr_legacy_irqs() is zero.
570 */
571 for (irq = 0; irq < NR_IRQS_LEGACY; irq++) {
572 int trigger, polarity;
573
574 if (acpi_get_override_irq(irq, &trigger, &polarity) == -1)
575 continue;
576
577 xen_register_pirq(irq,
578 trigger ? ACPI_LEVEL_SENSITIVE : ACPI_EDGE_SENSITIVE,
579 true /* Map GSI to PIRQ */);
580 }
581 if (0 == nr_ioapics) {
582 for (irq = 0; irq < nr_legacy_irqs(); irq++)
583 xen_bind_pirq_gsi_to_irq(irq, irq, 0, "xt-pic");
584 }
585 return 0;
586}
587#endif
588