xref: /freebsd/sys/x86/x86/msi.c (revision 2a01feab)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2006 Yahoo!, Inc.
5  * All rights reserved.
6  * Written by: John Baldwin <jhb@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the author nor the names of any co-contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 /*
34  * Support for PCI Message Signalled Interrupts (MSI).  MSI interrupts on
35  * x86 are basically APIC messages that the northbridge delivers directly
36  * to the local APICs as if they had come from an I/O APIC.
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_acpi.h"
43 
44 #include <sys/param.h>
45 #include <sys/bus.h>
46 #include <sys/kernel.h>
47 #include <sys/lock.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.h>
50 #include <sys/sx.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #include <x86/apicreg.h>
54 #include <machine/cputypes.h>
55 #include <machine/md_var.h>
56 #include <machine/frame.h>
57 #include <machine/intr_machdep.h>
58 #include <x86/apicvar.h>
59 #include <x86/iommu/iommu_intrmap.h>
60 #include <machine/specialreg.h>
61 #include <dev/pci/pcivar.h>
62 
63 /* Fields in address for Intel MSI messages. */
64 #define	MSI_INTEL_ADDR_DEST		0x000ff000
65 #define	MSI_INTEL_ADDR_RH		0x00000008
66 # define MSI_INTEL_ADDR_RH_ON		0x00000008
67 # define MSI_INTEL_ADDR_RH_OFF		0x00000000
68 #define	MSI_INTEL_ADDR_DM		0x00000004
69 # define MSI_INTEL_ADDR_DM_PHYSICAL	0x00000000
70 # define MSI_INTEL_ADDR_DM_LOGICAL	0x00000004
71 
72 /* Fields in data for Intel MSI messages. */
73 #define	MSI_INTEL_DATA_TRGRMOD		IOART_TRGRMOD	/* Trigger mode. */
74 # define MSI_INTEL_DATA_TRGREDG		IOART_TRGREDG
75 # define MSI_INTEL_DATA_TRGRLVL		IOART_TRGRLVL
76 #define	MSI_INTEL_DATA_LEVEL		0x00004000	/* Polarity. */
77 # define MSI_INTEL_DATA_DEASSERT	0x00000000
78 # define MSI_INTEL_DATA_ASSERT		0x00004000
79 #define	MSI_INTEL_DATA_DELMOD		IOART_DELMOD	/* Delivery mode. */
80 # define MSI_INTEL_DATA_DELFIXED	IOART_DELFIXED
81 # define MSI_INTEL_DATA_DELLOPRI	IOART_DELLOPRI
82 # define MSI_INTEL_DATA_DELSMI		IOART_DELSMI
83 # define MSI_INTEL_DATA_DELNMI		IOART_DELNMI
84 # define MSI_INTEL_DATA_DELINIT		IOART_DELINIT
85 # define MSI_INTEL_DATA_DELEXINT	IOART_DELEXINT
86 #define	MSI_INTEL_DATA_INTVEC		IOART_INTVEC	/* Interrupt vector. */
87 
88 /*
89  * Build Intel MSI message and data values from a source.  AMD64 systems
90  * seem to be compatible, so we use the same function for both.
91  */
92 #define	INTEL_ADDR(msi)							\
93 	(MSI_INTEL_ADDR_BASE | (msi)->msi_cpu << 12 |			\
94 	    MSI_INTEL_ADDR_RH_OFF | MSI_INTEL_ADDR_DM_PHYSICAL)
95 #define	INTEL_DATA(msi)							\
96 	(MSI_INTEL_DATA_TRGREDG | MSI_INTEL_DATA_DELFIXED | (msi)->msi_vector)
97 
98 static MALLOC_DEFINE(M_MSI, "msi", "PCI MSI");
99 
100 /*
101  * MSI sources are bunched into groups.  This is because MSI forces
102  * all of the messages to share the address and data registers and
103  * thus certain properties (such as the local APIC ID target on x86).
104  * Each group has a 'first' source that contains information global to
105  * the group.  These fields are marked with (g) below.
106  *
107  * Note that local APIC ID is kind of special.  Each message will be
108  * assigned an ID by the system; however, a group will use the ID from
109  * the first message.
110  *
111  * For MSI-X, each message is isolated.
112  */
113 struct msi_intsrc {
114 	struct intsrc msi_intsrc;
115 	device_t msi_dev;		/* Owning device. (g) */
116 	struct msi_intsrc *msi_first;	/* First source in group. */
117 	u_int msi_irq;			/* IRQ cookie. */
118 	u_int msi_msix;			/* MSI-X message. */
119 	u_int msi_vector:8;		/* IDT vector. */
120 	u_int msi_cpu;			/* Local APIC ID. (g) */
121 	u_int msi_count:8;		/* Messages in this group. (g) */
122 	u_int msi_maxcount:8;		/* Alignment for this group. (g) */
123 	u_int *msi_irqs;		/* Group's IRQ list. (g) */
124 	u_int msi_remap_cookie;
125 };
126 
127 static void	msi_create_source(void);
128 static void	msi_enable_source(struct intsrc *isrc);
129 static void	msi_disable_source(struct intsrc *isrc, int eoi);
130 static void	msi_eoi_source(struct intsrc *isrc);
131 static void	msi_enable_intr(struct intsrc *isrc);
132 static void	msi_disable_intr(struct intsrc *isrc);
133 static int	msi_vector(struct intsrc *isrc);
134 static int	msi_source_pending(struct intsrc *isrc);
135 static int	msi_config_intr(struct intsrc *isrc, enum intr_trigger trig,
136 		    enum intr_polarity pol);
137 static int	msi_assign_cpu(struct intsrc *isrc, u_int apic_id);
138 
139 struct pic msi_pic = {
140 	.pic_enable_source = msi_enable_source,
141 	.pic_disable_source = msi_disable_source,
142 	.pic_eoi_source = msi_eoi_source,
143 	.pic_enable_intr = msi_enable_intr,
144 	.pic_disable_intr = msi_disable_intr,
145 	.pic_vector = msi_vector,
146 	.pic_source_pending = msi_source_pending,
147 	.pic_suspend = NULL,
148 	.pic_resume = NULL,
149 	.pic_config_intr = msi_config_intr,
150 	.pic_assign_cpu = msi_assign_cpu,
151 	.pic_reprogram_pin = NULL,
152 };
153 
154 u_int first_msi_irq;
155 
156 #ifdef SMP
157 /**
158  * Xen hypervisors prior to 4.6.0 do not properly handle updates to
159  * enabled MSI-X table entries.  Allow migration of MSI-X interrupts
160  * to be disabled via a tunable. Values have the following meaning:
161  *
162  * -1: automatic detection by FreeBSD
163  *  0: enable migration
164  *  1: disable migration
165  */
166 int msix_disable_migration = -1;
167 SYSCTL_INT(_machdep, OID_AUTO, disable_msix_migration, CTLFLAG_RDTUN,
168     &msix_disable_migration, 0,
169     "Disable migration of MSI-X interrupts between CPUs");
170 #endif
171 
172 static int msi_enabled;
173 static u_int msi_last_irq;
174 static struct mtx msi_lock;
175 
176 static void
177 msi_enable_source(struct intsrc *isrc)
178 {
179 }
180 
181 static void
182 msi_disable_source(struct intsrc *isrc, int eoi)
183 {
184 
185 	if (eoi == PIC_EOI)
186 		lapic_eoi();
187 }
188 
189 static void
190 msi_eoi_source(struct intsrc *isrc)
191 {
192 
193 	lapic_eoi();
194 }
195 
196 static void
197 msi_enable_intr(struct intsrc *isrc)
198 {
199 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
200 
201 	apic_enable_vector(msi->msi_cpu, msi->msi_vector);
202 }
203 
204 static void
205 msi_disable_intr(struct intsrc *isrc)
206 {
207 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
208 
209 	apic_disable_vector(msi->msi_cpu, msi->msi_vector);
210 }
211 
212 static int
213 msi_vector(struct intsrc *isrc)
214 {
215 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
216 
217 	return (msi->msi_irq);
218 }
219 
220 static int
221 msi_source_pending(struct intsrc *isrc)
222 {
223 
224 	return (0);
225 }
226 
227 static int
228 msi_config_intr(struct intsrc *isrc, enum intr_trigger trig,
229     enum intr_polarity pol)
230 {
231 
232 	return (ENODEV);
233 }
234 
235 static int
236 msi_assign_cpu(struct intsrc *isrc, u_int apic_id)
237 {
238 	struct msi_intsrc *sib, *msi = (struct msi_intsrc *)isrc;
239 	int old_vector;
240 	u_int old_id;
241 	int i, vector;
242 
243 	/*
244 	 * Only allow CPUs to be assigned to the first message for an
245 	 * MSI group.
246 	 */
247 	if (msi->msi_first != msi)
248 		return (EINVAL);
249 
250 #ifdef SMP
251 	if (msix_disable_migration && msi->msi_msix)
252 		return (EINVAL);
253 #endif
254 
255 	/* Store information to free existing irq. */
256 	old_vector = msi->msi_vector;
257 	old_id = msi->msi_cpu;
258 	if (old_id == apic_id)
259 		return (0);
260 
261 	/* Allocate IDT vectors on this cpu. */
262 	if (msi->msi_count > 1) {
263 		KASSERT(msi->msi_msix == 0, ("MSI-X message group"));
264 		vector = apic_alloc_vectors(apic_id, msi->msi_irqs,
265 		    msi->msi_count, msi->msi_maxcount);
266 	} else
267 		vector = apic_alloc_vector(apic_id, msi->msi_irq);
268 	if (vector == 0)
269 		return (ENOSPC);
270 
271 	msi->msi_cpu = apic_id;
272 	msi->msi_vector = vector;
273 	if (msi->msi_intsrc.is_handlers > 0)
274 		apic_enable_vector(msi->msi_cpu, msi->msi_vector);
275 	if (bootverbose)
276 		printf("msi: Assigning %s IRQ %d to local APIC %u vector %u\n",
277 		    msi->msi_msix ? "MSI-X" : "MSI", msi->msi_irq,
278 		    msi->msi_cpu, msi->msi_vector);
279 	for (i = 1; i < msi->msi_count; i++) {
280 		sib = (struct msi_intsrc *)intr_lookup_source(msi->msi_irqs[i]);
281 		sib->msi_cpu = apic_id;
282 		sib->msi_vector = vector + i;
283 		if (sib->msi_intsrc.is_handlers > 0)
284 			apic_enable_vector(sib->msi_cpu, sib->msi_vector);
285 		if (bootverbose)
286 			printf(
287 		    "msi: Assigning MSI IRQ %d to local APIC %u vector %u\n",
288 			    sib->msi_irq, sib->msi_cpu, sib->msi_vector);
289 	}
290 	BUS_REMAP_INTR(device_get_parent(msi->msi_dev), msi->msi_dev,
291 	    msi->msi_irq);
292 
293 	/*
294 	 * Free the old vector after the new one is established.  This is done
295 	 * to prevent races where we could miss an interrupt.
296 	 */
297 	if (msi->msi_intsrc.is_handlers > 0)
298 		apic_disable_vector(old_id, old_vector);
299 	apic_free_vector(old_id, old_vector, msi->msi_irq);
300 	for (i = 1; i < msi->msi_count; i++) {
301 		sib = (struct msi_intsrc *)intr_lookup_source(msi->msi_irqs[i]);
302 		if (sib->msi_intsrc.is_handlers > 0)
303 			apic_disable_vector(old_id, old_vector + i);
304 		apic_free_vector(old_id, old_vector + i, msi->msi_irqs[i]);
305 	}
306 	return (0);
307 }
308 
309 void
310 msi_init(void)
311 {
312 
313 	/* Check if we have a supported CPU. */
314 	switch (cpu_vendor_id) {
315 	case CPU_VENDOR_INTEL:
316 	case CPU_VENDOR_AMD:
317 		break;
318 	case CPU_VENDOR_CENTAUR:
319 		if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
320 		    CPUID_TO_MODEL(cpu_id) >= 0xf)
321 			break;
322 		/* FALLTHROUGH */
323 	default:
324 		return;
325 	}
326 
327 #ifdef SMP
328 	if (msix_disable_migration == -1) {
329 		/* The default is to allow migration of MSI-X interrupts. */
330 		msix_disable_migration = 0;
331 	}
332 #endif
333 
334 	first_msi_irq = max(MINIMUM_MSI_INT, num_io_irqs);
335 	num_io_irqs = first_msi_irq + NUM_MSI_INTS;
336 
337 	msi_enabled = 1;
338 	intr_register_pic(&msi_pic);
339 	mtx_init(&msi_lock, "msi", NULL, MTX_DEF);
340 }
341 
342 static void
343 msi_create_source(void)
344 {
345 	struct msi_intsrc *msi;
346 	u_int irq;
347 
348 	mtx_lock(&msi_lock);
349 	if (msi_last_irq >= NUM_MSI_INTS) {
350 		mtx_unlock(&msi_lock);
351 		return;
352 	}
353 	irq = msi_last_irq + first_msi_irq;
354 	msi_last_irq++;
355 	mtx_unlock(&msi_lock);
356 
357 	msi = malloc(sizeof(struct msi_intsrc), M_MSI, M_WAITOK | M_ZERO);
358 	msi->msi_intsrc.is_pic = &msi_pic;
359 	msi->msi_irq = irq;
360 	intr_register_source(&msi->msi_intsrc);
361 	nexus_add_irq(irq);
362 }
363 
364 /*
365  * Try to allocate 'count' interrupt sources with contiguous IDT values.
366  */
367 int
368 msi_alloc(device_t dev, int count, int maxcount, int *irqs)
369 {
370 	struct msi_intsrc *msi, *fsrc;
371 	u_int cpu, domain, *mirqs;
372 	int cnt, i, vector;
373 #ifdef ACPI_DMAR
374 	u_int cookies[count];
375 	int error;
376 #endif
377 
378 	if (!msi_enabled)
379 		return (ENXIO);
380 
381 	if (bus_get_domain(dev, &domain) != 0)
382 		domain = 0;
383 
384 	if (count > 1)
385 		mirqs = malloc(count * sizeof(*mirqs), M_MSI, M_WAITOK);
386 	else
387 		mirqs = NULL;
388 again:
389 	mtx_lock(&msi_lock);
390 
391 	/* Try to find 'count' free IRQs. */
392 	cnt = 0;
393 	for (i = first_msi_irq; i < first_msi_irq + NUM_MSI_INTS; i++) {
394 		msi = (struct msi_intsrc *)intr_lookup_source(i);
395 
396 		/* End of allocated sources, so break. */
397 		if (msi == NULL)
398 			break;
399 
400 		/* If this is a free one, save its IRQ in the array. */
401 		if (msi->msi_dev == NULL) {
402 			irqs[cnt] = i;
403 			cnt++;
404 			if (cnt == count)
405 				break;
406 		}
407 	}
408 
409 	/* Do we need to create some new sources? */
410 	if (cnt < count) {
411 		/* If we would exceed the max, give up. */
412 		if (i + (count - cnt) >= first_msi_irq + NUM_MSI_INTS) {
413 			mtx_unlock(&msi_lock);
414 			free(mirqs, M_MSI);
415 			return (ENXIO);
416 		}
417 		mtx_unlock(&msi_lock);
418 
419 		/* We need count - cnt more sources. */
420 		while (cnt < count) {
421 			msi_create_source();
422 			cnt++;
423 		}
424 		goto again;
425 	}
426 
427 	/* Ok, we now have the IRQs allocated. */
428 	KASSERT(cnt == count, ("count mismatch"));
429 
430 	/* Allocate 'count' IDT vectors. */
431 	cpu = intr_next_cpu(domain);
432 	vector = apic_alloc_vectors(cpu, irqs, count, maxcount);
433 	if (vector == 0) {
434 		mtx_unlock(&msi_lock);
435 		free(mirqs, M_MSI);
436 		return (ENOSPC);
437 	}
438 
439 #ifdef ACPI_DMAR
440 	mtx_unlock(&msi_lock);
441 	error = iommu_alloc_msi_intr(dev, cookies, count);
442 	mtx_lock(&msi_lock);
443 	if (error == EOPNOTSUPP)
444 		error = 0;
445 	if (error != 0) {
446 		for (i = 0; i < count; i++)
447 			apic_free_vector(cpu, vector + i, irqs[i]);
448 		free(mirqs, M_MSI);
449 		return (error);
450 	}
451 	for (i = 0; i < count; i++) {
452 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
453 		msi->msi_remap_cookie = cookies[i];
454 	}
455 #endif
456 
457 	/* Assign IDT vectors and make these messages owned by 'dev'. */
458 	fsrc = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
459 	for (i = 0; i < count; i++) {
460 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
461 		msi->msi_cpu = cpu;
462 		msi->msi_dev = dev;
463 		msi->msi_vector = vector + i;
464 		if (bootverbose)
465 			printf(
466 		    "msi: routing MSI IRQ %d to local APIC %u vector %u\n",
467 			    msi->msi_irq, msi->msi_cpu, msi->msi_vector);
468 		msi->msi_first = fsrc;
469 		KASSERT(msi->msi_intsrc.is_handlers == 0,
470 		    ("dead MSI has handlers"));
471 	}
472 	fsrc->msi_count = count;
473 	fsrc->msi_maxcount = maxcount;
474 	if (count > 1)
475 		bcopy(irqs, mirqs, count * sizeof(*mirqs));
476 	fsrc->msi_irqs = mirqs;
477 	mtx_unlock(&msi_lock);
478 	return (0);
479 }
480 
481 int
482 msi_release(int *irqs, int count)
483 {
484 	struct msi_intsrc *msi, *first;
485 	int i;
486 
487 	mtx_lock(&msi_lock);
488 	first = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
489 	if (first == NULL) {
490 		mtx_unlock(&msi_lock);
491 		return (ENOENT);
492 	}
493 
494 	/* Make sure this isn't an MSI-X message. */
495 	if (first->msi_msix) {
496 		mtx_unlock(&msi_lock);
497 		return (EINVAL);
498 	}
499 
500 	/* Make sure this message is allocated to a group. */
501 	if (first->msi_first == NULL) {
502 		mtx_unlock(&msi_lock);
503 		return (ENXIO);
504 	}
505 
506 	/*
507 	 * Make sure this is the start of a group and that we are releasing
508 	 * the entire group.
509 	 */
510 	if (first->msi_first != first || first->msi_count != count) {
511 		mtx_unlock(&msi_lock);
512 		return (EINVAL);
513 	}
514 	KASSERT(first->msi_dev != NULL, ("unowned group"));
515 
516 	/* Clear all the extra messages in the group. */
517 	for (i = 1; i < count; i++) {
518 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
519 		KASSERT(msi->msi_first == first, ("message not in group"));
520 		KASSERT(msi->msi_dev == first->msi_dev, ("owner mismatch"));
521 #ifdef ACPI_DMAR
522 		iommu_unmap_msi_intr(first->msi_dev, msi->msi_remap_cookie);
523 #endif
524 		msi->msi_first = NULL;
525 		msi->msi_dev = NULL;
526 		apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
527 		msi->msi_vector = 0;
528 	}
529 
530 	/* Clear out the first message. */
531 #ifdef ACPI_DMAR
532 	mtx_unlock(&msi_lock);
533 	iommu_unmap_msi_intr(first->msi_dev, first->msi_remap_cookie);
534 	mtx_lock(&msi_lock);
535 #endif
536 	first->msi_first = NULL;
537 	first->msi_dev = NULL;
538 	apic_free_vector(first->msi_cpu, first->msi_vector, first->msi_irq);
539 	first->msi_vector = 0;
540 	first->msi_count = 0;
541 	first->msi_maxcount = 0;
542 	free(first->msi_irqs, M_MSI);
543 	first->msi_irqs = NULL;
544 
545 	mtx_unlock(&msi_lock);
546 	return (0);
547 }
548 
549 int
550 msi_map(int irq, uint64_t *addr, uint32_t *data)
551 {
552 	struct msi_intsrc *msi;
553 	int error;
554 #ifdef ACPI_DMAR
555 	struct msi_intsrc *msi1;
556 	int i, k;
557 #endif
558 
559 	mtx_lock(&msi_lock);
560 	msi = (struct msi_intsrc *)intr_lookup_source(irq);
561 	if (msi == NULL) {
562 		mtx_unlock(&msi_lock);
563 		return (ENOENT);
564 	}
565 
566 	/* Make sure this message is allocated to a device. */
567 	if (msi->msi_dev == NULL) {
568 		mtx_unlock(&msi_lock);
569 		return (ENXIO);
570 	}
571 
572 	/*
573 	 * If this message isn't an MSI-X message, make sure it's part
574 	 * of a group, and switch to the first message in the
575 	 * group.
576 	 */
577 	if (!msi->msi_msix) {
578 		if (msi->msi_first == NULL) {
579 			mtx_unlock(&msi_lock);
580 			return (ENXIO);
581 		}
582 		msi = msi->msi_first;
583 	}
584 
585 #ifdef ACPI_DMAR
586 	if (!msi->msi_msix) {
587 		for (k = msi->msi_count - 1, i = first_msi_irq; k > 0 &&
588 		    i < first_msi_irq + NUM_MSI_INTS; i++) {
589 			if (i == msi->msi_irq)
590 				continue;
591 			msi1 = (struct msi_intsrc *)intr_lookup_source(i);
592 			if (!msi1->msi_msix && msi1->msi_first == msi) {
593 				mtx_unlock(&msi_lock);
594 				iommu_map_msi_intr(msi1->msi_dev,
595 				    msi1->msi_cpu, msi1->msi_vector,
596 				    msi1->msi_remap_cookie, NULL, NULL);
597 				k--;
598 				mtx_lock(&msi_lock);
599 			}
600 		}
601 	}
602 	mtx_unlock(&msi_lock);
603 	error = iommu_map_msi_intr(msi->msi_dev, msi->msi_cpu,
604 	    msi->msi_vector, msi->msi_remap_cookie, addr, data);
605 #else
606 	mtx_unlock(&msi_lock);
607 	error = EOPNOTSUPP;
608 #endif
609 	if (error == EOPNOTSUPP) {
610 		*addr = INTEL_ADDR(msi);
611 		*data = INTEL_DATA(msi);
612 		error = 0;
613 	}
614 	return (error);
615 }
616 
617 int
618 msix_alloc(device_t dev, int *irq)
619 {
620 	struct msi_intsrc *msi;
621 	u_int cpu, domain;
622 	int i, vector;
623 #ifdef ACPI_DMAR
624 	u_int cookie;
625 	int error;
626 #endif
627 
628 	if (!msi_enabled)
629 		return (ENXIO);
630 
631 	if (bus_get_domain(dev, &domain) != 0)
632 		domain = 0;
633 
634 again:
635 	mtx_lock(&msi_lock);
636 
637 	/* Find a free IRQ. */
638 	for (i = first_msi_irq; i < first_msi_irq + NUM_MSI_INTS; i++) {
639 		msi = (struct msi_intsrc *)intr_lookup_source(i);
640 
641 		/* End of allocated sources, so break. */
642 		if (msi == NULL)
643 			break;
644 
645 		/* Stop at the first free source. */
646 		if (msi->msi_dev == NULL)
647 			break;
648 	}
649 
650 	/* Do we need to create a new source? */
651 	if (msi == NULL) {
652 		/* If we would exceed the max, give up. */
653 		if (i + 1 >= first_msi_irq + NUM_MSI_INTS) {
654 			mtx_unlock(&msi_lock);
655 			return (ENXIO);
656 		}
657 		mtx_unlock(&msi_lock);
658 
659 		/* Create a new source. */
660 		msi_create_source();
661 		goto again;
662 	}
663 
664 	/* Allocate an IDT vector. */
665 	cpu = intr_next_cpu(domain);
666 	vector = apic_alloc_vector(cpu, i);
667 	if (vector == 0) {
668 		mtx_unlock(&msi_lock);
669 		return (ENOSPC);
670 	}
671 
672 	msi->msi_dev = dev;
673 #ifdef ACPI_DMAR
674 	mtx_unlock(&msi_lock);
675 	error = iommu_alloc_msi_intr(dev, &cookie, 1);
676 	mtx_lock(&msi_lock);
677 	if (error == EOPNOTSUPP)
678 		error = 0;
679 	if (error != 0) {
680 		msi->msi_dev = NULL;
681 		apic_free_vector(cpu, vector, i);
682 		return (error);
683 	}
684 	msi->msi_remap_cookie = cookie;
685 #endif
686 
687 	if (bootverbose)
688 		printf("msi: routing MSI-X IRQ %d to local APIC %u vector %u\n",
689 		    msi->msi_irq, cpu, vector);
690 
691 	/* Setup source. */
692 	msi->msi_cpu = cpu;
693 	msi->msi_first = msi;
694 	msi->msi_vector = vector;
695 	msi->msi_msix = 1;
696 	msi->msi_count = 1;
697 	msi->msi_maxcount = 1;
698 	msi->msi_irqs = NULL;
699 
700 	KASSERT(msi->msi_intsrc.is_handlers == 0, ("dead MSI-X has handlers"));
701 	mtx_unlock(&msi_lock);
702 
703 	*irq = i;
704 	return (0);
705 }
706 
707 int
708 msix_release(int irq)
709 {
710 	struct msi_intsrc *msi;
711 
712 	mtx_lock(&msi_lock);
713 	msi = (struct msi_intsrc *)intr_lookup_source(irq);
714 	if (msi == NULL) {
715 		mtx_unlock(&msi_lock);
716 		return (ENOENT);
717 	}
718 
719 	/* Make sure this is an MSI-X message. */
720 	if (!msi->msi_msix) {
721 		mtx_unlock(&msi_lock);
722 		return (EINVAL);
723 	}
724 
725 	KASSERT(msi->msi_dev != NULL, ("unowned message"));
726 
727 	/* Clear out the message. */
728 #ifdef ACPI_DMAR
729 	mtx_unlock(&msi_lock);
730 	iommu_unmap_msi_intr(msi->msi_dev, msi->msi_remap_cookie);
731 	mtx_lock(&msi_lock);
732 #endif
733 	msi->msi_first = NULL;
734 	msi->msi_dev = NULL;
735 	apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
736 	msi->msi_vector = 0;
737 	msi->msi_msix = 0;
738 	msi->msi_count = 0;
739 	msi->msi_maxcount = 0;
740 
741 	mtx_unlock(&msi_lock);
742 	return (0);
743 }
744