xref: /freebsd/sys/x86/x86/msi.c (revision 0957b409)
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/limits.h>
48 #include <sys/lock.h>
49 #include <sys/malloc.h>
50 #include <sys/mutex.h>
51 #include <sys/sx.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 #include <x86/apicreg.h>
55 #include <machine/cputypes.h>
56 #include <machine/md_var.h>
57 #include <machine/frame.h>
58 #include <machine/intr_machdep.h>
59 #include <x86/apicvar.h>
60 #include <x86/iommu/iommu_intrmap.h>
61 #include <machine/specialreg.h>
62 #include <dev/pci/pcivar.h>
63 
64 /* Fields in address for Intel MSI messages. */
65 #define	MSI_INTEL_ADDR_DEST		0x000ff000
66 #define	MSI_INTEL_ADDR_RH		0x00000008
67 # define MSI_INTEL_ADDR_RH_ON		0x00000008
68 # define MSI_INTEL_ADDR_RH_OFF		0x00000000
69 #define	MSI_INTEL_ADDR_DM		0x00000004
70 # define MSI_INTEL_ADDR_DM_PHYSICAL	0x00000000
71 # define MSI_INTEL_ADDR_DM_LOGICAL	0x00000004
72 
73 /* Fields in data for Intel MSI messages. */
74 #define	MSI_INTEL_DATA_TRGRMOD		IOART_TRGRMOD	/* Trigger mode. */
75 # define MSI_INTEL_DATA_TRGREDG		IOART_TRGREDG
76 # define MSI_INTEL_DATA_TRGRLVL		IOART_TRGRLVL
77 #define	MSI_INTEL_DATA_LEVEL		0x00004000	/* Polarity. */
78 # define MSI_INTEL_DATA_DEASSERT	0x00000000
79 # define MSI_INTEL_DATA_ASSERT		0x00004000
80 #define	MSI_INTEL_DATA_DELMOD		IOART_DELMOD	/* Delivery mode. */
81 # define MSI_INTEL_DATA_DELFIXED	IOART_DELFIXED
82 # define MSI_INTEL_DATA_DELLOPRI	IOART_DELLOPRI
83 # define MSI_INTEL_DATA_DELSMI		IOART_DELSMI
84 # define MSI_INTEL_DATA_DELNMI		IOART_DELNMI
85 # define MSI_INTEL_DATA_DELINIT		IOART_DELINIT
86 # define MSI_INTEL_DATA_DELEXINT	IOART_DELEXINT
87 #define	MSI_INTEL_DATA_INTVEC		IOART_INTVEC	/* Interrupt vector. */
88 
89 /*
90  * Build Intel MSI message and data values from a source.  AMD64 systems
91  * seem to be compatible, so we use the same function for both.
92  */
93 #define	INTEL_ADDR(msi)							\
94 	(MSI_INTEL_ADDR_BASE | (msi)->msi_cpu << 12 |			\
95 	    MSI_INTEL_ADDR_RH_OFF | MSI_INTEL_ADDR_DM_PHYSICAL)
96 #define	INTEL_DATA(msi)							\
97 	(MSI_INTEL_DATA_TRGREDG | MSI_INTEL_DATA_DELFIXED | (msi)->msi_vector)
98 
99 static MALLOC_DEFINE(M_MSI, "msi", "PCI MSI");
100 
101 /*
102  * MSI sources are bunched into groups.  This is because MSI forces
103  * all of the messages to share the address and data registers and
104  * thus certain properties (such as the local APIC ID target on x86).
105  * Each group has a 'first' source that contains information global to
106  * the group.  These fields are marked with (g) below.
107  *
108  * Note that local APIC ID is kind of special.  Each message will be
109  * assigned an ID by the system; however, a group will use the ID from
110  * the first message.
111  *
112  * For MSI-X, each message is isolated.
113  */
114 struct msi_intsrc {
115 	struct intsrc msi_intsrc;
116 	device_t msi_dev;		/* Owning device. (g) */
117 	struct msi_intsrc *msi_first;	/* First source in group. */
118 	u_int msi_irq;			/* IRQ cookie. */
119 	u_int msi_msix;			/* MSI-X message. */
120 	u_int msi_vector:8;		/* IDT vector. */
121 	u_int msi_cpu;			/* Local APIC ID. (g) */
122 	u_int msi_count:8;		/* Messages in this group. (g) */
123 	u_int msi_maxcount:8;		/* Alignment for this group. (g) */
124 	u_int *msi_irqs;		/* Group's IRQ list. (g) */
125 	u_int msi_remap_cookie;
126 };
127 
128 static void	msi_create_source(void);
129 static void	msi_enable_source(struct intsrc *isrc);
130 static void	msi_disable_source(struct intsrc *isrc, int eoi);
131 static void	msi_eoi_source(struct intsrc *isrc);
132 static void	msi_enable_intr(struct intsrc *isrc);
133 static void	msi_disable_intr(struct intsrc *isrc);
134 static int	msi_vector(struct intsrc *isrc);
135 static int	msi_source_pending(struct intsrc *isrc);
136 static int	msi_config_intr(struct intsrc *isrc, enum intr_trigger trig,
137 		    enum intr_polarity pol);
138 static int	msi_assign_cpu(struct intsrc *isrc, u_int apic_id);
139 
140 struct pic msi_pic = {
141 	.pic_enable_source = msi_enable_source,
142 	.pic_disable_source = msi_disable_source,
143 	.pic_eoi_source = msi_eoi_source,
144 	.pic_enable_intr = msi_enable_intr,
145 	.pic_disable_intr = msi_disable_intr,
146 	.pic_vector = msi_vector,
147 	.pic_source_pending = msi_source_pending,
148 	.pic_suspend = NULL,
149 	.pic_resume = NULL,
150 	.pic_config_intr = msi_config_intr,
151 	.pic_assign_cpu = msi_assign_cpu,
152 	.pic_reprogram_pin = NULL,
153 };
154 
155 u_int first_msi_irq;
156 SYSCTL_UINT(_machdep, OID_AUTO, first_msi_irq, CTLFLAG_RD, &first_msi_irq, 0,
157     "Number of first IRQ reserved for MSI and MSI-X interrupts");
158 
159 u_int num_msi_irqs = 512;
160 SYSCTL_UINT(_machdep, OID_AUTO, num_msi_irqs, CTLFLAG_RDTUN, &num_msi_irqs, 0,
161     "Number of IRQs reserved for MSI and MSI-X interrupts");
162 
163 #ifdef SMP
164 /**
165  * Xen hypervisors prior to 4.6.0 do not properly handle updates to
166  * enabled MSI-X table entries.  Allow migration of MSI-X interrupts
167  * to be disabled via a tunable. Values have the following meaning:
168  *
169  * -1: automatic detection by FreeBSD
170  *  0: enable migration
171  *  1: disable migration
172  */
173 int msix_disable_migration = -1;
174 SYSCTL_INT(_machdep, OID_AUTO, disable_msix_migration, CTLFLAG_RDTUN,
175     &msix_disable_migration, 0,
176     "Disable migration of MSI-X interrupts between CPUs");
177 #endif
178 
179 static int msi_enabled;
180 static u_int msi_last_irq;
181 static struct mtx msi_lock;
182 
183 static void
184 msi_enable_source(struct intsrc *isrc)
185 {
186 }
187 
188 static void
189 msi_disable_source(struct intsrc *isrc, int eoi)
190 {
191 
192 	if (eoi == PIC_EOI)
193 		lapic_eoi();
194 }
195 
196 static void
197 msi_eoi_source(struct intsrc *isrc)
198 {
199 
200 	lapic_eoi();
201 }
202 
203 static void
204 msi_enable_intr(struct intsrc *isrc)
205 {
206 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
207 
208 	apic_enable_vector(msi->msi_cpu, msi->msi_vector);
209 }
210 
211 static void
212 msi_disable_intr(struct intsrc *isrc)
213 {
214 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
215 
216 	apic_disable_vector(msi->msi_cpu, msi->msi_vector);
217 }
218 
219 static int
220 msi_vector(struct intsrc *isrc)
221 {
222 	struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
223 
224 	return (msi->msi_irq);
225 }
226 
227 static int
228 msi_source_pending(struct intsrc *isrc)
229 {
230 
231 	return (0);
232 }
233 
234 static int
235 msi_config_intr(struct intsrc *isrc, enum intr_trigger trig,
236     enum intr_polarity pol)
237 {
238 
239 	return (ENODEV);
240 }
241 
242 static int
243 msi_assign_cpu(struct intsrc *isrc, u_int apic_id)
244 {
245 	struct msi_intsrc *sib, *msi = (struct msi_intsrc *)isrc;
246 	int old_vector;
247 	u_int old_id;
248 	int i, vector;
249 
250 	/*
251 	 * Only allow CPUs to be assigned to the first message for an
252 	 * MSI group.
253 	 */
254 	if (msi->msi_first != msi)
255 		return (EINVAL);
256 
257 #ifdef SMP
258 	if (msix_disable_migration && msi->msi_msix)
259 		return (EINVAL);
260 #endif
261 
262 	/* Store information to free existing irq. */
263 	old_vector = msi->msi_vector;
264 	old_id = msi->msi_cpu;
265 	if (old_id == apic_id)
266 		return (0);
267 
268 	/* Allocate IDT vectors on this cpu. */
269 	if (msi->msi_count > 1) {
270 		KASSERT(msi->msi_msix == 0, ("MSI-X message group"));
271 		vector = apic_alloc_vectors(apic_id, msi->msi_irqs,
272 		    msi->msi_count, msi->msi_maxcount);
273 	} else
274 		vector = apic_alloc_vector(apic_id, msi->msi_irq);
275 	if (vector == 0)
276 		return (ENOSPC);
277 
278 	msi->msi_cpu = apic_id;
279 	msi->msi_vector = vector;
280 	if (msi->msi_intsrc.is_handlers > 0)
281 		apic_enable_vector(msi->msi_cpu, msi->msi_vector);
282 	if (bootverbose)
283 		printf("msi: Assigning %s IRQ %d to local APIC %u vector %u\n",
284 		    msi->msi_msix ? "MSI-X" : "MSI", msi->msi_irq,
285 		    msi->msi_cpu, msi->msi_vector);
286 	for (i = 1; i < msi->msi_count; i++) {
287 		sib = (struct msi_intsrc *)intr_lookup_source(msi->msi_irqs[i]);
288 		sib->msi_cpu = apic_id;
289 		sib->msi_vector = vector + i;
290 		if (sib->msi_intsrc.is_handlers > 0)
291 			apic_enable_vector(sib->msi_cpu, sib->msi_vector);
292 		if (bootverbose)
293 			printf(
294 		    "msi: Assigning MSI IRQ %d to local APIC %u vector %u\n",
295 			    sib->msi_irq, sib->msi_cpu, sib->msi_vector);
296 	}
297 	BUS_REMAP_INTR(device_get_parent(msi->msi_dev), msi->msi_dev,
298 	    msi->msi_irq);
299 
300 	/*
301 	 * Free the old vector after the new one is established.  This is done
302 	 * to prevent races where we could miss an interrupt.
303 	 */
304 	if (msi->msi_intsrc.is_handlers > 0)
305 		apic_disable_vector(old_id, old_vector);
306 	apic_free_vector(old_id, old_vector, msi->msi_irq);
307 	for (i = 1; i < msi->msi_count; i++) {
308 		sib = (struct msi_intsrc *)intr_lookup_source(msi->msi_irqs[i]);
309 		if (sib->msi_intsrc.is_handlers > 0)
310 			apic_disable_vector(old_id, old_vector + i);
311 		apic_free_vector(old_id, old_vector + i, msi->msi_irqs[i]);
312 	}
313 	return (0);
314 }
315 
316 void
317 msi_init(void)
318 {
319 
320 	/* Check if we have a supported CPU. */
321 	switch (cpu_vendor_id) {
322 	case CPU_VENDOR_INTEL:
323 	case CPU_VENDOR_AMD:
324 		break;
325 	case CPU_VENDOR_CENTAUR:
326 		if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
327 		    CPUID_TO_MODEL(cpu_id) >= 0xf)
328 			break;
329 		/* FALLTHROUGH */
330 	default:
331 		return;
332 	}
333 
334 #ifdef SMP
335 	if (msix_disable_migration == -1) {
336 		/* The default is to allow migration of MSI-X interrupts. */
337 		msix_disable_migration = 0;
338 	}
339 #endif
340 
341 	if (num_msi_irqs == 0)
342 		return;
343 
344 	first_msi_irq = num_io_irqs;
345 	if (num_msi_irqs > UINT_MAX - first_msi_irq)
346 		panic("num_msi_irqs too high");
347 	num_io_irqs = first_msi_irq + num_msi_irqs;
348 
349 	msi_enabled = 1;
350 	intr_register_pic(&msi_pic);
351 	mtx_init(&msi_lock, "msi", NULL, MTX_DEF);
352 }
353 
354 static void
355 msi_create_source(void)
356 {
357 	struct msi_intsrc *msi;
358 	u_int irq;
359 
360 	mtx_lock(&msi_lock);
361 	if (msi_last_irq >= num_msi_irqs) {
362 		mtx_unlock(&msi_lock);
363 		return;
364 	}
365 	irq = msi_last_irq + first_msi_irq;
366 	msi_last_irq++;
367 	mtx_unlock(&msi_lock);
368 
369 	msi = malloc(sizeof(struct msi_intsrc), M_MSI, M_WAITOK | M_ZERO);
370 	msi->msi_intsrc.is_pic = &msi_pic;
371 	msi->msi_irq = irq;
372 	intr_register_source(&msi->msi_intsrc);
373 	nexus_add_irq(irq);
374 }
375 
376 /*
377  * Try to allocate 'count' interrupt sources with contiguous IDT values.
378  */
379 int
380 msi_alloc(device_t dev, int count, int maxcount, int *irqs)
381 {
382 	struct msi_intsrc *msi, *fsrc;
383 	u_int cpu, domain, *mirqs;
384 	int cnt, i, vector;
385 #ifdef ACPI_DMAR
386 	u_int cookies[count];
387 	int error;
388 #endif
389 
390 	if (!msi_enabled)
391 		return (ENXIO);
392 
393 	if (bus_get_domain(dev, &domain) != 0)
394 		domain = 0;
395 
396 	if (count > 1)
397 		mirqs = malloc(count * sizeof(*mirqs), M_MSI, M_WAITOK);
398 	else
399 		mirqs = NULL;
400 again:
401 	mtx_lock(&msi_lock);
402 
403 	/* Try to find 'count' free IRQs. */
404 	cnt = 0;
405 	for (i = first_msi_irq; i < first_msi_irq + num_msi_irqs; i++) {
406 		msi = (struct msi_intsrc *)intr_lookup_source(i);
407 
408 		/* End of allocated sources, so break. */
409 		if (msi == NULL)
410 			break;
411 
412 		/* If this is a free one, save its IRQ in the array. */
413 		if (msi->msi_dev == NULL) {
414 			irqs[cnt] = i;
415 			cnt++;
416 			if (cnt == count)
417 				break;
418 		}
419 	}
420 
421 	/* Do we need to create some new sources? */
422 	if (cnt < count) {
423 		/* If we would exceed the max, give up. */
424 		if (i + (count - cnt) > first_msi_irq + num_msi_irqs) {
425 			mtx_unlock(&msi_lock);
426 			free(mirqs, M_MSI);
427 			return (ENXIO);
428 		}
429 		mtx_unlock(&msi_lock);
430 
431 		/* We need count - cnt more sources. */
432 		while (cnt < count) {
433 			msi_create_source();
434 			cnt++;
435 		}
436 		goto again;
437 	}
438 
439 	/* Ok, we now have the IRQs allocated. */
440 	KASSERT(cnt == count, ("count mismatch"));
441 
442 	/* Allocate 'count' IDT vectors. */
443 	cpu = intr_next_cpu(domain);
444 	vector = apic_alloc_vectors(cpu, irqs, count, maxcount);
445 	if (vector == 0) {
446 		mtx_unlock(&msi_lock);
447 		free(mirqs, M_MSI);
448 		return (ENOSPC);
449 	}
450 
451 #ifdef ACPI_DMAR
452 	mtx_unlock(&msi_lock);
453 	error = iommu_alloc_msi_intr(dev, cookies, count);
454 	mtx_lock(&msi_lock);
455 	if (error == EOPNOTSUPP)
456 		error = 0;
457 	if (error != 0) {
458 		for (i = 0; i < count; i++)
459 			apic_free_vector(cpu, vector + i, irqs[i]);
460 		free(mirqs, M_MSI);
461 		return (error);
462 	}
463 	for (i = 0; i < count; i++) {
464 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
465 		msi->msi_remap_cookie = cookies[i];
466 	}
467 #endif
468 
469 	/* Assign IDT vectors and make these messages owned by 'dev'. */
470 	fsrc = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
471 	for (i = 0; i < count; i++) {
472 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
473 		msi->msi_cpu = cpu;
474 		msi->msi_dev = dev;
475 		msi->msi_vector = vector + i;
476 		if (bootverbose)
477 			printf(
478 		    "msi: routing MSI IRQ %d to local APIC %u vector %u\n",
479 			    msi->msi_irq, msi->msi_cpu, msi->msi_vector);
480 		msi->msi_first = fsrc;
481 		KASSERT(msi->msi_intsrc.is_handlers == 0,
482 		    ("dead MSI has handlers"));
483 	}
484 	fsrc->msi_count = count;
485 	fsrc->msi_maxcount = maxcount;
486 	if (count > 1)
487 		bcopy(irqs, mirqs, count * sizeof(*mirqs));
488 	fsrc->msi_irqs = mirqs;
489 	mtx_unlock(&msi_lock);
490 	return (0);
491 }
492 
493 int
494 msi_release(int *irqs, int count)
495 {
496 	struct msi_intsrc *msi, *first;
497 	int i;
498 
499 	mtx_lock(&msi_lock);
500 	first = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
501 	if (first == NULL) {
502 		mtx_unlock(&msi_lock);
503 		return (ENOENT);
504 	}
505 
506 	/* Make sure this isn't an MSI-X message. */
507 	if (first->msi_msix) {
508 		mtx_unlock(&msi_lock);
509 		return (EINVAL);
510 	}
511 
512 	/* Make sure this message is allocated to a group. */
513 	if (first->msi_first == NULL) {
514 		mtx_unlock(&msi_lock);
515 		return (ENXIO);
516 	}
517 
518 	/*
519 	 * Make sure this is the start of a group and that we are releasing
520 	 * the entire group.
521 	 */
522 	if (first->msi_first != first || first->msi_count != count) {
523 		mtx_unlock(&msi_lock);
524 		return (EINVAL);
525 	}
526 	KASSERT(first->msi_dev != NULL, ("unowned group"));
527 
528 	/* Clear all the extra messages in the group. */
529 	for (i = 1; i < count; i++) {
530 		msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
531 		KASSERT(msi->msi_first == first, ("message not in group"));
532 		KASSERT(msi->msi_dev == first->msi_dev, ("owner mismatch"));
533 #ifdef ACPI_DMAR
534 		iommu_unmap_msi_intr(first->msi_dev, msi->msi_remap_cookie);
535 #endif
536 		msi->msi_first = NULL;
537 		msi->msi_dev = NULL;
538 		apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
539 		msi->msi_vector = 0;
540 	}
541 
542 	/* Clear out the first message. */
543 #ifdef ACPI_DMAR
544 	mtx_unlock(&msi_lock);
545 	iommu_unmap_msi_intr(first->msi_dev, first->msi_remap_cookie);
546 	mtx_lock(&msi_lock);
547 #endif
548 	first->msi_first = NULL;
549 	first->msi_dev = NULL;
550 	apic_free_vector(first->msi_cpu, first->msi_vector, first->msi_irq);
551 	first->msi_vector = 0;
552 	first->msi_count = 0;
553 	first->msi_maxcount = 0;
554 	free(first->msi_irqs, M_MSI);
555 	first->msi_irqs = NULL;
556 
557 	mtx_unlock(&msi_lock);
558 	return (0);
559 }
560 
561 int
562 msi_map(int irq, uint64_t *addr, uint32_t *data)
563 {
564 	struct msi_intsrc *msi;
565 	int error;
566 #ifdef ACPI_DMAR
567 	struct msi_intsrc *msi1;
568 	int i, k;
569 #endif
570 
571 	mtx_lock(&msi_lock);
572 	msi = (struct msi_intsrc *)intr_lookup_source(irq);
573 	if (msi == NULL) {
574 		mtx_unlock(&msi_lock);
575 		return (ENOENT);
576 	}
577 
578 	/* Make sure this message is allocated to a device. */
579 	if (msi->msi_dev == NULL) {
580 		mtx_unlock(&msi_lock);
581 		return (ENXIO);
582 	}
583 
584 	/*
585 	 * If this message isn't an MSI-X message, make sure it's part
586 	 * of a group, and switch to the first message in the
587 	 * group.
588 	 */
589 	if (!msi->msi_msix) {
590 		if (msi->msi_first == NULL) {
591 			mtx_unlock(&msi_lock);
592 			return (ENXIO);
593 		}
594 		msi = msi->msi_first;
595 	}
596 
597 #ifdef ACPI_DMAR
598 	if (!msi->msi_msix) {
599 		for (k = msi->msi_count - 1, i = first_msi_irq; k > 0 &&
600 		    i < first_msi_irq + num_msi_irqs; i++) {
601 			if (i == msi->msi_irq)
602 				continue;
603 			msi1 = (struct msi_intsrc *)intr_lookup_source(i);
604 			if (!msi1->msi_msix && msi1->msi_first == msi) {
605 				mtx_unlock(&msi_lock);
606 				iommu_map_msi_intr(msi1->msi_dev,
607 				    msi1->msi_cpu, msi1->msi_vector,
608 				    msi1->msi_remap_cookie, NULL, NULL);
609 				k--;
610 				mtx_lock(&msi_lock);
611 			}
612 		}
613 	}
614 	mtx_unlock(&msi_lock);
615 	error = iommu_map_msi_intr(msi->msi_dev, msi->msi_cpu,
616 	    msi->msi_vector, msi->msi_remap_cookie, addr, data);
617 #else
618 	mtx_unlock(&msi_lock);
619 	error = EOPNOTSUPP;
620 #endif
621 	if (error == EOPNOTSUPP) {
622 		*addr = INTEL_ADDR(msi);
623 		*data = INTEL_DATA(msi);
624 		error = 0;
625 	}
626 	return (error);
627 }
628 
629 int
630 msix_alloc(device_t dev, int *irq)
631 {
632 	struct msi_intsrc *msi;
633 	u_int cpu, domain;
634 	int i, vector;
635 #ifdef ACPI_DMAR
636 	u_int cookie;
637 	int error;
638 #endif
639 
640 	if (!msi_enabled)
641 		return (ENXIO);
642 
643 	if (bus_get_domain(dev, &domain) != 0)
644 		domain = 0;
645 
646 again:
647 	mtx_lock(&msi_lock);
648 
649 	/* Find a free IRQ. */
650 	for (i = first_msi_irq; i < first_msi_irq + num_msi_irqs; i++) {
651 		msi = (struct msi_intsrc *)intr_lookup_source(i);
652 
653 		/* End of allocated sources, so break. */
654 		if (msi == NULL)
655 			break;
656 
657 		/* Stop at the first free source. */
658 		if (msi->msi_dev == NULL)
659 			break;
660 	}
661 
662 	/* Are all IRQs in use? */
663 	if (i == first_msi_irq + num_msi_irqs) {
664 		mtx_unlock(&msi_lock);
665 		return (ENXIO);
666 	}
667 
668 	/* Do we need to create a new source? */
669 	if (msi == NULL) {
670 		mtx_unlock(&msi_lock);
671 
672 		/* Create a new source. */
673 		msi_create_source();
674 		goto again;
675 	}
676 
677 	/* Allocate an IDT vector. */
678 	cpu = intr_next_cpu(domain);
679 	vector = apic_alloc_vector(cpu, i);
680 	if (vector == 0) {
681 		mtx_unlock(&msi_lock);
682 		return (ENOSPC);
683 	}
684 
685 	msi->msi_dev = dev;
686 #ifdef ACPI_DMAR
687 	mtx_unlock(&msi_lock);
688 	error = iommu_alloc_msi_intr(dev, &cookie, 1);
689 	mtx_lock(&msi_lock);
690 	if (error == EOPNOTSUPP)
691 		error = 0;
692 	if (error != 0) {
693 		msi->msi_dev = NULL;
694 		apic_free_vector(cpu, vector, i);
695 		return (error);
696 	}
697 	msi->msi_remap_cookie = cookie;
698 #endif
699 
700 	if (bootverbose)
701 		printf("msi: routing MSI-X IRQ %d to local APIC %u vector %u\n",
702 		    msi->msi_irq, cpu, vector);
703 
704 	/* Setup source. */
705 	msi->msi_cpu = cpu;
706 	msi->msi_first = msi;
707 	msi->msi_vector = vector;
708 	msi->msi_msix = 1;
709 	msi->msi_count = 1;
710 	msi->msi_maxcount = 1;
711 	msi->msi_irqs = NULL;
712 
713 	KASSERT(msi->msi_intsrc.is_handlers == 0, ("dead MSI-X has handlers"));
714 	mtx_unlock(&msi_lock);
715 
716 	*irq = i;
717 	return (0);
718 }
719 
720 int
721 msix_release(int irq)
722 {
723 	struct msi_intsrc *msi;
724 
725 	mtx_lock(&msi_lock);
726 	msi = (struct msi_intsrc *)intr_lookup_source(irq);
727 	if (msi == NULL) {
728 		mtx_unlock(&msi_lock);
729 		return (ENOENT);
730 	}
731 
732 	/* Make sure this is an MSI-X message. */
733 	if (!msi->msi_msix) {
734 		mtx_unlock(&msi_lock);
735 		return (EINVAL);
736 	}
737 
738 	KASSERT(msi->msi_dev != NULL, ("unowned message"));
739 
740 	/* Clear out the message. */
741 #ifdef ACPI_DMAR
742 	mtx_unlock(&msi_lock);
743 	iommu_unmap_msi_intr(msi->msi_dev, msi->msi_remap_cookie);
744 	mtx_lock(&msi_lock);
745 #endif
746 	msi->msi_first = NULL;
747 	msi->msi_dev = NULL;
748 	apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
749 	msi->msi_vector = 0;
750 	msi->msi_msix = 0;
751 	msi->msi_count = 0;
752 	msi->msi_maxcount = 0;
753 
754 	mtx_unlock(&msi_lock);
755 	return (0);
756 }
757