2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 2006 Yahoo!, Inc.
6 * Written by: John Baldwin <jhb@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
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.
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
44 #include <sys/param.h>
46 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.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>
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
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. */
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.
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)
98 static MALLOC_DEFINE(M_MSI, "msi", "PCI MSI");
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.
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
111 * For MSI-X, each message is isolated.
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;
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);
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,
149 .pic_config_intr = msi_config_intr,
150 .pic_assign_cpu = msi_assign_cpu,
151 .pic_reprogram_pin = NULL,
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:
162 * -1: automatic detection by FreeBSD
163 * 0: enable migration
164 * 1: disable migration
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");
172 static int msi_enabled;
173 static u_int msi_last_irq;
174 static struct mtx msi_lock;
177 msi_enable_source(struct intsrc *isrc)
182 msi_disable_source(struct intsrc *isrc, int eoi)
190 msi_eoi_source(struct intsrc *isrc)
197 msi_enable_intr(struct intsrc *isrc)
199 struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
201 apic_enable_vector(msi->msi_cpu, msi->msi_vector);
205 msi_disable_intr(struct intsrc *isrc)
207 struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
209 apic_disable_vector(msi->msi_cpu, msi->msi_vector);
213 msi_vector(struct intsrc *isrc)
215 struct msi_intsrc *msi = (struct msi_intsrc *)isrc;
217 return (msi->msi_irq);
221 msi_source_pending(struct intsrc *isrc)
228 msi_config_intr(struct intsrc *isrc, enum intr_trigger trig,
229 enum intr_polarity pol)
236 msi_assign_cpu(struct intsrc *isrc, u_int apic_id)
238 struct msi_intsrc *sib, *msi = (struct msi_intsrc *)isrc;
244 * Only allow CPUs to be assigned to the first message for an
247 if (msi->msi_first != msi)
251 if (msix_disable_migration && msi->msi_msix)
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)
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);
267 vector = apic_alloc_vector(apic_id, msi->msi_irq);
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);
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);
287 "msi: Assigning MSI IRQ %d to local APIC %u vector %u\n",
288 sib->msi_irq, sib->msi_cpu, sib->msi_vector);
290 BUS_REMAP_INTR(device_get_parent(msi->msi_dev), msi->msi_dev,
294 * Free the old vector after the new one is established. This is done
295 * to prevent races where we could miss an interrupt.
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]);
313 /* Check if we have a supported CPU. */
314 switch (cpu_vendor_id) {
315 case CPU_VENDOR_INTEL:
318 case CPU_VENDOR_CENTAUR:
319 if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
320 CPUID_TO_MODEL(cpu_id) >= 0xf)
328 if (msix_disable_migration == -1) {
329 /* The default is to allow migration of MSI-X interrupts. */
330 msix_disable_migration = 0;
334 MPASS(num_io_irqs > 0);
335 first_msi_irq = max(MINIMUM_MSI_INT, num_io_irqs);
336 num_io_irqs = first_msi_irq + NUM_MSI_INTS;
339 intr_register_pic(&msi_pic);
340 mtx_init(&msi_lock, "msi", NULL, MTX_DEF);
344 msi_create_source(void)
346 struct msi_intsrc *msi;
350 if (msi_last_irq >= NUM_MSI_INTS) {
351 mtx_unlock(&msi_lock);
354 irq = msi_last_irq + first_msi_irq;
356 mtx_unlock(&msi_lock);
358 msi = malloc(sizeof(struct msi_intsrc), M_MSI, M_WAITOK | M_ZERO);
359 msi->msi_intsrc.is_pic = &msi_pic;
361 intr_register_source(&msi->msi_intsrc);
366 * Try to allocate 'count' interrupt sources with contiguous IDT values.
369 msi_alloc(device_t dev, int count, int maxcount, int *irqs)
371 struct msi_intsrc *msi, *fsrc;
372 u_int cpu, domain, *mirqs;
375 u_int cookies[count];
382 if (bus_get_domain(dev, &domain) != 0)
386 mirqs = malloc(count * sizeof(*mirqs), M_MSI, M_WAITOK);
392 /* Try to find 'count' free IRQs. */
394 for (i = first_msi_irq; i < first_msi_irq + NUM_MSI_INTS; i++) {
395 msi = (struct msi_intsrc *)intr_lookup_source(i);
397 /* End of allocated sources, so break. */
401 /* If this is a free one, save its IRQ in the array. */
402 if (msi->msi_dev == NULL) {
410 /* Do we need to create some new sources? */
412 /* If we would exceed the max, give up. */
413 if (i + (count - cnt) >= first_msi_irq + NUM_MSI_INTS) {
414 mtx_unlock(&msi_lock);
418 mtx_unlock(&msi_lock);
420 /* We need count - cnt more sources. */
421 while (cnt < count) {
428 /* Ok, we now have the IRQs allocated. */
429 KASSERT(cnt == count, ("count mismatch"));
431 /* Allocate 'count' IDT vectors. */
432 cpu = intr_next_cpu(domain);
433 vector = apic_alloc_vectors(cpu, irqs, count, maxcount);
435 mtx_unlock(&msi_lock);
441 mtx_unlock(&msi_lock);
442 error = iommu_alloc_msi_intr(dev, cookies, count);
444 if (error == EOPNOTSUPP)
447 for (i = 0; i < count; i++)
448 apic_free_vector(cpu, vector + i, irqs[i]);
452 for (i = 0; i < count; i++) {
453 msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
454 msi->msi_remap_cookie = cookies[i];
458 /* Assign IDT vectors and make these messages owned by 'dev'. */
459 fsrc = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
460 for (i = 0; i < count; i++) {
461 msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
464 msi->msi_vector = vector + i;
467 "msi: routing MSI IRQ %d to local APIC %u vector %u\n",
468 msi->msi_irq, msi->msi_cpu, msi->msi_vector);
469 msi->msi_first = fsrc;
470 KASSERT(msi->msi_intsrc.is_handlers == 0,
471 ("dead MSI has handlers"));
473 fsrc->msi_count = count;
474 fsrc->msi_maxcount = maxcount;
476 bcopy(irqs, mirqs, count * sizeof(*mirqs));
477 fsrc->msi_irqs = mirqs;
478 mtx_unlock(&msi_lock);
483 msi_release(int *irqs, int count)
485 struct msi_intsrc *msi, *first;
489 first = (struct msi_intsrc *)intr_lookup_source(irqs[0]);
491 mtx_unlock(&msi_lock);
495 /* Make sure this isn't an MSI-X message. */
496 if (first->msi_msix) {
497 mtx_unlock(&msi_lock);
501 /* Make sure this message is allocated to a group. */
502 if (first->msi_first == NULL) {
503 mtx_unlock(&msi_lock);
508 * Make sure this is the start of a group and that we are releasing
511 if (first->msi_first != first || first->msi_count != count) {
512 mtx_unlock(&msi_lock);
515 KASSERT(first->msi_dev != NULL, ("unowned group"));
517 /* Clear all the extra messages in the group. */
518 for (i = 1; i < count; i++) {
519 msi = (struct msi_intsrc *)intr_lookup_source(irqs[i]);
520 KASSERT(msi->msi_first == first, ("message not in group"));
521 KASSERT(msi->msi_dev == first->msi_dev, ("owner mismatch"));
523 iommu_unmap_msi_intr(first->msi_dev, msi->msi_remap_cookie);
525 msi->msi_first = NULL;
527 apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
531 /* Clear out the first message. */
533 mtx_unlock(&msi_lock);
534 iommu_unmap_msi_intr(first->msi_dev, first->msi_remap_cookie);
537 first->msi_first = NULL;
538 first->msi_dev = NULL;
539 apic_free_vector(first->msi_cpu, first->msi_vector, first->msi_irq);
540 first->msi_vector = 0;
541 first->msi_count = 0;
542 first->msi_maxcount = 0;
543 free(first->msi_irqs, M_MSI);
544 first->msi_irqs = NULL;
546 mtx_unlock(&msi_lock);
551 msi_map(int irq, uint64_t *addr, uint32_t *data)
553 struct msi_intsrc *msi;
556 struct msi_intsrc *msi1;
561 msi = (struct msi_intsrc *)intr_lookup_source(irq);
563 mtx_unlock(&msi_lock);
567 /* Make sure this message is allocated to a device. */
568 if (msi->msi_dev == NULL) {
569 mtx_unlock(&msi_lock);
574 * If this message isn't an MSI-X message, make sure it's part
575 * of a group, and switch to the first message in the
578 if (!msi->msi_msix) {
579 if (msi->msi_first == NULL) {
580 mtx_unlock(&msi_lock);
583 msi = msi->msi_first;
587 if (!msi->msi_msix) {
588 for (k = msi->msi_count - 1, i = first_msi_irq; k > 0 &&
589 i < first_msi_irq + NUM_MSI_INTS; i++) {
590 if (i == msi->msi_irq)
592 msi1 = (struct msi_intsrc *)intr_lookup_source(i);
593 if (!msi1->msi_msix && msi1->msi_first == msi) {
594 mtx_unlock(&msi_lock);
595 iommu_map_msi_intr(msi1->msi_dev,
596 msi1->msi_cpu, msi1->msi_vector,
597 msi1->msi_remap_cookie, NULL, NULL);
603 mtx_unlock(&msi_lock);
604 error = iommu_map_msi_intr(msi->msi_dev, msi->msi_cpu,
605 msi->msi_vector, msi->msi_remap_cookie, addr, data);
607 mtx_unlock(&msi_lock);
610 if (error == EOPNOTSUPP) {
611 *addr = INTEL_ADDR(msi);
612 *data = INTEL_DATA(msi);
619 msix_alloc(device_t dev, int *irq)
621 struct msi_intsrc *msi;
632 if (bus_get_domain(dev, &domain) != 0)
638 /* Find a free IRQ. */
639 for (i = first_msi_irq; i < first_msi_irq + NUM_MSI_INTS; i++) {
640 msi = (struct msi_intsrc *)intr_lookup_source(i);
642 /* End of allocated sources, so break. */
646 /* Stop at the first free source. */
647 if (msi->msi_dev == NULL)
651 /* Do we need to create a new source? */
653 /* If we would exceed the max, give up. */
654 if (i + 1 >= first_msi_irq + NUM_MSI_INTS) {
655 mtx_unlock(&msi_lock);
658 mtx_unlock(&msi_lock);
660 /* Create a new source. */
665 /* Allocate an IDT vector. */
666 cpu = intr_next_cpu(domain);
667 vector = apic_alloc_vector(cpu, i);
669 mtx_unlock(&msi_lock);
675 mtx_unlock(&msi_lock);
676 error = iommu_alloc_msi_intr(dev, &cookie, 1);
678 if (error == EOPNOTSUPP)
682 apic_free_vector(cpu, vector, i);
685 msi->msi_remap_cookie = cookie;
689 printf("msi: routing MSI-X IRQ %d to local APIC %u vector %u\n",
690 msi->msi_irq, cpu, vector);
694 msi->msi_first = msi;
695 msi->msi_vector = vector;
698 msi->msi_maxcount = 1;
699 msi->msi_irqs = NULL;
701 KASSERT(msi->msi_intsrc.is_handlers == 0, ("dead MSI-X has handlers"));
702 mtx_unlock(&msi_lock);
709 msix_release(int irq)
711 struct msi_intsrc *msi;
714 msi = (struct msi_intsrc *)intr_lookup_source(irq);
716 mtx_unlock(&msi_lock);
720 /* Make sure this is an MSI-X message. */
721 if (!msi->msi_msix) {
722 mtx_unlock(&msi_lock);
726 KASSERT(msi->msi_dev != NULL, ("unowned message"));
728 /* Clear out the message. */
730 mtx_unlock(&msi_lock);
731 iommu_unmap_msi_intr(msi->msi_dev, msi->msi_remap_cookie);
734 msi->msi_first = NULL;
736 apic_free_vector(msi->msi_cpu, msi->msi_vector, msi->msi_irq);
740 msi->msi_maxcount = 0;
742 mtx_unlock(&msi_lock);