git: c28220d8661e - main - vmd: Bypass MSI/MSI-X remapping when possible.

From: Alexander Motin <mav_at_FreeBSD.org>
Date: Thu, 06 Oct 2022 16:20:24 UTC
The branch main has been updated by mav:

URL: https://cgit.FreeBSD.org/src/commit/?id=c28220d8661ee1928bff7d2c6bcd3e64c25f927b

commit c28220d8661ee1928bff7d2c6bcd3e64c25f927b
Author:     Alexander Motin <mav@FreeBSD.org>
AuthorDate: 2022-10-06 16:15:25 +0000
Commit:     Alexander Motin <mav@FreeBSD.org>
CommitDate: 2022-10-06 16:15:25 +0000

    vmd: Bypass MSI/MSI-X remapping when possible.
    
    By default all VMD devices remap children MSI/MSI-X interrupts into their
    own.  It creates additional isolation, but also complicates things due to
    sharing, etc.  Fortunately some VMD devices can bypass the remapping.
    Add tunable to control it for remap testing or if something go wrong.
    
    MFC after:      2 weeks
---
 share/man/man4/vmd.4 |  7 ++++++-
 sys/dev/vmd/vmd.c    | 59 ++++++++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 63 insertions(+), 3 deletions(-)

diff --git a/share/man/man4/vmd.4 b/share/man/man4/vmd.4
index 74419fd87ffe..6c49e2b83ec0 100644
--- a/share/man/man4/vmd.4
+++ b/share/man/man4/vmd.4
@@ -27,7 +27,7 @@
 .\"
 .\" $FreeBSD$
 .\"
-.Dd August 31, 2021
+.Dd October 6, 2022
 .Dt VMD 4
 .Os
 .Sh NAME
@@ -57,6 +57,11 @@ The following tunables are settable via
 or
 .Xr sysctl 8 :
 .Bl -tag -width indent
+.It Va hw.vmd.bypass_msi
+By default all VMD devices remap children MSI/MSI-X interrupts into their
+own.  It creates additional isolation, but also complicates things due to
+sharing, etc.  Fortunately some VMD devices can bypass the remapping.
+Defaults to 1.
 .It Va hw.vmd.max_msi
 Limits number of Message Signaled Interrupt (MSI) vectors allowed to each
 child device.
diff --git a/sys/dev/vmd/vmd.c b/sys/dev/vmd/vmd.c
index 208ea8a865f8..078101b58b2d 100644
--- a/sys/dev/vmd/vmd.c
+++ b/sys/dev/vmd/vmd.c
@@ -65,12 +65,14 @@ struct vmd_type {
 	int		flags;
 #define BUS_RESTRICT	1
 #define VECTOR_OFFSET	2
+#define CAN_BYPASS_MSI	4
 };
 
 #define VMD_CAP		0x40
 #define VMD_BUS_RESTRICT	0x1
 
 #define VMD_CONFIG	0x44
+#define VMD_BYPASS_MSI		0x2
 #define VMD_BUS_START(x)	((x >> 8) & 0x3)
 
 #define VMD_LOCK	0x70
@@ -78,6 +80,15 @@ struct vmd_type {
 SYSCTL_NODE(_hw, OID_AUTO, vmd, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
     "Intel Volume Management Device tuning parameters");
 
+/*
+ * By default all VMD devices remap children MSI/MSI-X interrupts into their
+ * own.  It creates additional isolation, but also complicates things due to
+ * sharing, etc.  Fortunately some VMD devices can bypass the remapping.
+ */
+static int vmd_bypass_msi = 1;
+SYSCTL_INT(_hw_vmd, OID_AUTO, bypass_msi, CTLFLAG_RWTUN, &vmd_bypass_msi, 0,
+    "Bypass MSI remapping on capable hardware");
+
 /*
  * All MSIs within a group share address, so VMD can't distinguish them.
  * It makes no sense to use more than one per device, only if required by
@@ -97,7 +108,7 @@ SYSCTL_INT(_hw_vmd, OID_AUTO, max_msix, CTLFLAG_RWTUN, &vmd_max_msix, 0,
 
 static struct vmd_type vmd_devs[] = {
         { 0x8086, 0x201d, "Intel Volume Management Device", 0 },
-        { 0x8086, 0x28c0, "Intel Volume Management Device", BUS_RESTRICT },
+        { 0x8086, 0x28c0, "Intel Volume Management Device", BUS_RESTRICT | CAN_BYPASS_MSI },
         { 0x8086, 0x467f, "Intel Volume Management Device", BUS_RESTRICT | VECTOR_OFFSET },
         { 0x8086, 0x4c3d, "Intel Volume Management Device", BUS_RESTRICT | VECTOR_OFFSET },
         { 0x8086, 0x9a0b, "Intel Volume Management Device", BUS_RESTRICT | VECTOR_OFFSET },
@@ -216,6 +227,19 @@ vmd_write_config(device_t dev, u_int b, u_int s, u_int f, u_int reg,
 	}
 }
 
+static void
+vmd_set_msi_bypass(device_t dev, bool enable)
+{
+	uint16_t val;
+
+	val = pci_read_config(dev, VMD_CONFIG, 2);
+	if (enable)
+		val |= VMD_BYPASS_MSI;
+	else
+		val &= ~VMD_BYPASS_MSI;
+	pci_write_config(dev, VMD_CONFIG, val, 2);
+}
+
 static int
 vmd_intr(void *arg)
 {
@@ -340,7 +364,10 @@ vmd_attach(device_t dev)
 	LIST_INIT(&sc->vmd_users);
 	sc->vmd_fist_vector = (t->flags & VECTOR_OFFSET) ? 1 : 0;
 	sc->vmd_msix_count = pci_msix_count(dev);
-	if (pci_alloc_msix(dev, &sc->vmd_msix_count) == 0) {
+	if (vmd_bypass_msi && (t->flags & CAN_BYPASS_MSI)) {
+		sc->vmd_msix_count = 0;
+		vmd_set_msi_bypass(dev, true);
+	} else if (pci_alloc_msix(dev, &sc->vmd_msix_count) == 0) {
 		sc->vmd_irq = malloc(sizeof(struct vmd_irq) *
 		    sc->vmd_msix_count, M_DEVBUF, M_WAITOK | M_ZERO);
 		for (i = 0; i < sc->vmd_msix_count; i++) {
@@ -362,6 +389,7 @@ vmd_attach(device_t dev)
 				goto fail;
 			}
 		}
+		vmd_set_msi_bypass(dev, false);
 	}
 
 	sc->vmd_dma_tag = bus_get_dma_tag(dev);
@@ -386,6 +414,8 @@ vmd_detach(device_t dev)
 	error = device_delete_children(dev);
 	if (error)
 		return (error);
+	if (sc->vmd_msix_count == 0)
+		vmd_set_msi_bypass(dev, false);
 	vmd_free(sc);
 	return (0);
 }
@@ -482,6 +512,11 @@ vmd_alloc_msi(device_t dev, device_t child, int count, int maxcount,
 	struct vmd_irq_user *u;
 	int i, ibest = 0, best = INT_MAX;
 
+	if (sc->vmd_msix_count == 0) {
+		return (PCIB_ALLOC_MSI(device_get_parent(device_get_parent(dev)),
+		    child, count, maxcount, irqs));
+	}
+
 	if (count > vmd_max_msi)
 		return (ENOSPC);
 	LIST_FOREACH(u, &sc->vmd_users, viu_link) {
@@ -513,6 +548,11 @@ vmd_release_msi(device_t dev, device_t child, int count, int *irqs)
 	struct vmd_softc *sc = device_get_softc(dev);
 	struct vmd_irq_user *u;
 
+	if (sc->vmd_msix_count == 0) {
+		return (PCIB_RELEASE_MSI(device_get_parent(device_get_parent(dev)),
+		    child, count, irqs));
+	}
+
 	LIST_FOREACH(u, &sc->vmd_users, viu_link) {
 		if (u->viu_child == child) {
 			sc->vmd_irq[u->viu_vector].vi_nusers -= count;
@@ -531,6 +571,11 @@ vmd_alloc_msix(device_t dev, device_t child, int *irq)
 	struct vmd_irq_user *u;
 	int i, ibest = 0, best = INT_MAX;
 
+	if (sc->vmd_msix_count == 0) {
+		return (PCIB_ALLOC_MSIX(device_get_parent(device_get_parent(dev)),
+		    child, irq));
+	}
+
 	i = 0;
 	LIST_FOREACH(u, &sc->vmd_users, viu_link) {
 		if (u->viu_child == child)
@@ -562,6 +607,11 @@ vmd_release_msix(device_t dev, device_t child, int irq)
 	struct vmd_softc *sc = device_get_softc(dev);
 	struct vmd_irq_user *u;
 
+	if (sc->vmd_msix_count == 0) {
+		return (PCIB_RELEASE_MSIX(device_get_parent(device_get_parent(dev)),
+		    child, irq));
+	}
+
 	LIST_FOREACH(u, &sc->vmd_users, viu_link) {
 		if (u->viu_child == child &&
 		    sc->vmd_irq[u->viu_vector].vi_irq == irq) {
@@ -580,6 +630,11 @@ vmd_map_msi(device_t dev, device_t child, int irq, uint64_t *addr, uint32_t *dat
 	struct vmd_softc *sc = device_get_softc(dev);
 	int i;
 
+	if (sc->vmd_msix_count == 0) {
+		return (PCIB_MAP_MSI(device_get_parent(device_get_parent(dev)),
+		    child, irq, addr, data));
+	}
+
 	for (i = sc->vmd_fist_vector; i < sc->vmd_msix_count; i++) {
 		if (sc->vmd_irq[i].vi_irq == irq)
 			break;