This patch aims at optimizing IRQ handling using irqfd framework.
Instead of handling the eventfds on user-side they are handled on
kernel side using
- the KVM irqfd framework,
- the VFIO driver virqfd framework.
the virtual IRQ completion is trapped at interrupt controller
This removes the need for fast/slow path swap.
Overall this brings significant performance improvements.
it depends on host kernel KVM irqfd.
Signed-off-by: Alvise Rigo <a.r...@virtualopensystems.com>
Signed-off-by: Eric Auger <eric.au...@linaro.org>
---
v5 -> v6
- rely on kvm_irqfds_enabled() and kvm_resamplefds_enabled()
- guard KVM code with #ifdef CONFIG_KVM
v3 -> v4:
[Alvise Rigo]
Use of VFIO Platform driver v6 unmask/virqfd feature and removal
of resamplefd handler. Physical IRQ unmasking is now done in
VFIO driver.
v3:
[Eric Auger]
initial support with resamplefd handled on QEMU side since the
unmask was not supported on VFIO platform driver v5.
Conflicts:
hw/vfio/platform.c
---
hw/vfio/platform.c | 96 +++++++++++++++++++++++++++++++++++++++++
include/hw/vfio/vfio-platform.h | 1 +
trace-events | 2 +
3 files changed, 99 insertions(+)
diff --git a/hw/vfio/platform.c b/hw/vfio/platform.c
index 41f8693..97d98bf 100644
--- a/hw/vfio/platform.c
+++ b/hw/vfio/platform.c
@@ -25,6 +25,7 @@
#include "hw/sysbus.h"
#include "trace.h"
#include "hw/platform-bus.h"
+#include "sysemu/kvm.h"
static void vfio_intp_interrupt(VFIOINTp *intp);
typedef void (*eventfd_user_side_handler_t)(VFIOINTp *intp);
@@ -236,6 +237,83 @@ static int vfio_start_eventfd_injection(VFIOINTp *intp)
}
/*
+ * Functions used for irqfd
+ */
+
+#ifdef CONFIG_KVM
+
+/**
+ * vfio_set_resample_eventfd - sets the resamplefd for an IRQ
+ * @intp: the IRQ struct pointer
+ * programs the VFIO driver to unmask this IRQ when the
+ * intp->unmask eventfd is triggered
+ */
+static int vfio_set_resample_eventfd(VFIOINTp *intp)
+{
+ VFIODevice *vbasedev = &intp->vdev->vbasedev;
+ struct vfio_irq_set *irq_set;
+ int argsz, ret;
+ int32_t *pfd;
+
+ argsz = sizeof(*irq_set) + sizeof(*pfd);
+ irq_set = g_malloc0(argsz);
+ irq_set->argsz = argsz;
+ irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
+ irq_set->index = intp->pin;
+ irq_set->start = 0;
+ irq_set->count = 1;
+ pfd = (int32_t *)&irq_set->data;
+ *pfd = event_notifier_get_fd(&intp->unmask);
+ qemu_set_fd_handler(*pfd, NULL, NULL, intp);
+ ret = ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
+ g_free(irq_set);
+ if (ret < 0) {
+ error_report("vfio: Failed to set resample eventfd: %m");
+ qemu_set_fd_handler(*pfd, NULL, NULL, NULL);
+ }
+ return ret;
+}
+
+/**
+ * vfio_start_irqfd_injection - starts irqfd injection for an IRQ
+ * programs VFIO driver with both the trigger and resamplefd
+ * programs KVM with the gsi, trigger & resample eventfds
+ */
+static int vfio_start_irqfd_injection(VFIOINTp *intp)
+{
+ struct kvm_irqfd irqfd = {
+ .fd = event_notifier_get_fd(&intp->interrupt),
+ .resamplefd = event_notifier_get_fd(&intp->unmask),
+ .gsi = intp->virtualID,
+ .flags = KVM_IRQFD_FLAG_RESAMPLE,
+ };
+
+ if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
+ error_report("vfio: Error: Failed to assign the irqfd: %m");
+ goto fail_irqfd;
+ }
+ if (vfio_set_trigger_eventfd(intp, NULL) < 0) {
+ goto fail_vfio;
+ }
+ if (vfio_set_resample_eventfd(intp) < 0) {
+ goto fail_vfio;
+ }
+
+ intp->kvm_accel = true;
+ trace_vfio_platform_start_irqfd_injection(intp->pin, intp->virtualID,
+ irqfd.fd, irqfd.resamplefd);
+ return 0;
+
+fail_vfio:
+ irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
+ kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
+fail_irqfd:
+ return -1;
+}
+
+#endif
+
+/*
* Functions used whatever the injection method
*/
@@ -314,6 +392,13 @@ static VFIOINTp *vfio_init_intp(VFIODevice *vbasedev,
unsigned int index)
error_report("vfio: Error: trigger event_notifier_init failed ");
return NULL;
}
+ /* Get an eventfd for resample/unmask */
+ ret = event_notifier_init(&intp->unmask, 0);
+ if (ret) {
+ g_free(intp);
+ error_report("vfio: Error: resample event_notifier_init failed eoi");
+ return NULL;
+ }
/* store the new intp in qlist */
QLIST_INSERT_HEAD(&vdev->intp_list, intp, next);
@@ -520,7 +605,17 @@ static void vfio_platform_realize(DeviceState *dev, Error
**errp)
vbasedev->type = VFIO_DEVICE_TYPE_PLATFORM;
vbasedev->ops = &vfio_platform_ops;
+
+#ifdef CONFIG_KVM
+ if (kvm_irqfds_enabled() && kvm_resamplefds_enabled() &&
+ vdev->irqfd_allowed) {
+ vdev->start_irq_fn = vfio_start_irqfd_injection;
+ } else {
+ vdev->start_irq_fn = vfio_start_eventfd_injection;
+ }
+#else
vdev->start_irq_fn = vfio_start_eventfd_injection;
+#endif
trace_vfio_platform_realize(vbasedev->name, vdev->compat);
@@ -598,6 +693,7 @@ static Property vfio_platform_dev_properties[] = {
DEFINE_PROP_STRING("host", VFIOPlatformDevice, vbasedev.name),
DEFINE_PROP_UINT32("mmap-timeout-ms", VFIOPlatformDevice,
mmap_timeout, 1100),
+ DEFINE_PROP_BOOL("x-irqfd", VFIOPlatformDevice, irqfd_allowed, true),
DEFINE_PROP_END_OF_LIST(),
};
diff --git a/include/hw/vfio/vfio-platform.h b/include/hw/vfio/vfio-platform.h
index 449c049..de0b5d5 100644
--- a/include/hw/vfio/vfio-platform.h
+++ b/include/hw/vfio/vfio-platform.h
@@ -58,6 +58,7 @@ typedef struct VFIOPlatformDevice {
QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
start_irq_fn_t start_irq_fn;
QemuMutex intp_mutex;
+ bool irqfd_allowed; /* debug option to force irqfd on/off */
} VFIOPlatformDevice;
diff --git a/trace-events b/trace-events
index 7a9e780..59a09f6 100644
--- a/trace-events
+++ b/trace-events
@@ -1441,6 +1441,8 @@ vfio_platform_base_device_init(char *name, int groupid)
"%s belongs to group #%d
vfio_platform_realize(char *name, char *compat) "vfio device %s, compat = %s"
vfio_intp_interrupt_set_pending(int index) "irq %d is set PENDING"
vfio_platform_eoi_handle_pending(int index) "handle PENDING IRQ %d"
+vfio_platform_start_irqfd_injection(int index, int gsi, int fd, int
resamplefd) "IRQ index=%d, gsi =%d, fd = %d, resamplefd = %d"
+vfio_start_eventfd_injection(int index, int fd) "IRQ index=%d, fd = %d"
#hw/acpi/memory_hotplug.c
mhp_acpi_invalid_slot_selected(uint32_t slot) "0x%"PRIx32
--
1.8.3.2
