Re: [PATCH V1 12/26] vfio-pci: preserve MSI

["Cédric Le Goater" <clg@redhat.com>]

On 1/29/25 15:43, Steve Sistare wrote:
> Save the MSI message area as part of vfio-pci vmstate, and preserve the
> interrupt and notifier eventfd's.  migrate_incoming loads the MSI data,
> then the vfio-pci post_load handler finds the eventfds in CPR state,
> rebuilds vector data structures, and attaches the interrupts to the new
> KVM instance.
> 
> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
> ---
>   hw/vfio/pci.c | 117 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
>   1 file changed, 116 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
> index fa77c36..df6e298 100644
> --- a/hw/vfio/pci.c
> +++ b/hw/vfio/pci.c
> @@ -56,11 +56,37 @@ static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
>   static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
>   static void vfio_msi_disable_common(VFIOPCIDevice *vdev);
>   
> +#define EVENT_FD_NAME(vdev, name)   \
> +    g_strdup_printf("%s_%s", (vdev)->vbasedev.name, (name))

hmm, this helper could lead to memory leaks if not used as done below.
Being explict would be safer.

> +static void save_event_fd(VFIOPCIDevice *vdev, const char *name, int nr,
> +                          EventNotifier *ev)
> +{
> +    int fd = event_notifier_get_fd(ev);
> +
> +    if (fd >= 0) {
> +        g_autofree char *fdname = EVENT_FD_NAME(vdev, name);
> +        cpr_resave_fd(fdname, nr, fd);
> +    }
> +}
> +
> +static int load_event_fd(VFIOPCIDevice *vdev, const char *name, int nr)
> +{
> +    g_autofree char *fdname = EVENT_FD_NAME(vdev, name);
> +    return cpr_find_fd(fdname, nr);
> +}
> +
> +static void delete_event_fd(VFIOPCIDevice *vdev, const char *name, int nr)
> +{
> +    g_autofree char *fdname = EVENT_FD_NAME(vdev, name);
> +    cpr_delete_fd(fdname, nr);
> +}
> +

please move these helpers to a cpr file. They are not strictly VFIO
related too. So they could me moved outside of hw/vfio.


>   /* Create new or reuse existing eventfd */
>   static int vfio_notifier_init(VFIOPCIDevice *vdev, EventNotifier *e,
>                                 const char *name, int nr)
>   {
> -    int fd = -1;   /* placeholder until a subsequent patch */
> +    int fd = load_event_fd(vdev, name, nr);


>       int ret = 0;
>   
>       if (fd >= 0) {
> @@ -71,6 +97,8 @@ static int vfio_notifier_init(VFIOPCIDevice *vdev, EventNotifier *e,
>               Error *err = NULL;
>               error_setg_errno(&err, -ret, "vfio_notifier_init %s failed", name);
>               error_report_err(err);
> +        } else {
> +            save_event_fd(vdev, name, nr, e);

I'd rather move the CPR related fd handling which is done in
vfio_notifier_init() in a cpr routine which vfio_notifier_init()
would call. This comment applies to all the series. Anything
related to CPR should be handled explicitely :

     if (cpr_in_progress) {
         cpr_do_cpr_related_stuff()
     }

It will ease reading and long term maintenance.



>           }
>       }
>       return ret;
> @@ -79,6 +107,7 @@ static int vfio_notifier_init(VFIOPCIDevice *vdev, EventNotifier *e,
>   static void vfio_notifier_cleanup(VFIOPCIDevice *vdev, EventNotifier *e,
>                                     const char *name, int nr)
>   {
> +    delete_event_fd(vdev, name, nr);
>       event_notifier_cleanup(e);
>   }
>   
> @@ -561,6 +590,15 @@ static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
>       int ret;
>       bool resizing = !!(vdev->nr_vectors < nr + 1);
>   
> +    /*
> +     * Ignore the callback from msix_set_vector_notifiers during resume.
> +     * The necessary subset of these actions is called from vfio_claim_vectors
> +     * during post load.
> +     */
> +    if (vdev->vbasedev.reused) {
> +        return 0;
> +    }

again, I would prefer some explicit CPR test. Same below.


>       trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
>   
>       vector = &vdev->msi_vectors[nr];
> @@ -2896,6 +2934,11 @@ static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
>       fd = event_notifier_get_fd(&vdev->err_notifier);
>       qemu_set_fd_handler(fd, vfio_err_notifier_handler, NULL, vdev);
>   
> +    /* Do not alter irq_signaling during vfio_realize for cpr */
> +    if (vdev->vbasedev.reused) {
> +        return;
> +    }
> +
>       if (!vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_ERR_IRQ_INDEX, 0,
>                                   VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
>           error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
> @@ -2960,6 +3003,12 @@ static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
>       fd = event_notifier_get_fd(&vdev->req_notifier);
>       qemu_set_fd_handler(fd, vfio_req_notifier_handler, NULL, vdev);
>   
> +    /* Do not alter irq_signaling during vfio_realize for cpr */
> +    if (vdev->vbasedev.reused) {
> +        vdev->req_enabled = true;
> +        return;
> +    }
> +
>       if (!vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_REQ_IRQ_INDEX, 0,
>                                   VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
>           error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
> @@ -3454,6 +3503,46 @@ static void vfio_pci_set_fd(Object *obj, const char *str, Error **errp)
>   }
>   #endif
>   
> +static void vfio_claim_vectors(VFIOPCIDevice *vdev, int nr_vectors, bool msix)
> +{
> +    int i, fd;
> +    bool pending = false;
> +    PCIDevice *pdev = &vdev->pdev;
> +
> +    vdev->nr_vectors = nr_vectors;
> +    vdev->msi_vectors = g_new0(VFIOMSIVector, nr_vectors);
> +    vdev->interrupt = msix ? VFIO_INT_MSIX : VFIO_INT_MSI;
> +
> +    vfio_prepare_kvm_msi_virq_batch(vdev);
> +
> +    for (i = 0; i < nr_vectors; i++) {
> +        VFIOMSIVector *vector = &vdev->msi_vectors[i];
> +
> +        fd = load_event_fd(vdev, "interrupt", i);
> +        if (fd >= 0) {
> +            vfio_vector_init(vdev, i);
> +            qemu_set_fd_handler(fd, vfio_msi_interrupt, NULL, vector);
> +        }
> +
> +        if (load_event_fd(vdev, "kvm_interrupt", i) >= 0) {
> +            vfio_add_kvm_msi_virq(vdev, vector, i, msix);
> +        } else {
> +            vdev->msi_vectors[i].virq = -1;
> +        }
> +
> +        if (msix && msix_is_pending(pdev, i) && msix_is_masked(pdev, i)) {
> +            set_bit(i, vdev->msix->pending);
> +            pending = true;
> +        }
> +    }
> +
> +    vfio_commit_kvm_msi_virq_batch(vdev);
> +
> +    if (msix) {
> +        memory_region_set_enabled(&pdev->msix_pba_mmio, pending);
> +    }
> +}

move to a cpr file please. We can have a vfio-pci lib/common file
for external users. It will take more work to get the interface right
but it will benefit other proposals. I think vfio-user as more or less
the same needs.


> +
>   /*
>    * The kernel may change non-emulated config bits.  Exclude them from the
>    * changed-bits check in get_pci_config_device.
> @@ -3472,13 +3561,39 @@ static int vfio_pci_pre_load(void *opaque)
>       return 0;
>   }
>   
> +static int vfio_pci_post_load(void *opaque, int version_id)
> +{
> +    VFIOPCIDevice *vdev = opaque;
> +    PCIDevice *pdev = &vdev->pdev;
> +    int nr_vectors;
> +
> +    if (msix_enabled(pdev)) {
> +        msix_set_vector_notifiers(pdev, vfio_msix_vector_use,
> +                                   vfio_msix_vector_release, NULL);
> +        nr_vectors = vdev->msix->entries;
> +        vfio_claim_vectors(vdev, nr_vectors, true);
> +
> +    } else if (msi_enabled(pdev)) {
> +        nr_vectors = msi_nr_vectors_allocated(pdev);
> +        vfio_claim_vectors(vdev, nr_vectors, false);
> +
> +    } else if (vfio_pci_read_config(pdev, PCI_INTERRUPT_PIN, 1)) {
> +        g_assert_not_reached();      /* completed in a subsequent patch */
> +    }
> +
> +    return 0;
> +}
> +
>   static const VMStateDescription vfio_pci_vmstate = {
>       .name = "vfio-pci",
>       .version_id = 0,
>       .minimum_version_id = 0,
>       .pre_load = vfio_pci_pre_load,
> +    .post_load = vfio_pci_post_load,
>       .needed = cpr_needed_for_reuse,
>       .fields = (VMStateField[]) {
> +        VMSTATE_PCI_DEVICE(pdev, VFIOPCIDevice),
> +        VMSTATE_MSIX_TEST(pdev, VFIOPCIDevice, vfio_msix_present),
>           VMSTATE_END_OF_LIST()
>       }
>   };


I think you can move vfio_pci_vmstate out of hw/vfio/pci.c too. Only
cpr needs it.

Thanks,

C.