Re: [Qemu-devel] [PATCH v2 2/2] ivshmem: use irqfd to interrupt among VMs

[Cam Macdonell <cam@cs.ualberta.ca>]

On Sat, Nov 24, 2012 at 8:51 PM, Liu Ping Fan <qemulist@gmail.com> wrote:
> From: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
>
> Using irqfd, so we can avoid switch between kernel and user when
> VMs interrupts each other.
>

Hi Liu Ping,

With this patch applied I was still seeing transitions to user-level
on the receipt of an msi interrupt.

uncomment DEBUG_IVSHMEM in hw/ivshmem.c (and fix one compile error in
the debug statement :) )

IVSHMEM: msix initialized (1 vectors)
...
IVSHMEM: interrupt on vector 0x7f2971b1d8d0 0
IVSHMEM: interrupt on vector 0x7f2971b1d8d0 0

if we're using irqfd, this should be avoided.  Here's my command-line:

-device ivshmem,chardev=nahanni,msi=on,ioeventfd=on,size=2048,use64=1,role=peer

There are two issues as I see it:
1) irqfd is not being enabled in my tests
2) the defaults handlers are still being added

One difference is that I'm using the UIO driver, which enables PCI
using pci_enable_msix as follows

pci_enable_msix(ivs_info->dev, ivs_info->msix_entries,
                    ivs_info->nvectors);

and succeeds

[    2.651253] uio_ivshmem 0000:00:04.0: irq 43 for MSI/MSI-X
[    2.651326] MSI-X enabled

(continued below)

> Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
> ---
>  hw/ivshmem.c |   54 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 files changed, 53 insertions(+), 1 deletions(-)
>
> diff --git a/hw/ivshmem.c b/hw/ivshmem.c
> index 7c8630c..5709e89 100644
> --- a/hw/ivshmem.c
> +++ b/hw/ivshmem.c
> @@ -19,6 +19,7 @@
>  #include "hw.h"
>  #include "pc.h"
>  #include "pci.h"
> +#include "msi.h"
>  #include "msix.h"
>  #include "kvm.h"
>  #include "migration.h"
> @@ -83,6 +84,7 @@ typedef struct IVShmemState {
>      uint32_t vectors;
>      uint32_t features;
>      EventfdEntry *eventfd_table;
> +    int *vector_virqs;
>
>      Error *migration_blocker;
>
> @@ -625,16 +627,62 @@ static int ivshmem_load(QEMUFile* f, void *opaque, int version_id)
>      return 0;
>  }
>
> +static int ivshmem_vector_use(PCIDevice *dev, unsigned vector,
> +                                     MSIMessage msg)
> +{
> +    IVShmemState *s = DO_UPCAST(IVShmemState, dev, dev);
> +    int virq;
> +    EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
> +
> +    virq = kvm_irqchip_add_msi_route(kvm_state, msg);
> +    if (virq >= 0 && kvm_irqchip_add_irqfd_notifier(kvm_state, n, virq) >= 0) {
> +        s->vector_virqs[vector] = virq;
> +        qemu_chr_add_handlers(s->eventfd_chr[vector], NULL, NULL, NULL, NULL);
> +    } else if (virq >= 0) {
> +        kvm_irqchip_release_virq(kvm_state, virq);
> +        error_report("ivshmem, can not setup irqfd\n");
> +        return -1;
> +    } else {
> +        error_report("ivshmem, no enough msi route to setup irqfd\n");
> +        return -1;
> +    }
> +
> +    return 0;
> +}
> +
> +static void ivshmem_vector_release(PCIDevice *dev, unsigned vector)
> +{
> +    IVShmemState *s = DO_UPCAST(IVShmemState, dev, dev);
> +    EventNotifier *n = &s->peers[s->vm_id].eventfds[vector];
> +    int virq = s->vector_virqs[vector];
> +
> +    if (s->vector_virqs[vector] >= 0) {
> +        kvm_irqchip_remove_irqfd_notifier(kvm_state, n, virq);
> +        kvm_irqchip_release_virq(kvm_state, virq);
> +        s->vector_virqs[vector] = -1;
> +    }
> +}
> +
>  static void ivshmem_write_config(PCIDevice *pci_dev, uint32_t address,
>                                  uint32_t val, int len)
>  {
> +    bool is_enabled, was_enabled = msi_enabled(pci_dev);
> +
>      pci_default_write_config(pci_dev, address, val, len);
> +    is_enabled = msi_enabled(pci_dev);

Problem 1)  in my tests is_enabled is always 0, so I don't think the
irqfds are getting setup

> +    if (!was_enabled && is_enabled) {
> +        msix_set_vector_notifiers(pci_dev, ivshmem_vector_use,
> +            ivshmem_vector_release);
> +    } else if (was_enabled && !is_enabled) {
> +        msix_unset_vector_notifiers(pci_dev);
> +    }
>  }
>
>  static int pci_ivshmem_init(PCIDevice *dev)
>  {
>      IVShmemState *s = DO_UPCAST(IVShmemState, dev, dev);
>      uint8_t *pci_conf;
> +    int i;
>
>      if (s->sizearg == NULL)
>          s->ivshmem_size = 4 << 20; /* 4 MB default */
> @@ -758,7 +806,10 @@ static int pci_ivshmem_init(PCIDevice *dev)
>      }
>
>      s->dev.config_write = ivshmem_write_config;
> -
> +    s->vector_virqs = g_new0(int, s->vectors);
> +    for (i = 0; i < s->vectors; i++) {
> +        s->vector_virqs[i] = -1;
> +    }
>      return 0;
>  }
>
> @@ -770,6 +821,7 @@ static void pci_ivshmem_uninit(PCIDevice *dev)
>          migrate_del_blocker(s->migration_blocker);
>          error_free(s->migration_blocker);
>      }
> +    g_free(s->vector_virqs);
>
>      memory_region_destroy(&s->ivshmem_mmio);
>      memory_region_del_subregion(&s->bar, &s->ivshmem);
> --
> 1.7.4.4
>

Problem 2)
We'll also have to not add the handlers as below if irqfd is present
otherwise we'll get double interrupts, so we'll have to add a check
here too.

    /* if MSI is supported we need multiple interrupts */
    if (!ivshmem_has_feature(s, IVSHMEM_MSI)) {
        s->eventfd_table[vector].pdev = &s->dev;
        s->eventfd_table[vector].vector = vector;

        qemu_chr_add_handlers(chr, ivshmem_can_receive, fake_irqfd,
                      ivshmem_event, &s->eventfd_table[vector]);
    } else {
        qemu_chr_add_handlers(chr, ivshmem_can_receive, ivshmem_receive,
                      ivshmem_event, s);
    }

Sincerely,
Cam