Re: [PATCH v7 8/8] vhost-user-device: Add shared memory BAR

[Stefan Hajnoczi <stefanha@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 8503 bytes --]

On Mon, Aug 18, 2025 at 12:03:53PM +0200, Albert Esteve wrote:
> Add shared memory BAR support to vhost-user-device-pci
> to enable direct file mapping for VIRTIO Shared
> Memory Regions.
> 
> The implementation creates a consolidated shared
> memory BAR that contains all VIRTIO Shared
> Memory Regions as subregions. Each region is
> configured with its proper shmid, size, and
> offset within the BAR. The number and size of
> regions are retrieved via VHOST_USER_GET_SHMEM_CONFIG
> message sent by vhost-user-base during realization
> after virtio_init().
> 
> Specifiically, it uses BAR 3 to avoid conflicts, as
> it is currently unused.
> 
> The shared memory BAR is only created when the
> backend supports VHOST_USER_PROTOCOL_F_SHMEM and
> has configured shared memory regions. This maintains
> backward compatibility with backends that do not
> support shared memory functionality.
> 
> Signed-off-by: Albert Esteve <aesteve@redhat.com>
> ---
>  hw/virtio/vhost-user-base.c       | 49 +++++++++++++++++++++++++++++--
>  hw/virtio/vhost-user-device-pci.c | 34 +++++++++++++++++++--
>  2 files changed, 78 insertions(+), 5 deletions(-)
> 
> diff --git a/hw/virtio/vhost-user-base.c b/hw/virtio/vhost-user-base.c
> index ff67a020b4..932f9b5596 100644
> --- a/hw/virtio/vhost-user-base.c
> +++ b/hw/virtio/vhost-user-base.c
> @@ -16,6 +16,7 @@
>  #include "hw/virtio/virtio-bus.h"
>  #include "hw/virtio/vhost-user-base.h"
>  #include "qemu/error-report.h"
> +#include "migration/blocker.h"
>  
>  static void vub_start(VirtIODevice *vdev)
>  {
> @@ -276,7 +277,9 @@ static void vub_device_realize(DeviceState *dev, Error **errp)
>  {
>      VirtIODevice *vdev = VIRTIO_DEVICE(dev);
>      VHostUserBase *vub = VHOST_USER_BASE(dev);
> -    int ret;
> +    uint64_t memory_sizes[VIRTIO_MAX_SHMEM_REGIONS];
> +    g_autofree char *name = NULL;
> +    int i, ret, nregions;
>  
>      if (!vub->chardev.chr) {
>          error_setg(errp, "vhost-user-base: missing chardev");
> @@ -319,7 +322,7 @@ static void vub_device_realize(DeviceState *dev, Error **errp)
>  
>      /* Allocate queues */
>      vub->vqs = g_ptr_array_sized_new(vub->num_vqs);
> -    for (int i = 0; i < vub->num_vqs; i++) {
> +    for (i = 0; i < vub->num_vqs; i++) {
>          g_ptr_array_add(vub->vqs,
>                          virtio_add_queue(vdev, vub->vq_size,
>                                           vub_handle_output));
> @@ -333,11 +336,51 @@ static void vub_device_realize(DeviceState *dev, Error **errp)
>                           VHOST_BACKEND_TYPE_USER, 0, errp);
>  
>      if (ret < 0) {
> -        do_vhost_user_cleanup(vdev, vub);
> +        goto err;
> +    }
> +
> +    ret = vub->vhost_dev.vhost_ops->vhost_get_shmem_config(&vub->vhost_dev,
> +                                                           &nregions,
> +                                                           memory_sizes,
> +                                                           errp);
> +
> +    if (ret < 0) {
> +        goto err;
> +    }
> +
> +    for (i = 0; i < nregions; i++) {
> +        if (memory_sizes[i]) {
> +            if (vub->vhost_dev.migration_blocker == NULL) {
> +                error_setg(&vub->vhost_dev.migration_blocker,
> +                       "Migration disabled: devices with VIRTIO Shared Memory "
> +                       "Regions do not support migration yet.");
> +                ret = migrate_add_blocker_normal(
> +                    &vub->vhost_dev.migration_blocker,
> +                    errp);
> +
> +                if (ret < 0) {
> +                    goto err;
> +                }
> +            }
> +
> +            if (memory_sizes[i] % qemu_real_host_page_size() != 0) {
> +                error_setg(errp, "Shared memory %d size must be a power of 2 "
> +                                 "no smaller than the page size", i);
> +                goto err;
> +            }
> +
> +            name = g_strdup_printf("vub-shm-%d", i);

name is leaked because it's scope extends until the end of the function
(after the loop) but a newly allocated string is assigned each time
around the loop. This can be fixed by moving the local variable
declaration inside the if statement body.

> +            memory_region_init(&virtio_new_shmem_region(vdev, i)->mr,
> +                               OBJECT(vdev), name,
> +                               memory_sizes[i]);

->mr is already initialized inside virtio_new_shmem_region(). I suggest
changing the definition of virtio_new_shmem_region() like this:

  void virtio_add_shmem_region(VirtIODevice *vdev, uint8_t shmid,
                               uint64_t size)

and then calling it like this:

  virtio_add_shmem_region(vdev, shmid, memory_sizes[i]);

("new" usually returns a new instance whereas "add" modifies an owner
object/container. I think "add" is more appropriate here.)

> +        }
>      }
>  
>      qemu_chr_fe_set_handlers(&vub->chardev, NULL, NULL, vub_event, NULL,
>                               dev, NULL, true);
> +    return;
> +err:
> +    do_vhost_user_cleanup(vdev, vub);
>  }
>  
>  static void vub_device_unrealize(DeviceState *dev)
> diff --git a/hw/virtio/vhost-user-device-pci.c b/hw/virtio/vhost-user-device-pci.c
> index f10bac874e..bac99e7c60 100644
> --- a/hw/virtio/vhost-user-device-pci.c
> +++ b/hw/virtio/vhost-user-device-pci.c
> @@ -8,14 +8,18 @@
>   */
>  
>  #include "qemu/osdep.h"
> +#include "qapi/error.h"
>  #include "hw/qdev-properties.h"
>  #include "hw/virtio/vhost-user-base.h"
>  #include "hw/virtio/virtio-pci.h"
>  
> +#define VIRTIO_DEVICE_PCI_SHMEM_BAR 3
> +
>  struct VHostUserDevicePCI {
>      VirtIOPCIProxy parent_obj;
>  
>      VHostUserBase vub;
> +    MemoryRegion shmembar;
>  };
>  
>  #define TYPE_VHOST_USER_DEVICE_PCI "vhost-user-device-pci-base"
> @@ -25,10 +29,36 @@ OBJECT_DECLARE_SIMPLE_TYPE(VHostUserDevicePCI, VHOST_USER_DEVICE_PCI)
>  static void vhost_user_device_pci_realize(VirtIOPCIProxy *vpci_dev, Error **errp)
>  {
>      VHostUserDevicePCI *dev = VHOST_USER_DEVICE_PCI(vpci_dev);
> -    DeviceState *vdev = DEVICE(&dev->vub);
> +    DeviceState *dev_state = DEVICE(&dev->vub);
> +    VirtIODevice *vdev = VIRTIO_DEVICE(dev_state);
> +    VirtioSharedMemory *shmem, *next;
> +    uint64_t offset = 0, shmem_size = 0;
>  
>      vpci_dev->nvectors = 1;
> -    qdev_realize(vdev, BUS(&vpci_dev->bus), errp);
> +    qdev_realize(dev_state, BUS(&vpci_dev->bus), errp);
> +
> +    QSIMPLEQ_FOREACH_SAFE(shmem, &vdev->shmem_list, entry, next) {

This is not specific to vhost-user-device-pci.c. All VIRTIO devices with
Shared Memory Regions need PCI BAR setup code. Since vdev->shmem_list is
part of the core hw/virtio/ code, it would make sense to move this into
into hw/virtio/virtio-pci.c.

> +        if (shmem->mr.size > UINT64_MAX - shmem_size) {
> +            error_setg(errp, "Total shared memory required overflow");
> +            return;
> +        }
> +        shmem_size = shmem_size + shmem->mr.size;
> +    }
> +    if (shmem_size) {
> +        memory_region_init(&dev->shmembar, OBJECT(vpci_dev),
> +                           "vhost-device-pci-shmembar", shmem_size);
> +        QSIMPLEQ_FOREACH_SAFE(shmem, &vdev->shmem_list, entry, next) {
> +            memory_region_add_subregion(&dev->shmembar, offset, &shmem->mr);
> +            virtio_pci_add_shm_cap(vpci_dev, VIRTIO_DEVICE_PCI_SHMEM_BAR,
> +                                   offset, shmem->mr.size, shmem->shmid);
> +            offset = offset + shmem->mr.size;
> +        }
> +        pci_register_bar(&vpci_dev->pci_dev, VIRTIO_DEVICE_PCI_SHMEM_BAR,
> +                        PCI_BASE_ADDRESS_SPACE_MEMORY |
> +                        PCI_BASE_ADDRESS_MEM_PREFETCH |
> +                        PCI_BASE_ADDRESS_MEM_TYPE_64,
> +                        &dev->shmembar);

This does not follow the same approach as virtio-gpu-pci.c and
virtio-vga.c. They config the VirtIOPCIProxy's BARs
(->modern_io_bar_idx, ->modern_mem_bar_idx, and ->msix_bar_idx) to
control the BAR layout first and then call qdev_realize().

Why does this patch do things differently? It looks like it's assuming
vpci_dev always has a specific BAR layout (it could change).

> +    }

>  }
>  
>  static void vhost_user_device_pci_class_init(ObjectClass *klass,
> -- 
> 2.49.0
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]