[Virtio-fs] False error message when DAX cache is disabled

[misono.tomohiro at fujitsu.com]

> * misono.tomohiro at fujitsu.com (misono.tomohiro at fujitsu.com) wrote:
> > Hi,
> 
> Hi Misono,
> 
> > I'm testing virtiofs on ARM enviroment.
> > Since current ARM does not support DAX, I use cache-size=0 for vhost-user-fs-pci parameter.
> >
> > This results in the following error message during unmount:
> > ```
> > vhost_user_fs_slave_unmap: unmap when DAX cache not present
> > lo_destroy: unmap during destroy failed ```
> 
> Thanks for reporting this,
> 
> > This is because lo_destroy() always calls fuse_virtio_unmap().
> > So, I want to fix this by not calling fuse_virtio_unmap() when cache
> > is not used, but is there any good way for virtiofsd to know vhost-user's cahce-size?
> 
> Unfortunately not; that's why I used the special ~0 as a 'unmap all'.
> It feels easier to fix the qemu side to be happy unmapping everything when everything is actually nothing :-)
> 
> Does the following (build tested only) patch fix it:
> 
> diff --git a/hw/virtio/vhost-user-fs.c b/hw/virtio/vhost-user-fs.c index 6cacdb24b0..07a04d9b21 100644
> --- a/hw/virtio/vhost-user-fs.c
> +++ b/hw/virtio/vhost-user-fs.c
> @@ -92,10 +92,6 @@ uint64_t vhost_user_fs_slave_unmap(struct vhost_dev *dev, VhostUserFSSlaveMsg *s
>          return -1;
>      }
>      size_t cache_size = fs->conf.cache_size;
> -    if (!cache_size) {
> -        fprintf(stderr, "%s: unmap when DAX cache not present\n", __func__);
> -        return (uint64_t)-1;
> -    }
>      void *cache_host = memory_region_get_ram_ptr(&fs->cache);
> 
>      unsigned int i;
> @@ -106,13 +102,19 @@ uint64_t vhost_user_fs_slave_unmap(struct vhost_dev *dev, VhostUserFSSlaveMsg *s
>       */
>      for (i = 0; i < VHOST_USER_FS_SLAVE_ENTRIES; i++) {
>          void *ptr;
> +        if (sm->len[i] == ~(uint64_t)0) {
> +            /* Special case meaning the whole arena */
> +            sm->len[i] = cache_size;
> +        }
> +
>          if (sm->len[i] == 0) {
>              continue;
>          }
> 
> -        if (sm->len[i] == ~(uint64_t)0) {
> -            /* Special case meaning the whole arena */
> -            sm->len[i] = cache_size;
> +        if (!cache_size) {
> +            fprintf(stderr, "%s: unmap when DAX cache not present\n", __func__);
> +            res = -1;
> +            break;
>          }
> 
>          if ((sm->c_offset[i] + sm->len[i]) < sm->len[i] ||
> 

Thanks for the patch, but I got qemu crash during umount:
  qemu/memory.c:2187: memory_region_get_ram_ptr: Assertion `mr->ram_block' failed.

Obviously we don't setup fs->cache when cache_size is zero...
So, maybe we can just do something like this?

diff --git a/hw/virtio/vhost-user-fs.c b/hw/virtio/vhost-user-fs.c
index 6cacdb24b0..459dce5bde 100644
--- a/hw/virtio/vhost-user-fs.c
+++ b/hw/virtio/vhost-user-fs.c
@@ -93,6 +93,14 @@ uint64_t vhost_user_fs_slave_unmap(struct vhost_dev *dev, VhostUserFSSlaveMsg *s
     }
     size_t cache_size = fs->conf.cache_size;
     if (!cache_size) {
+        /*
+         * Since dax cache is disabled, there should be no unmap request.
+         * Howerver we still receives whole range unmap request during umount
+         * for cleanup. Ignore it.
+         */
+        if (sm->len[0] == ~(uint64_t)0)
+            return 0;
+
         fprintf(stderr, "%s: unmap when DAX cache not present\n", __func__);
         return (uint64_t)-1;
     }

Thanks,
Misono