[LSF/MM TOPIC] filesystem virtualization

[LSF/MM TOPIC] filesystem virtualization

[Miklos Szeredi]

This is a joint topic proposal with Vivek Goyal.

Discuss remaining issues related to exporting a filesystem from host to
guest (see virtio-fs prototype posting[1] for more background):

* How to provide strong coherency guarantees?

   - Fsnotify synchronous post op notification
   - Extend lease (delegation) API to userspace
   - Multi-component path revalidation

* Issues related to c/m/atime handling:

   - Clock synchronization between host and guest (not an fs issue, but fs
     has to deal with lack of it)

   - Could we have O_NOCMTIME open flag?

* Shared access to file data:

   - Use DAX in guest to access host page cache (mmap/munmap sections of
     file in guest address space; happens dynamically on a need basis); is
     this design reasonable?

   - Avoids memory copies and cache duplication, but host page table setup
     may have high cost.  Can that be improved?  (E.g. fault-ahead)

   - Too many VMA’s on host per qemu instance?

* File locking:

   - Host API for POSIX lock sharing?

* Ideas for the future:

   - Buffering directory operations (e.g. "tar xfz ..." with as few
     synchronous ops as possible)

[1] https://lore.kernel.org/lkml/20181210171318.16998-1-vgoyal@redhat.com/

Re: [LSF/MM TOPIC] filesystem virtualization

[Amir Goldstein]

On Wed, Feb 6, 2019 at 3:06 PM Miklos Szeredi <miklos@szeredi.hu> wrote:
>
> This is a joint topic proposal with Vivek Goyal.
>
> Discuss remaining issues related to exporting a filesystem from host to
> guest (see virtio-fs prototype posting[1] for more background):
>
> * How to provide strong coherency guarantees?
>
>    - Fsnotify synchronous post op notification
>    - Extend lease (delegation) API to userspace

There are some commonalities here with userspace networking
file servers. I am interested in samba in particular and things
that we can do to improve Linux interfaces for samba.
SMB3 directory leases for example.

Thanks,
Amir.

>    - Multi-component path revalidation
>
> * Issues related to c/m/atime handling:
>
>    - Clock synchronization between host and guest (not an fs issue, but fs
>      has to deal with lack of it)
>
>    - Could we have O_NOCMTIME open flag?
>
> * Shared access to file data:
>
>    - Use DAX in guest to access host page cache (mmap/munmap sections of
>      file in guest address space; happens dynamically on a need basis); is
>      this design reasonable?
>
>    - Avoids memory copies and cache duplication, but host page table setup
>      may have high cost.  Can that be improved?  (E.g. fault-ahead)
>
>    - Too many VMA’s on host per qemu instance?
>
> * File locking:
>
>    - Host API for POSIX lock sharing?
>
> * Ideas for the future:
>
>    - Buffering directory operations (e.g. "tar xfz ..." with as few
>      synchronous ops as possible)
>
> [1] https://lore.kernel.org/lkml/20181210171318.16998-1-vgoyal@redhat.com/