[Bug 269328] nfs client: data corruption using fspacectl and mmap

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=269328

Konstantin Belousov <kib@FreeBSD.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pho@FreeBSD.org

--- Comment #3 from Konstantin Belousov <kib@FreeBSD.org> ---
It would be somewhat interesting to see where the corruption occurs, on client
or on server.  My intuition is that the problem happens on client, esp. because
server does not have fspacectl() at all.

From the code reading, nfs client does enough page flushing both in
nfs_deallocate() and in ncl_vinvalbuf().  At least, it is not obvious what
is missed and why msync(2) before deallocation helps. The sequence seems to
be, from nfs_deallocate() for NFSv4.2:
  ncl_vinvalbuf(V_SAVE)
    vm_object_page_clean(OBJPC_SYNC) <- sync pages synchronously
    vinvalbuf(V_SAVE)
      vm_object_page_remove(OBJPR_CLEANONLY) <- only remove clean pages
    vnode_pager_purge_range()
      vm_object_page_remove(0)
    nfsrpc_deallocate()

It might be that a clean page somehow avoided being removed from the queue,
but as I said, I do not see how this could occur.  Or it might be the handling
of the partially deallocated pages is somewhat wrong.

-- 
You are receiving this mail because:
You are the assignee for the bug.