Thank you very much for the feedback,
I'm testing with Ubuntu 18.04, nfs-common 1:1.3.4-2.1ubuntu5.2.
I think this means "nfs utils 1.3.4".
I tried explicitly listing all submounts in exports, and specifying an
fsid everywhere, and that worked, for example:
exports:
/home
*(fsid=4858dab5b4ac16ad2b7d274698c2532a,rw,async,crossmnt,no_subtree_check,no_root_squash,insecure)
/home/share
*(fsid=8c1748909cac2548372caead5bab9aa5,rw,async,no_subtree_check,no_root_squash,insecure)
...and the clients only mount /home, and then properly see the bindfs
share permissions.
But I'd really like to avoid that as in the real scenario there are many
submounts which are frequently added/removed, not just /home/share.
I'll try to follow your advice for debugging information.
On 10/11/19 7:47 PM, J. Bruce Fields wrote:
On Fri, Oct 11, 2019 at 08:24:14AM +0300, Alkis Georgopoulos wrote:
I'm not sure if this is an NFS issue, or a bindfs issue, or if I'm
not using the appropriate NFS options.
I export my /home via NFS with:
/home *(rw,async,crossmnt,no_subtree_check,no_root_squash,insecure)
Inside my /home I'm providing a shared folder with a bindfs mount:
bindfs -u 1000 --create-for-user=1000 -g 100
--create-for-group=100 -p 770,af-x /home/share /home/share
I.e. this just sets fixed permissions for anything under /home/share.
And finally I mount /home on some NFS client (or on localhost):
mount -t nfs server:/home /home
The problem is that /home/share on the client doesn't show the
bindfs permissions, but it shows the underlying file system of the
server's /home/share.
The crossmnt NFS option follows submounts with other file systems,
but not with bindfs.
On the other hand, if the bindfs source is on a different file
system than the bindfs target directory, everything works fine (i.e.
bindfs /other/filesystem/share /home/share).
Huh. I wonder if nfsd is for some reason determining the existence of a
mountpoint by comparing some kind of filesystem id and not seeing a
change. Looking at the code to remind myself how this works....
nfsd_mountpoint() is using d_mountpoint() and follow_down(), which
should be right. Then it's making an upcall to mountd. That's handled
by nfs-utils/mountd/cache.c:nfsd_export().
The is_mountpoint() check there is indeed going to return false in your
case because it's just comparing inode and device numbers.... But I
think that case is only for the "mountpoint" export option.
So I think all that matters is that export_matches() does the right
thing, and it certainly looks like it does--it should succeed as long as
there's a parent directory that's exported with crossmnt.
There's some debugging you could try by looking at
net/sunrpc/nfsd.*/content or using strace to watch rpc.mountd's reads
and writes of net/sunrpc/nfsd.*/channel.
What version of nfs-utils are you on?
--b.
Is there any way to configure either NFS or bindfs, so that this
works when I only have one partition, i.e. when the share is on the
same file system as /home?
If anyone answers, please Cc me as I'm not in the list.
Thank you very much,
Alkis Georgopoulos
LTSP developer
