NFS: Propagate 'fsc' mount option through automounts
RFC PATCH 1/1 NFS: Propagate 'fsc' mount option through automounts
Propagate the NFS 'fsc' mount option through NFS automounts of various types.
This is now required as commit:
commit c02d7adf8c5429727a98bad1d039bccad4c61c50 Author: Trond Myklebust <[email protected]> Date: Mon Jun 22 15:09:14 2009 -0400 NFSv4: Replace nfs4_path_walk() with VFS path lookup in a private namespace
now uses VFS-driven automounting to reach all submounts barring the root, thus preventing fscaching from being enabled on any submount other than the root.
This patch gets around that by propagating the NFS_OPTION_FSCACHE flag across automounts.
Note, however, that it doesn't propagate the uniquifier in the case of 'fsc=<xxx>' being passed to mount. This is probably wrong, and needs looking at.
---
fs/nfs/client.c | 1 + fs/nfs/super.c | 6 ++++++ 2 files changed, 7 insertions(+), 0 deletions(-)
fs/nfs/client.c +1 -0
@@ -964,6 +964,7 @@ static void nfs_server_copy_userdata(struct nfs_server *target, struct nfs_serve target->acdirmin = source->acdirmin; target->acdirmax = source->acdirmax; target->caps = source->caps; + target->options = source->options; }/*
fs/nfs/super.c +6 -0
@@ -2214,6 +2214,7 @@ static int nfs_xdev_get_sb(struct file_system_type *fs_type, int flags,struct nfs_server *server;struct dentry *mntroot;int (*compare_super)(struct super_block *,void *) = nfs_compare_super; +struct nfs_parsed_mount_data parsed_data = { .fscache_uniq =NULL , };struct nfs_sb_mountdata sb_mntdata = { .mntflags = flags, };
@@ -2251,6 +2252,7 @@ static int nfs_xdev_get_sb(struct file_system_type *fs_type, int flags,if (!s->s_root) {/* initial superblock/root creation */ nfs_clone_super (s, data->sb); +nfs_fscache_get_super_cookie (s, &parsed_data); } mntroot =nfs_get_root (s, data->fh);
@@ -2687,6 +2689,7 @@ static int nfs4_xdev_get_sb(struct file_system_type *fs_type, int flags,struct nfs_server *server;struct dentry *mntroot;int (*compare_super)(struct super_block *,void *) = nfs_compare_super; +struct nfs_parsed_mount_data parsed_data = { .fscache_uniq =NULL , };struct nfs_sb_mountdata sb_mntdata = { .mntflags = flags, };
@@ -2724,6 +2727,7 @@ static int nfs4_xdev_get_sb(struct file_system_type *fs_type, int flags,if (!s->s_root) {/* initial superblock/root creation */ nfs4_clone_super (s, data->sb); +nfs_fscache_get_super_cookie (s, &parsed_data); } mntroot =nfs4_get_root (s, data->fh);
@@ -2768,6 +2772,7 @@ static int nfs4_remote_referral_get_sb(struct file_system_type *fs_type,struct dentry *mntroot;struct nfs_fh mntfh;int (*compare_super)(struct super_block *,void *) = nfs_compare_super; +struct nfs_parsed_mount_data parsed_data = { .fscache_uniq =NULL , };struct nfs_sb_mountdata sb_mntdata = { .mntflags = flags, };
@@ -2805,6 +2810,7 @@ static int nfs4_remote_referral_get_sb(struct file_system_type *fs_type,if (!s->s_root) {/* initial superblock/root creation */ nfs4_fill_super (s); +nfs_fscache_get_super_cookie (s, &parsed_data); } mntroot =nfs4_get_root (s, &mntfh);
Quoted 13 lines · David Howells
Propagate the NFS 'fsc' mount option through NFS automounts of various types.
This is now required as commit:
commit c02d7adf8c5429727a98bad1d039bccad4c61c50 Author: Trond Myklebust <[email protected]> Date: Mon Jun 22 15:09:14 2009 -0400 NFSv4: Replace nfs4_path_walk() with VFS path lookup in a private namespacenow uses VFS-driven automounting to reach all submounts barring the root, thus preventing fscaching from being enabled on any submount other than the root.
Quoted 6 lines
This patch gets around that by propagating the
NFS_OPTION_FSCACHEflag across automounts.Note, however, that it doesn't propagate the uniquifier in the case of 'fsc=<xxx>' being passed to mount. This is probably wrong, and needs looking at.
Why not just use the mount path as the default uniquifier?
Quoted 13 lines
Signed-off-by David Howells <[email protected]>---
fs/nfs/client.c | 1 + fs/nfs/super.c | 6 ++++++ 2 files changed, 7 insertions(+), 0 deletions(-)fs/nfs/client.c +0
-0
fs/nfs/client.c static void nfs_server_copy_userdata(struct nfs_server *target, struct nfs_serve · line 967
Quoted 4 lines
fs/nfs/client.c +1
-0@@ -964,6 +964,7 @@ static void nfs_server_copy_userdata(struct nfs_server *target, struct nfs_serve target->acdirmin = source->acdirmin; target->acdirmax = source->acdirmax; target->caps = source->caps; + target->options = source->options;
BTW: Why does fscache require a private flag field?
fs/nfs/super.c static int nfs_xdev_get_sb(struct file_system_type *fs_type, int flags, · line 2217
Quoted 12 lines
fs/nfs/client.c +0
-0}/* fs/nfs/super.c +1
-0@@ -2214,6 +2214,7 @@ static int nfs_xdev_get_sb(struct file_system_type *fs_type, int flags,struct nfs_server *server;struct dentry *mntroot;int (*compare_super)(struct super_block *,void *) = nfs_compare_super; +struct nfs_parsed_mount_data parsed_data = { .fscache_uniq =NULL , };
Rather than wasting all this space on the stack, how about just changing the second argument of nfs_fscache_get_super_cookie()? You only use the pointer to data->fscache_uniq.
Quoted 3 lines
fs/nfs/super.c +0
-0struct nfs_sb_mountdata sb_mntdata = { .mntflags = flags, };
Signature · 6 lines
Trond Myklebust Linux NFS client maintainer NetApp [email protected] www.netapp.com
Quoted 1 line · Trond Myklebust
Why not just use the mount path as the default uniquifier?
Because:
(1) Which mount path? You're now constructing the mount tree in its own private namespace.
(2) Due to the new VFS-pathwalk-based mount, the uniquifier is applied to the root mount only, and if modified to inherit, would be applied to every automount subordinate to that, rather than just the actual subject of the mount.
(3) Due to the new VFS-pathwalk-based mount and the private namespace, the default uniquifier for all instances of any particular mount is always the same.
(4) If we work around some how the mounts now getting their mount path from the private namespace, there are still real namespaces, chroots and root-pivots to contend with.
Quoted 3 lines
fs/nfs/client.c static void nfs_server_copy_userdata(struct nfs_server *target, struct nfs_serve · line 967
Quoted 1 line
fs/nfs/client.c +1
-0+ target->options = source->options;BTW: Why does fscache require a private flag field?
I thought it best to avoid touching nfs_server::flags as that is at least partially visible to userspace.
Quoted 5 lines
Review
Quoted 1 line
(quoted diff) +1
-0+ struct nfs_parsed_mount_data parsed_data = { .fscache_uniq = NULL, };Rather than wasting all this space on the stack, how about just changing the second argument of
nfs_fscache_get_super_cookie()? You only use the pointer todata->fscache_uniq.
Yeah, that's probably wise.
Actually, I don't really want to get this data via mount at all (as there are automounts to locally configure for caching).
David
Quoted 8 lines · David Howells
Quoted 1 line · Trond Myklebust
Why not just use the mount path as the default uniquifier?
Because:
(1) Which mount path? You're now constructing the mount tree in its own private namespace.
The natural thing would be to use the mount path that is required to locate the super block's root dentry on the server (i.e. the mount path for sb->s_root).
In the private namespace (since that reflects the namespace of the server, and hence mnt->mnt_root == sb->s_root) this mount path will always be the same as the contents of mnt->mnt_devname.
(In the local namespace, however, sb->s_root != mnt->mnt_root, and so mnt->mnt_devname is fixed up to reflect that difference. See nfs_fix_devname()...)
Quoted 8 lines
(2) Due to the new VFS-pathwalk-based mount, the uniquifier is applied to the root mount only, and if modified to inherit, would be applied to every automount subordinate to that, rather than just the actual subject of the mount.
(3) Due to the new VFS-pathwalk-based mount and the private namespace, the default uniquifier for all instances of any particular mount is always the same.
...and is that sufficient? If it is, then fine, let's leave it...
Quoted 3 lines
(4) If we work around some how the mounts now getting their mount path from the private namespace, there are still real namespaces, chroots and root-pivots to contend with.
Real namespaces, chroots and root-pivots are all changes to the client only. They do not change the server's namespace, and hence the mount path is invariant.
Trond
Signature · 6 lines
Trond Myklebust Linux NFS client maintainer NetApp [email protected] www.netapp.com
Quoted 1 line · Trond Myklebust
Quoted 1 line
Quoted 1 line
Why not just use the mount path as the default uniquifier?
In fact, why change it from being blank? If I do this:
The natural thing would be to use the mount path that is required to locate the super block's root dentry on the server (i.e. the mount path for sb->s_root).
then it may as well be blank, since the FSID is a sufficient uniquifier, and that's also included in the key.
Where a uniquifier is needed is when someone sets up two mounts of the same thing, but with different network parameters as these aren't recorded in the key. Under such a circumstance, manual intervention is required in the form of a supplied uniquifier on one of the mounts.
That uniquifier should also only apply to the superblock actually being mounted, I think; but it's probably okay to apply it to all the intervening transient superblocks from root that nfs_follow_remote_path() causes to exist, and just accept that you might get a little bit of rubbish in the cache once we start caching directories.
Also, the uniquifier would be applied to all superblocks that get created by automounts under a superblock that has a uniquifier.
David