[PATCH] cifs: Allow mounts for paths for restricted intermediate paths

How does NFS handle this use case?

Post by s***@public.gmane.org
Allow cifs mounts for a prefixpath with intermediate paths without access,
so as to continue with the shared superblock model.
For the intermediate path entries that do not allow access, create "placeholder"
inodes and instantiate dentries with those inodes.
If and when such a path entry becomes accessible it is filled with correct
info.
Reference: Samba bugzilla 6950
---
fs/cifs/cifsfs.c | 17 +++++++++++++++--
fs/cifs/cifsproto.h | 3 +++
fs/cifs/inode.c | 16 ++++++++++++++++
3 files changed, 34 insertions(+), 2 deletions(-)
diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
index 849f613..bce185c 100644
--- a/fs/cifs/cifsfs.c
+++ b/fs/cifs/cifsfs.c
@@ -584,6 +584,9 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
char *full_path = NULL;
char *s, *p;
char sep;
+ struct inode *dir, *phinode;
+ struct dentry *child;
+ struct cifs_fattr phfattr;
full_path = cifs_build_path_to_root(vol, cifs_sb,
cifs_sb_master_tcon(cifs_sb));
@@ -592,13 +595,13 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
cifs_dbg(FYI, "Get root dentry for %s\n", full_path);
+ fill_phfattr(&phfattr, sb);
sep = CIFS_DIR_SEP(cifs_sb);
dentry = dget(sb->s_root);
p = s = full_path;
do {
- struct inode *dir = dentry->d_inode;
- struct dentry *child;
+ dir = dentry->d_inode;
if (!dir) {
dput(dentry);
@@ -626,6 +629,16 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
mutex_unlock(&dir->i_mutex);
dput(dentry);
dentry = child;
+ if (!IS_ERR(dentry)) {
+ if (*s) {
+ /* EACCESS on an intermediate dir */
+ if (!dentry->d_inode) {
+ phinode = cifs_iget(sb, &phfattr);
+ if (phinode)
+ d_instantiate(dentry, phinode);
+ }
+ }
+ }
} while (!IS_ERR(dentry));
kfree(full_path);
return dentry;
diff --git a/fs/cifs/cifsproto.h b/fs/cifs/cifsproto.h
index aa33976..9a84e5c 100644
--- a/fs/cifs/cifsproto.h
+++ b/fs/cifs/cifsproto.h
@@ -498,4 +498,7 @@ void cifs_writedata_release(struct kref *refcount);
int open_query_close_cifs_symlink(const unsigned char *path, char *pbuf,
unsigned int *pbytes_read, struct cifs_sb_info *cifs_sb,
unsigned int xid);
+
+extern void fill_phfattr(struct cifs_fattr *, struct super_block *);
+
#endif /* _CIFSPROTO_H */
diff --git a/fs/cifs/inode.c b/fs/cifs/inode.c
index 36f9ebb..4f5a09a 100644
--- a/fs/cifs/inode.c
+++ b/fs/cifs/inode.c
@@ -322,6 +322,22 @@ cifs_create_dfs_fattr(struct cifs_fattr *fattr, struct super_block *sb)
fattr->cf_flags |= CIFS_FATTR_DFS_REFERRAL;
}
+void
+fill_phfattr(struct cifs_fattr *cf, struct super_block *sb)
+{
+ struct cifs_sb_info *cifs_sb = CIFS_SB(sb);
+
+ memset(cf, 0, sizeof(*cf));
+ cf->cf_uniqueid = iunique(sb, ROOT_I);
+ cf->cf_nlink = 1;
+ cf->cf_atime = CURRENT_TIME;
+ cf->cf_ctime = CURRENT_TIME;
+ cf->cf_mtime = CURRENT_TIME;
+ cf->cf_mode = S_IFDIR | S_IXUGO | S_IRWXU;
+ cf->cf_uid = cifs_sb->mnt_uid;
+ cf->cf_gid = cifs_sb->mnt_gid;
+}
+
static int
cifs_get_file_info_unix(struct file *filp)
{
--
1.8.3.2

--
Thanks,

Steve

Shirish Pargaonkar

2013-11-19 16:48:29 UTC

Steve,

I have seen the nfs case mentioned during discussions but I am
not sure about the exact details.
Will dig into nfs code for details.

Regards,

Shirish

Post by Steve French
How does NFS handle this use case?

--
Thanks,
Steve

Pavel Shilovsky

2013-11-19 17:54:07 UTC

I thinks it's 8950.

Post by s***@public.gmane.org
---
fs/cifs/cifsfs.c | 17 +++++++++++++++--
fs/cifs/cifsproto.h | 3 +++
fs/cifs/inode.c | 16 ++++++++++++++++
3 files changed, 34 insertions(+), 2 deletions(-)
diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
index 849f613..bce185c 100644
--- a/fs/cifs/cifsfs.c
+++ b/fs/cifs/cifsfs.c
@@ -584,6 +584,9 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
char *full_path = NULL;
char *s, *p;
char sep;
+ struct inode *dir, *phinode;
+ struct dentry *child;
+ struct cifs_fattr phfattr;
full_path = cifs_build_path_to_root(vol, cifs_sb,
cifs_sb_master_tcon(cifs_sb));
@@ -592,13 +595,13 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
cifs_dbg(FYI, "Get root dentry for %s\n", full_path);
+ fill_phfattr(&phfattr, sb);
sep = CIFS_DIR_SEP(cifs_sb);
dentry = dget(sb->s_root);
p = s = full_path;
do {
- struct inode *dir = dentry->d_inode;
- struct dentry *child;
+ dir = dentry->d_inode;
if (!dir) {
dput(dentry);
@@ -626,6 +629,16 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
mutex_unlock(&dir->i_mutex);
dput(dentry);
dentry = child;
+ if (!IS_ERR(dentry)) {
+ if (*s) {
+ /* EACCESS on an intermediate dir */
+ if (!dentry->d_inode) {
+ phinode = cifs_iget(sb, &phfattr);
+ if (phinode)
+ d_instantiate(dentry, phinode);
+ }
+ }
+ }
} while (!IS_ERR(dentry));
kfree(full_path);
return dentry;
diff --git a/fs/cifs/cifsproto.h b/fs/cifs/cifsproto.h
index aa33976..9a84e5c 100644
--- a/fs/cifs/cifsproto.h
+++ b/fs/cifs/cifsproto.h
@@ -498,4 +498,7 @@ void cifs_writedata_release(struct kref *refcount);
int open_query_close_cifs_symlink(const unsigned char *path, char *pbuf,
unsigned int *pbytes_read, struct cifs_sb_info *cifs_sb,
unsigned int xid);
+
+extern void fill_phfattr(struct cifs_fattr *, struct super_block *);
+
#endif /* _CIFSPROTO_H */
diff --git a/fs/cifs/inode.c b/fs/cifs/inode.c
index 36f9ebb..4f5a09a 100644
--- a/fs/cifs/inode.c
+++ b/fs/cifs/inode.c
@@ -322,6 +322,22 @@ cifs_create_dfs_fattr(struct cifs_fattr *fattr, struct super_block *sb)
fattr->cf_flags |= CIFS_FATTR_DFS_REFERRAL;
}
+void
+fill_phfattr(struct cifs_fattr *cf, struct super_block *sb)
+{
+ struct cifs_sb_info *cifs_sb = CIFS_SB(sb);
+
+ memset(cf, 0, sizeof(*cf));
+ cf->cf_uniqueid = iunique(sb, ROOT_I);
+ cf->cf_nlink = 1;
+ cf->cf_atime = CURRENT_TIME;
+ cf->cf_ctime = CURRENT_TIME;
+ cf->cf_mtime = CURRENT_TIME;
+ cf->cf_mode = S_IFDIR | S_IXUGO | S_IRWXU;
+ cf->cf_uid = cifs_sb->mnt_uid;
+ cf->cf_gid = cifs_sb->mnt_gid;
+}
+
static int
cifs_get_file_info_unix(struct file *filp)
{
--
1.8.3.2
--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
More majordomo info at http://vger.kernel.org/majordomo-info.html

Looks ok at the first glance - will test further.

--
Best regards,
Pavel Shilovsky.

Shirish Pargaonkar

2013-11-19 18:00:24 UTC

Pavel, Thanks, yes, 8950.

Regards,

Shirish

Post by Pavel Shilovsky

I thinks it's 8950.

Looks ok at the first glance - will test further.
--
Best regards,
Pavel Shilovsky.

Pavel Shilovsky

2013-11-19 19:09:26 UTC

Post by Shirish Pargaonkar
Pavel, Thanks, yes, 8950.
Regards,
Shirish

Post by Pavel Shilovsky

I thinks it's 8950.

Looks ok at the first glance - will test further.
--
Best regards,
Pavel Shilovsky.

I tested it with the following setup:

//server/share/dir/dir1/dir2

The user has a full access to //server/share and
//server/share/dir/dir1/dir2 and not to //server/share/dir and
//server/share/dir1. The mount of //server/share/dir/dir1/dir2 failed
whenever the patch is applied or not (3.12 kernel is used). I checked
it and found out that lookup_one_len returned EACCESS pointer but your
patch works for the case when negative dentry is returned.

--
Best regards,
Pavel Shilovsky.

Jeff Layton

2013-11-19 20:59:23 UTC

On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...

How exactly do they end up being filled with the correct info once that
happens?

Post by s***@public.gmane.org
Reference: Samba bugzilla 6950
---
fs/cifs/cifsfs.c | 17 +++++++++++++++--
fs/cifs/cifsproto.h | 3 +++
fs/cifs/inode.c | 16 ++++++++++++++++
3 files changed, 34 insertions(+), 2 deletions(-)
diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
index 849f613..bce185c 100644
--- a/fs/cifs/cifsfs.c
+++ b/fs/cifs/cifsfs.c
@@ -584,6 +584,9 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
char *full_path = NULL;
char *s, *p;
char sep;
+ struct inode *dir, *phinode;
+ struct dentry *child;
+ struct cifs_fattr phfattr;
full_path = cifs_build_path_to_root(vol, cifs_sb,
cifs_sb_master_tcon(cifs_sb));
@@ -592,13 +595,13 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
cifs_dbg(FYI, "Get root dentry for %s\n", full_path);
+ fill_phfattr(&phfattr, sb);

Hmmm...so this function sets up a cifs_fattr struct, including the
cf_uniqueid...which won't be very unique when you reuse that struct
several times in the do...while loop below. Don't you need to call it
again after you use it once?

Post by s***@public.gmane.org
sep = CIFS_DIR_SEP(cifs_sb);
dentry = dget(sb->s_root);
p = s = full_path;
do {
- struct inode *dir = dentry->d_inode;
- struct dentry *child;
+ dir = dentry->d_inode;
if (!dir) {
dput(dentry);
@@ -626,6 +629,16 @@ cifs_get_root(struct smb_vol *vol, struct super_block *sb)
mutex_unlock(&dir->i_mutex);
dput(dentry);
dentry = child;
+ if (!IS_ERR(dentry)) {
+ if (*s) {
+ /* EACCESS on an intermediate dir */
+ if (!dentry->d_inode) {
+ phinode = cifs_iget(sb, &phfattr);
+ if (phinode)
+ d_instantiate(dentry, phinode);
+ }
+ }
+ }

This doesn't make any sense to me.

The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.

Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Post by s***@public.gmane.org
} while (!IS_ERR(dentry));
kfree(full_path);
return dentry;
diff --git a/fs/cifs/cifsproto.h b/fs/cifs/cifsproto.h
index aa33976..9a84e5c 100644
--- a/fs/cifs/cifsproto.h
+++ b/fs/cifs/cifsproto.h
@@ -498,4 +498,7 @@ void cifs_writedata_release(struct kref *refcount);
int open_query_close_cifs_symlink(const unsigned char *path, char *pbuf,
unsigned int *pbytes_read, struct cifs_sb_info *cifs_sb,
unsigned int xid);
+
+extern void fill_phfattr(struct cifs_fattr *, struct super_block *);
+
#endif /* _CIFSPROTO_H */
diff --git a/fs/cifs/inode.c b/fs/cifs/inode.c
index 36f9ebb..4f5a09a 100644
--- a/fs/cifs/inode.c
+++ b/fs/cifs/inode.c
@@ -322,6 +322,22 @@ cifs_create_dfs_fattr(struct cifs_fattr *fattr, struct super_block *sb)
fattr->cf_flags |= CIFS_FATTR_DFS_REFERRAL;
}
+void
+fill_phfattr(struct cifs_fattr *cf, struct super_block *sb)
+{
+ struct cifs_sb_info *cifs_sb = CIFS_SB(sb);
+
+ memset(cf, 0, sizeof(*cf));
+ cf->cf_uniqueid = iunique(sb, ROOT_I);

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...

Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.

One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

Post by s***@public.gmane.org
+ cf->cf_nlink = 1;
+ cf->cf_atime = CURRENT_TIME;
+ cf->cf_ctime = CURRENT_TIME;
+ cf->cf_mtime = CURRENT_TIME;
+ cf->cf_mode = S_IFDIR | S_IXUGO | S_IRWXU;
+ cf->cf_uid = cifs_sb->mnt_uid;
+ cf->cf_gid = cifs_sb->mnt_gid;
+}
+
static int
cifs_get_file_info_unix(struct file *filp)
{

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...

--
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+***@public.gmane.org>

Shirish Pargaonkar

2013-11-21 17:16:20 UTC

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

I was thinking when the access permissions for this directory change
on the server.

For example, if we had two mounts with shared superblocks
/mnt_m/a/b
/mnt_n
where a was inaccessible. If the permissions change on a such that
it becomes accessible, stat /mnt_n/a will fetch correct info.

That was silly. Will fix that.

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

If server returns ENOENT, lookup_one_len will return with a negative
dentry. This is the case I was handling. Need to access EACCESS.
Not sure why/when server returns ENOENT (object not found) even when
the (directory) entry exists.

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

Looking into this.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

Jeff Layton

2013-11-21 21:02:04 UTC

On Thu, 21 Nov 2013 11:16:20 -0600

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

I was thinking when the access permissions for this directory change
on the server.
For example, if we had two mounts with shared superblocks
/mnt_m/a/b
/mnt_n
where a was inaccessible. If the permissions change on a such that
it becomes accessible, stat /mnt_n/a will fetch correct info.

Right, but the dentry is already set at that point to use your fake
inode, and it has children so we can't just toss it out of the cache
and create a new one.

That was silly. Will fix that.

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

This seems backward to me. The ENOENT case is a case where we should
*fail*, not where we should fake up an inode. That's a pretty strong
indicator that the prefixpath is just bogus.

In the case of something like EACCES then we may want to create a
placeholder inode since that sort of implies that the directory is
present, but that we don't have permissions to fetch the info about it.

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

Looking into this.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...

Again, the above is really the place to start. You've been given a
mount request with a prefixpath that has components to which you don't
have access. What do you do now?

--
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+***@public.gmane.org>

Shirish Pargaonkar

2013-12-19 18:56:22 UTC

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I was thinking, we can have a fake inode per dentry for an
inaccessible path which has life only during the function
cifs_get_root, that too only just after lookup_one_len().
So use d_instantiate()/d_iput() pair for temporary creatation/delete
of these placeholder inodes!

So once a path is built and cifs_get_root is done, we should have
those intermediate path entries as negative but will have children
so will not be evicted from the dcache?

There is really no need for the inodes of the dentries for intermediate
inaccessible paths once a prefixpath mount completes successfully.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

Jeff Layton

2013-12-19 21:36:50 UTC

On Thu, 19 Dec 2013 12:56:22 -0600

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I was thinking, we can have a fake inode per dentry for an
inaccessible path which has life only during the function
cifs_get_root, that too only just after lookup_one_len().
So use d_instantiate()/d_iput() pair for temporary creatation/delete
of these placeholder inodes!
So once a path is built and cifs_get_root is done, we should have
those intermediate path entries as negative but will have children
so will not be evicted from the dcache?
There is really no need for the inodes of the dentries for intermediate
inaccessible paths once a prefixpath mount completes successfully.

Having a negative dentry with children is going to be very problematic.
That was why Pavel's patchset from a couple of years ago got shot down.

The dcache has all sorts of assumptions that that's not possible (for
good reasons). A negative dentry is a negative lookup -- it's a
shortcut that says "this dentry doesn't exist". How can something that
doesn't exist have children?

So, I think that trying to turn treat these things as negative dentries is
the wrong approach. The best approach is probably to do something
similar to what NFS does. Instantiate these dentries as disconnected,
and then splice them into the dentry tree during the mount process if
they suddenly become reachable from the new mount.

Not trivial to get that right, but that should work...

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+***@public.gmane.org>

Shirish Pargaonkar

2014-03-12 17:38:02 UTC

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

Trying to solve this using like this:

If during mounting a share, if the share path is accessbile but
if any of the intermediate paths to the share is inaccessible
i.e. returns error EACCES,
- get inode info for the share path using query path info
- look for a dentry and if not found, add - using d_obtain_alias()
- if server does not support unique ids/ inode numbers, add to the
begining of the list maintained in superblock, an element consisting of
full path, pointer to this dentry, and a pointer to the inode.

During lookup,
for a case where server does support unique_ids/inode numbers,
- obtain inode info using query path info
- splice (d_splice_alias()) the dentry corrosponding to this inode if any
if spliced, remove the element from the list maintained in the superblock.

for a case where server does not support unique id/inode number,
- search for an entry for inode based on the full path
if no such entry, obtain inode info using query path info
- splice (d_splice_alias()) the dentry corrosponding to this inode if any
if successful, remove the element from the list maintained in the superblock.

That is one way an element will come off of the list maintained in the
superblock. If the dentry does not get ever get spliced, whenever either
it is deleted with zero d_count (unmount e.g.) or it is released
(superblock is being deleted), cifs d_delete and cifs_d_release
respectively, whichever applies first, will take the element off the list
and free it.

If the superblock is intact and anonymous root dentry does not get deleted
during unmount (d_count is not zero i.e. has child dentries with referenece
to the parent) i.e. d_delete does not get called, the dentries would remain
till vfs code deletes them as needed (as their d_count starts approaching
zero).

Whenever we have the same mount again with a anonymous dentry, the
elements are added to the head of the list maintained in the superblock,
so stale elements will not be reachable.

Your comments are most welcome.

Regards,

Shirish

Jeff Layton

2014-03-12 19:04:36 UTC

On Wed, 12 Mar 2014 12:38:02 -0500

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

If during mounting a share, if the share path is accessbile but
if any of the intermediate paths to the share is inaccessible
i.e. returns error EACCES,
- get inode info for the share path using query path info
- look for a dentry and if not found, add - using d_obtain_alias()
- if server does not support unique ids/ inode numbers, add to the
begining of the list maintained in superblock, an element consisting of
full path, pointer to this dentry, and a pointer to the inode.
During lookup,
for a case where server does support unique_ids/inode numbers,
- obtain inode info using query path info
- splice (d_splice_alias()) the dentry corrosponding to this inode if any
if spliced, remove the element from the list maintained in the superblock.
for a case where server does not support unique id/inode number,
- search for an entry for inode based on the full path
if no such entry, obtain inode info using query path info
- splice (d_splice_alias()) the dentry corrosponding to this inode if any
if successful, remove the element from the list maintained in the superblock.
That is one way an element will come off of the list maintained in the
superblock. If the dentry does not get ever get spliced, whenever either
it is deleted with zero d_count (unmount e.g.) or it is released
(superblock is being deleted), cifs d_delete and cifs_d_release
respectively, whichever applies first, will take the element off the list
and free it.
If the superblock is intact and anonymous root dentry does not get deleted
during unmount (d_count is not zero i.e. has child dentries with referenece
to the parent) i.e. d_delete does not get called, the dentries would remain
till vfs code deletes them as needed (as their d_count starts approaching
zero).
Whenever we have the same mount again with a anonymous dentry, the
elements are added to the head of the list maintained in the superblock,
so stale elements will not be reachable.
Your comments are most welcome.
Regards,
Shirish

Really, the whole problem boils down to this:

Suppose we mount two cifs mounts:

//server/a on /mnt/a

...and...

//server/a/b on /mnt/b

Now suppose there is a directory under there:

/mnt/a/b/c

...so logically, this should be the same dentry:

/mnt/b/c

The problem is -- how do you know that that's the same dentry?

The reason that the shared superblock stuff works with NFS is that
there is a 1:1 correspondence between inodes and filehandles. So, even
if the file is renamed on the server, you end up with the same inode
and can tell that that dentry is the same as another.

That's more or less the case for CIFS too where you have a uniqueid,
but when you don't that whole model breaks down.

Hashing filenames sounds like a possibility too, but what happens if
someone does this on the directory above on the server?

$ mv c d

Since you can't guard against changes to lookup information on the
server that makes that method of uniquely id'ing an inode pretty
worthless.

I think that really leaves you with just two choices:

1) figure out some mechanism to hook things up at mount time.
Basically, instantiate "placeholder" inodes for directories above the
root of the mount. IOW, if someone mounts /mnt/b above and then you'd
need a placeholder inode for "a" that will be filled out when /mnt/a
mounted.

2) abandon the shared-superblock model altogether. That'll mean that in
the above mount setup that you don't have cache coherency between the
two mounts (since they'll be different superblocks), and you may need
to tweak how fscache works in order to ensure that collisions between
two different mounts isn't a problem.

At this point #2 doesn't seem altogether unreasonable. Most people
don't have pathological mount configurations like that so I'm not sure
it's worthwhile to jump through hoops to enable them.

--
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+***@public.gmane.org>

Shirish Pargaonkar

2014-03-13 15:35:59 UTC

Post by Jeff Layton
On Wed, 12 Mar 2014 12:38:02 -0500

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

//server/a on /mnt/a
...and...
//server/a/b on /mnt/b
/mnt/a/b/c
/mnt/b/c
The problem is -- how do you know that that's the same dentry?

So this will be a problem only if a dentry for c is not in dcache,
is that correct?

If so, the problem arises when two lookups for c for two different
paths go out on the wire because a dentry for c is not in the cache.
So in that case, in cifs code, can we serialize instantiating a dentry
i.e. check for a dentry in dcache just before instantiating and if it is,
discard the inode?
dentries would be unique with their d_parent and d_name combination.

Post by Jeff Layton
The reason that the shared superblock stuff works with NFS is that
there is a 1:1 correspondence between inodes and filehandles. So, even
if the file is renamed on the server, you end up with the same inode
and can tell that that dentry is the same as another.
That's more or less the case for CIFS too where you have a uniqueid,
but when you don't that whole model breaks down.
Hashing filenames sounds like a possibility too, but what happens if
someone does this on the directory above on the server?
$ mv c d

Assuming using hashing full path names to generate uniqueids, if above happens,
would not revalidation of the dentry (for c) fail on the client so we
will end up discarding
the inode and dentry and a creating a new dentry with correct d_name
and an inode with
hashed full pathname unique_id during lookup?

In above case, wonder how NFS updates d_name of a dentry with the same inode!

Post by Jeff Layton
Since you can't guard against changes to lookup information on the
server that makes that method of uniquely id'ing an inode pretty
worthless.
1) figure out some mechanism to hook things up at mount time.
Basically, instantiate "placeholder" inodes for directories above the
root of the mount. IOW, if someone mounts /mnt/b above and then you'd
need a placeholder inode for "a" that will be filled out when /mnt/a
mounted.
2) abandon the shared-superblock model altogether. That'll mean that in
the above mount setup that you don't have cache coherency between the
two mounts (since they'll be different superblocks), and you may need
to tweak how fscache works in order to ensure that collisions between
two different mounts isn't a problem.
At this point #2 doesn't seem altogether unreasonable. Most people
don't have pathological mount configurations like that so I'm not sure
it's worthwhile to jump through hoops to enable them.
--

Jeff Layton

2014-03-13 20:57:15 UTC

On Thu, 13 Mar 2014 10:35:59 -0500

Post by Jeff Layton
On Wed, 12 Mar 2014 12:38:02 -0500

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

//server/a on /mnt/a
...and...
//server/a/b on /mnt/b
/mnt/a/b/c
/mnt/b/c
The problem is -- how do you know that that's the same dentry?

So this will be a problem only if a dentry for c is not in dcache,
is that correct?

No, it's a problem either way. Let's say you mount /mnt/b first. You
don't have an inode for "a" because your creds don't allow you to look
it up. So, you have to either instantiate "b" as a disconnected dentry
(I may have the wrong terminology here), or fake up "a" somehow so it
can be set up as the parent of "b".

The problem comes in later when you go to mount /mnt/a, and then the
user walks into /mnt/a/b. At that point you don't want a new dentry for
"b", you want to get the "b" that got instantiated when you did the
first mount, but how do you know they're one in the same so you can
splice it correctly into the hierarchy?

Note that this example is relatively simple, but sprinkle in some
renames/recreates of directories on the server and dentries that are
farther down in the directory tree and things get nastier to handle.

Post by Shirish Pargaonkar
If so, the problem arises when two lookups for c for two different
paths go out on the wire because a dentry for c is not in the cache.
So in that case, in cifs code, can we serialize instantiating a dentry
i.e. check for a dentry in dcache just before instantiating and if it is,
discard the inode?
dentries would be unique with their d_parent and d_name combination.

I don't get it. I don't see how serializing things makes any difference
at all. Part of the problem is that the client has no control over
things that happen on other clients or on the server. Those kind of
changes are very difficult to guard against and we always have to be
prepared for them to occur.

Not necessarily. What would cause d_revalidate to fail in that case? In
any case cifs d_revalidate is pretty iffy...

Post by Shirish Pargaonkar
In above case, wonder how NFS updates d_name of a dentry with the same inode!

Sorry, I don't quite follow this question. With NFS, we basically look
at the parent directory and if it has changed since the dentry was
instantiated, we toss out the dentry and redo the lookup.

--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+***@public.gmane.org>

Shirish Pargaonkar

2014-03-15 20:31:31 UTC

Post by Jeff Layton
On Thu, 13 Mar 2014 10:35:59 -0500

Post by Jeff Layton
On Wed, 12 Mar 2014 12:38:02 -0500

Post by Jeff Layton
On Tue, 19 Nov 2013 10:33:33 -0600

I'm unclear on this last bit...
How exactly do they end up being filled with the correct info once that
happens?

This doesn't make any sense to me.
The do...while loop here is basically doing lookups and instantiating
dentries and inodes down from the root of the mount to the root of the
share. The lookup_one_len call is what does the single-level lookup in
that directory.
Why are you adding special handling when "dentry" is not an error?
Would it not make more sense to do so when lookup_one_len does return
an error?

Also, iunique is only guaranteed to return a number that hasn't yet
been used. So suppose we're going to instantiate one of these bogus
inodes. We give it the uniqueid of '3'. Then later, once the mount is
done we find another (legitimate) inode on the server and it also has a
uniqueid of '3'. Now you have a collision. Worse yet, if it's a
directory now it looks like we have a hardlinked directory...
Once you start using iunique, you can't really go back to using server
inode numbers since you can't guarantee that there won't be collisions.
One possibility to deal with that would be to tag these entries as
bogus and ensure that cifs_iget will never match them. Another
possibility might be just to never hash these inodes, but I'm not clear
on how kosher that is from a vfs standpoint.

I think the thing to do is to step back and describe (in English) how
you envision this working. Then once that's done, we can start
discussing what the code should look like...
--

//server/a on /mnt/a
...and...
//server/a/b on /mnt/b
/mnt/a/b/c
/mnt/b/c
The problem is -- how do you know that that's the same dentry?

So this will be a problem only if a dentry for c is not in dcache,
is that correct?

No, it's a problem either way. Let's say you mount /mnt/b first. You
don't have an inode for "a" because your creds don't allow you to look
it up. So, you have to either instantiate "b" as a disconnected dentry
(I may have the wrong terminology here), or fake up "a" somehow so it
can be set up as the parent of "b".
The problem comes in later when you go to mount /mnt/a, and then the
user walks into /mnt/a/b. At that point you don't want a new dentry for
"b", you want to get the "b" that got instantiated when you did the
first mount, but how do you know they're one in the same so you can
splice it correctly into the hierarchy?

This is what I had suggested (with little bit clearer verbiage here),
would happen in this case.
What I mean by superblock is cifs_sb_info structure.

- if the server does not support unique ids/ inode numbers, add to the
begining of the list maintained in superblock, an element consisting of
full path, pointer to this dentry, and a pointer to the inode.

for a case where server does not support unique id/inode number,
- search for an element for inode based on the full path in the superbock list
- if no such element, obtain an inode info using query path info
if such an element is found, use that inode
- now splice (d_splice_alias()) the dentry corrosponding to above inode
and if d_splice_alias() returned a dentry , then remove the element from
the list in cifs_sb_info.

Post by Jeff Layton
Note that this example is relatively simple, but sprinkle in some
renames/recreates of directories on the server and dentries that are
farther down in the directory tree and things get nastier to handle.

What I mean is, e.g. in cifs_lookup, take a lock and check for a
dentry in dcache
using d_lookup().
if not found, d_add().
if found, release the inode.
release the lock.

Not necessarily. What would cause d_revalidate to fail in that case? In
any case cifs d_revalidate is pretty iffy...

Post by Shirish Pargaonkar
In above case, wonder how NFS updates d_name of a dentry with the same inode!

Sorry, I don't quite follow this question. With NFS, we basically look
at the parent directory and if it has changed since the dentry was
instantiated, we toss out the dentry and redo the lookup.