NFSD: Pin to vfsmount instead of mntget for export cache

Without initialized, done in fs_pin at stack space may contains strange value.

Signed-off-by Kinglong Mee <[email protected]>

---
 include/linux/fs_pin.h | 1 +
 1 file changed, 1 insertion(+)

include/linux/fs_pin.h +1 -0

index 3886b3b..18fad53 100644
@@ -16,6 +16,7 @@ static inline void init_fs_pin(struct fs_pin *p, void (*kill)(struct fs_pin *))
 	INIT_HLIST_NODE(&p->s_list);
 	INIT_HLIST_NODE(&p->m_list);
 	p->kill = kill;
+	p->done = 0;
 }
 
 void pin_remove(struct fs_pin *);

Signature · 1 line

2.4.0

Quoted 2 lines · Kinglong Mee

Without initialized, done in fs_pin at stack space may contains strange value.

Looks like both init_fs_pin callers use some variation on kzalloc(), so I don't think there's any actual bug here.

Maybe there's some other reason for the belt-and-suspenders approach, that's Al's call, I think.

--b.

Quoted 17 lines

Signed-off-by Kinglong Mee <[email protected]>

---
 include/linux/fs_pin.h | 1 +
 1 file changed, 1 insertion(+)

include/linux/fs_pin.h +1 -0

index 3886b3b..18fad53 100644
@@ -16,6 +16,7 @@ static inline void init_fs_pin(struct fs_pin *p, void (*kill)(struct fs_pin *))
 	INIT_HLIST_NODE(&p->s_list);
 	INIT_HLIST_NODE(&p->m_list);
 	p->kill = kill;
+	p->done = 0;
 }
 

Quoted 4 lines

include/linux/fs_pin.h +0 -0

 void pin_remove(struct fs_pin *);

Signature · 1 line

2.4.0

If there are some mount points(not exported for nfs) under pseudo root, after client's operation of those entry under the root, anyone *can't* unmount those mount points until export cache expired.

# cat /etc/exports
/nfs/xfs        *(rw,insecure,no_subtree_check,no_root_squash)
/nfs/pnfs       *(rw,insecure,no_subtree_check,no_root_squash)
# ll /nfs/
total 0
drwxr-xr-x. 3 root root 84 Apr 21 22:27 pnfs
drwxr-xr-x. 3 root root 84 Apr 21 22:27 test
drwxr-xr-x. 2 root root  6 Apr 20 22:01 xfs
# mount /dev/sde /nfs/test
# df
Filesystem                      1K-blocks    Used Available Use% Mounted on
......
/dev/sdd                          1038336   32944   1005392   4% /nfs/pnfs
/dev/sdc                         10475520   32928  10442592   1% /nfs/xfs
/dev/sde                           999320    1284    929224   1% /nfs/test
# mount -t nfs 127.0.0.1:/nfs/ /mnt
# ll /mnt/*/
/mnt/pnfs/:
total 0
-rw-r--r--. 1 root root 0 Apr 21 22:23 attr
drwxr-xr-x. 2 root root 6 Apr 21 22:19 tmp

/mnt/xfs/:
total 0

# umount /nfs/test/
umount: /nfs/test/: target is busy
        (In some cases useful info about processes that
         use the device is found by lsof(8) or fuser(1).)

I don't think that's user expect, they want umount /nfs/test/.

It's caused by exports cache of nfsd holds the reference of the path (here is /nfs/test/), so, it can't be umounted.

The patch site using fs_pin instead of mntget for export cache, user at nfs server can unmount any mount points includes exported for nfs. Maybe, only umounted for unexported mount points is better?

Signed-off-by Kinglong Mee <[email protected]>

---
 fs/nfsd/export.c | 37 ++++++++++++++++++++++++++++++-------
 fs/nfsd/export.h | 10 +++++++++-
 2 files changed, 39 insertions(+), 8 deletions(-)

fs/nfsd/export.c +30 -7

index f79521a..80f82f5 100644
@@ -42,10 +42,12 @@ static void expkey_put(struct kref *ref)
 	struct svc_expkey *key = container_of(ref, struct svc_expkey, h.ref);
 
 	if (test_bit(CACHE_VALID, &key->h.flags) &&
-	    !test_bit(CACHE_NEGATIVE, &key->h.flags))
-		path_put(&key->ek_path);
+	    !test_bit(CACHE_NEGATIVE, &key->h.flags)) {
+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);
+	}
 	auth_domain_put(key->ek_client);
-	kfree(key);
+	kfree_rcu(key, rcu_head);
 }
 
 static void expkey_request(struct cache_detail *cd,

@@ -120,6 +122,7 @@ static int expkey_parse(struct cache_detail *cd, char *mesg, int mlen)
 		goto out;
 
 	key.ek_client = dom;	
+	key.cd = cd;
 	key.ek_fsidtype = fsidtype;
 	memcpy(key.ek_fsid, buf, len);
 

@@ -210,6 +213,13 @@ static inline void expkey_init(struct cache_head *cnew,
 	new->ek_fsidtype = item->ek_fsidtype;
 
 	memcpy(new->ek_fsid, item->ek_fsid, sizeof(new->ek_fsid));
+	new->cd = item->cd;
+}
+
+static void expkey_pin_kill(struct fs_pin *pin)
+{
+	struct svc_expkey *key = container_of(pin, struct svc_expkey, ek_pin);
+	cache_force_expire(key->cd, &key->h);
 }
 
 static inline void expkey_update(struct cache_head *cnew,

@@ -218,8 +228,10 @@ static inline void expkey_update(struct cache_head *cnew,
 	struct svc_expkey *new = container_of(cnew, struct svc_expkey, h);
 	struct svc_expkey *item = container_of(citem, struct svc_expkey, h);
 
+	init_fs_pin(&new->ek_pin, expkey_pin_kill);
 	new->ek_path = item->ek_path;
-	path_get(&item->ek_path);
+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);
 }
 
 static struct cache_head *expkey_alloc(void)

@@ -309,11 +321,13 @@ static void nfsd4_fslocs_free(struct nfsd4_fs_locations *fsloc)
 static void svc_export_put(struct kref *ref)
 {
 	struct svc_export *exp = container_of(ref, struct svc_export, h.ref);
-	path_put(&exp->ex_path);
+
+	dput(exp->ex_path.dentry);
+	pin_remove(&exp->ex_pin);
 	auth_domain_put(exp->ex_client);
 	nfsd4_fslocs_free(&exp->ex_fslocs);
 	kfree(exp->ex_uuid);
-	kfree(exp);
+	kfree_rcu(exp, rcu_head);
 }
 
 static void svc_export_request(struct cache_detail *cd,

@@ -694,15 +708,23 @@ static int svc_export_match(struct cache_head *a, struct cache_head *b)
 		path_equal(&orig->ex_path, &new->ex_path);
 }
 
+static void export_pin_kill(struct fs_pin *pin)
+{
+	struct svc_export *exp = container_of(pin, struct svc_export, ex_pin);
+	cache_force_expire(exp->cd, &exp->h);
+}
+
 static void svc_export_init(struct cache_head *cnew, struct cache_head *citem)
 {
 	struct svc_export *new = container_of(cnew, struct svc_export, h);
 	struct svc_export *item = container_of(citem, struct svc_export, h);
 
+	init_fs_pin(&new->ex_pin, export_pin_kill);
 	kref_get(&item->ex_client->ref);
 	new->ex_client = item->ex_client;
 	new->ex_path = item->ex_path;
-	path_get(&item->ex_path);
+	dget(item->ex_path.dentry);
+	pin_insert_group(&new->ex_pin, item->ex_path.mnt, NULL);
 	new->ex_fslocs.locations = NULL;
 	new->ex_fslocs.locations_count = 0;
 	new->ex_fslocs.migrated = 0;

@@ -811,6 +833,7 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
 
 	key.ek_client = clp;
 	key.ek_fsidtype = fsid_type;
+	key.cd = cd;
 	memcpy(key.ek_fsid, fsidv, key_len(fsid_type));
 
 	ek = svc_expkey_lookup(cd, &key);

fs/nfsd/export.h +9 -1

index 1f52bfc..1cf6ada 100644
@@ -4,6 +4,7 @@
 #ifndef NFSD_EXPORT_H
 #define NFSD_EXPORT_H
 
+#include <linux/fs_pin.h>
 #include <linux/sunrpc/cache.h>
 #include <uapi/linux/nfsd/export.h>
 

@@ -46,6 +47,8 @@ struct exp_flavor_info {
 
 struct svc_export {
 	struct cache_head	h;
+	struct cache_detail	*cd;
+
 	struct auth_domain *	ex_client;
 	int			ex_flags;
 	struct path		ex_path;

@@ -58,7 +61,9 @@ struct svc_export {
 	struct exp_flavor_info	ex_flavors[MAX_SECINFO_LIST];
 	enum pnfs_layouttype	ex_layout_type;
 	struct nfsd4_deviceid_map *ex_devid_map;
-	struct cache_detail	*cd;
+
+	struct fs_pin		ex_pin;
+	struct rcu_head		rcu_head;
 };
 
 /* an "export key" (expkey) maps a filehandlefragement to an

@@ -67,12 +72,15 @@ struct svc_export {
  */
 struct svc_expkey {
 	struct cache_head	h;
+	struct cache_detail	*cd;
 
 	struct auth_domain *	ek_client;
 	int			ek_fsidtype;
 	u32			ek_fsid[6];
 
 	struct path		ek_path;
+	struct fs_pin		ek_pin;
+	struct rcu_head		rcu_head;
 };
 
 #define EX_ISSYNC(exp)		(!((exp)->ex_flags & NFSEXP_ASYNC))

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

fs/nfsd/export.c static inline void expkey_update(struct cache_head *cnew, · line 234

Quoted 2 lines

fs/nfsd/export.c +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

fs/nfsd/export.c static void expkey_put(struct kref *ref) · line 47

Quoted 2 lines

fs/nfsd/export.c +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

Quoted 32 lines · NeilBrown

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

fs/nfsd/export.c static inline void expkey_update(struct cache_head *cnew, · line 234

Quoted 2 lines

fs/nfsd/export.c +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

fs/nfsd/export.c static void expkey_put(struct kref *ref) · line 47

Quoted 2 lines

fs/nfsd/export.c +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Quoted 3 lines

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

The undisputed bug here was negative cache entries preventing unmount. So most conservative might be just to purge negative entries.

Otherwise, the only guarantees I think we've really had is that we won't allow unmount if you hold any actual state on the filesystem (NLM locks, NFSv4 locks, opens, or delegations).

If a filesystem is exported but no clients hold state on it, then it's currently mostly chance whether the unmount succeeds or not. So we're probably free to change the behavior in this case. I'd be inclined to allow the unmount, but haven't thought this through carefully.

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

We might similarly be able to add some way for the kernel to distinguish explicit exports from crossmnt-found exports, but I'm not seeing the use case for that.

Quoted 33 lines · J. Bruce Fields

Quoted 32 lines · NeilBrown

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

fs/nfsd/export.c static inline void expkey_update(struct cache_head *cnew, · line 234

Quoted 2 lines

fs/nfsd/export.c +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

fs/nfsd/export.c static void expkey_put(struct kref *ref) · line 47

Quoted 2 lines

fs/nfsd/export.c +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Quoted 6 lines

Quoted 3 lines

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

The undisputed bug here was negative cache entries preventing unmount. So most conservative might be just to purge negative entries.

I'd like this,
if the cache is valid, user should not be allowed to umount the filesystem.

Quoted 4 lines

Otherwise, the only guarantees I think we've really had is that we won't allow unmount if you hold any actual state on the filesystem (NLM locks, NFSv4 locks, opens, or delegations).

Those resources hold the reference of vfsmnt.

Quoted 1 line

Quoted 4 lines

If a filesystem is exported but no clients hold state on it, then it's currently mostly chance whether the unmount succeeds or not. So we're probably free to change the behavior in this case. I'd be inclined to allow the unmount, but haven't thought this through carefully.

If client mount a nfsserver succeed without holds state, nfs server umounts the exported filesystem,
client also think the filesystem is valid, but it is umounted.

Quoted 5 lines

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

Quoted 57 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee <[email protected]> wrote:

Quoted 33 lines · J. Bruce Fields

Quoted 32 lines · NeilBrown

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

fs/nfsd/export.c static inline void expkey_update(struct cache_head *cnew, · line 234

Quoted 2 lines

fs/nfsd/export.c +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

fs/nfsd/export.c static void expkey_put(struct kref *ref) · line 47

Quoted 2 lines

fs/nfsd/export.c +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Quoted 6 lines

Quoted 3 lines

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

The undisputed bug here was negative cache entries preventing unmount. So most conservative might be just to purge negative entries.

I'd like this,
if the cache is valid, user should not be allowed to umount the filesystem.

Quoted 4 lines

Otherwise, the only guarantees I think we've really had is that we won't allow unmount if you hold any actual state on the filesystem (NLM locks, NFSv4 locks, opens, or delegations).

Those resources hold the reference of vfsmnt.

Quoted 5 lines

If a filesystem is exported but no clients hold state on it, then it's currently mostly chance whether the unmount succeeds or not. So we're probably free to change the behavior in this case. I'd be inclined to allow the unmount, but haven't thought this through carefully.

Quoted 3 lines

If client mount a nfsserver succeed without holds state, nfs server umounts the exported filesystem,
client also think the filesystem is valid, but it is umounted.

This is no different from "exportfs -au" being run on the server, thus unexporting the filesystem and making in unavailable to the client, even though the client has it mounted.

I think we need to give the server admin control of their filesystems, and assume they won't do something that they don't really want to do.

Quoted 4 lines

Quoted 4 lines

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

We already have a mechanism to forcibly drop any locks by writing some magic to /proc/fs/nfsd/unlock_{ip,filesystem}. I don't think we need any more.

Quoted 59 lines · NeilBrown

Quoted 57 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee…

Quoted 33 lines · J. Bruce Fields

Quoted 32 lines · NeilBrown

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

Review

Quoted 2 lines

(quoted diff) +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

Review

Quoted 2 lines

(quoted diff) +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Quoted 6 lines

Quoted 3 lines

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

The undisputed bug here was negative cache entries preventing unmount. So most conservative might be just to purge negative entries.

I'd like this,
if the cache is valid, user should not be allowed to umount the filesystem.

Quoted 4 lines

Otherwise, the only guarantees I think we've really had is that we won't allow unmount if you hold any actual state on the filesystem (NLM locks, NFSv4 locks, opens, or delegations).

Those resources hold the reference of vfsmnt.

Quoted 5 lines

If a filesystem is exported but no clients hold state on it, then it's currently mostly chance whether the unmount succeeds or not. So we're probably free to change the behavior in this case. I'd be inclined to allow the unmount, but haven't thought this through carefully.

Quoted 7 lines

Quoted 3 lines

If client mount a nfsserver succeed without holds state, nfs server umounts the exported filesystem,
client also think the filesystem is valid, but it is umounted.

This is no different from "exportfs -au" being run on the server, thus unexporting the filesystem and making in unavailable to the client, even though the client has it mounted.

No, I don't think so.
If user using "exportfs -au" to flush caches, I think he known what the influence of he does, but an umount of filesystem, maybe he doesn't known that contains flushing nfsd's exports cache.

For an using of nfsd exports, I'd like an error of an umount, because I don't realize the exports for nfsd.

I also think nfsd should allowing umount of unexported filesystem, because user has the right to umount it.

Quoted 4 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee…

Quoted 3 lines · J. Bruce Fields

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

We already have a mechanism to forcibly drop any locks by writing some magic to /proc/fs/nfsd/unlock_{ip,filesystem}. I don't think we need any more.

Quoted 7 lines · NeilBrown

Quoted 4 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee <[email protected]> wrote:

Quoted 3 lines · J. Bruce Fields

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

Quoted 2 lines

We already have a mechanism to forcibly drop any locks by writing some magic to /proc/fs/nfsd/unlock_{ip,filesystem}. I don't think we need any more.

Yeah, I remember thinking this sort of approach would have advantages, maybe I was wrong, I need to revisit it.

The unlock_{ip,filesystem} approach requires temporarily shutting down mountd, doesn't it?

Quoted 8 lines · On Fri, 15 May 2015 17:11:34 -0400 "J. Bruce Fields" <[email protected]>…

Quoted 7 lines · NeilBrown

Quoted 4 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee <[email protected]> wrote:

Quoted 3 lines · J. Bruce Fields

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

Quoted 8 lines

Quoted 2 lines

We already have a mechanism to forcibly drop any locks by writing some magic to /proc/fs/nfsd/unlock_{ip,filesystem}. I don't think we need any more.

Yeah, I remember thinking this sort of approach would have advantages, maybe I was wrong, I need to revisit it.

The unlock_{ip,filesystem} approach requires temporarily shutting down mountd, doesn't it?

Not necessarily.
It does require ensuring that new locks aren't suddenly taken though.

I imagine an early step in the migration process is to "ifconfig down" the virtual interface with the floating ID. Then you can safely "unlock" and unmount any filesystems are that only accessed via the IP.

But you are right that using the "unlock_*" interface and then unmounting is racy in a way that we are trying to make "unmount" not racy. So maybe an "allow_force_unmount" would have a place.

Quoted 27 lines · NeilBrown

Quoted 8 lines · On Fri, 15 May 2015 17:11:34 -0400 "J. Bruce Fields" <[email protected]>…

Quoted 7 lines · NeilBrown

Quoted 4 lines · On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee <[email protected]> wrote:

Quoted 3 lines · J. Bruce Fields

It could also be useful to have the ability to force an unmount even in the presence of locks. That's not a safe default, but an "allow_force_unmount" export option might be useful.

Quoted 8 lines

Quoted 2 lines

We already have a mechanism to forcibly drop any locks by writing some magic to /proc/fs/nfsd/unlock_{ip,filesystem}. I don't think we need any more.

Yeah, I remember thinking this sort of approach would have advantages, maybe I was wrong, I need to revisit it.

The unlock_{ip,filesystem} approach requires temporarily shutting down mountd, doesn't it?

Not necessarily.
It does require ensuring that new locks aren't suddenly taken though.

I imagine an early step in the migration process is to "ifconfig down" the virtual interface with the floating ID. Then you can safely "unlock" and unmount any filesystems are that only accessed via the IP.

Quoted 3 lines

But you are right that using the "unlock_*" interface and then unmounting is racy in a way that we are trying to make "unmount" not racy. So maybe an "allow_force_unmount" would have a place.

No, unlock_{ip,filesystem} are used for nlmlock, doesn't support nfsv4 resources.

Some other interfaces under /sys/kernel/debug/nfsd/forget_* support nfsv4 resources, without for an filesystem. It seems will be removed sometime.

Quoted 33 lines · J. Bruce Fields

Quoted 32 lines · NeilBrown

Thanks for this patch. It looks good!

My only comment on the code is that I would really like to see a "path_get_pin()" and "path_put_unpin()" rather than open coding:

fs/nfsd/export.c static inline void expkey_update(struct cache_head *cnew, · line 234

Quoted 2 lines

fs/nfsd/export.c +2 -0

+	dget(item->ek_path.dentry);
+	pin_insert_group(&new->ek_pin, item->ek_path.mnt, NULL);

and

fs/nfsd/export.c static void expkey_put(struct kref *ref) · line 47

Quoted 2 lines

fs/nfsd/export.c +2 -0

+		dput(key->ek_path.dentry);
+		pin_remove(&key->ek_pin);

But the question you raise is an important one: Exactly which filesystems should be allowed to be unmounted?
This is a change in behaviour - is it one that people uniformly would want?

The kernel doesn't currently know which file systems were explicitly listed in /etc/exports, and which were found by following a 'crossmnt'. It could guess and allow the unmounting of anything below a 'crossmnt', but I wouldn't be comfortable with that - it is error prone.

mountd does know what is in /etc/exports, and could tell the kernel. For the expkey cache, we could always use path_get_pin.
For the export cache (where flags are available) we could use path_get or path_get_pin depending on some new flag.

I'm not really sure it is worth it. I would rather the filesystems could always be unmounted. But doing that could possibly break someone's work flow. Maybe.

Quoted 6 lines

Quoted 3 lines

Or maybe I'm seeing problems where there aren't any.

Anyone else have an opinion?

The undisputed bug here was negative cache entries preventing unmount. So most conservative might be just to purge negative entries.

I'd like this,
if the cache is valid, user should not be allowed to umount the filesystem.

Quoted 4 lines

Otherwise, the only guarantees I think we've really had is that we won't allow unmount if you hold any actual state on the filesystem (NLM locks, NFSv4 locks, opens, or delegations).

Those resources hold the reference of vfsmnt.

Quoted 1 line

Quoted 4 lines

If a filesystem is exported but no clients hold state on it, then it's currently mostly chance whether the unmount succeeds or not. So we're probably free to change the behavior in this case. I'd be inclined to allow the unmount, but haven't thought this through carefully.

If client mount a nfsserver succeed without holds state, nfs server umounts the exported filesystem,
client also think the filesystem is valid, but it is umounted.