LinuxLists.cc - [PATCH 0/6 v9] NFSD: Pin to vfsmount for nfsd exports cache

2015-08-18 07:17:27

Subject: [PATCH 0/6 v9] NFSD: Pin to vfsmount for nfsd exports cache

2015-08-18 07:19:09

Subject: [PATCH 1/6 v9] fs-pin: allow pin_remove() to be called other than from->kill()

From: NeilBrown <[email protected]>

fs-pin currently assumes when either the vfsmount or the fs_pin wants
to unpin, pin_kill() will be called.
This requires that the ->kill() function can wait for any transient
references to the fs_pin to be released. If the structure containing
the fs_pin doesn't already have the ability to wait for references,
this can be a burden.

As the fs_pin already has infrastructure for waiting, that can be
leveraged to remove the burden.

In this alternate scenario, only the vfsmount calls pin_kill() when it
wants to unpin. The owner of the fs_pin() instead calls pin_remove().

The ->kill() function removes any long-term references, and then calls
pin_kill() (recursively).
When the last reference on (the structure containing) the fs_pin is
dropped, pin_remove() will be called and the (recursive) pin_kill()
call will complete.

For this to be safe, the final "put" must *not* free the structure if
pin_kill() has already been called, as that could leave ->kill()
accessing freed data.

So we provide a return value for pin_remove() which reports the old
->done value.

When final put calls pin_remove() it checks that value.
If it was 0, then pin_kill() has not called ->kill and will not,
so final put can free the data structure.
If it was -1, then pin_kill() has called ->kill, and ->kill will
free the data structure - final put must not touch it.

This makes the 'wait' infrastructure of fs_pin available to any
pinning client which wants to use it.

Signed-Off-By: NeilBrown <[email protected]>
---
fs/fs_pin.c | 18 +++++++++++++++++-
include/linux/fs_pin.h | 2 +-
2 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/fs/fs_pin.c b/fs/fs_pin.c
index 611b540..b7954a9 100644
--- a/fs/fs_pin.c
+++ b/fs/fs_pin.c
@@ -6,16 +6,32 @@

static DEFINE_SPINLOCK(pin_lock);

-void pin_remove(struct fs_pin *pin)
+/**
+ * pin_remove - disconnect an fs_pin from the pinned structure.
+ * @pin: The struct fs_pin which is pinning something.
+ *
+ * Detach a 'pin' which was added by pin_insert(). A return value
+ * of -1 implies that pin_kill() has already been called and that the
+ * ->kill() function now owns the data structure containing @pin.
+ * The function which called pin_remove() must not touch the data structure
+ * again (unless it is the ->kill() function itself).
+ * A return value of 0 implies an uneventful disconnect: pin_kill() has not called,
+ * and will not call, the ->kill() function on this @pin.
+ * Any other return value is a usage error - e.g. repeated call to pin_remove().
+ */
+int pin_remove(struct fs_pin *pin)
{
+ int ret;
spin_lock(&pin_lock);
hlist_del_init(&pin->m_list);
hlist_del_init(&pin->s_list);
spin_unlock(&pin_lock);
spin_lock_irq(&pin->wait.lock);
+ ret = pin->done;
pin->done = 1;
wake_up_locked(&pin->wait);
spin_unlock_irq(&pin->wait.lock);
+ return ret;
}

void pin_insert_group(struct fs_pin *pin, struct vfsmount *m, struct hlist_head *p)
diff --git a/include/linux/fs_pin.h b/include/linux/fs_pin.h
index 3886b3b..2fe9d3b 100644
--- a/include/linux/fs_pin.h
+++ b/include/linux/fs_pin.h
@@ -18,7 +18,7 @@ static inline void init_fs_pin(struct fs_pin *p, void (*kill)(struct fs_pin *))
p->kill = kill;
}

-void pin_remove(struct fs_pin *);
+int pin_remove(struct fs_pin *);
void pin_insert_group(struct fs_pin *, struct vfsmount *, struct hlist_head *);
void pin_insert(struct fs_pin *, struct vfsmount *);
void pin_kill(struct fs_pin *);
--
2.4.3

2015-08-18 07:19:45

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 2/6] fs_pin: Export functions for specific filesystem

Exports functions for others who want pin to vfsmount,
eg, nfsd's export cache.

These are needed for any module to participate in pinning.

mnt_pin_kill() and group_pin_kill() are just helper-functions for
unmount etc (and are in fs/internal.h) so don't need to be exported.

v9, same as v4
add exporting of pin_kill.

Signed-off-by: Kinglong Mee <[email protected]>
---
fs/fs_pin.c | 4 ++++
1 file changed, 4 insertions(+)

diff --git a/fs/fs_pin.c b/fs/fs_pin.c
index b7954a9..b1a9654 100644
--- a/fs/fs_pin.c
+++ b/fs/fs_pin.c
@@ -33,6 +33,7 @@ int pin_remove(struct fs_pin *pin)
spin_unlock_irq(&pin->wait.lock);
return ret;
}
+EXPORT_SYMBOL(pin_remove);

void pin_insert_group(struct fs_pin *pin, struct vfsmount *m, struct hlist_head *p)
{
@@ -42,11 +43,13 @@ void pin_insert_group(struct fs_pin *pin, struct vfsmount *m, struct hlist_head
hlist_add_head(&pin->m_list, &real_mount(m)->mnt_pins);
spin_unlock(&pin_lock);
}
+EXPORT_SYMBOL(pin_insert_group);

void pin_insert(struct fs_pin *pin, struct vfsmount *m)
{
pin_insert_group(pin, m, &m->mnt_sb->s_pins);
}
+EXPORT_SYMBOL(pin_insert);

void pin_kill(struct fs_pin *p)
{
@@ -88,6 +91,7 @@ void pin_kill(struct fs_pin *p)
}
rcu_read_unlock();
}
+EXPORT_SYMBOL(pin_kill);

void mnt_pin_kill(struct mount *m)
{
--
2.4.3

2015-08-18 07:20:44

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 3/6 v9] path: New helpers path_get_pin/path_put_unpin for path pin

Two helpers for filesystem pining to vfsmnt, not mntget.

v9, Update base on NeilBrown's new patch

Signed-off-by: Kinglong Mee <[email protected]>
---
fs/namei.c | 26 ++++++++++++++++++++++++++
include/linux/path.h | 4 ++++
2 files changed, 30 insertions(+)

diff --git a/fs/namei.c b/fs/namei.c
index 1c2105e..4f95e16 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -492,6 +492,32 @@ void path_put(const struct path *path)
}
EXPORT_SYMBOL(path_put);

+/**
+ * path_get_pin - get a reference to a path's dentry
+ * and pin to path's vfsmnt
+ * @path: path to get the reference to
+ * @p: the fs_pin pin to vfsmnt
+ */
+void path_get_pin(struct path *path, struct fs_pin *p)
+{
+ dget(path->dentry);
+ pin_insert_group(p, path->mnt, NULL);
+}
+EXPORT_SYMBOL(path_get_pin);
+
+/**
+ * path_put_unpin - put a reference to a path's dentry
+ * and remove pin to path's vfsmnt
+ * @path: path to put the reference to
+ * @p: the fs_pin removed from vfsmnt
+ */
+int path_put_unpin(struct path *path, struct fs_pin *p)
+{
+ dput(path->dentry);
+ return pin_remove(p);
+}
+EXPORT_SYMBOL(path_put_unpin);
+
#define EMBEDDED_LEVELS 2
struct nameidata {
struct path path;
diff --git a/include/linux/path.h b/include/linux/path.h
index d137218..40d376a 100644
--- a/include/linux/path.h
+++ b/include/linux/path.h
@@ -3,6 +3,7 @@

struct dentry;
struct vfsmount;
+struct fs_pin;

struct path {
struct vfsmount *mnt;
@@ -12,6 +13,9 @@ struct path {
extern void path_get(const struct path *);
extern void path_put(const struct path *);

+extern void path_get_pin(struct path *, struct fs_pin *);
+extern int path_put_unpin(struct path *, struct fs_pin *);
+
static inline int path_equal(const struct path *path1, const struct path *path2)
{
return path1->mnt == path2->mnt && path1->dentry == path2->dentry;
--
2.4.3

2015-08-18 07:22:13

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 4/6 v9] fs: New helper legitimize_mntget() for getting a legitimize mnt

New helper legitimize_mntget for getting a mnt without setting
MNT_SYNC_UMOUNT | MNT_UMOUNT | MNT_DOOMED, otherwise return NULL.

v9, Update using read_seqbegin_or_lock instead read_seqlock_excl

Signed-off-by: Kinglong Mee <[email protected]>
---
fs/namespace.c | 30 ++++++++++++++++++++++++++++++
include/linux/mount.h | 1 +
2 files changed, 31 insertions(+)

diff --git a/fs/namespace.c b/fs/namespace.c
index 2b8aa15..bf4a9f5 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -1153,6 +1153,36 @@ struct vfsmount *mntget(struct vfsmount *mnt)
}
EXPORT_SYMBOL(mntget);

+struct vfsmount *legitimize_mntget(struct vfsmount *vfsmnt)
+{
+ struct mount *mnt;
+ unsigned seq = 0;
+
+ if (vfsmnt == NULL)
+ return NULL;
+retry:
+ read_seqbegin_or_lock(&mount_lock, &seq);
+
+ if (vfsmnt->mnt_flags & (MNT_SYNC_UMOUNT | MNT_UMOUNT | MNT_DOOMED))
+ vfsmnt = NULL;
+ else if (need_seqretry(&mount_lock, seq))
+ goto retry;
+ else {
+ mnt = real_mount(vfsmnt);
+ mnt_add_count(mnt, 1);
+ if (need_seqretry(&mount_lock, seq) ||
+ vfsmnt->mnt_flags & (MNT_SYNC_UMOUNT | MNT_UMOUNT |
+ MNT_DOOMED)) {
+ mnt_add_count(mnt, -1);
+ goto retry;
+ }
+ }
+
+ done_seqretry(&mount_lock, seq);
+ return vfsmnt;
+}
+EXPORT_SYMBOL(legitimize_mntget);
+
struct vfsmount *mnt_clone_internal(struct path *path)
{
struct mount *p;
diff --git a/include/linux/mount.h b/include/linux/mount.h
index f822c3c..8ae9dc0 100644
--- a/include/linux/mount.h
+++ b/include/linux/mount.h
@@ -79,6 +79,7 @@ extern void mnt_drop_write(struct vfsmount *mnt);
extern void mnt_drop_write_file(struct file *file);
extern void mntput(struct vfsmount *mnt);
extern struct vfsmount *mntget(struct vfsmount *mnt);
+extern struct vfsmount *legitimize_mntget(struct vfsmount *vfsmnt);
extern struct vfsmount *mnt_clone_internal(struct path *path);
extern int __mnt_is_readonly(struct vfsmount *mnt);

--
2.4.3

2015-08-18 07:23:02

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 5/6 v9] sunrpc: New helper cache_delete_entry for deleting cache_head directly

A new helper cache_delete_entry() for delete cache_head from
cache_detail directly.

It will be used by pin_kill, so make sure the cache_detail is valid
before deleting is needed.

Because pin_kill is not many times, so the influence of performance
is accepted.

v9, delete duplicate checking of cache_detail

Signed-off-by: Kinglong Mee <[email protected]>
---
include/linux/sunrpc/cache.h | 1 +
net/sunrpc/cache.c | 20 ++++++++++++++++++++
2 files changed, 21 insertions(+)

diff --git a/include/linux/sunrpc/cache.h b/include/linux/sunrpc/cache.h
index 03d3b4c..2824db5 100644
--- a/include/linux/sunrpc/cache.h
+++ b/include/linux/sunrpc/cache.h
@@ -210,6 +210,7 @@ extern int cache_check(struct cache_detail *detail,
struct cache_head *h, struct cache_req *rqstp);
extern void cache_flush(void);
extern void cache_purge(struct cache_detail *detail);
+extern void cache_delete_entry(struct cache_detail *cd, struct cache_head *h);
#define NEVER (0x7FFFFFFF)
extern void __init cache_initialize(void);
extern int cache_register_net(struct cache_detail *cd, struct net *net);
diff --git a/net/sunrpc/cache.c b/net/sunrpc/cache.c
index 4a2340a..430b5fa 100644
--- a/net/sunrpc/cache.c
+++ b/net/sunrpc/cache.c
@@ -454,6 +454,26 @@ static int cache_clean(void)
return rv;
}

+void cache_delete_entry(struct cache_detail *detail, struct cache_head *h)
+{
+ if (!detail || !h)
+ return;
+
+ write_lock(&detail->hash_lock);
+ if (hlist_unhashed(&h->cache_list)) {
+ write_unlock(&detail->hash_lock);
+ return ;
+ }
+
+ hlist_del_init(&h->cache_list);
+ detail->entries--;
+ set_bit(CACHE_CLEANED, &h->flags);
+ write_unlock(&detail->hash_lock);
+
+ cache_put(h, detail);
+}
+EXPORT_SYMBOL_GPL(cache_delete_entry);
+
/*
* We want to regularly clean the cache, so we need to schedule some work ...
*/
--
2.4.3

2015-08-18 07:24:14

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 6/6 v9] nfsd: Allows user un-mounting filesystem where nfsd exports base on

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.

/nfs/xfs *(rw,insecure,no_subtree_check,no_root_squash)
/nfs/pnfs *(rw,insecure,no_subtree_check,no_root_squash)
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
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
/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/: target is busy
(In some cases useful info about processes that
use the device is found by lsof(8) or fuser(1).)

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

I don't think that's user expect, they want umount /nfs/test/.
Bruce think user can also umount /nfs/pnfs/ and /nfs/xfs.

Also, using kzalloc for all memory allocating without kmalloc.
Thanks for Al Viro's commets for the logic of fs_pin.

v3,
1. using path_get_pin/path_put_unpin for path pin
2. using kzalloc for memory allocating

v5, v4,
1. add a completion for pin_kill waiting the reference is decreased to zero.
2. add a work_struct for pin_kill decreases the reference indirectly.
3. free svc_export/svc_expkey in pin_kill, not svc_export_put/svc_expkey_put.
4. svc_export_put/svc_expkey_put go though pin_kill logic.

v6,
1. Pin vfsmnt to mount point at first, when reference increace (==2),
grab a reference to vfsmnt by mntget. When decreace (==1),
drop the reference to vfsmnt, left pin.
2. Delete cache_head directly from cache_detail.

v7, implement self reference increase and decrease for nfsd exports/expkey

v8, new method as,

1. There are only one outlet from each cache, exp_find_key() for expkey,
exp_get_by_name() for export.
2. Any fsid to export or filehandle to export will call the function.
3. exp_get()/exp_put() increase/decrease the reference of export.

Call legitimize_mntget() in the only outlet function exp_find_key()/
exp_get_by_name(), if fail return STALE, otherwise, any valid
expkey/export from the cache is validated (Have get the reference of vfsmnt).

Add mntget() in exp_get() and mntput() in exp_put(), because the export
passed to exp_get/exp_put are returned from exp_find_key/exp_get_by_name.

For expkey cache,
1. At first, a fsid is passed to exp_find_key, and lookup a cache
in svc_expkey_lookup, if success, ekey->ek_path is pined to mount.
2. Then call legitimize_mntget getting a reference of vfsmnt
before return from exp_find_key.
3. Any calling exp_find_key with valid cache must put the vfsmnt.

for export cache,
1. At first, a path (returned from exp_find_key) with validate vfsmnt
is passed to exp_get_by_name, if success, exp->ex_path is pined to mount.
2. Then call legitimize_mntget getting a reference of vfsmnt
before return from exp_get_by_name.
3. Any calling exp_get_by_name with valid cache must put the vfsmnt
by exp_put();
4. Any using the exp returned from exp_get_by_name must call exp_get(),
will increase the reference of vfsmnt.

So that,
a. After getting the reference in 2, any umount of filesystem will get -EBUSY.
b. After put all reference after 4, or before get the reference in 2,
any umount of filesystem will call pin_kill, and delete the cache directly,
also unpin the vfsmount.
c. Between 1 and 2, have get the reference of exp/key cache, with invalidate vfsmnt.
Umount of filesystem only wait exp_find_key/exp_get_by_name put the reference
of cache when legitimize_mntget fail.

v9, thanks for NeilBrown's comments and update
1. Fix two string formats of path name.
2. Based on Neil's patch of
"allow pin_remove() to be called other than from ->kill()"

Besides the logical of waiting the last reference, others are same as v8.

Signed-off-by: Kinglong Mee <[email protected]>
---
fs/nfsd/export.c | 96 +++++++++++++++++++++++++++++++++++++++++++++-----------
fs/nfsd/export.h | 14 +++++++--
2 files changed, 90 insertions(+), 20 deletions(-)

diff --git a/fs/nfsd/export.c b/fs/nfsd/export.c
index b4d84b5..1d70324 100644
--- a/fs/nfsd/export.c
+++ b/fs/nfsd/export.c
@@ -37,15 +37,24 @@
#define EXPKEY_HASHMAX (1 << EXPKEY_HASHBITS)
#define EXPKEY_HASHMASK (EXPKEY_HASHMAX -1)

+static void expkey_destroy(struct svc_expkey *key)
+{
+ auth_domain_put(key->ek_client);
+ kfree_rcu(key, rcu_head);
+}
+
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);
- auth_domain_put(key->ek_client);
- kfree(key);
+ !test_bit(CACHE_NEGATIVE, &key->h.flags)) {
+ rcu_read_lock();
+ if (path_put_unpin(&key->ek_path, &key->ek_pin))
+ return ;
+ }
+
+ expkey_destroy(key);
}

static void expkey_request(struct cache_detail *cd,
@@ -119,6 +128,7 @@ static int expkey_parse(struct cache_detail *cd, char *mesg, int mlen)
if (key.h.expiry_time == 0)
goto out;

+ key.cd = cd;
key.ek_client = dom;
key.ek_fsidtype = fsidtype;
memcpy(key.ek_fsid, buf, len);
@@ -181,7 +191,11 @@ static int expkey_show(struct seq_file *m,
if (test_bit(CACHE_VALID, &h->flags) &&
!test_bit(CACHE_NEGATIVE, &h->flags)) {
seq_printf(m, " ");
- seq_path(m, &ek->ek_path, "\\ \t\n");
+ if (legitimize_mntget(ek->ek_path.mnt)) {
+ seq_path(m, &ek->ek_path, "\\ \t\n");
+ mntput(ek->ek_path.mnt);
+ } else
+ seq_printf(m, "Dir-unmounting");
}
seq_printf(m, "\n");
return 0;
@@ -210,6 +224,16 @@ 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_delete_entry(key->cd, &key->h);
+ /* Must call pin_kill to wait the last reference be put */
+ pin_kill(&key->ek_pin);
+ expkey_destroy(key);
}

static inline void expkey_update(struct cache_head *cnew,
@@ -218,13 +242,14 @@ 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);
+ path_get_pin(&new->ek_path, &new->ek_pin);
}

static struct cache_head *expkey_alloc(void)
{
- struct svc_expkey *i = kmalloc(sizeof(*i), GFP_KERNEL);
+ struct svc_expkey *i = kzalloc(sizeof(*i), GFP_KERNEL);
if (i)
return &i->h;
else
@@ -306,14 +331,21 @@ static void nfsd4_fslocs_free(struct nfsd4_fs_locations *fsloc)
fsloc->locations = NULL;
}

-static void svc_export_put(struct kref *ref)
+static void svc_export_destroy(struct svc_export *exp)
{
- struct svc_export *exp = container_of(ref, struct svc_export, h.ref);
- path_put(&exp->ex_path);
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_put(struct kref *ref)
+{
+ struct svc_export *exp = container_of(ref, struct svc_export, h.ref);
+
+ rcu_read_lock();
+ if (!path_put_unpin(&exp->ex_path, &exp->ex_pin))
+ svc_export_destroy(exp);
}

static void svc_export_request(struct cache_detail *cd,
@@ -636,7 +668,7 @@ static int svc_export_parse(struct cache_detail *cd, char *mesg, int mlen)
if (expp == NULL)
err = -ENOMEM;
else
- exp_put(expp);
+ cache_put(&expp->h, expp->cd);
out4:
nfsd4_fslocs_free(&exp.ex_fslocs);
kfree(exp.ex_uuid);
@@ -664,7 +696,12 @@ static int svc_export_show(struct seq_file *m,
return 0;
}
exp = container_of(h, struct svc_export, h);
- seq_path(m, &exp->ex_path, " \t\n\\");
+ if (legitimize_mntget(exp->ex_path.mnt)) {
+ seq_path(m, &exp->ex_path, " \t\n\\");
+ mntput(exp->ex_path.mnt);
+ } else
+ seq_printf(m, "Dir-unmounting");
+
seq_putc(m, '\t');
seq_escape(m, exp->ex_client->name, " \t\n\\");
seq_putc(m, '(');
@@ -694,15 +731,25 @@ 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_delete_entry(exp->cd, &exp->h);
+ /* Must call pin_kill to wait the last reference be put */
+ pin_kill(&exp->ex_pin);
+ svc_export_destroy(exp);
+}
+
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);
+ path_get_pin(&new->ex_path, &new->ex_pin);
new->ex_fslocs.locations = NULL;
new->ex_fslocs.locations_count = 0;
new->ex_fslocs.migrated = 0;
@@ -740,7 +787,7 @@ static void export_update(struct cache_head *cnew, struct cache_head *citem)

static struct cache_head *svc_export_alloc(void)
{
- struct svc_export *i = kmalloc(sizeof(*i), GFP_KERNEL);
+ struct svc_export *i = kzalloc(sizeof(*i), GFP_KERNEL);
if (i)
return &i->h;
else
@@ -809,6 +856,7 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
if (!clp)
return ERR_PTR(-ENOENT);

+ key.cd = cd;
key.ek_client = clp;
key.ek_fsidtype = fsid_type;
memcpy(key.ek_fsid, fsidv, key_len(fsid_type));
@@ -819,6 +867,12 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
err = cache_check(cd, &ek->h, reqp);
if (err)
return ERR_PTR(err);
+
+ if (!legitimize_mntget(ek->ek_path.mnt)) {
+ cache_put(&ek->h, ek->cd);
+ return ERR_PTR(-ENOENT);
+ }
+
return ek;
}

@@ -842,6 +896,8 @@ exp_get_by_name(struct cache_detail *cd, struct auth_domain *clp,
err = cache_check(cd, &exp->h, reqp);
if (err)
return ERR_PTR(err);
+
+ mntget(exp->ex_path.mnt);
return exp;
}

@@ -858,6 +914,7 @@ exp_parent(struct cache_detail *cd, struct auth_domain *clp, struct path *path)
struct dentry *parent = dget_parent(path->dentry);
dput(path->dentry);
path->dentry = parent;
+ exp_put(exp);
exp = exp_get_by_name(cd, clp, path, NULL);
}
dput(path->dentry);
@@ -928,7 +985,10 @@ static struct svc_export *exp_find(struct cache_detail *cd,
return ERR_CAST(ek);

exp = exp_get_by_name(cd, clp, &ek->ek_path, reqp);
- cache_put(&ek->h, nn->svc_expkey_cache);
+
+ /* Put the mnt get in exp_find_key() */
+ mntput(ek->ek_path.mnt);
+ cache_put(&ek->h, ek->cd);

if (IS_ERR(exp))
return ERR_CAST(exp);
@@ -1195,10 +1255,10 @@ static int e_show(struct seq_file *m, void *p)
return 0;
}

- exp_get(exp);
+ cache_get(&exp->h);
if (cache_check(cd, &exp->h, NULL))
return 0;
- exp_put(exp);
+ cache_put(&exp->h, exp->cd);
return svc_export_show(m, cd, cp);
}

diff --git a/fs/nfsd/export.h b/fs/nfsd/export.h
index 1f52bfc..d186a0a 100644
--- a/fs/nfsd/export.h
+++ b/fs/nfsd/export.h
@@ -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>

@@ -46,9 +47,10 @@ 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;
kuid_t ex_anon_uid;
kgid_t ex_anon_gid;
int ex_fsid;
@@ -58,7 +60,10 @@ 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 path ex_path;
+ 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))
@@ -100,12 +108,14 @@ __be32 nfserrno(int errno);

static inline void exp_put(struct svc_export *exp)
{
+ mntput(exp->ex_path.mnt);
cache_put(&exp->h, exp->cd);
}

static inline struct svc_export *exp_get(struct svc_export *exp)
{
cache_get(&exp->h);
+ mntget(exp->ex_path.mnt);
return exp;
}
struct svc_export * rqst_exp_find(struct svc_rqst *, int, u32 *);
--
2.4.3

2015-08-18 07:25:05

by Kinglong Mee

[permalink] [raw]

Subject: [PATCH 2/6 v9] fs_pin: Export functions for specific filesystem

2015-08-19 03:55:07

by NeilBrown

[permalink] [raw]

Subject: Re: [PATCH 6/6 v9] nfsd: Allows user un-mounting filesystem where nfsd exports base on

On Tue, 18 Aug 2015 15:23:59 +0800 Kinglong Mee <[email protected]>
wrote:

> 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);
> - auth_domain_put(key->ek_client);
> - kfree(key);
> + !test_bit(CACHE_NEGATIVE, &key->h.flags)) {
> + rcu_read_lock();
> + if (path_put_unpin(&key->ek_path, &key->ek_pin))
> + return ;
> + }

That rcu_read_lock() is unbalanced and not needed.

Comment applies below in svc_export code too.

> +static void expkey_pin_kill(struct fs_pin *pin)
> +{
> + struct svc_expkey *key = container_of(pin, struct svc_expkey, ek_pin);
> + cache_delete_entry(key->cd, &key->h);
> + /* Must call pin_kill to wait the last reference be put */
> + pin_kill(&key->ek_pin);
> + expkey_destroy(key);
> }

This, on the other hand, needs rcu_read_lock(). pin_kill() expected
rcu to be readlocked, and will drop the lock.

Same comment for svc_export code.

Did you test with lockdep enabled? That should have caught these
issues.

Otherwise it looks good.

Reviewed-by: NeilBrown <[email protected]>

Thanks,
NeilBrown

2015-08-19 04:51:03

by Al Viro

[permalink] [raw]

Subject: Re: [PATCH 6/6 v9] nfsd: Allows user un-mounting filesystem where nfsd exports base on

On Tue, Aug 18, 2015 at 03:23:59PM +0800, Kinglong Mee wrote:
> @@ -181,7 +191,11 @@ static int expkey_show(struct seq_file *m,
> if (test_bit(CACHE_VALID, &h->flags) &&
> !test_bit(CACHE_NEGATIVE, &h->flags)) {
> seq_printf(m, " ");
> - seq_path(m, &ek->ek_path, "\\ \t\n");
> + if (legitimize_mntget(ek->ek_path.mnt)) {
> + seq_path(m, &ek->ek_path, "\\ \t\n");
> + mntput(ek->ek_path.mnt);
> + } else
> + seq_printf(m, "Dir-unmounting");

IDGI... What locking environment do you have here? Note that use
of mnt_add_count(mnt, -1) in MNT_SYNC_UMOUNT case in __legitimize_mnt()
is OK only because we
a) are under rcu_read_lock() and
b) have synchronize_rcu() in namespace_unlock().

You don't seem to be under rcu_read_lock() here, so what's to stop that
from racing with the final mntput()? IOW, suppose that on the first
pass through legitimize_mntget() you read mount_lock, bump refcount,
recheck the lock and notice that it has been touched. You proceed to
decrement refcount. Fine, except that what would've been the final
mntput() has just noticed that refcount hadn't reached 0 and buggered
off. And in the meanwhile, MNT_UMOUNT had been set. Now you decrement
the refcount to zero, notice that MNT_UMOUNT and go away. Have a nice
leak...

The only reason why we are able to get away with mnt_add_count(mnt, -1) in
that specific case in legitimize_mnt() is that MNT_SYNC_UMOUNT must have
been set after we'd got rcu_read_lock() (otherwise we would've either hit
a mismatch on mount_lock before incrementing refcount or wouldn't have
run into that vfsmount at all) and thus we _know_ that the process that
has set MNT_SYNC_UMOUNT couldn't have passed through synchronize_rcu()
in namespace_unlock(), so it couldn't reach mntput_no_expire() until our
caller does rcu_read_unlock() and we are free to decrement the refcount and
leave - we won't be dropping the last reference here.

Without MNT_SYNC_UMOUNT the callers of __legitimize_mnt() must use mntput()
to drop the mistakenly acquired reference. Exactly because they can't
rely on that syncronize_rcu() delaying the final mntput().

> @@ -664,7 +696,12 @@ static int svc_export_show(struct seq_file *m,
> return 0;
> }
> exp = container_of(h, struct svc_export, h);
> - seq_path(m, &exp->ex_path, " \t\n\\");
> + if (legitimize_mntget(exp->ex_path.mnt)) {
> + seq_path(m, &exp->ex_path, " \t\n\\");
> + mntput(exp->ex_path.mnt);
> + } else
> + seq_printf(m, "Dir-unmounting");

Ditto. And grabbing/dropping references here seems to be an overkill...

> @@ -819,6 +867,12 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
> err = cache_check(cd, &ek->h, reqp);
> if (err)
> return ERR_PTR(err);
> +
> + if (!legitimize_mntget(ek->ek_path.mnt)) {
> + cache_put(&ek->h, ek->cd);
> + return ERR_PTR(-ENOENT);
> + }

Ditto.

> @@ -842,6 +896,8 @@ exp_get_by_name(struct cache_detail *cd, struct auth_domain *clp,
> err = cache_check(cd, &exp->h, reqp);
> if (err)
> return ERR_PTR(err);
> +
> + mntget(exp->ex_path.mnt);

What's to make that mntget() legitimate?

> static inline struct svc_export *exp_get(struct svc_export *exp)
> {
> cache_get(&exp->h);
> + mntget(exp->ex_path.mnt);

Ditto.

2015-08-19 04:57:00

by NeilBrown

[permalink] [raw]

Subject: Re: [PATCH 4/6 v9] fs: New helper legitimize_mntget() for getting a legitimize mnt

On Tue, 18 Aug 2015 15:21:59 +0800 Kinglong Mee <[email protected]>
wrote:

> New helper legitimize_mntget for getting a mnt without setting
> MNT_SYNC_UMOUNT | MNT_UMOUNT | MNT_DOOMED, otherwise return NULL.
>
> v9, Update using read_seqbegin_or_lock instead read_seqlock_excl
>
> Signed-off-by: Kinglong Mee <[email protected]>
> ---
> fs/namespace.c | 30 ++++++++++++++++++++++++++++++
> include/linux/mount.h | 1 +
> 2 files changed, 31 insertions(+)
>
> diff --git a/fs/namespace.c b/fs/namespace.c
> index 2b8aa15..bf4a9f5 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -1153,6 +1153,36 @@ struct vfsmount *mntget(struct vfsmount *mnt)
> }
> EXPORT_SYMBOL(mntget);
>
> +struct vfsmount *legitimize_mntget(struct vfsmount *vfsmnt)
> +{
> + struct mount *mnt;
> + unsigned seq = 0;
> +
> + if (vfsmnt == NULL)
> + return NULL;
> +retry:
> + read_seqbegin_or_lock(&mount_lock, &seq);
> +
> + if (vfsmnt->mnt_flags & (MNT_SYNC_UMOUNT | MNT_UMOUNT | MNT_DOOMED))
> + vfsmnt = NULL;
> + else if (need_seqretry(&mount_lock, seq))
> + goto retry;
> + else {
> + mnt = real_mount(vfsmnt);
> + mnt_add_count(mnt, 1);
> + if (need_seqretry(&mount_lock, seq) ||
> + vfsmnt->mnt_flags & (MNT_SYNC_UMOUNT | MNT_UMOUNT |
> + MNT_DOOMED)) {
> + mnt_add_count(mnt, -1);
> + goto retry;
> + }
> + }
> +
> + done_seqretry(&mount_lock, seq);
> + return vfsmnt;
> +}
> +EXPORT_SYMBOL(legitimize_mntget);

I'm struggling with this. I know I wrote the above, but I'm no longer
sure it is correct .. at all.

All three of those flags are only set while mount_lock is held for
writing, so if the second need_seqretry() returns zero, then we can
be certain that none of the flags have changed yet, so there is no point
testing them again.

Ahh... looking back at __legitimize_mnt I think I see the important
point now - or at least I now see that point which I saw before more
clearly.

After the mnt_add_count, if need_seqretry() returns zero, we are done.
If not, we need to clean up and try again.
To do this we need to check MNT_SYNC_UMOUNT.
If that is clear it is safe and best to call mntput() on the mnt.
If it is set, then we just do the mnt_add_count(mnt, -1) and give up
completely.

Also, the usage for read_seqbegin_or_lock is all wrong. I'm sure I
copied something but I got it badly wrong.
When you retry, you need to set 'seq' to 1, otherwise it never locks.
I thought it would be more "clever" than at.

So maybe:

retry:
read_seqbegin_or_lock(&mount_lock, &seq);

if (vfsmnt->mnt_flags & (....))
vfs_mnt = NULL;
else {
mnt = real_mount(vfsmnt);
mnt_add_count(mnt, 1);
if (need_seqretry(&mount_lock, seq)) {
/* lost the race, need to try again */
if (vfsmnt->mnt_flags & MNT_SYNC_UMOUNT) {
/* no point trying... */
mnt_add_count(mnt, -1):
vfsmnt = NULL;
} else {
mntput(vfsmnt);
seq = 1;
goto retry;
}
}
}
done_seqretry(&mount_lock, seq);
return vfsmnt;

NeilBrown

2015-08-19 05:02:49

by Al Viro

[permalink] [raw]

Subject: Re: [PATCH 4/6 v9] fs: New helper legitimize_mntget() for getting a legitimize mnt

On Wed, Aug 19, 2015 at 02:56:50PM +1000, NeilBrown wrote:
> If not, we need to clean up and try again.
> To do this we need to check MNT_SYNC_UMOUNT.
> If that is clear it is safe and best to call mntput() on the mnt.
> If it is set, then we just do the mnt_add_count(mnt, -1) and give up
> completely.

It's more subtle, actually. See my reply upthread for details, but
basically we rely on rcu_read_lock() held since before that MNT_SYNC_UMOUNT
had been set, making synchronize_rcu() in namespace_unlock() a guaranteed
delay of the final mntput until after we have returned.

2015-08-27 23:02:53

by Kinglong Mee

[permalink] [raw]

Subject: Re: [PATCH 4/6 v9] fs: New helper legitimize_mntget() for getting a legitimize mnt

On 8/19/2015 13:02, Al Viro wrote:
> On Wed, Aug 19, 2015 at 02:56:50PM +1000, NeilBrown wrote:
>> If not, we need to clean up and try again.
>> To do this we need to check MNT_SYNC_UMOUNT.
>> If that is clear it is safe and best to call mntput() on the mnt.
>> If it is set, then we just do the mnt_add_count(mnt, -1) and give up
>> completely.
>
> It's more subtle, actually. See my reply upthread for details, but
> basically we rely on rcu_read_lock() held since before that MNT_SYNC_UMOUNT
> had been set, making synchronize_rcu() in namespace_unlock() a guaranteed
> delay of the final mntput until after we have returned.
>

Thanks for the comments.
I will update this patch according to Neil's comments and __legitimize_mnt.

thanks,
Kinglong Mee

2015-08-27 23:15:53

by Kinglong Mee

[permalink] [raw]

Subject: Re: [PATCH 6/6 v9] nfsd: Allows user un-mounting filesystem where nfsd exports base on

On 8/19/2015 12:50, Al Viro wrote:
> On Tue, Aug 18, 2015 at 03:23:59PM +0800, Kinglong Mee wrote:
>> @@ -181,7 +191,11 @@ static int expkey_show(struct seq_file *m,
>> if (test_bit(CACHE_VALID, &h->flags) &&
>> !test_bit(CACHE_NEGATIVE, &h->flags)) {
>> seq_printf(m, " ");
>> - seq_path(m, &ek->ek_path, "\\ \t\n");
>> + if (legitimize_mntget(ek->ek_path.mnt)) {
>> + seq_path(m, &ek->ek_path, "\\ \t\n");
>> + mntput(ek->ek_path.mnt);
>> + } else
>> + seq_printf(m, "Dir-unmounting");
>
> IDGI... What locking environment do you have here? Note that use
> of mnt_add_count(mnt, -1) in MNT_SYNC_UMOUNT case in __legitimize_mnt()
> is OK only because we
> a) are under rcu_read_lock() and
> b) have synchronize_rcu() in namespace_unlock().

There are not any locking exist is this patch site.
Thanks for your comments about the mnt_add_count() following.

I want add rcu_read_lock in legitimize_mntget as,

struct vfsmount *legitimize_mntget(struct vfsmount *vfsmnt)
{
rcu_read_lock();
......
rcu_read_unlock();
return vfsmnt;
}

Is it OK?

>
> You don't seem to be under rcu_read_lock() here, so what's to stop that
> from racing with the final mntput()? IOW, suppose that on the first
> pass through legitimize_mntget() you read mount_lock, bump refcount,
> recheck the lock and notice that it has been touched. You proceed to
> decrement refcount. Fine, except that what would've been the final
> mntput() has just noticed that refcount hadn't reached 0 and buggered
> off. And in the meanwhile, MNT_UMOUNT had been set. Now you decrement
> the refcount to zero, notice that MNT_UMOUNT and go away. Have a nice
> leak...
>
> The only reason why we are able to get away with mnt_add_count(mnt, -1) in
> that specific case in legitimize_mnt() is that MNT_SYNC_UMOUNT must have
> been set after we'd got rcu_read_lock() (otherwise we would've either hit
> a mismatch on mount_lock before incrementing refcount or wouldn't have
> run into that vfsmount at all) and thus we _know_ that the process that
> has set MNT_SYNC_UMOUNT couldn't have passed through synchronize_rcu()
> in namespace_unlock(), so it couldn't reach mntput_no_expire() until our
> caller does rcu_read_unlock() and we are free to decrement the refcount and
> leave - we won't be dropping the last reference here.
>
> Without MNT_SYNC_UMOUNT the callers of __legitimize_mnt() must use mntput()
> to drop the mistakenly acquired reference. Exactly because they can't
> rely on that syncronize_rcu() delaying the final mntput().
>
>> @@ -664,7 +696,12 @@ static int svc_export_show(struct seq_file *m,
>> return 0;
>> }
>> exp = container_of(h, struct svc_export, h);
>> - seq_path(m, &exp->ex_path, " \t\n\\");
>> + if (legitimize_mntget(exp->ex_path.mnt)) {
>> + seq_path(m, &exp->ex_path, " \t\n\\");
>> + mntput(exp->ex_path.mnt);
>> + } else
>> + seq_printf(m, "Dir-unmounting");
>
> Ditto. And grabbing/dropping references here seems to be an overkill...

Do you mean that calling seq_path without legitimize_mntget here ?
But the mnt is without reference to vfsmnt, only a fs_pin many times.

>
>> @@ -819,6 +867,12 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
>> err = cache_check(cd, &ek->h, reqp);
>> if (err)
>> return ERR_PTR(err);
>> +
>> + if (!legitimize_mntget(ek->ek_path.mnt)) {
>> + cache_put(&ek->h, ek->cd);
>> + return ERR_PTR(-ENOENT);
>> + }
>
> Ditto.
>
>> @@ -842,6 +896,8 @@ exp_get_by_name(struct cache_detail *cd, struct auth_domain *clp,
>> err = cache_check(cd, &exp->h, reqp);
>> if (err)
>> return ERR_PTR(err);
>> +
>> + mntget(exp->ex_path.mnt);
>
> What's to make that mntget() legitimate?

The mnt has the reference to vfsmnt here, so just using mntget is safe.

>
>> static inline struct svc_export *exp_get(struct svc_export *exp)
>> {
>> cache_get(&exp->h);
>> + mntget(exp->ex_path.mnt);
>
> Ditto.
>

The mnt has the reference, Same as above.

thanks,
Kinglong Mee