Skip to contentSkip to conversation

kjournald keeps reference to namespace

PATCH Herbert Poetzl 5 participants 10 messages Feb 18, 2006 – Feb 24, 2006

lore
Share & tools
lore
#0 lore raw

Hi Folks!

when creating a private namespace (CLONE_NS) and then mounting an ext3 filesystem, a new kernel thread (kjournald) is created, which keeps a reference to the namespace, which after the the process exits, remains and blocks access to the block device, as it is still bd_claim-ed.

this leaves a private namespace behind and a block device which cannot be opened exclusively. unmount is not an option, as the namespace is not longer reachable.

this behaviour seems to be there since ever, well since namespaces and kjournald exists :)

the following 'cruel' hack 'solves' this issue

best,
Herbert

jbd/journal.c renamed from jbd/journal.c.orig +8 -0
@@ -33,6 +33,7 @@
 #include <linux/mm.h>
 #include <linux/suspend.h>
 #include <linux/pagemap.h>
+#include <linux/namespace.h>
 #include <asm/uaccess.h>
 #include <asm/page.h>
 #include <linux/proc_fs.h>
@@ -116,6 +117,13 @@ static int kjournald(void *arg)
 	struct timer_list timer;
 
 	daemonize("kjournald");
+	{
+		struct namespace *ns = current->namespace;
+
+		current->namespace = NULL;
+		put_namespace(ns);
+	}
+
 
 	/* Set up an interval timer which can be used to trigger a
            commit wakeup after the commit interval expires */
#1 lore raw
Quoted 10 lines · Herbert Poetzl

Hi Folks!

when creating a private namespace (CLONE_NS) and then mounting an ext3 filesystem, a new kernel thread (kjournald) is created, which keeps a reference to the namespace, which after the the process exits, remains and blocks access to the block device, as it is still bd_claim-ed.

There are numerous ways in which user processes can parent kernel threads.

bix:/usr/src/linux-2.6.16-rc4> grep -rl kernel_thread drivers net fs | wc

     64      64    1657
Quoted 26 lines

this leaves a private namespace behind and a block device which cannot be opened exclusively. unmount is not an option, as the namespace is not longer reachable.

this behaviour seems to be there since ever, well since namespaces and kjournald exists :)

the following 'cruel' hack 'solves' this issue

best,
Herbert

jbd/journal.c renamed from jbd/journal.c.orig +1 -0
@@ -33,6 +33,7 @@
 #include <linux/mm.h>
 #include <linux/suspend.h>
 #include <linux/pagemap.h>
+#include <linux/namespace.h>
 #include <asm/uaccess.h>
 #include <asm/page.h>
 #include <linux/proc_fs.h>
@@ -116,6 +117,13 @@ static int kjournald(void *arg)
 	struct timer_list timer;

jbd/journal.c static int kjournald(void *arg) · line 127

Quoted 10 lines
jbd/journal.c renamed from jbd/journal.c.orig +7 -0
 
 	daemonize("kjournald");
+	{
+		struct namespace *ns = current->namespace;
+
+		current->namespace = NULL;
+		put_namespace(ns);
+	}
+
 

I think it'd be better to convert ext3 to use the kthread API which appears to accidentally not have this problem, because such threads are parented by keventd, which were parented by init.

That being said, perhaps we should do a put_namespace() in kernel_thread(), too.

I'm kinda surprised that your patch didn't oops over a NULL ->namespace when the kernel internally mounted the root filesystem.

#2 lore raw
Quoted 44 lines · Andrew Morton
Quoted 10 lines · Herbert Poetzl

Hi Folks!

when creating a private namespace (CLONE_NS) and then mounting an ext3 filesystem, a new kernel thread (kjournald) is created, which keeps a reference to the namespace, which after the the process exits, remains and blocks access to the block device, as it is still bd_claim-ed.

There are numerous ways in which user processes can parent kernel threads.

bix:/usr/src/linux-2.6.16-rc4> grep -rl kernel_thread drivers net fs | wc

     64      64    1657
Quoted 27 lines

this leaves a private namespace behind and a block device which cannot be opened exclusively. unmount is not an option, as the namespace is not longer reachable.

this behaviour seems to be there since ever, well since namespaces and kjournald exists :)

the following 'cruel' hack 'solves' this issue

best,
Herbert

jbd/journal.c renamed from jbd/journal.c.orig +1 -0
@@ -33,6 +33,7 @@
 #include <linux/mm.h>
 #include <linux/suspend.h>
 #include <linux/pagemap.h>
+#include <linux/namespace.h>
 #include <asm/uaccess.h>
 #include <asm/page.h>
 #include <linux/proc_fs.h>
@@ -116,6 +117,13 @@ static int kjournald(void *arg)
 	struct timer_list timer;
 
Quoted 13 lines

jbd/journal.c static int kjournald(void *arg) · line 127

Quoted 9 lines
jbd/journal.c renamed from jbd/journal.c.orig +7 -0
 	daemonize("kjournald");
+	{
+		struct namespace *ns = current->namespace;
+
+		current->namespace = NULL;
+		put_namespace(ns);
+	}
+
 

I think it'd be better to convert ext3 to use the kthread API which appears to accidentally not have this problem, because such threads are parented by keventd, which were parented by init.

sounds like a plan!

Quoted 2 lines

That being said, perhaps we should do a put_namespace() in kernel_thread(), too.

hmm, keep the reference but put_namespace()?

Quoted 2 lines

I'm kinda surprised that your patch didn't oops over a NULL ->namespace when the kernel internally mounted the root filesystem.

nope, booted just fine, but I was worried too, the interesting detail is that the kjournald thread vanishes with this 'hack' when the namespace is disposed, which doesn't happen without it ...

anyway, will look into it, of course, input is welcome ...

best,
Herbert

#6 lore raw
In reply to Herbert Poetzl
Quoted 15 lines · Herbert Poetzl

Hi Folks!

when creating a private namespace (CLONE_NS) and then mounting an ext3 filesystem, a new kernel thread (kjournald) is created, which keeps a reference to the namespace, which after the the process exits, remains and blocks access to the block device, as it is still bd_claim-ed.

this leaves a private namespace behind and a block device which cannot be opened exclusively. unmount is not an option, as the namespace is not longer reachable.

Quoted 4 lines

this behaviour seems to be there since ever, well since namespaces and kjournald exists :)

the following 'cruel' hack 'solves' this issue

In daemonize() a new thread gets cleaned up and 'merged' with init_task. The current fs_struct is handled there, but not the current namespace. The following patch adds the namespace part.

---

kernel/exit.c +2 -0
diff -NurpP --minimal linux-2.6.16-rc4/kernel/exit.c linux-2.6.16-rc4-ns/kernel/exit.c
@@ -360,6 +360,8 @@ void daemonize(const char *name, ...)
 	fs = init_task.fs;
 	current->fs = fs;
 	atomic_inc(&fs->count);
+	exit_namespace(current);
+	current->namespace = init_task.namespace;
  	exit_files(current);
 	current->files = init_task.files;
 	atomic_inc(&current->files->count);
#7 lore raw
Quoted 6 lines

In daemonize() a new thread gets cleaned up and 'merged' with init_task. The current fs_struct is handled there, but not the current namespace. The following patch adds the namespace part.

---

Oops, forgot the increment the namespace usage count...
---

kernel/exit.c +3 -0
diff -NurpP --minimal linux-2.6.16-rc4/kernel/exit.c linux-2.6.16-rc4-ns/kernel/exit.c
@@ -360,6 +360,9 @@ void daemonize(const char *name, ...)
 	fs = init_task.fs;
 	current->fs = fs;
 	atomic_inc(&fs->count);
+	exit_namespace(current);
+	current->namespace = init_task.namespace;
+	atomic_inc(&current->namespace->count);
  	exit_files(current);
 	current->files = init_task.files;
 	atomic_inc(&current->files->count);
#8 lore raw
Quoted 8 lines · Björn Steinbrink
Quoted 6 lines

In daemonize() a new thread gets cleaned up and 'merged' with init_task. The current fs_struct is handled there, but not the current namespace. The following patch adds the namespace part.

---

Oops, forgot the increment the namespace usage count...

Ok, this time with the get_namespace wrapper, thanks to Eric Biederman for pointing that out to me.
---

kernel/exit.c +3 -0
diff -NurpP --minimal linux-2.6.16-rc4/kernel/exit.c linux-2.6.16-rc4-ns/kernel/exit.c
@@ -360,6 +360,9 @@ void daemonize(const char *name, ...)
 	fs = init_task.fs;
 	current->fs = fs;
 	atomic_inc(&fs->count);
+	exit_namespace(current);
+	current->namespace = init_task.namespace;
+	get_namespace(current->namespace);
  	exit_files(current);
 	current->files = init_task.files;
 	atomic_inc(&current->files->count);
#9 lore raw
Quoted 5 lines · Björn Steinbrink
Quoted 4 lines · Björn Steinbrink
Quoted 4 lines

In daemonize() a new thread gets cleaned up and 'merged' with init_task. The current fs_struct is handled there, but not the current namespace. The following patch adds the namespace part.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

---

Oops, forgot the increment the namespace usage count...

Ok, this time with the get_namespace wrapper, thanks to Eric Biederman for pointing that out to me.

Note we can't ever count on using our parents namespace because we already have called exit_fs(), which is the only way to the namespace from a process.

Quoted 17 lines

---

diff -NurpP --minimal linux-2.6.16-rc4/kernel/exit.c linux-2.6.16-rc4-ns/kernel/exit.c

kernel/exit.c +3 -0
@@ -360,6 +360,9 @@ void daemonize(const char *name, ...)
 	fs = init_task.fs;
 	current->fs = fs;
 	atomic_inc(&fs->count);
+	exit_namespace(current);
+	current->namespace = init_task.namespace;
+	get_namespace(current->namespace);
  	exit_files(current);
 	current->files = init_task.files;
 	atomic_inc(&current->files->count);
#3 lore raw
In reply to Herbert Poetzl
Quoted 6 lines · Herbert Poetzl
Quoted 3 lines · Andrew Morton

I think it'd be better to convert ext3 to use the kthread API which appears to accidentally not have this problem, because such threads are parented by keventd, which were parented by init.

sounds like a plan!

Here's my attempt at such a conversion. Since jbd doesn't seem to want to collect an exit status, I didn't bother with kthread_stop().

I got overexcited and also embedded the journal device in the process name, but that's probably useless churn. Looks nice in pstree though:

     |-kthread-+-kblockd/0
     |         |-khubd
     |         |-2*[pdflush]
     |         |-aio/0
     |         |-v9fs/0
     |         |-cqueue/0
     |         |-kfand
     |         |-kcryptd/0
     |         |-kjournald/3:3
     |         |-kjournald/3:8
     |         |-kjournald/3:4
     |         |-kjournald/3:5
     |         `-kjournald/254:1
fs/jbd/journal.c +6 -4
index e4b516a..e33d993 100644
@@ -33,6 +33,8 @@
 #include <linux/mm.h>
 #include <linux/suspend.h>
 #include <linux/pagemap.h>
+#include <linux/kthread.h>
+#include <linux/kdev_t.h>
 #include <asm/uaccess.h>
 #include <asm/page.h>
 #include <linux/proc_fs.h>
@@ -115,8 +117,6 @@ static int kjournald(void *arg)
 	transaction_t *transaction;
 	struct timer_list timer;
 
-	daemonize("kjournald");
-
 	/* Set up an interval timer which can be used to trigger a
            commit wakeup after the commit interval expires */
 	init_timer(&timer);
@@ -207,12 +207,14 @@ end_loop:
 	journal->j_task = NULL;
 	wake_up(&journal->j_wait_done_commit);
 	jbd_debug(1, "Journal thread exiting.\n");
-	return 0;
+	do_exit(0);
 }
 
 static void journal_start_thread(journal_t *journal)
 {
-	kernel_thread(kjournald, journal, CLONE_VM|CLONE_FS|CLONE_FILES);
+	dev_t jdev = journal->j_dev->bd_dev;
+	kthread_run(kjournald, journal, "kjournald/%d:%d",
+		    MAJOR(jdev), MINOR(jdev), NULL);
 	wait_event(journal->j_wait_done_commit, journal->j_task != 0);
 }
 
Signature · 1 line
Dag vijandelijk luchtschip de huismeester is dood
#4 lore raw
Quoted 4 lines · Paul Collins
Quoted 1 line · Herbert Poetzl

On Fri, Feb 17, 2006 at 05:54:28PM -0800, Andrew Morton wrote:

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

I think it'd be better to convert ext3 to use the kthread API which appears to accidentally not have this problem, because such threads are parented by keventd, which were parented by init.

sounds like a plan!

Here's my attempt at such a conversion. Since jbd doesn't seem to want to collect an exit status, I didn't bother with kthread_stop().

Ah. I already did something similar. ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.16-rc4/2.6.16-rc4-mm2/broken-out/jbd-convert-kjournald-to-kthread-api.patch

Quoted 16 lines

I got overexcited and also embedded the journal device in the process name, but that's probably useless churn. Looks nice in pstree though:

      |-kthread-+-kblockd/0
      |         |-khubd
      |         |-2*[pdflush]
      |         |-aio/0
      |         |-v9fs/0
      |         |-cqueue/0
      |         |-kfand
      |         |-kcryptd/0
      |         |-kjournald/3:3
      |         |-kjournald/3:8
      |         |-kjournald/3:4
      |         |-kjournald/3:5
      |         `-kjournald/254:1

We only have 15 chars for that string - the final one you have there is on the raggedy edge.

#5 lore raw
Quoted 9 lines · Andrew Morton
Quoted 8 lines · Paul Collins
Quoted 5 lines · Herbert Poetzl
Quoted 3 lines · Andrew Morton

I think it'd be better to convert ext3 to use the kthread API which appears to accidentally not have this problem, because such threads are parented by keventd, which were parented by init.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

sounds like a plan!

Here's my attempt at such a conversion. Since jbd doesn't seem to want to collect an exit status, I didn't bother with kthread_stop().

Ah. I already did something similar. ftp://ftp.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.16-rc4/2.6.16-rc4-mm2/broken-out/jbd-convert-kjournald-to-kthread-api.patch

Sweet, I managed to get it right at least.

Quoted 7 lines
Quoted 4 lines

I got overexcited and also embedded the journal device in the process name, but that's probably useless churn. Looks nice in pstree though:

      |         `-kjournald/254:1

We only have 15 chars for that string - the final one you have there is on the raggedy edge.

Oh well, it's not that useful anyway. I can count the number of times I've needed to kill a particular kjournald on the fingers of one head.

Signature · 1 line
Dag vijandelijk luchtschip de huismeester is dood