cgroup: deprecate remount option changes and "name=" mount option

* "name=" mount option: This has been broken for years before Li fixed it recently. It has very marginal usefulness (plain bind mount can mostly replace it) and is quite unusual. Take the chance and mark for deprecation.

Hmm, this is not clear to me. Are you intending to remove the name= mount option entirely, or just intend to remove that you can change it in remounts?

systemd uses name=systemd to get its own private cgroup tree, that has no controller bound, hence we kinda need this option. Or are you suggesting there was an other way to get a private cgroup tree that systemd can use for grouping and labelling things?

Hello,

Quoted 8 lines · Lennart Poettering

Hmm, this is not clear to me. Are you intending to remove the name= mount option entirely, or just intend to remove that you can change it in remounts?

systemd uses name=systemd to get its own private cgroup tree, that has no controller bound, hence we kinda need this option. Or are you suggesting there was an other way to get a private cgroup tree that systemd can use for grouping and labelling things?

Eh, you're right. I want to remove support for duplicate mounting by name or subsys as it just isn't necessary, so what I want is,

* "mount -t cgroupfs -o SUBSYS" will fail if SUBSYS is already mounted somewhere else.

* "mount -t cgroupfs -o none" will always create a new cgroup hierarchy w/o any controller attached.

I didn't realize that the mount code didn't allow none w/o name and now I don't see a non-intrusive way to remove the name option. :( Eh well, I'll drop that part.

Hello,

Quoted 8 lines · Lennart Poettering

Hmm, this is not clear to me. Are you intending to remove the name= mount option entirely, or just intend to remove that you can change it in remounts?

systemd uses name=systemd to get its own private cgroup tree, that has no controller bound, hence we kinda need this option. Or are you suggesting there was an other way to get a private cgroup tree that systemd can use for grouping and labelling things?

Eh, you're right. I want to remove support for duplicate mounting by name or subsys as it just isn't necessary, so what I want is,

* "mount -t cgroupfs -o SUBSYS" will fail if SUBSYS is already mounted somewhere else.

* "mount -t cgroupfs -o none" will always create a new cgroup hierarchy w/o any controller attached.

I didn't realize that the mount code didn't allow none w/o name and now I don't see a non-intrusive way to remove the name option. :( Eh well, I'll drop that part.

Hmm, this is not clear to me. Are you intending to remove the name= mount option entirely, or just intend to remove that you can change it in remounts?

systemd uses name=systemd to get its own private cgroup tree, that has no controller bound, hence we kinda need this option. Or are you suggesting there was an other way to get a private cgroup tree that systemd can use for grouping and labelling things?

* "name=" mount option: This has been broken for years before Li fixed it recently. It has very marginal usefulness (plain bind mount can mostly replace it) and is quite unusual. Take the chance and mark for deprecation.

This patch removes the following features.

* Rebinding subsys by remount: Never reached useful state - only works on empty hierarchies.

* release_agent update by remount: release_agent itself will be replaced with conventional fsnotify notification.

This patch marks the following features for deprecation.

* Rebinding subsys by remount: Never reached useful state - only works on empty hierarchies.

* release_agent update by remount: release_agent itself will be replaced with conventional fsnotify notification.

Quoted 2 lines

* release_agent update by remount: release_agent itself will be replaced with conventional fsnotify notification.

Just wondering, what kind of event do you plan to emit if a cgroup runs empty? To userspace, with fanotify, will this be FAN_MODIFY? I'd very much prefer if we only get events really when a group runs empty, and not each time a PID is added/removed from a cgroup, which FAN_NOTIFY might suggest?

This patch marks the following features for deprecation.

* Rebinding subsys by remount: Never reached useful state - only works on empty hierarchies.

* release_agent update by remount: release_agent itself will be replaced with conventional fsnotify notification.

v2: Lennart pointed out that "name=" is necessary for mounts w/o any
controller attached. Drop "name=" deprecation.

Signed-off-by Tejun Heo <[email protected]>

Cc Li Zefan <[email protected]>

Cc Lennart Poettering <[email protected]>

---
 Documentation/feature-removal-schedule.txt |   10 ++++++++++
 kernel/cgroup.c                            |    5 +++++
 2 files changed, 15 insertions(+)

Index: work/Documentation/feature-removal-schedule.txt
===================================================================

Documentation/feature-removal-schedule.txt +10 -0

@@ -510,3 +510,13 @@ Why:	The pci_scan_bus_parented() interfa
 	convert to using pci_scan_root_bus() so they can supply a list of
 	bus resources when the bus is created.
 Who:	Bjorn Helgaas <[email protected]>
+
+----------------------------
+
+What:	cgroup option updates via remount
+When:	March 2013
+Why:	Remount currently allows changing bound subsystems and
+	release_agent.  Rebinding is hardly useful as it only works
+	when the hierarchy is empty and release_agent itself should be
+	replaced with conventional fsnotify.
+Who:	Tejun Heo <[email protected]>

Index: work/kernel/cgroup.c
===================================================================

kernel/cgroup.c +5 -0

@@ -1294,6 +1294,11 @@ static int cgroup_remount(struct super_b
 	if (ret)
 		goto out_unlock;
 
+	/* See feature-removal-schedule.txt */
+	if (opts.subsys_bits != root->actual_subsys_bits || opts.release_agent)
+		pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
+			   task_tgid_nr(current), current->comm);
+
 	/* Don't allow flags or name to change at remount */
 	if (opts.flags != root->flags ||
 	    (opts.name && strcmp(opts.name, root->name))) {

This patch marks the following features for deprecation.

* Rebinding subsys by remount: Never reached useful state - only works on empty hierarchies.

* release_agent update by remount: release_agent itself will be replaced with conventional fsnotify notification.

v2: Lennart pointed out that "name=" is necessary for mounts w/o any
controller attached. Drop "name=" deprecation.

Signed-off-by Tejun Heo <[email protected]>

Cc Li Zefan <[email protected]>

Cc Lennart Poettering <[email protected]>

---
 Documentation/feature-removal-schedule.txt |   10 ++++++++++
 kernel/cgroup.c                            |    5 +++++
 2 files changed, 15 insertions(+)

Index: work/Documentation/feature-removal-schedule.txt
===================================================================

Documentation/feature-removal-schedule.txt +10 -0

@@ -510,3 +510,13 @@ Why:	The pci_scan_bus_parented() interfa
 	convert to using pci_scan_root_bus() so they can supply a list of
 	bus resources when the bus is created.
 Who:	Bjorn Helgaas <[email protected]>
+
+----------------------------
+
+What:	cgroup option updates via remount
+When:	March 2013
+Why:	Remount currently allows changing bound subsystems and
+	release_agent.  Rebinding is hardly useful as it only works
+	when the hierarchy is empty and release_agent itself should be
+	replaced with conventional fsnotify.
+Who:	Tejun Heo <[email protected]>

Index: work/kernel/cgroup.c
===================================================================

kernel/cgroup.c +5 -0

@@ -1294,6 +1294,11 @@ static int cgroup_remount(struct super_b
 	if (ret)
 		goto out_unlock;
 
+	/* See feature-removal-schedule.txt */
+	if (opts.subsys_bits != root->actual_subsys_bits || opts.release_agent)
+		pr_warning("cgroup: option changes via remount are deprecated (pid=%d comm=%s)\n",
+			   task_tgid_nr(current), current->comm);
+
 	/* Don't allow flags or name to change at remount */
 	if (opts.flags != root->flags ||
 	    (opts.name && strcmp(opts.name, root->name))) {