containers (V7): Generic Process Containers
Subsystems (11)
cover containers (V7): Generic Process Containers
--
This is an update to my multi-hierarchy multi-subsystem generic process containers patch. Changes since V6 (22nd December) include:
- updated to 2.6.20
- added more details about multiple hierarchy support in the documentation
- reduced the per-task memory overhead to one pointer (previously it was one pointer for each hierarchy). Now each task has a pointer to a container_group, which holds the pointers to the containers (one per active hierarchy) that the task is attached to and the associated per-subsystem state (one per active subsystem). This container group is shared (with reference counts) between all tasks that have the same set of container mappings.
- added API support for binding/unbinding subsystems to/from active hierarchies, by remounting with -oremount,<new-subsys-list>. Currently this fails with EBUSY if the hierarchy has a child containers; full implementation support is left to a later patch.
- added a bind() subsystem callback to indicate when a subsystem is moved between hierarchies
- added container_clone(subsys, task), which creates a child container for the hierarchy that the specified subsystem is bound to, and moves the given task into that container. An example use of this would be in sys_unshare, which could, if the namespace container subsystem is active, create a child container when the new namespace is created.
- temporarily removed the "release agent" support. It's only currently used by CPUsets, and intrudes somewhat on the per-container reference counting. If necessary it can be re-added, either as a generic subsystem feature or a CPUset-specific feature, via a kernel thread that periodically polls containers that have been designated as notify_on_release to see if they are releasable
Generic Process Containers
--------------------------
There have recently been various proposals floating around for resource management/accounting and other task grouping subsystems in the kernel, including ResGroups, User BeanCounters, NSProxy containers, and others. These all need the basic abstraction of being able to group together multiple processes in an aggregate, in order to track/limit the resources permitted to those processes, or control other behaviour of the processes, and all implement this grouping in different ways.
Already existing in the kernel is the cpuset subsystem; this has a process grouping mechanism that is mature, tested, and well documented (particularly with regards to synchronization rules).
This patchset extracts the process grouping code from cpusets into a generic container system, and makes the cpusets code a client of the container system.
It also provides several example clients of the container system, including ResGroups, BeanCounters and namespace proxy.
The change is implemented in three stages, plus four example subsystems that aren't necessarily intended to be merged as part of this patch set, but demonstrate the applicability of the framework.
1) extract the process grouping code from cpusets into a standalone system
2) remove the process grouping code from cpusets and hook into the container system
3) convert the container system to present a generic multi-hierarchy API, and make cpusets a client of that API
4) example of a simple CPU accounting container subsystem
5) example of implementing ResGroups and its numtasks controller over generic containers
6) example of implementing BeanCounters and its numfiles counter over generic containers
7) example of integrating the namespace isolation code (sys_unshare() or various clone flags) with generic containers, allowing virtual servers to take advantage of other resource control efforts.
The intention is that the various resource management and virtualization efforts can also become container clients, with the result that:
- the userspace APIs are (somewhat) normalised
- it's easier to test out e.g. the ResGroups CPU controller in conjunction with the BeanCounters memory controller, or use either of
them as the resource-control portion of a virtual server system.
- the additional kernel footprint of any of the competing resource management systems is substantially reduced, since it doesn't need to provide process grouping/containment, hence improving their chances of getting into the kernel
PATCH 1/7 containers (V7): Generic container system abstracted from cpusets code
This patch creates a generic process container system based on (and parallel top) the cpusets code. At a coarse level it was created by copying kernel/cpuset.c, doing s/cpuset/container/g, and stripping out any code that was cpuset-specific rather than applicable to any process container subsystem.
--- Documentation/containers.txt | 229 +++++++ fs/proc/base.c | 7 include/linux/container.h | 96 +++ include/linux/sched.h | 5 init/Kconfig | 9 init/main.c | 3 kernel/Makefile | 1 kernel/container.c | 1343 +++++++++++++++++++++++++++++++++++++++++++ kernel/exit.c | 2 kernel/fork.c | 3 10 files changed, 1697 insertions(+), 1 deletion(-)
fs/proc/base.c +7 -0
@@ -68,6 +68,7 @@#include <linux/security.h> #include <linux/ptrace.h> #include <linux/seccomp.h> +#include <linux/container.h> #include <linux/cpuset.h> #include <linux/audit.h> #include <linux/poll.h>
@@ -1870,6 +1871,9 @@ static struct pid_entry tgid_base_stuff[#ifdef CONFIG_CPUSETS REG("cpuset", S_IRUGO, cpuset),#endif +#ifdef CONFIG_CONTAINERS +REG ("container" ,S_IRUGO , container), +#endifINF ("oom_score" ,S_IRUGO , oom_score),REG ("oom_adj" ,S_IRUGO |S_IWUSR , oom_adjust), #ifdef CONFIG_AUDITSYSCALL
@@ -2151,6 +2155,9 @@ static struct pid_entry tid_base_stuff[]#ifdef CONFIG_CPUSETS REG("cpuset", S_IRUGO, cpuset),#endif +#ifdef CONFIG_CONTAINERS +REG ("container" ,S_IRUGO , container), +#endifINF ("oom_score" ,S_IRUGO , oom_score),REG ("oom_adj" ,S_IRUGO |S_IWUSR , oom_adjust), #ifdef CONFIG_AUDITSYSCALL
include/linux/container.h new file +96 -0
96 lines not on this page · raw message
include/linux/sched.h +4 -1
@@ -743,8 +743,8 @@ extern unsigned int max_cache_size;struct io_context ;/* See blkdev.h */ +struct container ;struct cpuset ; -#define NGROUPS_SMALL 32#define NGROUPS_PER_BLOCK ((int)(PAGE_SIZE / sizeof(gid_t)))struct group_info {
@@ -1031,6 +1031,9 @@ struct task_struct {int cpuset_mems_generation;int cpuset_mem_spread_rotor;#endif +#ifdef CONFIG_CONTAINERS +struct container *container; +#endif struct robust_list_head __user *robust_list;#ifdef CONFIG_COMPAT struct compat_robust_list_head __user *compat_robust_list;
init/Kconfig +9 -0
@@ -238,6 +238,15 @@ config IKCONFIG_PROC This option enables access to the kernel configuration file through /proc/config.gz. +config CONTAINERS + bool "Container support" + help + This option will let you create and manage process containers, + which can be used to aggregate multiple processes, e.g. for + the purposes of resource tracking. + + Say N if unsure + config CPUSETS bool "Cpuset support" depends on SMP
init/main.c +3 -0
@@ -39,6 +39,7 @@#include <linux/writeback.h> #include <linux/cpu.h> #include <linux/cpuset.h> +#include <linux/container.h> #include <linux/efi.h> #include <linux/taskstats_kern.h> #include <linux/delayacct.h>
@@ -485,6 +486,7 @@ asmlinkage void __init start_kernel(voidchar * command_line;extern struct kernel_param __start___param[], __stop___param[]; +container_init_early ();smp_setup_processor_id ();/*
@@ -608,6 +610,7 @@ asmlinkage void __init start_kernel(void#ifdef CONFIG_PROC_FS proc_root_init ();#endif +container_init ();cpuset_init ();taskstats_init_early ();delayacct_init ();
kernel/container.c new file +1343 -0
1,343 lines not on this page · raw message
kernel/exit.c +2 -0
@@ -31,6 +31,7 @@#include <linux/taskstats_kern.h> #include <linux/delayacct.h> #include <linux/cpuset.h> +#include <linux/container.h> #include <linux/syscalls.h> #include <linux/signal.h> #include <linux/posix-timers.h>
@@ -927,6 +928,7 @@ fastcall NORET_TYPE void do_exit(long co__exit_fs (tsk);exit_thread ();cpuset_exit (tsk); +container_exit (tsk);exit_keys (tsk);if (group_dead && tsk->signal->leader)
kernel/fork.c +3 -0
@@ -31,6 +31,7 @@#include <linux/capability.h> #include <linux/cpu.h> #include <linux/cpuset.h> +#include <linux/container.h> #include <linux/security.h> #include <linux/swap.h> #include <linux/syscalls.h>
@@ -1058,6 +1059,7 @@ static struct task_struct *copy_process( p->io_context =NULL ; p->io_wait =NULL ; p->audit_context =NULL ; +container_fork (p);cpuset_fork (p);#ifdef CONFIG_NUMA p->mempolicy =mpol_copy (p->mempolicy);
@@ -1291,6 +1293,7 @@ bad_fork_cleanup_policy: bad_fork_cleanup_cpuset:#endif cpuset_exit (p); +container_exit (p); bad_fork_cleanup_delays_binfmt:delayacct_tsk_free (p);if (p->binfmt)
kernel/Makefile +1 -0
@@ -35,6 +35,7 @@ obj-$(CONFIG_PM) += power/ obj-$(CONFIG_BSD_PROCESS_ACCT) += acct.o obj-$(CONFIG_KEXEC) += kexec.o obj-$(CONFIG_COMPAT) += compat.o +obj-$(CONFIG_CONTAINERS) += container.o obj-$(CONFIG_CPUSETS) += cpuset.o obj-$(CONFIG_IKCONFIG) += configs.o obj-$(CONFIG_STOP_MACHINE) += stop_machine.o
Documentation/containers.txt new file +229 -0
229 lines not on this page · raw message
--
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
kernel/container.c line 1174
Quoted 3 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
kernel/container.c new file +3
-0+void container_fork (struct task_struct *child) +{ +task_lock (current);
Can't this be just rcu_read_lock()?
Quoted 4 lines
kernel/container.c new file +4
-0+ child->container = current->container; +atomic_inc (&child->container->count); +task_unlock (current); +}
Signature · 2 lines
Regards, vatsa
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 7 lines · Srivatsa Vaddagiri
kernel/container.c line 1174
Quoted 3 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
kernel/container.c new file +3
-0+void container_fork (struct task_struct *child) +{ +task_lock (current);Can't this be just
rcu_read_lock()?
In this particular patch (which is an almost verbatim extraction/renaming of the generic bits of the cpusets code into container.c) it probably could - but the main patch that adds the container_group support would lose it since we use kref to refcount container_group objects, and hence they're freed when their refcount reaches zero. RCU is still fine for reading the container_group pointers, but it's no good for updating them, since by the time you update it it may no longer be your container_group structure, and may instead be about to be deleted as soon as the other thread's rcu_synchronize() completes.
Paul
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 5 lines · Paul Menage
reaches zero. RCU is still fine for reading the
container_grouppointers, but it's no good for updating them, since by the time you update it it may no longer be yourcontainer_groupstructure, and may instead be about to be deleted as soon as the other thread'srcu_synchronize()completes.
On further reflection, this probably would be safe after all. Since we don't call put_container_group() in attach_task() until after synchronize_rcu() completes, that implies that a container_group_get() from the RCU section would have already completed. So we should be fine.
Paul
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 5 lines · Paul Menage
On further reflection, this probably would be safe after all. Since we don't call
put_container_group()inattach_task()until aftersynchronize_rcu()completes, that implies that acontainer_group_get()from the RCU section would have already completed. So we should be fine.
Right.
Which make me wonder why we need task_lock() at all ..I can understand the need for a lock like that if we are reading/updating multiple words in task_struct under the lock. In this case, it is used to read/write just one pointer, isnt it? I think it can be eliminated all-together with the use of RCU.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 5 lines · Srivatsa Vaddagiri
Which make me wonder why we need
task_lock()at all ..I can understand the need for a lock like that if we are reading/updating multiple words intask_structunder the lock. In this case, it is used to read/write just one pointer, isnt it? I think it can be eliminated all-together with the use of RCU.
I see that cpuset.c uses task_lock to read/write multiple words (cpuset_update_task_memory_state) ..So yes it is necessary in attach_task() ..
Signature · 2 lines
Regards, vatsa
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to [email protected]Quoted 6 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
+/**
+ *container_lock- lock out any changes to container structures
+ *
+ * The out of memory (oom) code needs tomutex_lockcontainers
+ * from being changed while it scans the tasklist looking for a
+ * task in an overlapping container.
Which specific portion of oom code cares abt container structure being intact?
If I understand correctly, only cpuset requires this double locking. More specifically, cpusets cares about walking cpuset->parent list safely with callback_mutex held correct?
If that is the case, I think we can push container_lock entirely inside cpuset.c and not have others exposed to this double-lock complexity. This is possible because cpuset.c (build on top of containers) still has cpuset->parent and walking cpuset->parent list safely can be made possible with a second lock which is local to only cpuset.c.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 5 lines · Srivatsa Vaddagiri
If that is the case, I think we can push
container_lockentirely insidecpuset.cand not have others exposed to this double-lock complexity. This is possible becausecpuset.c(build on top of containers) still hascpuset->parentand walkingcpuset->parentlist safely can be made possible with a second lock which is local to onlycpuset.c.
Hope I am not shooting in the dark here!
If we can indeed avoid container_lock in generic code, it will simplify code like attach_task (for ex: post_attach/attach can be clubbed into one single callback).
Signature · 2 lines
Regards, vatsa
Re: [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to Srivatsa VaddagiriQuoted 6 lines · Srivatsa Vaddagiri
If that is the case, I think we can push
container_lockentirely insidecpuset.cand not have others exposed to this double-lock complexity. This is possible becausecpuset.c(build on top of containers) still hascpuset->parentand walkingcpuset->parentlist safely can be made possible with a second lock which is local to onlycpuset.c.
The callback mutex (which is what container_lock() actually locks) is also used to synchronize fork/exit against subsystem additions, in the event that some subsystem has registered fork or exit callbacks. We could probably have a separate subsystem_mutex for that instead.
Apart from that, yes, it may well be possible to move callback lock entirely inside cpusets.
Paul
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 4 lines · Paul Menage
The callback mutex (which is what
container_lock()actually locks) is also used to synchronize fork/exit against subsystem additions, in the event that some subsystem has registered fork or exit callbacks. We could probably have a separatesubsystem_mutexfor that instead.
Why can't manage_mutex itself be used there (to serialize fork/exit callbacks against modification to hierarchy)?
Quoted 2 lines
Apart from that, yes, it may well be possible to move callback lock entirely inside cpusets.
Yes, that way only the hierarchy hosting cpusets takes the hit of double-locking. cpuset_subsys->create/destroy can take this additional lock inside cpuset.c.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 8 lines · Srivatsa Vaddagiri
Quoted 4 lines · Paul Menage
The callback mutex (which is what
container_lock()actually locks) is also used to synchronize fork/exit against subsystem additions, in the event that some subsystem has registered fork or exit callbacks. We could probably have a separatesubsystem_mutexfor that instead.Why can't
manage_mutexitself be used there (to serialize fork/exit callbacks against modification to hierarchy)?
Because manage_mutex can be held for very long periods of time. I think that a combination of a new lock that's only taken by fork/exit and register_subsys, plus task_lock (which prevents the current task from being moved) would be more lightweight.
Paul
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to Srivatsa VaddagiriQuoted 3 lines · vatsa
Yes, that way only the hierarchy hosting cpusets takes the hit of double-locking. cpuset_subsys->create/destroy can take this additional lock inside
cpuset.c.
The primary reason for the cpuset double locking, as I recall, was because cpusets needs to access cpusets inside the memory allocator. A single, straight forward, cpuset lock failed under the following common scenario:
1) user does cpuset system call (writes some file below /dev/cpuset, e.g.)
2) kernel cpuset code locks its lock
3) cpuset code asks to allocate some memory for some cpuset structure
4) memory allocator tries to lock the cpuset lock - deadlock!
The reason that the memory allocator needs the cpuset lock is to check whether the memory nodes the current task is allowed to use have changed, due to changes in the current tasks cpuset.
A secondary reason that the cpuset code needs two locks is because the main cpuset lock is a long held, system wide lock, and various low level bits of performance critical code sometimes require quick, read-only access to cpusets.
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 2 lines · Paul Jackson
The primary reason for the cpuset double locking, as I recall, was because cpusets needs to access cpusets inside the memory allocator.
"needs to access cpusets" - can you be more specific?
Being able to safely walk cpuset->parent list - is it the only access you are talking of or more?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to [email protected]kernel/container.c line 238
Quoted 3 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
kernel/container.c new file +3
-0+static void remove_dir (struct dentry *d) +{ +struct dentry *parent =dget (d->d_parent);
Don't we need to lock parent inode's mutex here? sysfs seems to take that lock.
Quoted 5 lines
kernel/container.c new file +5
-0+ +d_delete (d); +simple_rmdir (parent->d_inode, d); +dput (parent); +}
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 7 lines · Srivatsa Vaddagiri
kernel/container.c line 238
Quoted 3 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
kernel/container.c new file +3
-0+static void remove_dir (struct dentry *d) +{ +struct dentry *parent =dget (d->d_parent);Don't we need to lock parent inode's mutex here? sysfs seems to take that lock.
Never mind. Maneesh pointed me to do_rmdir() in VFS which takes that mutex as well.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to [email protected]Quoted 5 lines · On Mon, Feb 12, 2007 at 12:15:22AM -0800, [email protected] wrote:
kernel/container.c new file +5
-0+static int attach_task (struct container *cont,char *pidbuf,char **ppathbuf) +{ +pid_t pid; +struct task_struct *tsk; +struct container *oldcont;
kernel/container.c line 498
Quoted 10 lines
kernel/container.c new file +10
-0+int retval; + +if (sscanf (pidbuf,"%d" , &pid) !=1 ) +return -EIO ; + +if (pid) { +read_lock (&tasklist_lock); + + tsk =find_task_by_pid (pid); +if (!tsk || tsk->flags &PF_EXITING ) {
This is probably carrying over code from cpuset.c, but :
/me thinks that there is a ugly race here with 'tsk' exiting. What happens if the tsk is marked PF_EXITING just after this check? If that happens, then:
Quoted 25 lines
kernel/container.c new file +25
-0+read_unlock (&tasklist_lock); +return -ESRCH ; + } + +get_task_struct (tsk); +read_unlock (&tasklist_lock); + +if ((current->euid) && (current->euid != tsk->uid) + && (current->euid != tsk->suid)) { +put_task_struct (tsk); +return -EACCES ; + } + }else { + tsk = current; +get_task_struct (tsk); + } + + retval =security_task_setscheduler (tsk,0 ,NULL ); +if (retval) { +put_task_struct (tsk); +return retval; + } + +mutex_lock (&callback_mutex); +
kernel/container.c line 533
Quoted 10 lines
kernel/container.c new file +10
-0+task_lock (tsk); + oldcont = tsk->container; +if (!oldcont) { +task_unlock (tsk); +mutex_unlock (&callback_mutex); +put_task_struct (tsk); +return -ESRCH ; + } +atomic_inc (&cont->count); +rcu_assign_pointer (tsk->container, cont);
Above assignment A1 can race with below assignment A2 in container_exit() :
tsk->container = &top_container; /* the_top_container_hack - see above */
What happens if A1 follows after A2? I feel very uncomfortable abt it.
IMO, we need to use task_lock() in container_exit() to avoid this race.
(I think this race already exists in mainline cpuset.c?)
P.S : cpuset.c checks for PF_EXITING twice in attach_task(), while this patch seems to be checking only once. Is that fine?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 6 lines
IMO, we need to use
task_lock()incontainer_exit()to avoid this race.(I think this race already exists in mainline
cpuset.c?)
P.S:cpuset.cchecks forPF_EXITINGtwice inattach_task(), while this patch seems to be checking only once. Is that fine?
I think the cpuset code is ok, because, as you note, it locks the task, picks off the cpuset pointer, and then checks a second time that the task still does not have PF_EXITING set:
In the kernel/cpuset.c code for attach_task():
task_lock(tsk);
oldcs = tsk->cpuset;
/*
* After getting 'oldcs' cpuset ptr, be sure still not exiting.
* If 'oldcs' might be the top_cpuset due to the_top_cpuset_hack
* then fail this attach_task(), to avoid breaking top_cpuset.count.
*/
if (tsk->flags & PF_EXITING) {
task_unlock(tsk);
mutex_unlock(&callback_mutex);
put_task_struct(tsk);
return -ESRCH;
}Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 6 lines · Paul Jackson
Quoted 2 lines
P.S:cpuset.cchecks forPF_EXITINGtwice inattach_task(), while this patch seems to be checking only once. Is that fine?I think the cpuset code is ok, because, as you note, it locks the task, picks off the cpuset pointer, and then checks a second time that the task still does not have
PF_EXITINGset:
Well afaics, PF_EXITING is set for the exiting task w/o taking any lock, which makes this racy always.
Quoted 10 lines
In the
kernel/cpuset.ccode forattach_task():task_lock(tsk); oldcs = tsk->cpuset; /* * After getting 'oldcs' cpuset ptr, be sure still not exiting. * If 'oldcs' might be the top_cpuset due to the_top_cpuset_hack * then fail this attach_task(), to avoid breaking top_cpuset.count. */ if (tsk->flags & PF_EXITING) {
What if PF_EXITING is set after this check? If that happens then,
Quoted 5 lines
task_unlock(tsk);
mutex_unlock(&callback_mutex);
put_task_struct(tsk);
return -ESRCH;
}the following code becomes racy with cpuset_exit() ...
atomic_inc(&cs->count);
rcu_assign_pointer(tsk->cpuset, cs);
task_unlock(tsk);Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 10 lines · vatsa
Quoted 1 line
if (tsk->flags & PF_EXITING) {What if
PF_EXITINGis set after this check? If that happens then,Quoted 5 lines
task_unlock(tsk); mutex_unlock(&callback_mutex); put_task_struct(tsk); return -ESRCH; }
Quoted 5 lines
the following code becomes racy with
cpuset_exit()...atomic_inc(&cs->count); rcu_assign_pointer(tsk->cpuset, cs); task_unlock(tsk);
eh ... so ... ?
I don't know of any sequence where that causes any problem.
Do you see one?
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 11 lines · Paul Jackson
Quoted 5 lines
the following code becomes racy with
cpuset_exit()...atomic_inc(&cs->count); rcu_assign_pointer(tsk->cpuset, cs); task_unlock(tsk);eh ... so ... ?
I don't know of any sequence where that causes any problem.
Do you see one?
Let's say we had two cpusets CS1 amd CS2 (both different from top_cpuset). CS1 has just one task T1 in it (CS1->count = 0) while CS2 has no tasks in it (CS2->count = 0).
Now consider:
--------------------------------------------------------------------
CPU0 (attach_task T1 to CS2) CPU1 (T1 is exiting)
--------------------------------------------------------------------
task_lock(T1);
oldcs = tsk->cpuset;
[oldcs = CS1]
T1->flags & PF_EXITING? (No)
T1->flags = PF_EXITING;
atomic_inc(&CS2->count);
cpuset_exit() cs = tsk->cpuset; (cs = CS1)
T1->cpuset = CS2;
T1->cpuset = &top_cpuset;
task_unlock(T1);
CS2 has one bogus count now (with no tasks in it), which may prevent it from being removed/freed forever.
Not just this, continuing further we have more trouble:
--------------------------------------------------------------------
CPU0 (attach_task T1 to CS2) CPU1 (T1 is exiting)
--------------------------------------------------------------------
synchronize_rcu()
atomic_dec(&CS1->count); [CS1->count = 0]
if atomic_dec_and_test(&oldcs->count))
[CS1->count = -1]
We now have CS1->count negative. Is that good? I am uncomfortable ..
We need a task_lock() in cpuset_exit to avoid this race.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 10 lines · Srivatsa Vaddagiri
Not just this, continuing further we have more trouble:
--------------------------------------------------------------------
CPU0 (attach_taskT1 to CS2) CPU1 (T1 is exiting)
--------------------------------------------------------------------
synchronize_rcu()atomic_dec(&CS1->count); [CS1->count = 0]
Quoted 8 lines
if
atomic_dec_and_test(&oldcs->count))[CS1->count = -1]We now have
CS1->countnegative. Is that good? I am uncomfortable ..We need a
task_lock()incpuset_exitto avoid this race.
2nd race is tricky. We probably need to do this to avoid it:
task_lock(tsk);
/* Check if tsk->cpuset is still same. We may have raced with
* cpuset_exit changing tsk->cpuset again under our feet.
*/
if (tsk->cpuset == cs && atomic_dec_and_test(&oldcs->count)) {
task_unlock(tsk);
check_for_release(oldcs, ppathbuf);
goto done;
}
task_unlock(tsk);done:
return 0;
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 15 lines · vatsa
Not just this, continuing further we have more trouble:
--------------------------------------------------------------------
CPU0 (attach_taskT1 to CS2) CPU1 (T1 is exiting)
--------------------------------------------------------------------
synchronize_rcu()atomic_dec(&CS1->count); [CS1->count = 0]if
atomic_dec_and_test(&oldcs->count))[CS1->count = -1]...
2nd race is tricky. We probably need to do this to avoid it:
Quoted 6 lines
task_lock(tsk);
/* Check if tsk->cpuset is still same. We may have raced with
* cpuset_exit changing tsk->cpuset again under our feet.
*/
if (tsk->cpuset == cs && atomic_dec_and_test(&oldcs->count)) {I'm unsure here, but this 'tsk->cpuset == cs' test feels fragile to me.
How about a bit earlier in attach_task(), right at the point we overwrite the victim tasks cpuset pointer, we decrement the count on the old cpuset, and if it went to zero, remember that we'll need to release it, once we've dropped some locks:
static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
{
...
struct cpuset *oldcs;
struct cpuset *oldcs_tobe_released;
...
task_lock(tsk);
oldcs = tsk->cpuset;
...
if (tsk->flags & PF_EXITING) {
...
}
atomic_inc(&cs->count);
rcu_assign_pointer(tsk->cpuset, cs);
oldcs_tobe_released = NULL;
if (atomic_dec_and_test(&oldcs->count))
oldcs_tobe_released = oldcs;
task_unlock(tsk);
...
put_task_struct(tsk);
synchronize_rcu();
if (oldcs_tobe_released)
check_for_release(oldcs_tobe_released, ppathbuf);
return 0;
}Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 6 lines · Paul Jackson
I'm unsure here, but this '
tsk->cpuset== cs' test feels fragile to me.How about a bit earlier in
attach_task(), right at the point we overwrite the victim tasks cpuset pointer, we decrement the count on the old cpuset, and if it went to zero, remember that we'll need to release it, once we've dropped some locks:
Looks neater. Will adopt this when I send the patch. Thanks.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
In reply to Srivatsa VaddagiriQuoted 1 line · vatsa
Now consider:
Nice work - thanks. Yes, both an extra cpuset count and a negative cpuset count are bad news, opening the door to the usual catastrophes.
Would you like the honor of submitting the patch to add a task_lock to cpuset_exit()? If you do, be sure to fix, or at least remove, the cpuset_exit comment lines:
* We don't need to task_lock() this reference to tsk->cpuset,
* because tsk is already marked PF_EXITING, so attach_task() won't
* mess with it, or task is a failed fork, never visible to attach_task.
I guess that taking task_lock() in cpuset_exit() should not be a serious performance issue. It's taking a spinlock that is in the current exiting tasks task struct, so it should be a cache hot memory line and a rarely contested lock.
And I guess I've not see this race in real life, as one side of it has to execute quite a bit of code in the task exit path, from when it sets PF_EXITING until it gets into the cpuset_exit() call, while the other side does the three lines:
if (tsk->flags & PF_EXITING) ... atomic_inc(&cs->count); rcu_assign_pointer(tsk->cpuset, cs);
So, in real life, this would be a difficult race to trigger.
Thanks for finding this.
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 6 lines · Paul Jackson
Nice work - thanks. Yes, both an extra cpuset count and a negative cpuset count are bad news, opening the door to the usual catastrophes.
Would you like the honor of submitting the patch to add a
task_locktocpuset_exit()? If you do, be sure to fix, or at least remove, thecpuset_exitcomment lines:
I will try to send out a patch later today to fix this bug in mainline cpuset code. I happened to notice this race with my rcfs patch and observed same is true with cpuset/container code also.
Quoted 3 lines
* We don't need to
task_lock()this reference totsk->cpuset,
* because tsk is already markedPF_EXITING, soattach_task()won't
* mess with it, or task is a failed fork, never visible toattach_task.
Sure, I had seen that.
Quoted 1 line
So, in real life, this would be a difficult race to trigger.
Agreed, but good to keep code clean isn't it? :)
Quoted 1 line
Thanks for finding this.
Wellcome!
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 1/7] containers (V7): Generic container system abstracted from cpusets code
Quoted 1 line
I will try to send out a patch later today to fix
Thanks!
Quoted 1 line
Agreed, but good to keep code clean isn't it? :)
Definitely.
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401PATCH 2/7 containers (V7): Cpusets hooked into containers
In reply to [email protected]This patch removes the process grouping code from the cpusets code, instead hooking it into the generic container system. This temporarily adds cpuset-specific code in kernel/container.c, which is removed by the next patch in the series.
--- Documentation/cpusets.txt | 81 +- fs/proc/base.c | 4 fs/super.c | 5 include/linux/container.h | 7 include/linux/cpuset.h | 25 include/linux/fs.h | 2 include/linux/mempolicy.h | 2 include/linux/sched.h | 4 init/Kconfig | 14 kernel/container.c | 107 +++ kernel/cpuset.c | 1269 +++++----------------------------------------- kernel/exit.c | 2 kernel/fork.c | 7 mm/oom_kill.c | 6 14 files changed, 319 insertions(+), 1216 deletions(-)
include/linux/container.h +7 -0
@@ -47,6 +47,10 @@ struct container {struct container *parent;/* my parent */ struct dentry *dentry;/* container fs entry */ + +#ifdef CONFIG_CPUSETS +struct cpuset *cpuset; +#endif };/* struct cftype:
@@ -79,6 +83,9 @@ struct cftype {int container_add_file (struct container *cont,const struct cftype *cft);int container_is_removed (const struct container *cont); +void container_set_release_agent_path (const char *path); + +int container_path (const struct container *cont,char *buf,int buflen);#else /* !CONFIG_CONTAINERS */
include/linux/cpuset.h +13 -12
@@ -11,16 +11,15 @@#include <linux/sched.h> #include <linux/cpumask.h> #include <linux/nodemask.h> +#include <linux/container.h> #ifdef CONFIG_CPUSETS -extern int number_of_cpusets;/* How many cpusets are defined in system? */ +extern int number_of_cpusets;/* How many cpusets are defined in system? */ extern int cpuset_init_early (void );extern int cpuset_init (void );extern void cpuset_init_smp (void ); -extern void cpuset_fork (struct task_struct *p); -extern void cpuset_exit (struct task_struct *p);extern cpumask_t cpuset_cpus_allowed (struct task_struct *p);extern nodemask_t cpuset_mems_allowed (struct task_struct *p);#define cpuset_current_mems_allowed (current->mems_allowed)
@@ -57,10 +56,6 @@ extern void __cpuset_memory_pressure_bumextern struct file_operations proc_cpuset_operations;extern char *cpuset_task_status_allowed (struct task_struct *task,char *buffer); - -extern void cpuset_lock (void ); -extern void cpuset_unlock (void ); -extern int cpuset_mem_spread_node (void );static inline int cpuset_do_page_mem_spread (void )
@@ -75,13 +70,22 @@ static inline int cpuset_do_slab_mem_sprextern void cpuset_track_online_nodes (void ); +extern int cpuset_can_attach_task (struct container *cont, +struct task_struct *tsk); +extern void cpuset_attach_task (struct container *cont, +struct task_struct *tsk); +extern void cpuset_post_attach_task (struct container *cont, +struct container *oldcont, +struct task_struct *tsk); +extern int cpuset_populate_dir (struct container *cont); +extern int cpuset_create (struct container *cont); +extern void cpuset_destroy (struct container *cont); +#else /* !CONFIG_CPUSETS */ static inline int cpuset_init_early (void ) {return 0 ; }static inline int cpuset_init (void ) {return 0 ; }static inline void cpuset_init_smp (void ) {} -static inline void cpuset_fork (struct task_struct *p) {} -static inline void cpuset_exit (struct task_struct *p) {}static inline cpumask_t cpuset_cpus_allowed (struct task_struct *p) {
@@ -126,9 +130,6 @@ static inline char *cpuset_task_status_areturn buffer; } -static inline void cpuset_lock (void ) {} -static inline void cpuset_unlock (void ) {} -static inline int cpuset_mem_spread_node (void ) {return 0 ;
kernel/exit.c +0 -2
@@ -30,7 +30,6 @@#include <linux/mempolicy.h> #include <linux/taskstats_kern.h> #include <linux/delayacct.h> -#include <linux/cpuset.h> #include <linux/container.h> #include <linux/syscalls.h> #include <linux/signal.h>
@@ -927,7 +926,6 @@ fastcall NORET_TYPE void do_exit(long co__exit_files (tsk);__exit_fs (tsk);exit_thread (); -cpuset_exit (tsk);container_exit (tsk);exit_keys (tsk);
kernel/fork.c +2 -5
@@ -30,7 +30,6 @@#include <linux/nsproxy.h> #include <linux/capability.h> #include <linux/cpu.h> -#include <linux/cpuset.h> #include <linux/container.h> #include <linux/security.h> #include <linux/swap.h>
@@ -1060,13 +1059,12 @@ static struct task_struct *copy_process( p->io_wait =NULL ; p->audit_context =NULL ;container_fork (p); -cpuset_fork (p);#ifdef CONFIG_NUMA p->mempolicy =mpol_copy (p->mempolicy);if (IS_ERR (p->mempolicy)) { retval =PTR_ERR (p->mempolicy); p->mempolicy =NULL ; - goto bad_fork_cleanup_cpuset; + goto bad_fork_cleanup_container; }mpol_fix_fork_child_flag (p);#endif
@@ -1290,9 +1288,8 @@ bad_fork_cleanup_security: bad_fork_cleanup_policy:#ifdef CONFIG_NUMA mpol_free (p->mempolicy); -bad_fork_cleanup_cpuset: +bad_fork_cleanup_container:#endif -cpuset_exit (p);container_exit (p); bad_fork_cleanup_delays_binfmt:delayacct_tsk_free (p);
kernel/container.c +100 -7
221 lines not on this page · raw message
kernel/cpuset.c +150 -1119
1,756 lines not on this page · raw message
init/Kconfig +7 -7
@@ -239,17 +239,12 @@ config IKCONFIG_PROC through /proc/config.gz. config CONTAINERS - bool "Container support" - help - This option will let you create and manage process containers, - which can be used to aggregate multiple processes, e.g. for - the purposes of resource tracking. - - Say N if unsure + bool config CPUSETS bool "Cpuset support" depends on SMP + select CONTAINERS help This option will let you create and manage CPUSETs which allow dynamically partitioning a system into sets of CPUs and
@@ -278,6 +273,11 @@ config SYSFS_DEPRECATED If you are using a distro that was released in 2006 or later, it should be safe to say N here. +config PROC_PID_CPUSET + bool "Include legacy /proc/<pid>/cpuset file" + depends on CPUSETS + default y + config RELAY bool "Kernel->user space relay support (formerly relayfs)" help
mm/oom_kill.c +3 -3
@@ -409,7 +409,7 @@ void out_of_memory(struct zonelist *zoneshow_mem (); } -cpuset_lock (); +container_lock ();read_lock (&tasklist_lock);/*
@@ -443,7 +443,7 @@ retry:/* Found nothing?!?! Either we hang forever, or we panic. */ if (!p) {read_unlock (&tasklist_lock); -cpuset_unlock (); +container_unlock ();panic ("Out of memory and no killable processes...\n" ); }
@@ -455,7 +455,7 @@ retry: out:read_unlock (&tasklist_lock); -cpuset_unlock (); +container_unlock ();/* * Give "p" a good chance of killing itself before we
include/linux/sched.h +1 -3
@@ -744,7 +744,6 @@ extern unsigned int max_cache_size;struct io_context ;/* See blkdev.h */ struct container ; -struct cpuset ;#define NGROUPS_SMALL 32#define NGROUPS_PER_BLOCK ((int)(PAGE_SIZE / sizeof(gid_t)))struct group_info {
@@ -1026,7 +1025,6 @@ struct task_struct {short il_next;#endif #ifdef CONFIG_CPUSETS -struct cpuset *cpuset;nodemask_t mems_allowed;int cpuset_mems_generation;int cpuset_mem_spread_rotor;
@@ -1471,7 +1469,7 @@ static inline int thread_group_empty(str/* * Protects ->fs, ->files, ->mm, ->group_info, ->comm, keyring * subscriptions and synchronises with wait4(). Also used in procfs. Also -* pins the final release of task.io_context. Also protects ->cpuset. +* pins the final release of task.io_context. Also protects ->container. * * Nests both inside and outside of read_lock(&tasklist_lock). * It must not be nested with write_lock_irq(&tasklist_lock),
Documentation/cpusets.txt +31 -50
166 lines not on this page · raw message
fs/super.c +0 -5
@@ -39,11 +39,6 @@#include <linux/mutex.h> #include <asm/uaccess.h> - -void get_filesystem (struct file_system_type *fs); -void put_filesystem (struct file_system_type *fs); -struct file_system_type *get_fs_type (const char *name); -LIST_HEAD (super_blocks);DEFINE_SPINLOCK (sb_lock);
include/linux/fs.h +2 -0
@@ -1841,6 +1841,8 @@ extern int vfs_fstat(unsigned int, strucextern int vfs_ioctl (struct file *,unsigned int ,unsigned int ,unsigned long ); +extern void get_filesystem (struct file_system_type *fs); +extern void put_filesystem (struct file_system_type *fs);extern struct file_system_type *get_fs_type (const char *name);extern struct super_block *get_super (struct block_device *);extern struct super_block *user_get_super (dev_t );
include/linux/mempolicy.h +1 -1
@@ -152,7 +152,7 @@ extern void mpol_fix_fork_child_flag(str#ifdef CONFIG_CPUSETS #define current_cpuset_is_being_rebound () \ - (cpuset_being_rebound == current->cpuset) + (cpuset_being_rebound == current->container->cpuset)#else #define current_cpuset_is_being_rebound () 0#endif
fs/proc/base.c +2 -2
@@ -1868,7 +1868,7 @@ static struct pid_entry tgid_base_stuff[#ifdef CONFIG_SCHEDSTATS INF("schedstat", S_IRUGO, pid_schedstat),#endif -#ifdef CONFIG_CPUSETS +#ifdef CONFIG_PROC_PID_CPUSET REG("cpuset", S_IRUGO, cpuset),#endif #ifdef CONFIG_CONTAINERS
@@ -2152,7 +2152,7 @@ static struct pid_entry tid_base_stuff[]#ifdef CONFIG_SCHEDSTATS INF("schedstat", S_IRUGO, pid_schedstat),#endif -#ifdef CONFIG_CPUSETS +#ifdef CONFIG_PROC_PID_CPUSET REG("cpuset", S_IRUGO, cpuset),#endif #ifdef CONFIG_CONTAINERS
--
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 6 lines · Quoting [email protected] ([email protected]):
Index: container-2.6.20/init/Kconfig
===================================================================init/Kconfig +0
-0@@ -239,17 +239,12 @@ config IKCONFIG_PROC through /proc/config.gz.
init/Kconfig config IKCONFIG_PROC · line 242
Quoted 10 lines
init/Kconfig +1
-7config CONTAINERS - bool "Container support" - help - This option will let you create and manage process containers, - which can be used to aggregate multiple processes, e.g. for - the purposes of resource tracking. - - Say N if unsure + bool
Hi,
why did you do this?
thanks,
-serge
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 11 lines · Serge E. Hallyn
init/Kconfig config IKCONFIG_PROC · line 242
Quoted 10 lines
config CONTAINERS
init/Kconfig +1
-7- bool "Container support" - help - This option will let you create and manage process containers, - which can be used to aggregate multiple processes, e.g. for - the purposes of resource tracking. - - Say N if unsure + bool
Quoted 4 lines
Hi,
why did you do this?
Containers on their own aren't all that useful, other than for process tracking, so it seemed reasonable to only build containers if there was some subsystem using them. If people wanted to use them for process tracking alone, it could certainly be made an individually-selectable option.
Paul
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
In reply to [email protected]kernel/cpuset.c static int update_nodemask(struct cpuset · line 540
Quoted 6 lines · On Mon, Feb 12, 2007 at 12:15:23AM -0800, [email protected] wrote:
/*
kernel/cpuset.c +1
-0@@ -913,12 +537,14 @@ static int update_nodemask(struct cpusetint migrate;int fudge;int retval; +struct container *cont;
This seems to be redundant?
Signature · 2 lines
Regards, vatsa
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 9 lines · Srivatsa Vaddagiri
kernel/cpuset.c static int update_nodemask(struct cpuset · line 540
Quoted 6 lines · On Mon, Feb 12, 2007 at 12:15:23AM -0800, [email protected] wrote:
/*
kernel/cpuset.c +1
-0@@ -913,12 +537,14 @@ static int update_nodemask(struct cpusetint migrate;int fudge;int retval; +struct container *cont;This seems to be redundant?
It gets used in the lower loop checking for processes whose memory policies we should be rebinding.
Paul
Re: [ckrm-tech] [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 10 lines · Paul Menage
Quoted 7 lines
kernel/cpuset.c static int update_nodemask(struct cpuset · line 540
Quoted 5 lines
kernel/cpuset.c +1
-0@@ -913,12 +537,14 @@ static int update_nodemask(struct cpusetint migrate;int fudge;int retval; +struct container *cont;This seems to be redundant?
It gets used in the lower loop checking for processes whose memory policies we should be rebinding.
It makes sense in the first cpuset patch (cpusets_using_containers.patch), but should be removed in the second cpuset patch (multiuser_container.patch). In the 2nd patch, we use this comparison:
if (task_cs(p) != cs) continue;
cont variable introduced in the 1st patch essentially becomes unused after 2nd patch.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 11 lines · Srivatsa Vaddagiri
It makes sense in the first cpuset patch (cpusets_using_containers.patch), but should be removed in the second cpuset patch (multiuser_container.patch). In the 2nd patch, we use this comparison:
if (task_cs(p) != cs) continue;cont variable introduced in the 1st patch essentially becomes unused after 2nd patch.
Yes, you're right - it could be removed in the 3rd patch.
Paul
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
In reply to [email protected]kernel/cpuset.c static int cpuset_populate_dir(struct de · line 1903
Quoted 6 lines · On Mon, Feb 12, 2007 at 12:15:23AM -0800, [email protected] wrote:
kernel/cpuset.c +2
-3-mutex_lock (&callback_mutex); -list_add (&cs->sibling, &cs->parent->children); + cont->cpuset = cs; + cs->container = cont; number_of_cpusets++; -mutex_unlock (&callback_mutex);
What's the rule to read/write number_of_cpusets? The earlier cpuset code was incrementing/decrementing under callback_mutex, but now we aren't. How safe is that?
The earlier cpuset code also was reading number_of_cpusets w/o the callback_mutex held (atleast in cpuset_zone_allowed_softwall). Is that safe?
Signature · 2 lines
Regards, vatsa
Re: [PATCH 2/7] containers (V7): Cpusets hooked into containers
Quoted 11 lines · Srivatsa Vaddagiri
kernel/cpuset.c static int cpuset_populate_dir(struct de · line 1903
Quoted 6 lines · On Mon, Feb 12, 2007 at 12:15:23AM -0800, [email protected] wrote:
kernel/cpuset.c +2
-3-mutex_lock (&callback_mutex); -list_add (&cs->sibling, &cs->parent->children); + cont->cpuset = cs; + cs->container = cont; number_of_cpusets++; -mutex_unlock (&callback_mutex);What's the rule to read/write
number_of_cpusets? The earlier cpuset code was incrementing/decrementing undercallback_mutex, but now we aren't. How safe is that?
We're still inside manage_mutex, so we guarantee that no-one else is changing it.
Quoted 3 lines
The earlier cpuset code also was reading
number_of_cpusetsw/o thecallback_mutexheld (atleast incpuset_zone_allowed_softwall). Is that safe?
Yes, I think so. Unless every memory allocator was to hold a lock for the duration of alloc_pages(), number_of_cpusets can theoretically be out of date by the time you're using it. But since the process could have allocated just before you created the first cpuset and moved it into that cpuset anywa, it's not really a race (and the consequences are inconsequential).
Paul
PATCH 4/7 containers (V7): Simple CPU accounting container subsystem
In reply to [email protected]This demonstrates how to use the generic container subsystem for a simple resource tracker that counts the total CPU time used by all processes in a container, during the time that they're members of the container.
--- include/linux/cpu_acct.h | 14 +++ init/Kconfig | 7 + kernel/Makefile | 1 kernel/cpu_acct.c | 213 +++++++++++++++++++++++++++++++++++++++++++++++ kernel/sched.c | 14 ++- 5 files changed, 246 insertions(+), 3 deletions(-)
include/linux/cpu_acct.h new file +14 -0
@@ -0,0 +1,14 @@ + +#ifndef _LINUX_CPU_ACCT_H +#define _LINUX_CPU_ACCT_H + +#include <linux/container.h> +#include <asm/cputime.h> + +#ifdef CONFIG_CONTAINER_CPUACCT +extern void cpuacct_charge (struct task_struct *,cputime_t cputime); +#else +static void inline cpuacct_charge (struct task_struct *p,cputime_t cputime) {} +#endif + +#endif
init/Kconfig +7 -0
@@ -290,6 +290,13 @@ config PROC_PID_CPUSET depends on CPUSETS default y +config CONTAINER_CPUACCT + bool "Simple CPU accounting container subsystem" + select CONTAINERS + help + Provides a simple Resource Controller for monitoring the + total CPU consumed by the tasks in a container + config RELAY bool "Kernel->user space relay support (formerly relayfs)" help
kernel/cpu_acct.c new file +213 -0
213 lines not on this page · raw message
kernel/Makefile +1 -0
@@ -37,6 +37,7 @@ obj-$(CONFIG_KEXEC) += kexec.o obj-$(CONFIG_COMPAT) += compat.o obj-$(CONFIG_CONTAINERS) += container.o obj-$(CONFIG_CPUSETS) += cpuset.o +obj-$(CONFIG_CONTAINER_CPUACCT) += cpu_acct.o obj-$(CONFIG_IKCONFIG) += configs.o obj-$(CONFIG_STOP_MACHINE) += stop_machine.o obj-$(CONFIG_AUDIT) += audit.o auditfilter.o
kernel/sched.c +11 -3
@@ -52,6 +52,7 @@#include <linux/tsacct_kern.h> #include <linux/kprobes.h> #include <linux/delayacct.h> +#include <linux/cpu_acct.h> #include <asm/tlb.h> #include <asm/unistd.h>
@@ -3066,9 +3067,13 @@ void account_user_time(struct task_struc {struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;cputime64_t tmp; +struct rq *rq =this_rq (); p->utime =cputime_add (p->utime, cputime); +if (p != rq->idle) +cpuacct_charge (p, cputime); +/* Add user time to cpustat. */ tmp =cputime_to_cputime64 (cputime);if (TASK_NICE (p) >0 )
@@ -3098,9 +3103,10 @@ void account_system_time(struct task_str cpustat->irq =cputime64_add (cpustat->irq, tmp); else if (softirq_count()) cpustat->softirq = cputime64_add(cpustat->softirq, tmp); - else if (p != rq->idle) + else if (p != rq->idle) { cpustat->system =cputime64_add (cpustat->system, tmp); - else if (atomic_read(&rq->nr_iowait) > 0) +cpuacct_charge (p, cputime); + } else if (atomic_read(&rq->nr_iowait) > 0) cpustat->iowait =cputime64_add (cpustat->iowait, tmp); else cpustat->idle = cputime64_add(cpustat->idle, tmp);
@@ -3125,8 +3131,10 @@ void account_steal_time(struct task_stru cpustat->iowait =cputime64_add (cpustat->iowait, tmp); else cpustat->idle = cputime64_add(cpustat->idle, tmp); - }else + }else { cpustat->steal =cputime64_add (cpustat->steal, tmp); +cpuacct_charge (p, -tmp); + } }static void task_running_tick (struct rq *rq,struct task_struct *p)
--
PATCH 5/7 containers (V7): Resource Groups over generic containers
In reply to [email protected]This patch provides the RG core and numtasks controller as container subsystems, intended as an example of how to implement a more complex resource control system over generic process containers. The changes to the core involve primarily removing the group management, task membership and configfs support and adding interface layers to talk to the generic container layer instead.
Each resource controller becomes an independent container subsystem; the RG core is essentially a library that the resource controllers can use to provide the RG API to userspace. Rather than a single shares and stats file in each group, there's a <controller>_shares and a <controller>_stats file, each linked to the appropriate resource controller.
include/linux/moduleparam.h | 12 - include/linux/numtasks.h | 28 ++ include/linux/res_group.h | 87 ++++++++ include/linux/res_group_rc.h | 97 ++++++++ init/Kconfig | 22 ++ kernel/Makefile | 1 kernel/fork.c | 7 kernel/res_group/Makefile | 2 kernel/res_group/local.h | 38 +++ kernel/res_group/numtasks.c | 467 +++++++++++++++++++++++++++++++++++++++++++ kernel/res_group/res_group.c | 160 ++++++++++++++ kernel/res_group/rgcs.c | 302 +++++++++++++++++++++++++++ kernel/res_group/shares.c | 228 ++++++++++++++++++++ 13 files changed, 1447 insertions(+), 4 deletions(-)
include/linux/moduleparam.h +9 -3
@@ -78,11 +78,17 @@ struct kparam_array/* Helper functions: type is byte, short, ushort, int, uint, long, ulong, charp, bool or invbool, or XXX if you define param_get_XXX, param_set_XXX and param_check_XXX. */ -#define module_param_named (name, value, type, perm) \ - param_check_##type(name, &(value)); \ - module_param_call(name, param_set_##type, param_get_##type, &value, perm); \ +#define module_param_named_call (name, value, type, set, perm) \ + param_check_##type(name, &(value)); \ + module_param_call(name, set, param_get_##type, &(value), perm); \ __MODULE_PARM_TYPE(name, #type) +#define module_param_named (name, value, type, perm) \ + module_param_named_call(name, value, type, param_set_##type, perm) + +#define module_param_set_call (name, type, setfn, perm) \ + module_param_named_call(name, name, type, setfn, perm) +#define module_param (name, type, perm) \ module_param_named(name, name, type, perm)
include/linux/numtasks.h new file +28 -0
@@ -0,0 +1,28 @@ +/* numtasks.h - No. of tasks resource controller for Resource Groups +* +* Copyright (C) Chandra Seetharaman, IBM Corp. 2003, 2004, 2005 +* +* Provides No. of tasks resource controller for Resource Groups +* +* Latest version, more details at http://ckrm.sf.net +* +* This program is free software; you can redistribute it and/or modify +* it under the terms of the GNU General Public License as published by +* the Free Software Foundation; either version 2 of the License, or +* (at your option) any later version. +* +*/ +#ifndef _LINUX_NUMTASKS_H +#define _LINUX_NUMTASKS_H + +#ifdef CONFIG_RES_GROUPS_NUMTASKS +#include <linux/res_group_rc.h> + +extern int numtasks_allow_fork (struct task_struct *); + +#else /* CONFIG_RES_GROUPS_NUMTASKS */ + +#define numtasks_allow_fork (task) (0) + +#endif /* CONFIG_RES_GROUPS_NUMTASKS */ +#endif /* _LINUX_NUMTASKS_H */
include/linux/res_group.h new file +87 -0
87 lines not on this page · raw message
include/linux/res_group_rc.h new file +97 -0
97 lines not on this page · raw message
init/Kconfig +22 -0
@@ -341,6 +341,28 @@ config TASK_IO_ACCOUNTING Say N if unsure. +menu "Resource Groups" + +config RES_GROUPS + bool "Resource Groups" + depends on EXPERIMENTAL + select CONTAINERS + help + Resource Groups is a framework for controlling and monitoring + resource allocation of user-defined groups of tasks. For more + information, please visit http://ckrm.sf.net. + +config RES_GROUPS_NUMTASKS + bool "Number of Tasks Resource Controller" + depends on RES_GROUPS + default y + help + Provides a Resource Controller for Resource Groups that allows + limiting number of tasks a resource group can have. + + Say N if unsure, Y to use the feature. + +endmenu config SYSCTL bool
kernel/Makefile +1 -0
@@ -52,6 +52,7 @@ obj-$(CONFIG_RELAY) += relay.o obj-$(CONFIG_UTS_NS) += utsname.o obj-$(CONFIG_TASK_DELAY_ACCT) += delayacct.o obj-$(CONFIG_TASKSTATS) += taskstats.o tsacct.o +obj-$(CONFIG_RES_GROUPS) += res_group/ ifneq ($(CONFIG_SCHED_NO_NO_OMIT_FRAME_POINTER),y) # According to Alan Modra <[email protected]>, the -fno-omit-frame-pointer is
kernel/fork.c +6 -1
@@ -49,6 +49,7 @@#include <linux/delayacct.h> #include <linux/taskstats_kern.h> #include <linux/random.h> +#include <linux/numtasks.h> #include <asm/pgtable.h> #include <asm/pgalloc.h>
@@ -1355,7 +1356,7 @@ long do_fork(unsigned long clone_flags, int __user *child_tidptr) {struct task_struct *p; -int trace =0 ; +int trace =0 , rc;struct pid *pid =alloc_pid ();long nr;
@@ -1368,6 +1369,10 @@ long do_fork(unsigned long clone_flags, clone_flags |=CLONE_PTRACE ; } + rc =numtasks_allow_fork (current); +if (rc) +return rc; + p =copy_process (clone_flags, stack_start, regs, stack_size, parent_tidptr, child_tidptr, nr);/* * Do this prior waking up the new thread - the thread pointer
kernel/res_group/Makefile new file +2 -0
@@ -0,0 +1,2 @@ +obj-y = res_group.o shares.o rgcs.o +obj-$(CONFIG_RES_GROUPS_NUMTASKS) += numtasks.o
kernel/res_group/local.h new file +38 -0
@@ -0,0 +1,38 @@ +/* +* Contains function definitions that are local to the Resource Groups. +* NOT to be included by controllers. +*/ + +#include <linux/res_group_rc.h> + +extern struct res_controller *get_controller_by_name (const char *); +extern struct res_controller *get_controller_by_id (unsigned int ); +extern void put_controller (struct res_controller *); +extern struct resource_group *alloc_res_group (struct resource_group *, +const char *); +extern int free_res_group (struct resource_group *); +extern void release_res_group (struct kref *); +extern int set_controller_shares (struct resource_group *, +struct res_controller *,const struct res_shares *); +/* Set shares for the given resource group and resource to default values */ +extern void set_shares_to_default (struct resource_group *, +struct res_controller *); +extern void res_group_teardown (void ); +extern int set_res_group (pid_t ,struct resource_group *); +extern void move_tasks_to_parent (struct resource_group *); + +ssize_t res_group_file_read (struct container *cont, +struct cftype *cft, +struct file *file, +char __user *buf, +size_t nbytes,loff_t *ppos); +ssize_t res_group_file_write (struct container *cont, +struct cftype *cft, +struct file *file, +const char __user *userbuf, +size_t nbytes,loff_t *ppos); + +enum { +RG_FILE_SHARES , +RG_FILE_STATS , +};
kernel/res_group/numtasks.c new file +467 -0
467 lines not on this page · raw message
kernel/res_group/res_group.c new file +160 -0
160 lines not on this page · raw message
kernel/res_group/rgcs.c new file +302 -0
302 lines not on this page · raw message
kernel/res_group/shares.c new file +228 -0
228 lines not on this page · raw message
--
PATCH 6/7 containers (V7): BeanCounters over generic process containers
In reply to [email protected]This patch implements the BeanCounter resource control abstraction over generic process containers. It contains the beancounter core code, plus the numfiles resource counter. It doesn't currently contain any of the memory tracking code or the code for switching beancounter context in interrupts.
Currently all the beancounters resource counters are lumped into a single hierarchy; ideally it would be possible for each resource counter to be a separate container subsystem, allowing them to be connected to different hierarchies.
--- fs/file_table.c | 11 + include/bc/beancounter.h | 192 ++++++++++++++++++++++++ include/bc/misc.h | 27 +++ include/linux/fs.h | 3 init/Kconfig | 4 init/main.c | 3 kernel/Makefile | 1 kernel/bc/Kconfig | 17 ++ kernel/bc/Makefile | 7 kernel/bc/beancounter.c | 371 +++++++++++++++++++++++++++++++++++++++++++++++ kernel/bc/misc.c | 56 +++++++ 11 files changed, 691 insertions(+), 1 deletion(-)
init/Kconfig +4 -0
@@ -619,6 +619,10 @@ config STOP_MACHINE Need stop_machine() primitive. endmenu +menu "Beancounters" +source "kernel/bc/Kconfig" +endmenu + menu "Block layer" source "block/Kconfig" endmenu
kernel/Makefile +1 -0
@@ -12,6 +12,7 @@ obj-y = sched.o fork.o exec_domain.o obj-$(CONFIG_STACKTRACE) += stacktrace.o obj-y += time/ +obj-$(CONFIG_BEANCOUNTERS) += bc/ obj-$(CONFIG_DEBUG_MUTEXES) += mutex-debug.o obj-$(CONFIG_LOCKDEP) += lockdep.o ifeq ($(CONFIG_PROC_FS),y)
kernel/bc/Kconfig new file +17 -0
@@ -0,0 +1,17 @@ +config BEANCOUNTERS + bool "Enable resource accounting/control" + default n + select CONTAINERS + help + When Y this option provides accounting and allows configuring + limits for user's consumption of exhaustible system resources. + The most important resource controlled by this patch is unswappable + memory (either mlock'ed or used by internal kernel structures and + buffers). The main goal of this patch is to protect processes + from running short of important resources because of accidental + misbehavior of processes or malicious activity aiming to ``kill'' + the system. It's worth mentioning that resource limits configured + by setrlimit(2) do not give an acceptable level of protection + because they cover only a small fraction of resources and work on a + per-process basis. Per-process accounting doesn't prevent malicious + users from spawning a lot of resource-consuming processes.
kernel/bc/Makefile new file +7 -0
@@ -0,0 +1,7 @@ +# +# kernel/bc/Makefile +# +# Copyright (C) 2006 OpenVZ SWsoft Inc. +# + +obj-y = beancounter.o misc.o
include/bc/beancounter.h new file +192 -0
192 lines not on this page · raw message
init/main.c +3 -0
@@ -54,6 +54,8 @@#include <linux/pid_namespace.h> #include <linux/device.h> +#include <bc/beancounter.h> +#include <asm/io.h> #include <asm/bugs.h> #include <asm/setup.h>
@@ -487,6 +489,7 @@ asmlinkage void __init start_kernel(voidextern struct kernel_param __start___param[], __stop___param[];container_init_early (); +bc_init_early ();smp_setup_processor_id ();/*
kernel/bc/beancounter.c new file +371 -0
371 lines not on this page · raw message
include/bc/misc.h new file +27 -0
@@ -0,0 +1,27 @@ +/* +* include/bc/misc.h +* +* Copyright (C) 2006 OpenVZ SWsoft Inc +* +*/ + +#ifndef __BC_MISC_H__ +#define __BC_MISC_H__ + +struct file ; + +#ifdef CONFIG_BEANCOUNTERS +int __must_check bc_file_charge (struct file *); +void bc_file_uncharge (struct file *); +#else +static inline int __must_check bc_file_charge (struct file *f) +{ +return 0 ; +} + +static inline void bc_file_uncharge (struct file *f) +{ +} +#endif + +#endif
kernel/bc/misc.c new file +56 -0
56 lines not on this page · raw message
fs/file_table.c +10 -1
@@ -22,6 +22,8 @@#include <linux/sysctl.h> #include <linux/percpu_counter.h> +#include <bc/misc.h> +#include <asm/atomic.h> /* sysctl tunables... */
@@ -43,6 +45,7 @@ static inline void file_free_rcu(structstatic inline void file_free (struct file *f) {percpu_counter_dec (&nr_files); +bc_file_uncharge (f);call_rcu (&f->f_u.fu_rcuhead, file_free_rcu); }
@@ -107,8 +110,10 @@ struct file *get_empty_filp(void)if (f ==NULL ) goto fail; -percpu_counter_inc (&nr_files);memset (f,0 ,sizeof (*f)); +if (bc_file_charge (f)) + goto fail_charge; +percpu_counter_inc (&nr_files);if (security_file_alloc (f)) goto fail_sec;
@@ -135,6 +140,10 @@ fail_sec:file_free (f); fail:return NULL ; + + fail_charge: +kmem_cache_free (filp_cachep, f); +return NULL ; }EXPORT_SYMBOL (get_empty_filp);
include/linux/fs.h +3 -0
@@ -739,6 +739,9 @@ struct file {spinlock_t f_ep_lock;#endif /* #ifdef CONFIG_EPOLL */ struct address_space *f_mapping; +#ifdef CONFIG_BEANCOUNTERS +struct beancounter *f_bc; +#endif };extern spinlock_t files_lock;#define file_list_lock () spin_lock(&files_lock);
--
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 2 lines · On Mon, Feb 12, 2007 at 12:15:27AM -0800, [email protected] wrote:
This patch implements the BeanCounter resource control abstraction over generic process containers.
Forgive my confusion, but do we really need two-levels of resource control abstraction here? Why can't resource controllers directly work with containers (just like cpu accounting does)?
Signature · 2 lines
Regards, vatsa
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 8 lines · Srivatsa Vaddagiri
Quoted 2 lines · On Mon, Feb 12, 2007 at 12:15:27AM -0800, [email protected] wrote:
This patch implements the BeanCounter resource control abstraction over generic process containers.
Forgive my confusion, but do we really need two-levels of resource control abstraction here? Why can't resource controllers directly work with containers (just like cpu accounting does)?
The generic containers patch represents a pretty low-level view of task grouping - it doesn't try to prescribe how to do accounting, nor exactly what API to present to the user (beyond providing a filesystem-based interface).
Resource controllers certainly can be written directly over it, but equally having additional abstractions to provide a common user API and kernel API for multiple resources is a reasonable goal.
I would imagine that each different resource being controlled would be represented as a container subsystem, which is how I structured the ResGroups example patch - ResGroups becomes a library that provides a common set of file manipulations for different resource controllers, each of which is a containers subsystem. The same could potentially be done for BeanCounters if people wanted.
But the main point of the latter four patches in this series is to illustrate to the various folks writing resource controller systems (and other observers) that this patch provides sufficient features to act as a base for their work. I don't presume to claim that one higher-level resource control abstraction is better than another.
Paul
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
In reply to [email protected]Quoted 5 lines · [email protected] wrote:
This patch implements the BeanCounter resource control abstraction over generic process containers. It contains the beancounter core code, plus the numfiles resource counter. It doesn't currently contain any of the memory tracking code or the code for switching beancounter context in interrupts.
Numfiles is not the most interesting place in beancounters. Kmemsize accounting is much more important actually.
Quoted 19 lines
Currently all the beancounters resource counters are lumped into a single hierarchy; ideally it would be possible for each resource counter to be a separate container subsystem, allowing them to be connected to different hierarchies.
--- fs/file_table.c | 11 + include/bc/beancounter.h | 192 ++++++++++++++++++++++++ include/bc/misc.h | 27 +++ include/linux/fs.h | 3 init/Kconfig | 4 init/main.c | 3 kernel/Makefile | 1 kernel/bc/Kconfig | 17 ++ kernel/bc/Makefile | 7 kernel/bc/beancounter.c | 371 +++++++++++++++++++++++++++++++++++++++++++++++ kernel/bc/misc.c | 56 +++++++ 11 files changed, 691 insertions(+), 1 deletion(-)
[snip]
Quoted 20 lines
Index:
container-2.6.20/kernel/bc/misc.c
===================================================================kernel/bc/misc.c new file +15
-0@@ -0,0 +1,56 @@ + +#include <linux/fs.h> +#include <bc/beancounter.h> + +int bc_file_charge (struct file *file) +{ +int sev; +struct beancounter *bc; + +task_lock (current); + bc =task_bc (current); +css_get_current (&bc->css); +task_unlock (current); + + sev = (capable (CAP_SYS_ADMIN ) ?BC_LIMIT :BC_BARRIER );
kernel/bc/misc.c line 25
Quoted 10 lines
kernel/bc/misc.c new file +10
-0+ +if (bc_charge (bc,BC_NUMFILES ,1 , sev)) { +css_put (&bc->css); +return -EMFILE ; + } + + file->f_bc = bc; +return 0 ; +} +
I have already pointed out the fact that this place will hurt performance too much. If we have some context on task this context must
1. be get-ed without any locking
2. be settable to some temporary one without locking as well
Unfortunately current containers implementation doesn't allow all of the above which blocks the rest implementation of beancounters over them.
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 9 lines · Pavel Emelianov
Quoted 5 lines · [email protected] wrote:
This patch implements the BeanCounter resource control abstraction over generic process containers. It contains the beancounter core code, plus the numfiles resource counter. It doesn't currently contain any of the memory tracking code or the code for switching beancounter context in interrupts.
Numfiles is not the most interesting place in beancounters. Kmemsize accounting is much more important actually.
Right, but the memory accouting was a much bigger and more intrusive patch than I wanted to include as an example.
Quoted 5 lines
I have already pointed out the fact that this place will hurt performance too much. If we have some context on task this context must
1. be get-ed without any locking
Would you also be happy with the restriction that a task couldn't be moved in/out of a beancounter container by any task other than itself? If so, the beancounter can_attach() method could simply return false if current != tsk, and then you'd not need to worry about locking in this situation.
Quoted 2 lines
2. be settable to some temporary one without
locking as well
I thought that we solved that problem by having a tmp_bc field in the task_struct that would take precedence over the main bc if it was non-null?
Paul
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 8 lines · Paul Menage
Quoted 7 lines · Pavel Emelianov
Quoted 5 lines · [email protected] wrote:
This patch implements the BeanCounter resource control abstraction over generic process containers. It contains the beancounter core code, plus the numfiles resource counter. It doesn't currently contain any of the memory tracking code or the code for switching beancounter context in interrupts.
Quoted 5 lines
Quoted 2 lines
Numfiles is not the most interesting place in beancounters. Kmemsize accounting is much more important actually.
Right, but the memory accouting was a much bigger and more intrusive patch than I wanted to include as an example.
I know it, but numfile doesn't show how good this
infrastructure is.
Quoted 8 lines
Quoted 5 lines
I have already pointed out the fact that this place will hurt performance too much. If we have some context on task this context must
1. be get-ed without any lockingWould you also be happy with the restriction that a task couldn't be moved in/out of a beancounter container by any task other than itself?
I have implementation that moves arbitrary task :) May be we can do context (container-on-task) handling lockless?
Quoted 3 lines
If so, the beancounter
can_attach()method could simply return false if current != tsk, and then you'd not need to worry about locking in this situation.
I may not, but this patch contains locking that is not good even for example.
Quoted 6 lines
Quoted 2 lines
2. be settable to some temporary one without
locking as wellI thought that we solved that problem by having a
tmp_bcfield in thetask_structthat would take precedence over the main bc if it was non-null?
Of course, but I'm commenting this patchset which doesn't have this facility.
Quoted 2 lines
Paul
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 2 lines · Pavel Emelianov
I have implementation that moves arbitrary task :)
Is that the one that calls stop_machine() in order to move a task around? That seemed a little heavyweight ...
Quoted 1 line
May be we can do context (container-on-task) handling lockless?
What did you have in mind?
Quoted 6 lines
Quoted 3 lines
I thought that we solved that problem by having a
tmp_bcfield in thetask_structthat would take precedence over the main bc if it was non-null?Of course, but I'm commenting this patchset which doesn't have this facility.
OK, I can add the concept in to the example too.
Paul
Re: [PATCH 6/7] containers (V7): BeanCounters over generic process containers
Quoted 6 lines · Paul Menage
Quoted 2 lines · Pavel Emelianov
I have implementation that moves arbitrary task :)
Is that the one that calls
stop_machine()in order to move a task around? That seemed a little heavyweight ...
Nope :) I've rewritten it completely.
Quoted 3 lines
Quoted 1 line
May be we can do context (container-on-task) handling lockless?
What did you have in mind?
The example patch is attached. Fits 2.6.20-rc6-mm3.
Quoted 11 lines
Quoted 6 lines
Quoted 3 lines
I thought that we solved that problem by having a
tmp_bcfield in thetask_structthat would take precedence over the main bc if it was non-null?Of course, but I'm commenting this patchset which doesn't have this facility.
OK, I can add the concept in to the example too.
Paul
kernel/bc/misc.c renamed from kernel/bc/misc.c.bcctx +63 -0
@@ -0,0 +1,63 @@ +/* +* kernel/bc/misc.c +* +* Copyright (C) 2007 OpenVZ SWsoft Inc +* +*/ + +#include <linux/sched.h> +#include <linux/stop_machine.h> +#include <linux/module.h> + +#include <bc/beancounter.h> +#include <bc/task.h> +#include <bc/misc.h> + +static DEFINE_MUTEX (task_move_mutex); + +int copy_beancounter (struct task_struct *tsk,struct task_struct *parent) +{ +struct beancounter *bc; + + bc = parent->exec_bc; + tsk->exec_bc =bc_get (bc); +BUG_ON (tsk->tmp_exec_bc !=NULL ); +return 0 ; +} + +void free_beancounter (struct task_struct *tsk) +{ +struct beancounter *bc; + +BUG_ON (tsk->tmp_exec_bc !=NULL ); + bc = tsk->exec_bc; +bc_put (bc); +} + +int bc_task_move (int pid,struct beancounter *bc) +{ +struct task_struct *tsk; +struct beancounter *old_bc; + +read_lock (&tasklist_lock); + tsk =find_task_by_pid (pid); +if (tsk) +get_task_struct (tsk); +read_unlock (&tasklist_lock); +if (tsk ==NULL ) +return -ESRCH ; + +mutex_lock (&task_move_mutex); + old_bc = tsk->exec_bc; + +bc_get (bc); +rcu_assign_pointer (tsk->exec_bc, bc); + +/* wait for all users if any get this beancounter */ +synchronize_rcu (); +mutex_unlock (&task_move_mutex); +bc_put (old_bc); + +return err; +} +EXPORT_SYMBOL (bc_task_move);
kernel/fork.c renamed from kernel/fork.c.bcctx +16 -2
@@ -51,6 +51,8 @@#include <linux/random.h> #include <linux/user_namespace.h> +#include <bc/task.h> +#include <asm/pgtable.h> #include <asm/pgalloc.h> #include <asm/uaccess.h>
@@ -105,12 +107,18 @@ struct kmem_cache *vm_area_cachep;/* SLAB cache for mm_struct structures (tsk->mm) */ static struct kmem_cache *mm_cachep; -void free_task (struct task_struct *tsk) +static void __free_task (struct task_struct *tsk) {free_thread_info (tsk->thread_info);rt_mutex_debug_task_free (tsk);free_task_struct (tsk); } + +void free_task (struct task_struct *tsk) +{ +free_beancounter (tsk); +__free_task (tsk); +}EXPORT_SYMBOL (free_task);void __put_task_struct (struct task_struct *tsk)
@@ -999,6 +1007,10 @@ static struct task_struct *copy_process(rt_mutex_init_task (p); + retval =copy_beancounter (p, current); +if (retval <0 ) + goto bad_fork_bc; +#ifdef CONFIG_TRACE_IRQFLAGS DEBUG_LOCKS_WARN_ON (!p->hardirqs_enabled);DEBUG_LOCKS_WARN_ON (!p->softirqs_enabled);
@@ -1321,7 +1333,9 @@ bad_fork_cleanup_count:atomic_dec (&p->user->processes);free_uid (p->user); bad_fork_free: -free_task (p); +free_beancounter (p); +bad_fork_bc: +__free_task (p); fork_out:return ERR_PTR (retval); }
kernel/softirq.c renamed from kernel/softirq.c.bcctx +5 -0
@@ -19,6 +19,8 @@#include <linux/smp.h> #include <linux/tick.h> +#include <bc/task.h> +#include <asm/irq.h> /* - No shared variables, all the data are CPU local.
@@ -210,6 +212,7 @@ asmlinkage void __do_softirq(void)__u32 pending;int max_restart =MAX_SOFTIRQ_RESTART ;int cpu; +struct beancounter *bc; pending =local_softirq_pending ();account_system_vtime (current);
@@ -226,6 +229,7 @@ restart: h = softirq_vec; + bc =set_exec_bc (&init_bc); do { if (pending & 1) { h->action(h);
@@ -234,6 +238,7 @@ restart: h++; pending >>=1 ; }while (pending); +reset_exec_bc (bc, &init_bc);local_irq_disable ();
include/linux/sched.h renamed from include/linux/sched.h.bcctx +4 -0
@@ -1082,6 +1082,10 @@ struct task_struct {#ifdef CONFIG_FAULT_INJECTION int make_it_fail;#endif +#ifdef CONFIG_BEANCOUNTERS +struct beancounter *exec_bc; +struct beancounter *tmp_exec_bc; +#endif };static inline pid_t process_group (struct task_struct *tsk)
include/bc/task.h renamed from include/bc/task.h.bcctx +68 -0
68 lines not on this page · raw message
PATCH 7/7 containers (V7): Container interface to nsproxy subsystem
In reply to [email protected]When a task enters a new namespace via a clone() or unshare(), a new container is created and the task moves into it. Developed by Serge Hallyn <[email protected]>, adapted by Paul Menage <[email protected]>
--- include/linux/nsproxy.h | 6 ++ init/Kconfig | 9 +++ kernel/Makefile | 1 kernel/fork.c | 4 + kernel/ns_container.c | 110 ++++++++++++++++++++++++++++++++++++++++++++++++ kernel/nsproxy.c | 6 ++ 6 files changed, 136 insertions(+)
include/linux/nsproxy.h +6 -0
@@ -53,4 +53,10 @@ static inline void exit_task_namespaces( put_nsproxy(ns); } } +#ifdef CONFIG_CONTAINER_NS +int ns_container_clone (struct task_struct *tsk); +#else +static inline int ns_container_clone (struct task_struct *tsk) {return 0 ; } +#endif +#endif
init/Kconfig +9 -0
@@ -297,6 +297,15 @@ config CONTAINER_CPUACCT Provides a simple Resource Controller for monitoring the total CPU consumed by the tasks in a container +config CONTAINER_NS + bool "Namespace container subsystem" + select CONTAINERS + help + Provides a simple namespace container subsystem to + provide hierarchical naming of sets of namespaces, + for instance virtual servers and checkpoint/restart + jobs. + config RELAY bool "Kernel->user space relay support (formerly relayfs)" help
kernel/Makefile +1 -0
@@ -39,6 +39,7 @@ obj-$(CONFIG_COMPAT) += compat.o obj-$(CONFIG_CONTAINERS) += container.o obj-$(CONFIG_CPUSETS) += cpuset.o obj-$(CONFIG_CONTAINER_CPUACCT) += cpu_acct.o +obj-$(CONFIG_CONTAINER_NS) += ns_container.o obj-$(CONFIG_IKCONFIG) += configs.o obj-$(CONFIG_STOP_MACHINE) += stop_machine.o obj-$(CONFIG_AUDIT) += audit.o auditfilter.o
kernel/fork.c +4 -0
@@ -1661,6 +1661,9 @@ asmlinkage long sys_unshare(unsigned lon err = -ENOMEM ; goto bad_unshare_cleanup_ipc; } + err =ns_container_clone (current); +if (err) + goto bad_unshare_cleanup_dupns; } if (new_fs || new_ns || new_mm || new_fd || new_ulist ||
@@ -1715,6 +1718,7 @@ asmlinkage long sys_unshare(unsigned lontask_unlock (current); } + bad_unshare_cleanup_dupns:if (new_nsproxy)put_nsproxy (new_nsproxy);
kernel/ns_container.c new file +110 -0
110 lines not on this page · raw message
kernel/nsproxy.c +6 -0
@@ -116,10 +116,16 @@ int copy_namespaces(int flags, struct taif (err) goto out_pid; + err =ns_container_clone (tsk); +if (err) + goto out_container; out:put_nsproxy (old_ns);return err; + out_container: +if (new_ns->pid_ns) +put_pid_ns (new_ns->pid_ns); out_pid:if (new_ns->ipc_ns)put_ipc_ns (new_ns->ipc_ns);
--
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
kernel/ns_container.c line 72
Quoted 11 lines · On Mon, Feb 12, 2007 at 12:15:28AM -0800, [email protected] wrote:
kernel/ns_container.c new file +11
-0+/* +* Rules: you can only create a container if +* 1. you are capable(CAP_SYS_ADMIN) +* 2. the target container is a descendant of your own container +*/ +static int ns_create (struct container_subsys *ss,struct container *cont) +{ +struct nscont *ns; + +if (!capable (CAP_SYS_ADMIN )) +return -EPERM ;
Does this check break existing namespace semantics in a subtle way? It now requires that unshare() of namespaces by any task requires CAP_SYS_ADMIN capabilities.
clone(.., CLONE_NEWUTS, ..)->copy_namespaces()->ns_container_clone()->
->container_clone()-> .. -> container_create() -> ns_create()
Earlier, one could unshare his uts namespace w/o CAP_SYS_ADMIN capabilities. Now it is required. Is that fine? Don't know.
I feel we can avoid this check totally and let the directory permissions take care of these checks.
Serge, what do you think?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 10 lines · Srivatsa Vaddagiri
kernel/ns_container.c line 72
Quoted 6 lines
kernel/ns_container.c new file +6
-0+static int ns_create (struct container_subsys *ss,struct container *cont) +{ +struct nscont *ns; + +if (!capable (CAP_SYS_ADMIN )) +return -EPERM ;Does this check break existing namespace semantics in a subtle way? It now requires that
unshare()of namespaces by any task requiresCAP_SYS_ADMINcapabilities.
I should clarify that I am referring to unshare thr' clone here (and not thr' sys_unshare)
Quoted 10 lines
clone(.., CLONE_NEWUTS, ..)->copy_namespaces()->ns_container_clone()->->container_clone()-> .. -> container_create() -> ns_create()Earlier, one could unshare his uts namespace w/o
CAP_SYS_ADMINcapabilities. Now it is required. Is that fine? Don't know.I feel we can avoid this check totally and let the directory permissions take care of these checks.
Serge, what do you think?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 12 lines · Srivatsa Vaddagiri
Quoted 10 lines · Srivatsa Vaddagiri
kernel/ns_container.c line 72
Quoted 6 lines
kernel/ns_container.c new file +6
-0+static int ns_create (struct container_subsys *ss,struct container *cont) +{ +struct nscont *ns; + +if (!capable (CAP_SYS_ADMIN )) +return -EPERM ;Does this check break existing namespace semantics in a subtle way? It now requires that
unshare()of namespaces by any task requiresCAP_SYS_ADMINcapabilities.
Quoted 2 lines
I should clarify that I am referring to unshare thr' clone here (and not thr'
sys_unshare)
That is still not true, see kernel/utsname:copy_utsname().
Now you might have run a userspace testcase in a kernel with CONFIG_UTS_NS=n, which at the moment erroneously returns 0 rather than -EINVAL when you clone(CLONE_NEWUTS). But you didn't get a new uts namespace, you were just lied to :)
-serge
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 6 lines · Serge E. Hallyn
That is still not true, see kernel/utsname:copy_utsname().
Now you might have run a userspace testcase in a kernel with
CONFIG_UTS_NS=n, which at the moment erroneously returns 0 rather than-EINVALwhen youclone(CLONE_NEWUTS). But you didn't get a new uts namespace, you were just lied to :)
I think you are right here, in that CONFIG_UTS_NS was not turned on, although I was thinking it was on.
However as a result of this experiment, I found this anomaly:
- On a kernel with CONFIG_UTS_NS=n, a test which does clone(CLONE_NEWUTS) works fine. clone() succeeds and the child starts running with no error.
- On the same kernel, if ns container hierarchy is mounted, then the test fails. clone() returns failure and child is never created. As soon as the ns container hierarchy is unmounted, the test works again.
I would have expected a consistent behavior here, irrespective of whether ns hierarchy is mounted or not. Is this difference in behavior acceptable? Returning -EINVAL in copy_utsname() when CONFIG_UTS_NS=n, as you say above, would fix this anomaly.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 21 lines · Srivatsa Vaddagiri
Quoted 6 lines · Serge E. Hallyn
That is still not true, see kernel/utsname:copy_utsname().
Now you might have run a userspace testcase in a kernel with
CONFIG_UTS_NS=n, which at the moment erroneously returns 0 rather than-EINVALwhen youclone(CLONE_NEWUTS). But you didn't get a new uts namespace, you were just lied to :)I think you are right here, in that
CONFIG_UTS_NSwas not turned on, although I was thinking it was on.However as a result of this experiment, I found this anomaly:
- On a kernel with
CONFIG_UTS_NS=n, a test which doesclone(CLONE_NEWUTS)works fine.clone()succeeds and the child starts running with no error.
- On the same kernel, if ns container hierarchy is mounted, then the test fails.clone()returns failure and child is never created. As soon as the ns container hierarchy is unmounted, the test works again.
Quoted 4 lines
I would have expected a consistent behavior here, irrespective of whether ns hierarchy is mounted or not. Is this difference in behavior acceptable? Returning
-EINVALincopy_utsname()whenCONFIG_UTS_NS=n, as you say above, would fix this anomaly.
Will, not would, fix the anomaly :)
2.6.21-rc5-mm2 has the correct behavior. Returning 0 was a bug.
thanks,
-serge
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 17 lines · Srivatsa Vaddagiri
kernel/ns_container.c line 72
Quoted 11 lines · On Mon, Feb 12, 2007 at 12:15:28AM -0800, [email protected] wrote:
kernel/ns_container.c new file +11
-0+/* +* Rules: you can only create a container if +* 1. you are capable(CAP_SYS_ADMIN) +* 2. the target container is a descendant of your own container +*/ +static int ns_create (struct container_subsys *ss,struct container *cont) +{ +struct nscont *ns; + +if (!capable (CAP_SYS_ADMIN )) +return -EPERM ;Does this check break existing namespace semantics in a subtle way? It now requires that
unshare()of namespaces by any task requiresCAP_SYS_ADMINcapabilities.
Quoted 5 lines
clone(.., CLONE_NEWUTS, ..)->copy_namespaces()->ns_container_clone()->->container_clone()-> .. -> container_create() -> ns_create()Earlier, one could unshare his uts namespace w/o
CAP_SYS_ADMINcapabilities.
Not so, CAP_SYS_ADMIN is required. Depending on your tree, check kernel/utsname.c or kernel/nsproxy.c.
Mind you, whether there is a reason to require CAP_SYS_ADMIN for utsname I'm not sure. But if there is any way of finding privileged software which would behave differently based on utsname information, then we should.
Quoted 6 lines
Now it is required. Is that fine? Don't know.
I feel we can avoid this check totally and let the directory permissions take care of these checks.
Serge, what do you think?
There is no way we can think about doing that until we determine that every kernel will have the ns container subsystem compiled in and mounted, right?
-serge
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to [email protected]kernel/ns_container.c line 27
Quoted 4 lines · On Mon, Feb 12, 2007 at 12:15:28AM -0800, [email protected] wrote:
kernel/ns_container.c new file +4
-0+int ns_container_clone (struct task_struct *tsk) +{ +return container_clone (tsk, &ns_subsys); +}
This function is a no-op if ns hierarchy is not mounted at this point. This would mean that we will miss out on some directories in ns hierarchy if it happened to be mounted later. It would be nice to recreate such missing directories upon mount. However I suspect it would not be easy ..Maybe we need to scan the task list and (re-)invoke ns_container_clone() for every new tsk->nsproxy we find in the list. Alternately perhaps we could auto mount (kern_mount) ns hierarchy very early at bootup? On the flip side that would require remount support so that additional controllers (like cpuset, mem) can be bound to (non-empty) ns hierarchy after bootup.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 16 lines · Srivatsa Vaddagiri
kernel/ns_container.c line 27
Quoted 4 lines · On Mon, Feb 12, 2007 at 12:15:28AM -0800, [email protected] wrote:
kernel/ns_container.c new file +4
-0+int ns_container_clone (struct task_struct *tsk) +{ +return container_clone (tsk, &ns_subsys); +}This function is a no-op if ns hierarchy is not mounted at this point. This would mean that we will miss out on some directories in ns hierarchy if it happened to be mounted later. It would be nice to recreate such missing directories upon mount. However I suspect it would not be easy ..Maybe we need to scan the task list and (re-)invoke
ns_container_clone()for every newtsk->nsproxywe find in the list. Alternately perhaps we could auto mount (kern_mount) ns hierarchy very early at bootup? On the flip side that would require remount support so that additional controllers (like cpuset, mem) can be bound to (non-empty) ns hierarchy after bootup.
Losing the directory isn't a big deal though. And both unsharing a namespace (which causes a ns_container_clone) and mounting the hierarchy are done by userspace, so if for some installation it is a big deal, the init scripts on initrd can mount the hierarchy before doing any unsharing.
Can you think of a reason why losing a few directories matters?
-serge
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Serge E. Hallyn
Losing the directory isn't a big deal though. And both unsharing a namespace (which causes a
ns_container_clone) and mounting the hierarchy are done by userspace, so if for some installation it is a big deal, the init scripts on initrd can mount the hierarchy before doing any unsharing.
Yes I thought of that after posting this (that initrd can mount the hierarchies), so this should not be an issue in practice ..
Quoted 1 line
Can you think of a reason why losing a few directories matters?
If we loose directories, then we don't have a way to manage the task-group it represents thr' the filesystem interface, so I consider that bad. As we agree, this will not be an issue if initrd mounts the ns hierarchy atleast at bootup.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 10 lines · Srivatsa Vaddagiri
Quoted 5 lines · Serge E. Hallyn
Losing the directory isn't a big deal though. And both unsharing a namespace (which causes a
ns_container_clone) and mounting the hierarchy are done by userspace, so if for some installation it is a big deal, the init scripts on initrd can mount the hierarchy before doing any unsharing.Yes I thought of that after posting this (that initrd can mount the hierarchies), so this should not be an issue in practice ..
Quoted 6 lines
Quoted 1 line
Can you think of a reason why losing a few directories matters?
If we loose directories, then we don't have a way to manage the task-group it represents thr' the filesystem interface, so I consider that bad. As we agree, this will not be an issue if initrd mounts the ns hierarchy atleast at bootup.
I suspect that could be a problem if we have recursive containers. Even by having a separate mount namespace for isolation you really don't want to share the mount. If you can see all of the processes you do want to be able to see and control everything.
I guess I want to ask before this gets to far. Why are all of the namespaces lumped into one group? I would think it would make much more sense to treat each namespace individually (at least for the user space interface).
Eric
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 9 lines · Eric W. Biederman
Quoted 4 lines
If we loose directories, then we don't have a way to manage the task-group it represents thr' the filesystem interface, so I consider that bad. As we agree, this will not be an issue if initrd mounts the ns hierarchy atleast at bootup.
I suspect that could be a problem if we have recursive containers. Even by having a separate mount namespace for isolation you really don't want to share the mount. If you can see all of the processes you do want to be able to see and control everything.
Won't there be some master (VFS) namespace which can see everything? The idea would be then to list all containers in that namespace. I am visualizing that a master namespace listing all containers like that will be like a management console, from which you can monitor/control resource consumption of all containers.
I agree the individual containers themselves should not be able to mount and view other containers in this container/resource-control filesystem. I presume existing VFS namespace mechanism would enforce that restriction.
Quoted 2 lines
I guess I want to ask before this gets to far. Why are all of the namespaces lumped into one group?
I don't think they are. From Serge's patches, a new group (or a directory in container filesystem) is created everytime a new nsproxy is created (copy_namespaces/sys_unshare).
Quoted 3 lines
I would think it would make much
more sense to treat each namespace individually (at least for the user space interface).
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Eric W. BiedermanQuoted 24 lines · Eric W. Biederman
Quoted 10 lines · Srivatsa Vaddagiri
Quoted 5 lines · Serge E. Hallyn
Losing the directory isn't a big deal though. And both unsharing a namespace (which causes a
ns_container_clone) and mounting the hierarchy are done by userspace, so if for some installation it is a big deal, the init scripts on initrd can mount the hierarchy before doing any unsharing.Yes I thought of that after posting this (that initrd can mount the hierarchies), so this should not be an issue in practice ..
Quoted 6 lines
Quoted 1 line
Can you think of a reason why losing a few directories matters?
If we loose directories, then we don't have a way to manage the task-group it represents thr' the filesystem interface, so I consider that bad. As we agree, this will not be an issue if initrd mounts the ns hierarchy atleast at bootup.
I suspect that could be a problem if we have recursive containers. Even by having a separate mount namespace for isolation you really don't want to share the mount. If you can see all of the processes you do want to be able to see and control everything.
Quoted 4 lines
I guess I want to ask before this gets to far. Why are all of the namespaces lumped into one group? I would think it would make much more sense to treat each namespace individually (at least for the user space interface).
You mean why is there one subsystem for all namespaces?
Convenience and no reason not to. What advantage would there be for the userspace tools if we have them split up?
There's no real reason other than memory waste not to split them up - vserver users could always just mount all namespace subsystems under one hierarchy.
But frankly I don't know where we stand right now wrt the containers patches. Do most people want to go with Vatsa's latest version moving containers into nsproxy? Has any other development been going on? Paul, have you made any updates?
-serge
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Serge E. Hallyn
But frankly I don't know where we stand right now wrt the containers patches. Do most people want to go with Vatsa's latest version moving containers into nsproxy? Has any other development been going on? Paul, have you made any updates?
I've not made major changes since the last patch post, just some small optimizations and fixes - I've been too tied up with other stuff.
Whilst I've got no objection in general to using nsproxy rather than the container_group object that I introduced in my latest patches, I think that Vatsa's approach of losing the general container object is flawed, since it loses any kind of per-group generic state (e.g. "this container is being deleted") and last time I saw it, I think it would tend to lose processes so that they didn't show up in any directory in the container fs.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 7 lines · Paul Menage
Quoted 5 lines · Serge E. Hallyn
But frankly I don't know where we stand right now wrt the containers patches. Do most people want to go with Vatsa's latest version moving containers into nsproxy? Has any other development been going on? Paul, have you made any updates?
Quoted 5 lines
I've not made major changes since the last patch post, just some small optimizations and fixes - I've been too tied up with other stuff.
Whilst I've got no objection in general to using nsproxy rather than the
container_groupobject that I introduced in my latest patches, I
Hmm, my largest objection had been that the nsproxy as a container structure would end up pointing to nsproxy as a namespace proxy.
But if we do as Eric suggests and have one subsystem per namespace type, rather than one subsystem for all namespaces, I guess that is no longer a problem.
That still leaves yours.
Quoted 12 lines
think that Vatsa's approach of losing the general container object is flawed, since it loses any kind of per-group generic state (e.g. "this container is being deleted") and last time I saw it, I think it would tend to lose processes so that they didn't show up in any directory in the container fs.
Paul
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Paul MenageQuoted 5 lines · Paul Menage
Whilst I've got no objection in general to using nsproxy rather than the
container_groupobject that I introduced in my latest patches, I think that Vatsa's approach of losing the general container object is flawed, since it loses any kind of per-group generic state (e.g. "this container is being deleted")
I think I should post the updated version what I have soon. It handles the "container/task group being deleted" case well, by setting task/ns_proxy->count = 0. The only problem I need to resolve is notify_on_release support (and also I haven't looked at the resource conters patches yet).
Quoted 3 lines
and last time I saw it, I think it would
tend to lose processes so that they didn't show up in any directory in the container fs.
I think these are fixed in the latest version I have. Will send out to you later this week (as soon as I drag myself off a higher prio task!).
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Paul MenageQuoted 11 lines · Paul Menage
Quoted 4 lines · Serge E. Hallyn
But frankly I don't know where we stand right now wrt the containers patches. Do most people want to go with Vatsa's latest version moving containers into nsproxy? Has any other development been going on? Paul, have you made any updates?
I've not made major changes since the last patch post, just some small optimizations and fixes - I've been too tied up with other stuff.
Quoted 7 lines
Whilst I've got no objection in general to using nsproxy rather than the
container_groupobject that I introduced in my latest patches, I think that Vatsa's approach of losing the general container object is flawed, since it loses any kind of per-group generic state (e.g. "this container is being deleted") and last time I saw it, I think it would tend to lose processes so that they didn't show up in any directory in the container fs.
Sounds reasonable.
Pavel is preparing new RSS patches on top of your patches which take into account other comments and Andrew objections. Can we cooperate to send it altogher then?
Thanks,
Kirill
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Paul MenageQuoted 2 lines · Paul Menage
Whilst I've got no objection in general to using nsproxy rather than the
container_groupobject that I introduced in my latest patches,
So are you saying lets (re-)use tsk->nsproxy but also retain 'struct container' to store general per-group state? If so, I think that would address my main concern of redundant/avoidable new pointers in task_struct introduced in the container patches ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 8 lines · Srivatsa Vaddagiri
Quoted 2 lines · Paul Menage
Whilst I've got no objection in general to using nsproxy rather than the
container_groupobject that I introduced in my latest patches,So are you saying lets (re-)use
tsk->nsproxybut also retain 'struct container' to store general per-group state? If so, I think that would address my main concern of redundant/avoidable new pointers intask_structintroduced in the container patches ..
I'm not saying "let's use nsproxy" - I'm not yet convinced that the lifetime/mutation/correlation rate of a pointer in an nsproxy is likely to be the same as for a container subsystem; if not, then reusing nsproxy could actually increase space overheads (since you'd end up with more, larger nsproxy objects, compared to smaller numbers of smaller nsproxy objects and smaller numbers of smaller container_group objects), even though it saved (just) one pointer per task_struct.
Reusing nsproxy versus using a separate container_group object as the aggregator is mostly an implementation detail, i think, and it would be pretty easy to switch from one to the other without any user-visible changes. So I'm inclined to keep them separate at this point.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 8 lines · Paul Menage
I'm not saying "let's use nsproxy" - I'm not yet convinced that the lifetime/mutation/correlation rate of a pointer in an nsproxy is likely to be the same as for a container subsystem; if not, then reusing nsproxy could actually increase space overheads (since you'd end up with more, larger nsproxy objects, compared to smaller numbers of smaller nsproxy objects and smaller numbers of smaller
container_groupobjects), even though it saved (just) one pointer pertask_struct.
Even if nsproxy objects are made larger a bit, the number of such object will be -much- lesser compared to number of task_structs I would think, so the win/lose in space savings would need to take that into account.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 11 lines · Srivatsa Vaddagiri
Quoted 8 lines · Paul Menage
I'm not saying "let's use nsproxy" - I'm not yet convinced that the lifetime/mutation/correlation rate of a pointer in an nsproxy is likely to be the same as for a container subsystem; if not, then reusing nsproxy could actually increase space overheads (since you'd end up with more, larger nsproxy objects, compared to smaller numbers of smaller nsproxy objects and smaller numbers of smaller
container_groupobjects), even though it saved (just) one pointer pertask_struct.Even if nsproxy objects are made larger a bit, the number of such object will
You're not making them "a bit" larger, you're adding N+M pointers where N is the number of container hierarchies and M is the number of subsystem slots.
Basically, it means that anyone that uses containers without namespaces or vice versa ends up paying the space overheads for both.
Quoted 2 lines
be -much- lesser compared to number of
task_structsI would think, so the win/lose in space savings would need to take that into account.
Agreed. So I'm not saying it's fundamentally a bad idea - just that merging container_group and nsproxy is a fairly simple space optimization that could easily be done later.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Paul Menage
Quoted 1 line
Even if nsproxy objects are made larger a bit, the number of such object will
You're not making them "a bit" larger, you're adding N+M pointers where N is the number of container hierarchies and M is the number of subsystem slots.
Hmm no .. I currently have nsproxy having just M additional pointers, where M is the maximum number of resource controllers and a single dentry pointer. As tasks get moved around in different hierarchies, new nsproxy's get created to point to proper objects in those hierarchies (which can be done with just M pointers in nsproxy). I have not considered the case of "search/reuse nsproxy that has the combination you want" just to keep it simple, but we should at some point in time do that.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 4 lines · Srivatsa Vaddagiri
Hmm no .. I currently have nsproxy having just M additional pointers, where M is the maximum number of resource controllers and a single dentry pointer.
So how do you implement something like the /proc/<PID>/container info file in my patches? (Or more generally, tell which container a task is in for a given hierarchy?)
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 2 lines · Paul Menage
So how do you implement something like the /proc/<PID>/container info file in my patches?
I havent implemented that yet, so I will look at your next question:
Quoted 2 lines
(Or more generally, tell which container a task is in for a given hierarchy?)
Why is the hierarchy bit important here? Usually controllers need to know "tell me what cpuset this task belongs to", which is answered by tsk->nsproxy->ctlr_data[CPUSET_ID].
The only usecase of storing hierarchy I have found is in container_clone() where you need to create a directory in the right hierarchy, so knowing the hierarchy to which a controller is bound becomes important.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 6 lines · Srivatsa Vaddagiri
Quoted 2 lines
(Or more generally, tell which container a task is in for a given hierarchy?)
Why is the hierarchy bit important here? Usually controllers need to know "tell me what cpuset this task belongs to", which is answered by tsk->nsproxy->ctlr_data[CPUSET_ID].
I was thinking of queries from userspace.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Paul Menage
Quoted 3 lines
Why is the hierarchy bit important here? Usually controllers need to know "tell me what cpuset this task belongs to", which is answered by tsk->nsproxy->ctlr_data[CPUSET_ID].
I was thinking of queries from userspace.
User space queries like "what is the cpuset to which this task belongs", where the answer needs to be something of the form "/dev/cpuset/C1"? The patches address that requirement atm by having a dentry pointer in struct cpuset itself.
Do you see similar queries coming in for every resource controller object (show me the path of cpu_acct, cpu_ctl, rss_ctl ... objects to which this task belongs)? IMO that will not be the case, in which case we can avoid adding N pointers (N = max hierarchies) in nsproxy just to support queries of those sort.
If additional resource objects need to support such queries in future, we could add a dentry pointer in those objects as and when the need arises.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 2 lines · vatsa
User space queries like "what is the cpuset to which this task belongs", where the answer needs to be something of the form "
/dev/cpuset/C1"?
I think that answer should be of the form "/C1", and not include the cpuset file system mount point ... though for the purposes of the present thread, this is probably not an interesting detail.
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 4 lines · Srivatsa Vaddagiri
User space queries like "what is the cpuset to which this task belongs", where the answer needs to be something of the form "
/dev/cpuset/C1"? The patches address that requirement atm by having a dentry pointer in struct cpuset itself.
Have you posted the cpuset implementation over your system yet?
The drawback to that is that every subsystem has to add a dentry to its state, and handle the processing.
Quoted 6 lines
Do you see similar queries coming in for every resource controller object (show me the path of
cpu_acct,cpu_ctl,rss_ctl... objects to which this task belongs)? IMO that will not be the case, in which case we can avoid adding N pointers (N = max hierarchies) in nsproxy just to support queries of those sort.
OK, I see your argument that putting it in the aggregator probably isn't the best thing to do from a space point of view in the case when the number of aggregators
This seems like a place where my container_subsys_state object is useful - it can store a pointer to the container object (and be maintained by the generic container system), at a space cost of 1 pointer per subsystem grouping, rather than N pointers per aggregator.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 1 line · Paul Menage
Have you posted the cpuset implementation over your system yet?
Yep, here:
http://lists.linux-foundation.org/pipermail/containers/2007-March/001497.html
For some reason, the above mail didnt make it into lkml (maybe it exceeded the max size allowed). I also have a updated version of that which I hope to post as soon as I am done with something else I am working on (sigh ..)
Quoted 2 lines
The drawback to that is that every subsystem has to add a dentry to its state, and handle the processing.
Again this depends on whether every subsystem need to be able to support the user-space query you pointed out.
Quoted 10 lines
Quoted 6 lines
Do you see similar queries coming in for every resource controller object (show me the path of
cpu_acct,cpu_ctl,rss_ctl... objects to which this task belongs)? IMO that will not be the case, in which case we can avoid adding N pointers (N = max hierarchies) in nsproxy just to support queries of
those sort.OK, I see your argument that putting it in the aggregator probably isn't the best thing to do from a space point of view in the case when the number of aggregators
Sorry that sentence seems to be garbled by some mail router :)
Did you mean to say "when the number of aggregators sharing the same container object are more" ?
I agree ..Putting N pointers in container_group object just to support queries isn't justified at this point, because we don't know whether all subsystems need to support such queries.
Quoted 4 lines
This seems like a place where my
container_subsys_stateobject is useful - it can store a pointer to the container object (and be maintained by the generic container system), at a space cost of 1 pointer per subsystem grouping, rather than N pointers per aggregator.
Yes that would be better than having N pointers in aggregator. From supporting purely user-space query pov, I think that is roughly same as having a 'dentry pointer' per resource object (what I mentioned earlier).
IMO we should add a dentry/container_subsys_state pointer only for those subsystems which need to support such queries ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 11 lines · Srivatsa Vaddagiri
Quoted 1 line · Paul Menage
Have you posted the cpuset implementation over your system yet?
Yep, here:
http://lists.linux-foundation.org/pipermail/containers/2007-March/001497.html
For some reason, the above mail didnt make it into lkml (maybe it exceeded the max size allowed). I also have a updated version of that which I hope to post as soon as I am done with something else I am working on (sigh ..)
OK, looking at that, I see a few problems related to the use of nsproxy and lack of a container object:
- your find_nsproxy() function can return an nsproxy object that's correct in its set of resource controllers but not in its other nsproxy pointers.
- rcfs_rmdir() checks the count on the dentry's nsproxy pointer. But it doesn't check for any of the other nsproxy objects that tasks in the same grouping in this hierarchy might have.
- rcfs_rmdir() sets ns->count to 0. But it doesn't stop anyone else from picking up a reference to that nsproxy from the list. Which could happen if someone else has opened the container's tasks file and is trying to write into it (but is blocked on manage_mutex). You can possibly get around this by completely freeing the namespace and setting dentry->fsdata to NULL before you release manage_mutex (and treat a NULL fsdata as a dead container).
- how do you handle additional reference counts on subsystems? E.g. beancounters wants to be able to associate each file with the container that owns it. You need to be able to lock out subsystems from taking new reference counts on an unreferenced container that you're deleting, without making the refcount operation too heavyweight.
- I think there's value in being able to mount a containerfs with no attached subsystems, simply for secure task grouping (e.g. job tracking). My latest patch set didn't support that, but it's a trivial small change to allow it. How would you do that with no container-like object?
You could have a null subsystem that does nothing other than let you attach processes to it, but then you'd be limited to one such grouping. (Since a given subsystem can only be instantiated in one hierarchy).
Quoted 6 lines
Quoted 3 lines
The drawback to that is that every subsystem has to add a dentry to its state, and handle the processing.
Again this depends on whether every subsystem need to be able to support the user-space query you pointed out.
Well, if more than one wants to support it, it means duplicating code that could equally easily be generically provided.
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Do you see similar queries coming in for every resource controller object (show me the path of
cpu_acct,cpu_ctl,rss_ctl... objects to which this task belongs)? IMO that will not be the case, in which case we can avoid adding N pointers (N = max hierarchies) in nsproxy just to support queries of
those sort.
Quoted 5 lines
Quoted 3 lines
OK, I see your argument that putting it in the aggregator probably isn't the best thing to do from a space point of view in the case when the number of aggregators
Sorry that sentence seems to be garbled by some mail router :)
Nope, it got garbled because I didn't proof-read my email sufficiently ...
Quoted 3 lines
Did you mean to say "when the number of aggregators sharing the same container object are more" ?
Yes. Although having thought about the possibility of null groupings that I described above, I'm no longer convinced that argument is valid. It would depend a lot on how people used containers in practice
- whether the number of aggregators was very small (when all subsystems are in one hierarchy, or in different hierarchies with isomorphic groupings) or very large (when all subsystems are in different hierarchies, with orthogonal groupings).
Quoted 4 lines
I agree ..Putting N pointers in
container_groupobject just to support queries isn't justified at this point, because we don't know whether all subsystems need to support such queries.
It's not just to support such queries - that's just one example. There are definitely other uses to having the container pointer directly in the aggregator, including those that I mentioned above.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 2 lines · Paul Menage
OK, looking at that, I see a few problems related to the use of nsproxy and lack of a container object:
Before we (and everyone else!) gets lost in this thread, let me say somethings upfront:
- From my understanding, task_struct is considered very pristine and any changes to it will need to have a good reason, from performance and/or memory footprint pov.
Coming from that understanding, my main intention of doing the rcfs patch was to see if it makes technical sense in reusing the nsproxy pointer in task_struct for task-group based resource management purpose.
The experience I have gained so far doing rcfs leads me to think that it is technically possible to reuse tsk->nsproxy for both task-group based resource management purpose, w/o breaking its existing user : containers (although rcfs patches that I have posted may not be the best way to exploit/reuse nsproxy pointer).
- If reusing nsproxy may not be a bad idea and is technically feasible for use in res mgmt, even if later down the lane, then IMHO it deserves some attention right at the begining, because it is so centric to task-grouping function and because we haven't merged anything yet.
That said, now onto some replies :)
Quoted 3 lines
- your
find_nsproxy()function can return an nsproxy object that's correct in its set of resource controllers but not in its other nsproxy pointers.
Very true. That's a bug and can be rectified. Atm however in my patch stack, I have dropped this whole find_nsproxy() and instead create a new nsproxy whenever tasks move ..Not the best I agree on long run.
Quoted 3 lines
-
rcfs_rmdir()checks the count on the dentry's nsproxy pointer. But it doesn't check for any of the other nsproxy objects that tasks in the same grouping in this hierarchy might have.
This is rectified in my latest stack. I do a rcfs_task_count() [same routine which is used in rcfs_tasks_open] to count tasks sharing a resource object attached to the task_proxy. Not the most optimal solution, but works and is probably ok if we consider rmdir to be infrequent operation. If we have a 'struct container' or some object to have shared state, as you mentioned in a diff thread, in addition to just reusing nsproxy, it can be made more optimal than that.
Quoted 7 lines
-
rcfs_rmdir()setsns->countto 0. But it doesn't stop anyone else from picking up a reference to that nsproxy from the list. Which could happen if someone else has opened the container's tasks file and is trying to write into it (but is blocked onmanage_mutex). You can possibly get around this by completely freeing the namespace and settingdentry->fsdatato NULL before you releasemanage_mutex(and treat a NULL fsdata as a dead container).
Isnt that handled already by the patch in rcfs_common_file_write()?
mutex_lock(&manage_mutex);
ns = __d_ns(file->f_dentry);
if (!atomic_read(&ns->count)) {
retval = -ENODEV;
goto out2;
}Quoted 6 lines
- how do you handle additional reference counts on subsystems? E.g. beancounters wants to be able to associate each file with the container that owns it. You need to be able to lock out subsystems from taking new reference counts on an unreferenced container that you're deleting, without making the refcount operation too heavyweight.
Firstly, this is not a unique problem introduced by using ->nsproxy. Secondly we have discussed this to some extent before
(http://lkml.org/lkml/2007/2/13/122). Essentially if we see zero tasks sharing a resource object pointed to by ->nsproxy, then we can't be racing with a function like bc_file_charge(), which simplifies the problem quite a bit. In other words, seeing zero tasks in xxx_rmdir() after taking manage_mutex is permission to kill nsproxy and associated objects. Correct me if I am wrong here.
Quoted 10 lines
- I think there's value in being able to mount a containerfs with no attached subsystems, simply for secure task grouping (e.g. job tracking). My latest patch set didn't support that, but it's a trivial small change to allow it. How would you do that with no container-like object?
You could have a null subsystem that does nothing other than let you attach processes to it, but then you'd be limited to one such grouping. (Since a given subsystem can only be instantiated in one hierarchy).
Again note that I am not hell-bent on avoiding a container-like object to store shared state of a group. My main desire was to avoid additional pointer in task_struct and reuse nsproxy if possible. If the grand scheme of things requires a 'struct container' object in addition to reusing ->nsproxy, to store shared state like 'notify_on_release', 'hierarchical information' then I have absolutely no objection.
Having said that, I don't know if there is practical use for what you describe above.
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
The drawback to that is that every subsystem has to add a dentry to its state, and handle the processing.
Again this depends on whether every subsystem need to be able to support the user-space query you pointed out.
Well, if more than one wants to support it, it means duplicating code that could equally easily be generically provided.
Why would it mean duplicating code? A generic function which takes a dentry pointer and returns its vfs path will avoid this code duplication?
Quoted 6 lines
Quoted 2 lines
Did you mean to say "when the number of aggregators sharing the same container object are more" ?
Yes. Although having thought about the possibility of null groupings that I described above, I'm no longer convinced that argument is valid.
I think the null grouping as defined so far is very fuzzy ..How would the kernel use this grouping, which would require reserving N pointers in 'struct container_group'/'struct nsproxy'?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 3 lines · Srivatsa Vaddagiri
I think the null grouping as defined so far is very fuzzy ..How would the kernel use this grouping, which would require reserving N pointers in 'struct container_group'/'struct nsproxy'?
aside from needing to support 'cat tasks' that is ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 10 lines · Srivatsa Vaddagiri
Quoted 2 lines · Paul Menage
OK, looking at that, I see a few problems related to the use of nsproxy and lack of a container object:
Before we (and everyone else!) gets lost in this thread, let me say somethings upfront:
- From my understanding,
task_structis considered very pristine and any changes to it will need to have a good reason, from performance and/or memory footprint pov.
Well, my latest set of patches replaced a struct cpuset pointer with a struct container_group pointer - so no net increase there.
Quoted 10 lines
Coming from that understanding, my main intention of doing the rcfs patch was to see if it makes technical sense in reusing the nsproxy pointer in
task_structfor task-group based resource management purpose.The experience I have gained so far doing rcfs leads me to think that it is technically possible to reuse
tsk->nsproxyfor both task-group based resource management purpose, w/o breaking its existing user : containers (although rcfs patches that I have posted may not be the best way to exploit/reuse nsproxy pointer).
Agreed - I think that it is technically possible, with a bunch of changes to nsproxy to support this.
Quoted 5 lines
- If reusing nsproxy may not be a bad idea and is technically feasible for use in res mgmt, even if later down the lane, then IMHO it deserves some attention right at the begining, because it is so centric to task-grouping function and because we haven't merged anything yet.
Agreed, it's worth thinking about it - but reimplementing/renaming the whole set of patches just to change container_group to nsproxy seems a bit extreme.
I'd much rather get something that works with container_group, and then work with the folks who have a lot invested in nsproxy to see about merging the two.
Quoted 10 lines
That said, now onto some replies :)
Quoted 3 lines
- your
find_nsproxy()function can return an nsproxy object that's correct in its set of resource controllers but not in its other nsproxy pointers.Very true. That's a bug and can be rectified. Atm however in my patch stack, I have dropped this whole
find_nsproxy()and instead create a new nsproxy whenever tasks move ..Not the best I agree on long run.
Or even short term, I think - although it won't hurt too much for cases where a single process enters a container and all children stay in the container, it'll hurt a lot in cases where you ever move the contents of one container into another. (Possibly in an orthogonal hierarchy).
BTW, what does rcfs do if a subsystem is registered when there are already mounted hierarchies? It looks as though it leaves the relevant pointers as NULLs.
Quoted 11 lines
Quoted 8 lines
-
rcfs_rmdir()setsns->countto 0. But it doesn't stop anyone else from picking up a reference to that nsproxy from the list. Which could happen if someone else has opened the container's tasks file and is trying to write into it (but is blocked onmanage_mutex). You can possibly get around this by completely freeing the namespace and settingdentry->fsdatato NULL before you releasemanage_mutex(and treat a NULL fsdata as a dead container).Isnt that handled already by the patch in
rcfs_common_file_write()?
Quoted 7 lines
mutex_lock(&manage_mutex);
ns = __d_ns(file->f_dentry);
if (!atomic_read(&ns->count)) {
retval = -ENODEV;
goto out2;
}You're right, I missed that, sorry.
Quoted 2 lines
Firstly, this is not a unique problem introduced by using ->nsproxy.
True - but it is one that I think needs to be addressed.
Quoted 7 lines
Secondly we have discussed this to some extent before
(http://lkml.org/lkml/2007/2/13/122). Essentially if we see zero tasks sharing a resource object pointed to by ->nsproxy, then we can't be racing with a function likebc_file_charge(), which simplifies the problem quite a bit. In other words, seeing zero tasks inxxx_rmdir()after takingmanage_mutexis permission to kill nsproxy and associated objects. Correct me if I am wrong here.
I think you're only right if you're going to overload the nsproxy count field as the total refcount, rather than the number of tasks. If you do that then you risk having to allocate way more memory than you need in any routine that tries to allocate an array with one pointer per task in the container (e.g. when reading the tasks file, or moving a task into a new cpuset). My patch separates out the refcounts from the tasks counts for exactly that reason - there's a per-subsys_state refcount which can be cheaply manipulated via the subsystem. (I have a version with lower locking requirements that I've not yet posted).
Quoted 12 lines
Quoted 11 lines
- I think there's value in being able to mount a containerfs with no attached subsystems, simply for secure task grouping (e.g. job tracking). My latest patch set didn't support that, but it's a trivial small change to allow it. How would you do that with no container-like object?
You could have a null subsystem that does nothing other than let you attach processes to it, but then you'd be limited to one such grouping. (Since a given subsystem can only be instantiated in one hierarchy).
Quoted 6 lines
Again note that I am not hell-bent on avoiding a container-like object to store shared state of a group. My main desire was to avoid additional pointer in
task_structand reuse nsproxy if possible. If the grand scheme of things requires a 'struct container' object in addition to reusing ->nsproxy, to store shared state like 'notify_on_release', 'hierarchical information' then I have absolutely no objection.
OK, but at that point you're basically back at my patches, but using nsproxy rather than container_group.
Quoted 4 lines
Why would it mean duplicating code? A generic function which takes a dentry pointer and returns its vfs path will avoid this code duplication?
It's still adding boilerplate (extra pointers, extra /proc files, logic to hook them together) to every subsystem that needs this, that could be avoided. But I guess this aspect of it isn't a huge issue.
Quoted 11 lines
Quoted 7 lines
Quoted 3 lines
Did you mean to say "when the number of aggregators sharing the same container object are more" ?
Yes. Although having thought about the possibility of null groupings that I described above, I'm no longer convinced that argument is valid.
I think the null grouping as defined so far is very fuzzy ..How would the kernel use this grouping, which would require reserving N pointers in 'struct container_group'/'struct nsproxy'?
The kernel wouldn't use it, other than to act as an inescapable task grouping whose members could always be easily listed from userspace.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Next time I have a moment I will try and take a closer look. However currently these approaches feel like there is some unholy coupling going on between different things.
In addition there appear to be some weird assumptions (an array with one member per task_struct) in the group. The pid limit allows
us millions of task_structs if the user wants it. A several megabyte
array sounds like a completely unsuitable data structure. What
is wrong with just traversing task_list at least initially?
What happened to the unix philosophy of starting with simple and correct designs?
Further we have several different questions that are all mixed up in this thread.
- What functionality do we want to provide.
- How do we want to export that functionality to user space.
You can share code by having an embedded structure instead of a magic subsystem things have to register with, and I expect that would be preferable. Libraries almost always are easier to work with then a subsystem with strict rules that doesn't give you choices.
Why do we need a subsystem id? Do we expect any controllers to be provided as modules? I think the code is so in the core that modules of any form are a questionable assumption.
.....
There is a real issue to be solved here that we can't add controls/limits for a group of processes if we don't have a user space interface for it.
Are the issues of building a user space interface so very hard, and the locking so very nasty or is this a case of over engineering?
I'm inclined to the rcfs variant and using nsproxy (from what I have seen in passing) because it is more of an exercise in minimalism, and I am strongly inclined to be minimalistic. The straight cpuset derivative seems to start with everything but the kitchen sink and then add on to it. Which at first glance seems unhealthy.
Eric
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 6 lines · Eric W. Biederman
In addition there appear to be some weird assumptions (an array with one member per
task_struct) in the group. The pid limit allows
us millions oftask_structsif the user wants it. A several megabyte
array sounds like a completely unsuitable data structure. What
is wrong with just traversingtask_listat least initially?
The current code creates such arrays when it needs an atomic snapshot of the set of tasks in the container (e.g. for reporting them to userspace or updating the mempolicies of all the tasks in the case of cpusets). It may be possible to do it by traversing tasklist and dropping the lock to operate on each task where necessary - I'll take a look at that.
Quoted 3 lines
- What functionality do we want to provide.
- How do we want to export that functionality to user space.
I'm working on a paper for OLS that sets out a lot of these issues - I'll send you a draft copy hopefully tonight.
Quoted 9 lines
You can share code by having an embedded structure instead of a magic subsystem things have to register with, and I expect that would be preferable. Libraries almost always are easier to work with then a subsystem with strict rules that doesn't give you choices.
Why do we need a subsystem id? Do we expect any controllers to be provided as modules? I think the code is so in the core that modules of any form are a questionable assumption.
One of the things that I'd really like to be able to do is allow userspace to mount one filesystem instance with several controllers/subsystems attached to it, so that they can define a single task->container mapping to be used for multiple different subsystems.
I think that would be hard to do without a central registration of subsystems.
I'm not particularly expecting modules to register as subsystems, since as you say they'd probably need to be somewhat embedded in the kernel in order to operate. And there are definite performance advantages to be had from allowing in-kernel subsystems to have a constant subsystem id.
It would be nice if subsystems could optionally have a dynamic subsystem id so that new subsystem patches didn't have to clash with one another in some global enum, but maybe that's not worth worrying about at this point.
Quoted 4 lines
I'm inclined to the rcfs variant and using nsproxy (from what I have seen in passing) because it is more of an exercise in minimalism, and I am strongly inclined to be minimalistic.
There's not a whole lot of difference between the two patch sets, other than the names, and whether they piggyback on nsproxy or on a separate object (container_group) that could be merged with nsproxy later if it seemed appropriate. Plus mine have the ability to support subsystem callbacks on fork/exit, with no significant overhead in the event that no configured subsystems actually want those callbacks.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 6 lines · Paul Menage
The current code creates such arrays when it needs an atomic snapshot of the set of tasks in the container (e.g. for reporting them to userspace or updating the mempolicies of all the tasks in the case of cpusets). It may be possible to do it by traversing tasklist and dropping the lock to operate on each task where necessary - I'll take a look at that.
Just to clarify this - the cases that currently need an array of task pointers *do* already traverse tasklist in order to locate those tasks as needed - its when they want to be able to operate on those tasks outside of the tasklist lock that the array is needed - lock tasklist_lock, fill the array with tasks (with added refcounts), drop tasklist_lock, do stuff.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Paul MenageQuoted 10 lines · Paul Menage
Quoted 3 lines
Very true. That's a bug and can be rectified. Atm however in my patch stack, I have dropped this whole
find_nsproxy()and instead create a new nsproxy whenever tasks move ..Not the best I agree on long run.Or even short term, I think - although it won't hurt too much for cases where a single process enters a container and all children stay in the container, it'll hurt a lot in cases where you ever move the contents of one container into another. (Possibly in an orthogonal hierarchy).
Quoted 3 lines
BTW, what does rcfs do if a subsystem is registered when there are already mounted hierarchies? It looks as though it leaves the relevant pointers as NULLs.
The approach I am on currently doesnt deal with dynamically loaded modules ..Partly because it allows subsystem ids to be compile-time decided and also init_nsproxy.ctlr_data[] can be initialised to default values at compile time itself.
For ex: init_nsproxy.ctlr_data[CPUSET_ID] can be set to &top_cpuset at compile time.
If there is demand later, I agree we need to address the shortcoming you point out ..
Quoted 9 lines
I think you're only right if you're going to overload the nsproxy count field as the total refcount, rather than the number of tasks. If you do that then you risk having to allocate way more memory than you need in any routine that tries to allocate an array with one pointer per task in the container (e.g. when reading the tasks file, or moving a task into a new cpuset). My patch separates out the refcounts from the tasks counts for exactly that reason - there's a per-subsys_state refcount which can be cheaply manipulated via the subsystem. (I have a version with lower locking requirements that I've not yet posted).
We have been thr' this in a diff thread. I trust we have discussed all that need to be discussed here?
Quoted 9 lines
Quoted 6 lines
Again note that I am not hell-bent on avoiding a container-like object to store shared state of a group. My main desire was to avoid additional pointer in
task_structand reuse nsproxy if possible. If the grand scheme of things requires a 'struct container' object in addition to reusing ->nsproxy, to store shared state like 'notify_on_release', 'hierarchical information' then I have absolutely no objection.OK, but at that point you're basically back at my patches, but using nsproxy rather than
container_group.
Ok ..by posting rcfs patches, I didn't mean to introduce a "yours" and "mine" rift ..honestly. In fact you would notice that they have your (sole) copyright still on them! It took me just two days to convert over the patches to use nsproxy and come up with the rcfs patches and obviously I couldnt have done that without your excellent patches to start with.
So if there is is larger interest in using nsproxy at some point, I hope they serve as some reference patches to do the conversion. I would go ahead and post what I have now (which incorporates several bug fixes) which could be used as you deem necessary at a later point if/when considering moving to nsproxy.
Quoted 7 lines
Quoted 3 lines
Why would it mean duplicating code? A generic function which takes a dentry pointer and returns its vfs path will avoid this code duplication?
It's still adding boilerplate (extra pointers, extra /proc files, logic to hook them together) to every subsystem that needs this, that could be avoided. But I guess this aspect of it isn't a huge issue.
I agree that is not a big issue now (considering there are larger things to go after!).
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
Quoted 2 lines
Did you mean to say "when the number of aggregators sharing the same container object are more" ?
Yes. Although having thought about the possibility of null groupings that I described above, I'm no longer convinced that argument is valid.
Quoted 6 lines
Quoted 3 lines
I think the null grouping as defined so far is very fuzzy ..How would the kernel use this grouping, which would require reserving N pointers in 'struct container_group'/'struct nsproxy'?
The kernel wouldn't use it, other than to act as an inescapable task grouping whose members could always be easily listed from userspace.
I am still trying to come to terms with this null groupings and how they would be used in real life.
- Can you list a real world use of it?
- If they are "inescapable" task groups, how does the first task enter such a group, using just the filesystem interface?
- If there is no real kernel use for such groups, can this be implemented in userspace (user ids, session ids etc)
- If userspace soln is not feasible, why would you want such null groupings in multiple hierarchies and not in one hierarchy? If having them in one hierarchy satisfies the requirements, then we could create empty cpusets at top level (or at an appropriate subcpuset level) and use them as null groups? By defining permissionss of the null-group cpuset directories, we could restrict movement of tasks across groups ?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Srivatsa Vaddagiri
Ok ..by posting rcfs patches, I didn't mean to introduce a "yours" and "mine" rift ..honestly. In fact you would notice that they have your (sole) copyright still on them! It took me just two days to convert over the patches to use nsproxy and come up with the rcfs patches and obviously I couldnt have done that without your excellent patches to start with.
By doing rcfs, I have been able to do a much better review of your patches than by just browsing thr' the code. I hope you would find usefull some of the feedback I have sent so far on the container patches ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 4 lines · Srivatsa Vaddagiri
The approach I am on currently doesnt deal with dynamically loaded modules ..Partly because it allows subsystem ids to be compile-time decided
Yes, that part is definitely a good idea, since it removes one of the potential performance complaints that people have compared to hard-coded pointers in a structure.
I've reworked my patches to require subsystems to be declared at compile time.
Quoted 8 lines
and also init_nsproxy.ctlr_data[] can be initialised to default values at compile time itself.
Ok ..by posting rcfs patches, I didn't mean to introduce a "yours" and "mine" rift ..honestly. In fact you would notice that they have your (sole) copyright still on them! It took me just two days to convert over the patches to use nsproxy and come up with the rcfs patches and obviously I couldnt have done that without your excellent patches to start with.
OK, sorry if I came across as possessive :-) There are definitely some great ideas in your patches, some of which I've incorporated in my patches as you'll see when I send them out later this afternoon
Quoted 5 lines
I am still trying to come to terms with this null groupings and how they would be used in real life.
- Can you list a real world use of it?
As a simple job tracking mechanism, without any other implications.
Quoted 3 lines
- If they are "inescapable" task groups, how does the first task enter such a group, using just the filesystem interface?
Root would be able to move tasks around between containers, as normal.
Quoted 3 lines
- If there is no real kernel use for such groups, can this be implemented in userspace (user ids, session ids etc)
Not cleanly. (Multiple jobs with the same user, session ids can be changed by the user). Currently in the job-control system I'm working on here, I was tagging any processes introduced in a job with a job-unique extra gid, so we could identify which job a process was in by looking at its group list. But that's a bit ugly.
In a more modern kernel we can just use cpusets without bothering to make distinctions between the memory and cpus in different cpusets, but it seems ugly to have to use a more heavyweight solution than you really need.
In practice, this would be more of a toy/example, since anyone doing job control probably is interested in at least some rudimentary kind of resource tracking/control.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 7 lines · Srivatsa Vaddagiri
Quoted 6 lines
- how do you handle additional reference counts on subsystems? E.g. beancounters wants to be able to associate each file with the container that owns it. You need to be able to lock out subsystems from taking new reference counts on an unreferenced container that you're deleting, without making the refcount operation too heavyweight.
Quoted 9 lines
Firstly, this is not a unique problem introduced by using ->nsproxy. Secondly we have discussed this to some extent before
(http://lkml.org/lkml/2007/2/13/122). Essentially if we see zero tasks sharing a resource object pointed to by ->nsproxy, then we can't be racing with a function likebc_file_charge(), which simplifies the problem quite a bit. In other words, seeing zero tasks inxxx_rmdir()after takingmanage_mutexis permission to kill nsproxy and associated objects. Correct me if I am wrong here.
OK, I've managed to reconstruct my reasoning remembered why it's important to have the refcounts associated with the subsystems, and why the simple use of the nsproxy count doesn't work. Essentially, there's no way to figure out which underlying subsystem the refcount refers to:
1) Assume the system has a single task T, and two subsystems, A and B
2) Mount hierarchy H1, with subsystem A and root subsystem state A0, and hierarchy H2 with subsystem B and root subsystem state B0. Both H1/ and H2/ share a single nsproxy N0, with refcount 3 (including the reference from T), pointing at A0 and B0.
3) Create directory H1/foo, which creates subsystem state A1 (nsproxy N1, refcount 1, pointing at A1 and B0)
4) Create directory H2/bar, which creates subsystem state B1 (nsproxy N2, refcount 1, pointing at A0 and B1)
5) Move T into H1/foo/tasks and then H2/bar/tasks. It ends up with nsproxy N3, refcount 1, pointing at A1 and B1.
6) T creates an object that is charged to A1 and hence needs to take a reference on A1 in order to uncharge it later when it's released. So N3 now has a refcount of 2
7) Move T back to H1/tasks and H2/tasks; assume it picks up nsproxy N0 again; N3 has a refcount of 1 now. (Assume that the object created in step 6 isn't one that's practical/desirable to relocate when the task that created it moves to a different container)
In this particular case the extra refcount on N3 is intended to keep A1 alive (which prevents H1/foo being deleted), but there's no way to tell from the structures in use whether it was taken on A1 or on B1. Neither H1/foo nor H2/bar can be deleted, even though nothing is intending to have a reference count on H2/bar.
Putting the extra refcount explicitly either in A1, or else in a container object associated with H1/foo makes this more obvious.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 8 lines · Paul Menage
Quoted 8 lines
Firstly, this is not a unique problem introduced by using ->nsproxy. Secondly we have discussed this to some extent before
(http://lkml.org/lkml/2007/2/13/122). Essentially if we see zero tasks sharing a resource object pointed to by ->nsproxy, then we can't be racing with a function likebc_file_charge(), which simplifies the problem quite a bit. In other words, seeing zero tasks inxxx_rmdir()after takingmanage_mutexis permission to kill nsproxy and associated objects. Correct me if I am wrong here.
Let me clarify first that I wasn't proposing an extra ref count in nsproxy to account for non-task references to a resource object pointed to by nsproxy (say nsproxy->ctlr_data[BC_ID]). Refcounts needed on beancounter because a non-task object is pointing to it (like struct file) will be put in the beancounter itself.
What I did want to say was this (sorry about the verbose rant):
mount -t container -obeancounter none /dev/bean mkdir /dev/bean/foo
echo some_pid > /dev/bean/foo
Associated with foo is a beancounter object A1 which contains (among other things) max files that can be opened by tasks in foo. Also upon successful file open, file->f_bc will point to A1.
Now lets say that someone is doing
rmdir /dev/bean/foo
while will lead us to xxx_rmdir() doing this:
mutex_lock(&manage_mutex); count = rcfs_task_count(foo's dentry);
rcfs_task_count will essentially return number of tasks pointing to A1 thr' their nsproxy->ctlr_data[BC_ID].
IF (note that /if/ again) the count returned is zero, then my point was we can destroy nsproxy behind foo and also B1, not worrying about a 'struct file' still pointing to B1. This stems from the fact that you cannot have a task's file->f_bc pointing to B1 w/o the task itself pointing to B1 also (task->nsproxy->ctlr_data[BC_ID] == B1). I also assume f_bc will get migrated with its owner task across beancounters (which seems reasonable to me atleast from 'struct file' context).
If there was indeed a file object still pointing to B1, then that can only be true if rcfs_task_count() returns non-zero value. Correct?
This is what I had in mind when I said this above : "In other words, seeing zero tasks in xxx_rmdir() after taking manage_mutex is permission to kill nsproxy and associated objects".
OT : In your posting of beancounter patches on top of containers, f_bc isnt being migrated upon task movements. Is that on intention?
Quoted 3 lines
OK, I've managed to reconstruct my reasoning remembered why it's important to have the refcounts associated with the subsystems, and why the simple use of the nsproxy count doesn't work.
I didn't mean to have non-task objects add refcounts to nsproxy. See above.
Quoted 6 lines
1) Assume the system has a single task T, and two subsystems, A and B
2) Mount hierarchy H1, with subsystem A and root subsystem state A0, and hierarchy H2 with subsystem B and root subsystem state B0. Both H1/ and H2/ share a single nsproxy N0, with refcount 3 (including the reference from T), pointing at A0 and B0.
Why refcount 3? I can only be 1 (from T) ..
Quoted 2 lines
3) Create directory H1/foo, which creates subsystem state A1 (nsproxy N1, refcount 1, pointing at A1 and B0)
right. At this point A1.count should be 1 (because N1 is pointing to it)
Quoted 2 lines
4) Create directory H2/bar, which creates subsystem state B1 (nsproxy N2, refcount 1, pointing at A0 and B1)
right. B1.count = 1 also.
Quoted 2 lines
5) Move T into H1/foo/tasks and then H2/bar/tasks. It ends up with nsproxy N3, refcount 1, pointing at A1 and B1.
right. A1.count = 2 (N1, N3) and B1.count = 2 (N2, N3)
Quoted 3 lines
6) T creates an object that is charged to A1 and hence needs to take a reference on A1 in order to uncharge it later when it's released. So N3 now has a refcount of 2
no ..N3 can continue to have 1 while A1.count becomes 3 (N1, N3 and file->f_bc)
Quoted 4 lines
7) Move T back to H1/tasks and H2/tasks; assume it picks up nsproxy N0 again; N3 has a refcount of 1 now. (Assume that the object created in step 6 isn't one that's practical/desirable to relocate when the task that created it moves to a different container)
The object was created by the task, so I would expect it should get migrated too to the new task's context (which should be true in case of f_bc atleast?). Can you give a practical example where you want to migrate the task and not the object it created?
Anyway, coming down to the impact of all this for a nsproxy based solution, I would imagine this is what will happen when T moves back to H1/tasks and H2/tasks:
- N3.count becomes zero
- We invoke free_nsproxy(N3), which drops refcounts on
all objects it is pointing to i.e
free_nsproxy()
{
if (N3->mnt_ns)
put_mnt_ns(N3->mnt_ns);
...
if (N3->ctlr_data[BC_ID])
put_bc(N3->ctlr_data[BC_ID]);
}put/get_bc() manages refcounts on beancounters. It will drop A1.count to 2 (if f_bc wasnt migrated) and not finding it zero will not destroy A1.
Essentially, in the nsproxy based approach, I am having individual controllers maintain their own refcount mechanism (just like mnt_ns or uts_ns are doing today).
Quoted 8 lines
In this particular case the extra refcount on N3 is intended to keep A1 alive (which prevents H1/foo being deleted), but there's no way to tell from the structures in use whether it was taken on A1 or on B1. Neither H1/foo nor H2/bar can be deleted, even though nothing is intending to have a reference count on H2/bar.
Putting the extra refcount explicitly either in A1, or else in a container object associated with H1/foo makes this more obvious.
Hope the above description resolves these points ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 10 lines · Srivatsa Vaddagiri
IF (note that /if/ again) the count returned is zero, then my point was we can destroy nsproxy behind foo and also B1, not worrying about a 'struct file' still pointing to B1. This stems from the fact that you cannot have a task's
file->f_bcpointing to B1 w/o the task itself pointing to B1 also (task->nsproxy->ctlr_data[BC_ID] == B1). I also assumef_bcwill get migrated with its owner task across beancounters (which seems reasonable to me atleast from 'struct file' context).If there was indeed a file object still pointing to B1, then that can only be true if
rcfs_task_count()returns non-zero value. Correct?
arghh .. s/B1/A1 pls
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 30 lines · Srivatsa Vaddagiri
Quoted 8 lines · Paul Menage
Quoted 8 lines
Firstly, this is not a unique problem introduced by using ->nsproxy. Secondly we have discussed this to some extent before
(http://lkml.org/lkml/2007/2/13/122). Essentially if we see zero tasks sharing a resource object pointed to by ->nsproxy, then we can't be racing with a function likebc_file_charge(), which simplifies the problem quite a bit. In other words, seeing zero tasks inxxx_rmdir()after takingmanage_mutexis permission to kill nsproxy and associated objects. Correct me if I am wrong here.Let me clarify first that I wasn't proposing an extra ref count in nsproxy to account for non-task references to a resource object pointed to by nsproxy (say nsproxy->ctlr_data[BC_ID]). Refcounts needed on beancounter because a non-task object is pointing to it (like struct file) will be put in the beancounter itself.
What I did want to say was this (sorry about the verbose rant):
mount -t container -obeancounter none
/dev/beanmkdir/dev/bean/foo
echosome_pid>/dev/bean/fooAssociated with foo is a beancounter object A1 which contains (among other things) max files that can be opened by tasks in foo. Also upon successful file open,
file->f_bcwill point to A1.Now lets say that someone is doing
rmdir
/dev/bean/foo
Quoted 8 lines
while will lead us to
xxx_rmdir()doing this:mutex_lock(&manage_mutex); count = rcfs_task_count(foo's dentry);
rcfs_task_countwill essentially return number of tasks pointing to A1 thr' their nsproxy->ctlr_data[BC_ID].
One small issue with the (last posted) version of your patch is that it doesn't take into account the refcounts from the directories themselves - I think you probably need to subtract one for each active subsystem.
Quoted 8 lines
IF (note that /if/ again) the count returned is zero, then my point was we can destroy nsproxy behind foo and also B1, not worrying about a 'struct file' still pointing to B1. This stems from the fact that you cannot have a task's
file->f_bcpointing to B1 w/o the task itself pointing to B1 also (task->nsproxy->ctlr_data[BC_ID] == B1). I also assumef_bcwill get migrated with its owner task across beancounters (which seems reasonable to me atleast from 'struct file' context).
I don't think that's a reasonable assumption. A task can have thousands of file handles open - having to scan and move every file that the task has open would make a move operation incredibly expensive. Additionally, tasks can share many of those file handles with other tasks. So what happens if one task that has a file open moves out of the container, but another stays behind? It's cleaner and more efficient, and conceptually desirable, IMO, just to keep the file associated with the container.
Quoted 3 lines
OT : In your posting of beancounter patches on top of containers,
f_bcisnt being migrated upon task movements. Is that on intention?
Yes.
Quoted 8 lines
Quoted 4 lines
OK, I've managed to reconstruct my reasoning remembered why it's important to have the refcounts associated with the subsystems, and why the simple use of the nsproxy count doesn't work.
I didn't mean to have non-task objects add refcounts to nsproxy. See above.
Quoted 8 lines
Quoted 6 lines
1) Assume the system has a single task T, and two subsystems, A and B
2) Mount hierarchy H1, with subsystem A and root subsystem state A0, and hierarchy H2 with subsystem B and root subsystem state B0. Both H1/ and H2/ share a single nsproxy N0, with refcount 3 (including the reference from T), pointing at A0 and B0.
Why refcount 3? I can only be 1 (from T) ..
Plus the refcounts from the two filesystem roots.
Quoted 28 lines
Quoted 3 lines
3) Create directory H1/foo, which creates subsystem state A1 (nsproxy N1, refcount 1, pointing at A1 and B0)
right. At this point A1.count should be 1 (because N1 is pointing to it)
Quoted 2 lines
4) Create directory H2/bar, which creates subsystem state B1 (nsproxy N2, refcount 1, pointing at A0 and B1)
right. B1.count = 1 also.
Quoted 2 lines
5) Move T into H1/foo/tasks and then H2/bar/tasks. It ends up with nsproxy N3, refcount 1, pointing at A1 and B1.
right. A1.count = 2 (N1, N3) and B1.count = 2 (N2, N3)
Quoted 3 lines
6) T creates an object that is charged to A1 and hence needs to take a reference on A1 in order to uncharge it later when it's released. So N3 now has a refcount of 2
no ..N3 can continue to have 1 while A1.count becomes 3 (N1, N3 and
file->f_bc)Quoted 4 lines
7) Move T back to H1/tasks and H2/tasks; assume it picks up nsproxy N0 again; N3 has a refcount of 1 now. (Assume that the object created in step 6 isn't one that's practical/desirable to relocate when the task that created it moves to a different container)
Quoted 4 lines
The object was created by the task, so I would expect it should get migrated too to the new task's context (which should be true in case of
f_bcatleast?). Can you give a practical example where you want to migrate the task and not the object it created?
I gave one above, for files; others could include pages (do you want to have to migrate every page when a task switches container? what about shared pages?)
Obviously this fundamental difference of opinion means that we're going to end up disagreeing on whether the scenario I presented is a problem or not ...
Quoted 18 lines
Anyway, coming down to the impact of all this for a nsproxy based solution, I would imagine this is what will happen when T moves back to H1/tasks and H2/tasks:
- N3.count becomes zero - We invoke free_nsproxy(N3), which drops refcounts on all objects it is pointing to i.e free_nsproxy() { if (N3->mnt_ns) put_mnt_ns(N3->mnt_ns); ... if (N3->ctlr_data[BC_ID]) put_bc(N3->ctlr_data[BC_ID]); }
Quoted 6 lines
put/get_bc() manages refcounts on beancounters. It will drop A1.count to 2 (if
f_bcwasnt migrated) and not finding it zero will not destroy A1.Essentially, in the nsproxy based approach, I am having individual controllers maintain their own refcount mechanism (just like
mnt_nsoruts_nsare doing today).
The problem with that is that (given the assumption that some subsystems might not want to migrate objects) you can then end up with a subsystem state object that has refcounts on it from active objects like files, but which is unreachable via any container filesystem mechanism. Better IMO to be able to fail the rmdir() in that situation so that the subsystem object remains accessible (so you can see where the resources are being used up).
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 6 lines · Paul Menage
Quoted 2 lines
rcfs_task_countwill essentially return number of tasks pointing to A1 thr' their nsproxy->ctlr_data[BC_ID].One small issue with the (last posted) version of your patch is that it doesn't take into account the refcounts from the directories themselves
You mean dentry->d_fsdata pointing to nsproxy should take a ref count on nsproxy? afaics it is not needed as long as you first drop the dentry before freeing associated nsproxy.
Quoted 2 lines
- I think you probably need to subtract one for each active subsystem.
I don't understand this.
Quoted 10 lines
I don't think that's a reasonable assumption. A task can have thousands of file handles open - having to scan and move every file that the task has open would make a move operation incredibly expensive.
Additionally, tasks can share many of those file handles with other tasks. So what happens if one task that has a file open moves out of the container, but another stays behind? It's cleaner and more efficient, and conceptually desirable, IMO, just to keep the file associated with the container.
I don't have a authoritative view here on whether open file count should be migrated or not, but from a layman perspective consider this:
- Task T1 is in Container C1, whose max open files can be 100
- T1 opens all of those 100 files
- T1 migrates to Container C2, but its open file count is not migrated
- T2 is migrated to container C1 and tries opening a file but is denied. T2 looks for "who is in my container who has opened all files" and doesn't find anyone.
Isn't that a bit abnormal from an end-user pov?
Quoted 3 lines
Quoted 1 line
Why refcount 3? I can only be 1 (from T) ..
Plus the refcounts from the two filesystem roots.
Filesystem root dentry's are special case. They will point to init_nsproxy which is never deleted and hence they need not add additional ref counts.
For other directories created, say H1/foo, foo's dentry will point to N1 but need not take additional refcount. N1 won't be deleted w/o dropping foo's dentry first. I think this is very similar to cpuset case, where dentry->d_fsdata = cs doesnt take additional ref counts on cpuset.
Quoted 5 lines
Quoted 4 lines
The object was created by the task, so I would expect it should get migrated too to the new task's context (which should be true in case of
f_bcatleast?). Can you give a practical example where you want to migrate the task and not the object it created?
Quoted 7 lines
I gave one above, for files; others could include pages (do you want to have to migrate every page when a task switches container? what about shared pages?)
Obviously this fundamental difference of opinion means that we're going to end up disagreeing on whether the scenario I presented is a problem or not ...
Again I am not a VM expert to say whether pages should get migrated or not. But coming to the impact of this discussion on xxx_rmdir() ..
Quoted 7 lines
The problem with that is that (given the assumption that some subsystems might not want to migrate objects) you can then end up with a subsystem state object that has refcounts on it from active objects like files, but which is unreachable via any container filesystem mechanism. Better IMO to be able to fail the
rmdir()in that situation so that the subsystem object remains accessible (so you can see where the resources are being used up).
I agree we shouldn't delete a dir going by just the task count. How abt a (optional) ->can_destroy callback which will return -EBUSY if additional non-task objects are pointing to a subsyste's resource object?
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 4 lines · Srivatsa Vaddagiri
You mean
dentry->d_fsdatapointing to nsproxy should take a ref count on nsproxy? afaics it is not needed as long as you first drop the dentry before freeing associated nsproxy.
You get the nsproxy object from dup_namespaces(), which will give you an nsproxy with a refcount of 1. So yes, your d_fsdata is already holding a refcount.
Quoted 5 lines
Quoted 3 lines
- I think you probably need to subtract one for each active subsystem.
I don't understand this.
I meant for each active hierarchy, sorry. But thinking about this further, I think you're right - since every container directory points to an nsproxy that contains its own subsystems groups for subsystems bound to that hierarchy, plus the root subsystem groups for subsystems not bound to that hierarchy, no other container directory can have a refcount on an nsproxy that matches this container directories bound subsystem groups. So subtracting 1 as you do at the moment is fine. It would be more of a problem if we were trying to rmdir the root directories, but that's not an issue.
Quoted 11 lines
I don't have a authoritative view here on whether open file count should be migrated or not, but from a layman perspective consider this:
- Task T1 is in Container C1, whose max open files can be 100
- T1 opens all of those 100 files
- T1 migrates to Container C2, but its open file count is not migrated
- T2 is migrated to container C1 and tries opening a file but is denied. T2 looks for "who is in my container who has opened all files" and doesn't find anyone.
Quoted 1 line
Isn't that a bit abnormal from an end-user pov?
Possibly. But isn't it equally abnormal if T1 opens a bunch of files, forks, and its child is moved into a different container. Then C1 has no open file count? I'm not sure that there's a universally right answer to this, so we'd need to support both behaviours.
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
Why refcount 3? I can only be 1 (from T) ..
Plus the refcounts from the two filesystem roots.
Filesystem root dentry's are special case. They will point to
init_nsproxywhich is never deleted and hence they need not add additional ref counts.
But you are taking an extra ref count - the call to get_task_namespaces(&init_task).
Quoted 2 lines
N1 won't be deleted w/o dropping
foo's dentry first.
If the container directory were to have no refcount on the nsproxy, so the initial refcount was 0, then surely moving a task in and out of the container would push the refcount up to 1; moving it away would cause the nsproxy to be freed when you call put_nsproxy(oldns) at the end of attach_task(), since that would bring the refcount back down to
0. Similarly, the task exiting would have the same effect. But that's not what's happening, since you are taking a ref count.
Quoted 2 lines
I think this is very similar to cpuset case, where
dentry->d_fsdata= cs doesnt take additional ref counts on cpuset.
That's different - the refcount on a cpuset falling to 0 doesn't free the cpuset, it just makes it eligible to be freed by someone holding the manage_mutex. the refcount on an nsproxy falling to 0 (via put_nsproxy()) does cause the nsproxy to be freed).
Quoted 3 lines
I agree we shouldn't delete a dir going by just the task count. How abt a (optional) ->can_destroy callback which will return
-EBUSYif additional non-task objects are pointing to a subsyste's resource object?
Possibly - there are two choices:
1) expose a refcount to them directly, and just interrogate the refcount from the generic code to see if it's safe to delete the directory
2) have a can_destroy() callback with well defined semantics about when it's safe to take refcounts again - it's quite possible that one subsystem can return true from can_destroy() but others don't, in which case the subsystem can become active again.
My patches solve this with an exposed refcount; to take a refcount on a subsystem object that you don't already know to be still live, you atomically bump the refcount if it's not -1; if it is -1 you wait for it either to return to 0 or for the "dead" flag to be set by the generic code. (This is all handled by inline functions so the subsystem doesn't have to worry about the complexity).
If we were to say that if at any time the refcount and the task count are both zero, then the refcount isn't allowed to increment until the task count also increments, then it would probably be possible to simplify these semantics to just check that all the refcounts were zero. This probably isn't an unreasonable restriction to make. The analog of this for option 2 would be to require that if ever the task count is 0 and the resut of can_destroy() is true, the subsystem can't get into a state where can_destroy() would return false without the task count being incremented first (by a task moving into the container).
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 9 lines · Paul Menage
Quoted 1 line
I don't understand this.
I meant for each active hierarchy, sorry. But thinking about this further, I think you're right - since every container directory points to an nsproxy that contains its own subsystems groups for subsystems bound to that hierarchy, plus the root subsystem groups for subsystems not bound to that hierarchy, no other container directory can have a refcount on an nsproxy that matches this container directories bound subsystem groups.
Yep.
Quoted 9 lines
So subtracting 1 as you do at the moment is fine. It would be more of a problem if we were trying to rmdir the root directories, but that's not an issue.
Quoted 1 line
Isn't that a bit abnormal from an end-user pov?
Possibly. But isn't it equally abnormal if T1 opens a bunch of files, forks, and its child is moved into a different container. Then C1 has no open file count?
Yep, that would be abnormal too. I think the issues are similar here wrt shared pages too? So whatever is the consensus on shared page accounting (is there one?), should apply here maybe ..
Quoted 6 lines
Quoted 3 lines
Filesystem root dentry's are special case. They will point to
init_nsproxywhich is never deleted and hence they need not add additional ref counts.But you are taking an extra ref count - the call to
get_task_namespaces(&init_task).
/me looks at his current patch stack and doesnt find that anymore.
Quoted 2 lines
If the container directory were to have no refcount on the nsproxy, so the initial refcount was 0,
No it should be 1.
mkdir H1/foo rcfs_create() ns = dup_namespaces(parent); .... dentry->d_fsdata = ns;
ns should have a refcount of 1 to begin with.
Quoted 2 lines
then surely moving a task in and out of
the container would push the refcount up to 1;
it should push the recount to 2.
Quoted 4 lines
moving it away would
cause the nsproxy to be freed when you callput_nsproxy(oldns)at the end ofattach_task(), since that would bring the refcount back down to 0.
It should bring the refcount back to 1 and hence put_nsproxy() shouldnt be called.
Quoted 2 lines
Similarly, the task exiting would have the same effect. But that's not what's happening, since you are taking a ref count.
Yes, dup_namespaces returns with a refcount of 1 (which should cover dentry reference to it).
Quoted 9 lines
Quoted 2 lines
I think this is very similar to cpuset case, where
dentry->d_fsdata= cs doesnt take additional ref counts on cpuset.That's different - the refcount on a cpuset falling to 0 doesn't free the cpuset, it just makes it eligible to be freed by someone holding the
manage_mutex. the refcount on an nsproxy falling to 0 (viaput_nsproxy()) does cause the nsproxy to be freed).From what I described above:
- refcount of a nsproxy attached to a directory dentry can never fall to zero because of tasks coming in and out. The only way for the refcount of such nsproxies to fall to zero and hence trigger their destruction is thr' the rmdir i/f.
- New nsproxies derived from the base directory nsproxy can have their's refcount go to zero as tasks exit or move around and hence they will be destroyed.
Does that sound like correct behavior?
Quoted 10 lines
Possibly - there are two choices:
1) expose a refcount to them directly, and just interrogate the refcount from the generic code to see if it's safe to delete the directory
2) have a
can_destroy()callback with well defined semantics about when it's safe to take refcounts again - it's quite possible that one subsystem can return true fromcan_destroy()but others don't, in which case the subsystem can become active again.
Lets go back to the f_bc example here for a moment. Lets say T1 was in C1 and opened file f1. f1->f_bc points to C1->beancounter.
T1 moves from C1 -> C2, but f1 is not migrated. C1->beancounter.count stays at 1 (to account for f1->f_bc).
File f1 is closed. C1->beancounter.count becomes zero.
Now user issues rmdir C1. If rmdir finds (after taking manage_mutex that is)
- zero tasks in C1
- zero refcount in C1->beancounter
why is it not safe to assume that C1->beancounter.count will continue to stay zero?
Basically I am struggling to answer "How can a zero refcount (beancounter) object go non-zero when zero tasks are attached to it" ..
If that is the case, ->can_destroy can simply return 0 if it finds ->beancounter.count == 0 else returns -EBUSY. Provided we had checked for zero tasks prior to calling ->can_destroy, it should be safe to assume no new ref counts will take place on ->beancounter.count ??
Quoted 7 lines
My patches solve this with an exposed refcount; to take a refcount on a subsystem object that you don't already know to be still live, you atomically bump the refcount if it's not -1; if it is -1 you wait for it either to return to 0 or for the "dead" flag to be set by the generic code. (This is all handled by inline functions so the subsystem doesn't have to worry about the complexity).
Quoted 10 lines
If we were to say that if at any time the refcount and the task count are both zero, then the refcount isn't allowed to increment until the task count also increments, then it would probably be possible to simplify these semantics to just check that all the refcounts were zero. This probably isn't an unreasonable restriction to make. The analog of this for option 2 would be to require that if ever the task count is 0 and the resut of
can_destroy()is true, the subsystem can't get into a state wherecan_destroy()would return false without the task count being incremented first (by a task moving into the container).
Precisely. This is what I outlined above in the example. I personally prefer ->can_destroy because it is simpler and only those subsystems that have this problem need to support it. On a hierarchy where all subsystems attached don't have a can_destroy callback, then rmdir just collapses to a "zero task count check" ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Srivatsa Vaddagiri
Lets go back to the
f_bcexample here for a moment. Lets say T1 was in C1 and opened file f1.f1->f_bcpoints toC1->beancounter.T1 moves from C1 -> C2, but f1 is not migrated.
C1->beancounter.countstays at 1 (to account forf1->f_bc).
Actually C1->beancounter.count should be at 2 (C1->beancounter and f1->f_bc are pointing to it).
Quoted 1 line
File f1 is closed.
C1->beancounter.countbecomes zero.
C1->beancounter.count should go to 1 ..
Quoted 5 lines
Now user issues rmdir C1. If rmdir finds (after taking
manage_mutexthat is)- zero tasks in C1
- zero refcount inC1->beancounter
s/zero refcount in C1->beancounter/exactly 1 refcount in C1->beancounter
Quoted 2 lines
why is it not safe to assume that
C1->beancounter.countwill continue to stay zero?
s/zero/at one
Quoted 2 lines
Basically I am struggling to answer "How can a zero refcount (beancounter) object go non-zero when zero tasks are attached to it" ..
s/zero/one and s/non-zero/>1
Essentially bc_subsys->can_attach(struct bean_counter *b) can return -EBUSY if (atomic_read(&b->count) > 1) ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Srivatsa VaddagiriQuoted 5 lines · Srivatsa Vaddagiri
Quoted 2 lines
If the container directory were to have no refcount on the nsproxy, so the initial refcount was 0,
No it should be 1.
Quoted 8 lines
mkdir H1/foo rcfs_create() ns = dup_namespaces(parent); .... dentry->d_fsdata = ns;ns should have a refcount of 1 to begin with.
Right - that's my point, you're effectively passing the initial refcount of the nsproxy to the container directory's d_fsdata reference.
Quoted 11 lines
- refcount of a nsproxy attached to a directory dentry can never fall to zero because of tasks coming in and out. The only way for the refcount of such nsproxies to fall to zero and hence trigger their destruction is thr' the rmdir i/f.
- New nsproxies derived from the base directory nsproxy can have their's refcount go to zero as tasks exit or move around and hence they will be destroyed.
Does that sound like correct behavior?
Sounds good.
Quoted 23 lines
Quoted 11 lines
Possibly - there are two choices:
1) expose a refcount to them directly, and just interrogate the refcount from the generic code to see if it's safe to delete the directory
2) have a
can_destroy()callback with well defined semantics about when it's safe to take refcounts again - it's quite possible that one subsystem can return true fromcan_destroy()but others don't, in which case the subsystem can become active again.Lets go back to the
f_bcexample here for a moment. Lets say T1 was in C1 and opened file f1.f1->f_bcpoints toC1->beancounter.T1 moves from C1 -> C2, but f1 is not migrated.
C1->beancounter.countstays at 1 (to account forf1->f_bc).File f1 is closed.
C1->beancounter.countbecomes zero.Now user issues rmdir C1. If rmdir finds (after taking
manage_mutexthat is)
Quoted 8 lines
- zero tasks in C1
- zero refcount inC1->beancounterwhy is it not safe to assume that
C1->beancounter.countwill continue to stay zero?Basically I am struggling to answer "How can a zero refcount (beancounter) object go non-zero when zero tasks are attached to it" ..
In that case, I think you're fine. Your last posted patches didn't provide a way to check for that, though, as far as I could see.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 5 lines · Paul Menage
Quoted 1 line
ns should have a refcount of 1 to begin with.
Right - that's my point, you're effectively passing the initial refcount of the nsproxy to the container directory's
d_fsdatareference.
sure ..
Quoted 5 lines
Quoted 2 lines
Basically I am struggling to answer "How can a zero refcount (beancounter) object go non-zero when zero tasks are attached to it" ..
In that case, I think you're fine. Your last posted patches didn't provide a way to check for that, though, as far as I could see.
Yes, the last patch I posted is very primitive considering all the discussions we had ..For ex: attach_task() in the last patch has more bugs [apart from the exit race I found recently with cpusets/containers] Each time I stopped and stared at attach_task() I could find a bug ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Paul MenageQuoted 3 lines · Paul Menage
Agreed. So I'm not saying it's fundamentally a bad idea - just that merging
container_groupand nsproxy is a fairly simple space optimization that could easily be done later.
IMHO, if we agree that space optimization is important, then its better we tackle it right at design phase, rather than ripping it out later ..
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
Quoted 8 lines · Srivatsa Vaddagiri
Quoted 3 lines · Paul Menage
Agreed. So I'm not saying it's fundamentally a bad idea - just that merging
container_groupand nsproxy is a fairly simple space optimization that could easily be done later.IMHO, if we agree that space optimization is important, then its better we tackle it right at design phase, rather than ripping it out later ..
"Premature optimization is the root of all evil".
I think we're in agreement that there's value in having an aggregator to avoid every task having to have a forest of pointers. But it wouldn't be a major design change to switch that aggregator from being container_group to being nsproxy at some later point.
Paul
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Serge E. HallynQuoted 5 lines · Serge E. Hallyn
But frankly I don't know where we stand right now wrt the containers patches. Do most people want to go with Vatsa's latest version moving containers into nsproxy?
Has any other development been going on?
I have another update to the rcfs patches more or less ready to go. I hope to post them out this week.
Note that from user or controller perspective, there should be very little difference between the two patches. rcfs is, I would say, same as container patches to the extent of 70-80%. The rest of the difference is mainly avoiding a new pointer in task_struct and new structures to represent a task_group.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 7/7] containers (V7): Container interface to nsproxy subsystem
In reply to Eric W. BiedermanQuoted 19 lines · Eric W. Biederman
Quoted 10 lines · Srivatsa Vaddagiri
Quoted 5 lines · Serge E. Hallyn
Losing the directory isn't a big deal though. And both unsharing a namespace (which causes a
ns_container_clone) and mounting the hierarchy are done by userspace, so if for some installation it is a big deal, the init scripts on initrd can mount the hierarchy before doing any unsharing.Yes I thought of that after posting this (that initrd can mount the hierarchies), so this should not be an issue in practice ..
Quoted 6 lines
Quoted 1 line
Can you think of a reason why losing a few directories matters?
If we loose directories, then we don't have a way to manage the task-group it represents thr' the filesystem interface, so I consider that bad. As we agree, this will not be an issue if initrd mounts the ns hierarchy atleast at bootup.
Quoted 4 lines
I suspect that could be a problem if we have recursive containers. Even by having a separate mount namespace for isolation you really don't want to share the mount. If you can see all of the processes you do want to be able to see and control everything.
Are you asking about what happens if initrd in a vserver asks to
mount -t container -o ns,cpusets /containers
?
I suppose in that case it would make sense to allow a separate mount instance with it's own superblock, with s_root pointing to the dentry for the container the vserver is in?
Quoted 6 lines
I guess I want to ask before this gets to far. Why are all of the namespaces lumped into one group? I would think it would make much more sense to treat each namespace individually (at least for the user space interface).
Eric
Quoted 1 line
- temporarily removed the "release agent" support.
ouch
Quoted 1 line
... it can be re-added ... via a kernel thread that periodically polls containers ...
double ouch.
You'll have a rough time selling me on the idea that some kernel thread should be waking up every few seconds, grabbing system-wide locks, on a big honkin NUMA box, for the few times per hour, or less, that a cpuset is abandoned.
Offhand, that sounds mildly insane to me.
And how would this get the edge-triggered, rather than level-triggered, release? In other words, if a new cpuset is created, and marked with the notify_on_release flag, but otherwise not yet used (no child cpusets and no tasks in it) then it is not to be released (removed.) Only children and/or tasks are added, then later removed, is it a candidate for release. I guess you'll need yet another state bit, set when the cpuset is abandoned (last child removed or last pid exits/leaves), and cleared when this kernel thread visits the cpuset to see if it should be removed.
Can you explain to me how this intruded on the reference counting?
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401Quoted 5 lines · Paul Jackson
You'll have a rough time selling me on the idea that some kernel thread should be waking up every few seconds, grabbing system-wide locks, on a big honkin NUMA box, for the few times per hour, or less, that a cpuset is abandoned.
I think it could be made smarter than that, e.g. have a workqueue task that's only woken when a refcount does actually reach zero. (I think that waking a workqueue task is something that can be done without too much worry about locks)
Quoted 3 lines
Can you explain to me how this intruded on the reference counting?
Essentially, it means that anything that releases a reference count on a container needs to be able to trigger a call to the release agent. The reference count is often released at a point when important locks are held, so you end up having to pass buffers into any function that might drop a ref count, in order to store a path to a release agent to be invoked.
In particular, the new container_clone() function can be called during the task fork path; at which point forking a new release_agent process would be impossible, or at least nasty. Additionally, if containers are potentially going to be used for virtual servers, having the release agent run from a top-level process rather than the process context that released the refcount sounds like a sane option.
Paul
Paul M, responding to Paul J:
Quoted 9 lines
I think it could be made smarter than that, e.g. have a workqueue task that's only woken when a refcount does actually reach zero. (I think that waking a workqueue task is something that can be done without too much worry about locks)
Quoted 3 lines
Can you explain to me how this intruded on the reference counting?
Quoted 6 lines
Essentially, it means that anything that releases a reference count on a container needs to be able to trigger a call to the release agent. The reference count is often released at a point when important locks are held, so you end up having to pass buffers into any function that might drop a ref count, in order to store a path to a release agent to be invoked.
Ok - now that you put it like that - it's much more persuasive.
Consider me sold on this aspect of your proposal, until and unless I protest otherwise, which is not likely.
Thanks.
Signature · 3 lines
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <[email protected]> 1.925.600.0401A message that is not in the archive (sent off-list or not yet received): [email protected]
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Review
Quoted 8 lines · On Mon, Feb 12, 2007 at 12:15:24AM -0800, [email protected] wrote:
(quoted diff) +8
-0+/* + * Call css_get() to hold a reference on the container; following a + * return of 0, this container subsystem state object is guaranteed + * not to be destroyed until css_put() is called on it. A non-zero + * return code indicates that a reference could not be taken. + * + */ +
Why can't we reuse container->count (or container_group->ref) to refcount the per-subsystem object attached to a container? I think that is how it is done for cpusets? That would make css_get/put unnecessary?
Quoted 16 lines
(quoted diff) +16
-0+static inline int css_get(struct container_subsys_state *css) +{ + int retval = 0; + unsigned long flags; + /* Synchronize with container_rmdir() */ + spin_lock_irqsave(&css->refcnt_lock, flags); + if (atomic_read(&css->refcnt) >= 0) { + /* Container is still alive */ + atomic_inc(&css->refcnt); + } else { + /* Container removal is in progress */ + retval = -EINVAL; + } + spin_unlock_irqrestore(&css->refcnt_lock, flags); + return retval; +}
Signature · 2 lines
Regards, vatsa
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Quoted 14 lines · Srivatsa Vaddagiri
Review
Quoted 8 lines · On Mon, Feb 12, 2007 at 12:15:24AM -0800, [email protected] wrote:
(quoted diff) +8
-0+/* + * Call css_get() to hold a reference on the container; following a + * return of 0, this container subsystem state object is guaranteed + * not to be destroyed until css_put() is called on it. A non-zero + * return code indicates that a reference could not be taken. + * + */ +Why can't we reuse
container->count(orcontainer_group->ref) to refcount the per-subsystem object attached to a container? I think that is how it is done for cpusets? That would make css_get/put unnecessary?
I did consider that approach at one point. The reason I rejected it was that then container->count would no longer even vaguely represent the number of processes in a container. Now that we have the container_group object, we have to use that for counting the number of processes in a container anyway, so that objection goes away.
However, I think it's important to be able to provide some kind of a reference count that subsystems can grab (e.g. to store a reference in a non-task object such as a file struct) without taking manage_mutex or callback_mutex (since that would be excessively heavyweight) but which can still be "frozen" at zero at the point when you're trying to destroy a container. Additionally, having it per subsystem will be important for when we implement arbitrary binding/unbinding of subsystems from hierarchies - at that point we need to be able know which subsystems have external reference counts, and hence aren't removeable.
Paul
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Quoted 5 lines · Paul Menage
I did consider that approach at one point. The reason I rejected it was that then
container->countwould no longer even vaguely represent the number of processes in a container. Now that we have thecontainer_groupobject, we have to use that for counting the number of processes in a container anyway, so that objection goes away.
Yep.
Quoted 6 lines
However, I think it's important to be able to provide some kind of a reference count that subsystems can grab (e.g. to store a reference in a non-task object such as a file struct) without taking
manage_mutexorcallback_mutex(since that would be excessively heavyweight) but which can still be "frozen" at zero at the point when you're trying to destroy a container.
Well, we already bump up reference count in fork() w/o grabbing those mutexes don't we? Also if rmdir() sees container->count to be zero, then it means no task is attached to the container. How will then a function like bc_file_charge() bump up the reference count to such a container (presuming it wanted to do so w/o manage/callback mutexes -and- that the container pointer in bc_file_charge is derived from some task in that container). I think it is safe to bump up container->count in bc_file_charge w/o grabbing manage/callback mutexes.
Quoted 5 lines
Additionally, having it per subsystem will be
important for when we implement arbitrary binding/unbinding of subsystems from hierarchies - at that point we need to be able know which subsystems have external reference counts, and hence aren't removeable.
Are you talking about (un)bind of subsystem to/from hierararchies that have non-zero containers in them? That sounds very icky. Anyway that doesnt seem to be supported in current patches.
Basically I felt we should defer introducing css_get/put until we find a good user for it, (and bc_file_(un)charge dont seem to be good users of it- see above).
Signature · 2 lines
Regards, vatsa
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Quoted 9 lines · Srivatsa Vaddagiri
Well, we already bump up reference count in
fork()w/o grabbing those mutexes don't we? Also ifrmdir()seescontainer->countto be zero, then it means no task is attached to the container. How will then a function likebc_file_charge()bump up the reference count to such a container (presuming it wanted to do so w/o manage/callback mutexes -and- that the container pointer inbc_file_chargeis derived from some task in that container). I think it is safe to bump upcontainer->countinbc_file_chargew/o grabbing manage/callback mutexes.
Right, I was never suggesting that we take either of those mutexes for this operation. The spin lock in css_get() was an attempt to avoid that. But I think you're right that it was too heavyweight, and can be avoided with atomic operations. See my other email to Pavel.
Quoted 4 lines
Are you talking about (un)bind of subsystem to/from hierararchies that have non-zero containers in them? That sounds very icky. Anyway that doesnt seem to be supported in current patches.
The bind/unbind from active hierarchies is supported in the user-space API, and it's implemented for hierarchies that have no child containers. Hence it's important, at least conceptually, for the reference count to be held by the subsystem state rather than the container.
Implementing a full bind/unbind for arbitrary subsystems and hierarchies will indeed be a lot of work, which is why I'm not trying to do it at this point.
Paul
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
In reply to a message not in the archiveQuoted 2,019 lines · Quoting [email protected] ([email protected]):
This patch removes all cpuset-specific knowlege from the container system, replacing it with a generic API that can be used by multiple subsystems. Cpusets is adapted to be a container subsystem.
Signed-off-by Paul Menage <[email protected]>--- Documentation/containers.txt | 415 +++++++++-- Documentation/cpusets.txt | 20 include/linux/container.h | 178 ++++ include/linux/cpuset.h | 16 include/linux/mempolicy.h | 12 include/linux/sched.h | 4 init/Kconfig | 12 kernel/container.c | 1601 ++++++++++++++++++++++++++++++------------- kernel/cpuset.c | 170 ++-- mm/mempolicy.c | 2 10 files changed, 1808 insertions(+), 622 deletions(-)Index:
container-2.6.20/include/linux/container.h
===================================================================include/linux/container.h +169
-9@@ -9,13 +9,12 @@*/ #include <linux/sched.h> +#include <linux/kref.h> #include <linux/cpumask.h> #include <linux/nodemask.h> #ifdef CONFIG_CONTAINERS -extern int number_of_containers;/* How many containers are defined in system? */ -extern int container_init_early (void );extern int container_init (void );extern void container_init_smp (void );@@ -30,13 +29,105 @@ extern void container_unlock(void);extern void container_manage_lock (void );extern void container_manage_unlock (void ); +struct containerfs_root ; + +/* Per-subsystem/per-container state maintained by the system. */ +struct container_subsys_state { +/* The container that this subsystem is attached to. Useful +* for subsystems that want to know about the container +* hierarchy structure */ +struct container *container; + +/* State maintained by the container system to allow +* subsystems to be "busy". Should be accessed via css_get() +* and css_put() */ +spinlock_t refcnt_lock; +atomic_t refcnt; +}; + +/* A container_group is a structure holding pointers to a set of +* containers. This saves space in the task struct object and speeds +* up fork()/exit(), since a single inc/dec can bump the reference +* count on the entire container set for a task. */ + +struct container_group { + +/* Reference count */ +struct kref ref; + +/* List running through all container groups */ +struct list_head list; + +/* Set of containers, one for each hierarchy. These are +* immutable once the container group has been created */ +struct container *container[CONFIG_MAX_CONTAINER_HIERARCHIES ]; + +/* Set of subsystem states, one for each subsystem. NULL for +* subsystems that aren't part of this hierarchy. These +* pointers reduce the number of dereferences required to get +* from a task to its state for a given container, but result +* in increased space usage if tasks are in wildly different +* groupings across different hierarchies. This array is +* mostly immutable after creation - a newly registered +* subsystem can result in a pointer in this array +* transitioning from NULL to non-NULL */ +struct container_subsys_state *subsys[CONFIG_MAX_CONTAINER_SUBSYS ]; +}; + +/* +* Call css_get() to hold a reference on the container; following a +* return of 0, this container subsystem state object is guaranteed +* not to be destroyed until css_put() is called on it. A non-zero +* return code indicates that a reference could not be taken. +* +*/ + +static inline int css_get (struct container_subsys_state *css) +{ +int retval =0 ; +unsigned long flags; +/* Synchronize with container_rmdir() */ +spin_lock_irqsave (&css->refcnt_lock, flags); +if (atomic_read (&css->refcnt) >=0 ) { +/* Container is still alive */ +atomic_inc (&css->refcnt); + }else { +/* Container removal is in progress */ + retval = -EINVAL ; + } +spin_unlock_irqrestore (&css->refcnt_lock, flags); +return retval; +} + +/* +* If you are holding current->alloc_lock then it's impossible for you +* to be moved out of your container, and hence it's impossible for +* your container to be destroyed. Therefore doing a simple +* atomic_inc() on a css is safe. +*/ + +static inline void css_get_current (struct container_subsys_state *css) +{ +atomic_inc (&css->refcnt); +} + +/* +* css_put() should be called to release a reference taken by +* css_get() or css_get_current() +*/ + +static inline void css_put (struct container_subsys_state *css) { +atomic_dec (&css->refcnt); +} +struct container {unsigned long flags;/* "unsigned long" so bitops work */ /* * Count is atomic so can incr (fork) or decr (exit) without a lock. */ -atomic_t count;/* count tasks using this container */ +atomic_t count;/* count of container groups +* using this container*/ /* * We link our 'sibling' struct into our parent's 'children'. @@ -46,11 +137,15 @@ struct container {struct list_head children;/* my children */ struct container *parent;/* my parent */ -struct dentry *dentry;/* container fs entry */ +struct dentry *dentry;/* container fs entry */ -#ifdef CONFIG_CPUSETS -struct cpuset *cpuset; -#endif +/* Private pointers for each registered subsystem */ +struct container_subsys_state *subsys[CONFIG_MAX_CONTAINER_SUBSYS ]; + +int hierarchy; + +struct containerfs_root *root; +struct container *top_container; };/* struct cftype: @@ -67,8 +162,11 @@ struct container {*/ struct inode ; +#define MAX_CFTYPE_NAME 64struct cftype { -char *name; +/* By convention, the name should begin with the name of the +* subsystem, followed by a period */ +char name[MAX_CFTYPE_NAME ];int private;int (*open) (struct inode *inode,struct file *file); ssize_t (*read) (struct container *cont, struct cftype *cft,@@ -80,10 +178,72 @@ struct cftype {int (*release) (struct inode *inode,struct file *file); }; +/* Add a new file to the given container directory. Should only be +* called by subsystems from within a populate() method */ int container_add_file (struct container *cont,const struct cftype *cft);int container_is_removed (const struct container *cont); -void container_set_release_agent_path (const char *path); + +int container_path (const struct container *cont,char *buf,int buflen); + +int container_task_count (const struct container *cont); + +/* Return true if the container is a descendant of the current container */ +int container_is_descendant (const struct container *cont); + +/* Container subsystem type. See Documentation/containers.txt for details */ + +struct container_subsys { +int (*create)(struct container_subsys *ss, +struct container *cont); +void (*destroy)(struct container_subsys *ss,struct container *cont); +int (*can_attach)(struct container_subsys *ss, +struct container *cont,struct task_struct *tsk); +void (*attach)(struct container_subsys *ss,struct container *cont, +struct container *old_cont,struct task_struct *tsk); +void (*post_attach)(struct container_subsys *ss, +struct container *cont, +struct container *old_cont, +struct task_struct *tsk); +void (*fork)(struct container_subsys *ss,struct task_struct *task); +void (*exit)(struct container_subsys *ss,struct task_struct *task); +int (*populate)(struct container_subsys *ss, +struct container *cont); +void (*bind)(struct container_subsys *ss,struct container *root); +int subsys_id; +int active; + +#define MAX_CONTAINER_TYPE_NAMELEN 32 +const char *name; + +/* Protected by RCU */ +int hierarchy; + +struct list_head sibling; +}; + +int container_register_subsys (struct container_subsys *subsys); +int container_clone (struct task_struct *tsk,struct container_subsys *ss); + +static inline struct container_subsys_state *container_subsys_state ( +struct container *cont, +struct container_subsys *ss) +{ +return cont->subsys[ss->subsys_id]; +} + +static inline struct container *task_container (struct task_struct *task, +struct container_subsys *ss) +{ +return rcu_dereference (task->containers->container[ss->hierarchy]); +} + +static inline struct container_subsys_state *task_subsys_state ( +struct task_struct *task, +struct container_subsys *ss) +{ +return rcu_dereference (task->containers->subsys[ss->subsys_id]); +}int container_path (const struct container *cont,char *buf,int buflen);Index:
container-2.6.20/include/linux/cpuset.h
===================================================================include/linux/cpuset.h +6
-10@@ -70,16 +70,7 @@ static inline int cpuset_do_slab_mem_sprextern void cpuset_track_online_nodes (void ); -extern int cpuset_can_attach_task (struct container *cont, -struct task_struct *tsk); -extern void cpuset_attach_task (struct container *cont, -struct task_struct *tsk); -extern void cpuset_post_attach_task (struct container *cont, -struct container *oldcont, -struct task_struct *tsk); -extern int cpuset_populate_dir (struct container *cont); -extern int cpuset_create (struct container *cont); -extern void cpuset_destroy (struct container *cont); +extern int current_cpuset_is_being_rebound (void );#else /* !CONFIG_CPUSETS */ @@ -147,6 +138,11 @@ static inline int cpuset_do_slab_mem_sprstatic inline void cpuset_track_online_nodes (void ) {} +static inline int current_cpuset_is_being_rebound (void ) +{ +return 0 ; +} +#endif /* !CONFIG_CPUSETS */ #endif /* _LINUX_CPUSET_H */ Index:
container-2.6.20/kernel/container.c
===================================================================kernel/container.c +917
-403@@ -55,7 +55,6 @@#include <linux/time.h> #include <linux/backing-dev.h> #include <linux/sort.h> -#include <linux/cpuset.h> #include <asm/uaccess.h> #include <asm/atomic.h> @@ -63,17 +62,56 @@#define CONTAINER_SUPER_MAGIC 0x27e0eb -/* -* Tracks how many containers are currently defined in system. -* When there is only one container (the root container) we can -* short circuit some hooks. +static struct container_subsys *subsys[CONFIG_MAX_CONTAINER_SUBSYS ]; +static int subsys_count =0 ; + +/* A containerfs_root represents the root of a container hierarchy, +* and may be associated with a superblock to form an active +* hierarchy */ +struct containerfs_root { +struct super_block *sb; + +/* The bitmask of subsystems attached to this hierarchy */ +unsigned long subsys_bits; + +/* A list running through the attached subsystems */ +struct list_head subsys_list; + +/* The root container for this hierarchy */ +struct container top_container; + +/* Tracks how many containers are currently defined in hierarchy.*/ +int number_of_containers; + +}; + +/* The set of hierarchies in use. Hierarchy 0 is the "dummy +* container", reserved for the subsystems that are otherwise +* unattached - it never has more than a single container, and all +* tasks are part of that container. */ + +static struct containerfs_root rootnode[CONFIG_MAX_CONTAINER_HIERARCHIES ]; + +/* dummytop is a shorthand for the dummy hierarchy's top container */ +#define dummytop (&rootnode[0].top_container) + +/* This flag indicates whether tasks in the fork and exit paths should +* take callback_mutex and check for fork/exit handlers to call. This +* avoids us having to take locks in the fork/exit path if none of the +* subsystems need to be called. +* +* It is protected via RCU, with the invariant that a process in an +* rcu_read_lock() section will never see this as 0 if there are +* actually registered subsystems with a fork or exit +* handler. (Sometimes it may be 1 without there being any registered +* subsystems with such a handler, but such periods are safe and of +* short duration). */ -int number_of_containers__read_mostly ; +static int need_forkexit_callback =0 ;/* bits in struct container flags field */ typedef enum {CONT_REMOVED , -CONT_NOTIFY_ON_RELEASE , }container_flagbits_t ;/* convenient tests for these bits */ @@ -82,31 +120,144 @@ inline int container_is_removed(const streturn test_bit (CONT_REMOVED , &cont->flags); } -static inline int notify_on_release (const struct container *cont) -{ -return test_bit (CONT_NOTIFY_ON_RELEASE , &cont->flags); +/* for_each_subsys() allows you to act on each subsystem attached to +* an active hierarchy */ +#define for_each_subsys (_hierarchy, _ss) \ +list_for_each_entry(_ss, &rootnode[_hierarchy].subsys_list, sibling) + +/* The default container group - used by init and its children prior +* to any hierarchies being mounted. It contains a pointer to the top +* container in each hierarchy. Also used to anchor the list of +* container groups */ +static struct container_group init_container_group; +static DEFINE_SPINLOCK (container_group_lock); +static int container_group_count; + +static void release_container_group (struct kref *k) { +struct container_group *cg = + container_of(k, struct container_group, ref); +int i; +spin_lock (&container_group_lock); +/* Release reference counts on all the containers pointed to +* by this container_group */ +for (i =0 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { +struct container *cont = cg->container[i]; +if (!cont)continue ; +atomic_dec (&cont->count); + } +list_del (&cg->list); + container_group_count--; +spin_unlock (&container_group_lock); +kfree (cg); } -static struct container top_container = { - .count =ATOMIC_INIT (0 ), - .sibling =LIST_HEAD_INIT (top_container.sibling), - .children =LIST_HEAD_INIT (top_container.children), -}; +static inline void get_container_group (struct container_group *cg) { +kref_get (&cg->ref); +} -/* The path to use for release notifications. No locking between -* setting and use - so if userspace updates this while subcontainers -* exist, you could miss a notification */ -static char release_agent_path[PATH_MAX ] ="/sbin/container_release_agent" ; +static inline void put_container_group (struct container_group *cg) { +kref_put (&cg->ref, release_container_group); +} -void container_set_release_agent_path (const char *path) -{ -container_manage_lock (); -strcpy (release_agent_path, path); -container_manage_unlock (); +/* +* find_existing_container_group() is a helper for +* find_container_group(), and checks to see whether an existing +* container_group is suitable. This currently walks a linked-list for +* simplicity; a later patch will use a hash table for better +* performance +*/ + +static struct container_group *find_existing_container_group ( +struct container_group *oldcg, +struct container *cont) +{ +int h = cont->hierarchy; +struct list_head *l = &init_container_group.list; +do { +int i; +struct container_group *cg = + list_entry(l, struct container_group, list); + +/* A container matches what we want if its container +* set is the same as "oldcg", except for the +* hierarchy for "cont" which should match "cont" */ +for (i =0 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { +if (i == h) { +if (cg->container[i] != cont) +break ; + }else { +if (cg->container[i] != oldcg->container[i]) +break ; + } + } +if (i ==CONFIG_MAX_CONTAINER_HIERARCHIES ) { +/* All hierarchies matched what we want - success */ +return cg; + } +/* Try the next container group */ + l = l->next; + }while (l != &init_container_group.list); + +/* No existing container group matched */ +return NULL ; } -static struct vfsmount *container_mount; -static struct super_block *container_sb; +/* +* find_container_group() takes an existing container group and a +* container object, and returns a container_group object that's +* equivalent to the old group, but with the given container +* substituted into the appropriate hierarchy. Must be called with +* manage_mutex held +*/ + +static struct container_group *find_container_group ( +struct container_group *oldcg,struct container *cont) +{ +struct container_group *res; +struct container_subsys *ss; +int h = cont->hierarchy; +int i; + +BUG_ON (oldcg->container[h] == cont); +/* First see if we already have a container group that matches +* the desired set */ +spin_lock (&container_group_lock); + res =find_existing_container_group (oldcg, cont); +if (res) +get_container_group (res); +spin_unlock (&container_group_lock); + +if (res) +return res; + + res =kmalloc (sizeof (*res),GFP_KERNEL ); +if (!res) +return NULL ; + +/* Copy the old container group into the new one but overwrite +* the appropriate hierarchy with the new container object and +* subsystem states and reset the reference count. */ + *res = *oldcg; +kref_init (&res->ref); + res->container[h] = cont; +for_each_subsys (h, ss) { + res->subsys[ss->subsys_id] = cont->subsys[ss->subsys_id]; + } +/* Take reference counts on all the referenced containers, +* including the new one */ +for (i =0 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { +BUG_ON (!res->container[i]); +atomic_inc (&res->container[i]->count); + } + +/* Link this container group into the list */ +spin_lock (&container_group_lock); +list_add (&res->list, &init_container_group.list); + container_group_count++; +spin_unlock (&container_group_lock); + +return res; +}/* * We have two global container mutexes below. They can nest. @@ -156,44 +307,109 @@ static struct super_block *container_sb;* small pieces of code, such as when reading out possibly multi-word * cpumasks and nodemasks. * -* The fork and exit callbacks container_fork() and container_exit(), don't -* (usually) take either mutex. These are the two most performance -* critical pieces of code here. The exception occurs on container_exit(), -* when a task in a notify_on_release container exits. Then manage_mutex -* is taken, and if the container count is zero, a usermode call made -* to /sbin/container_release_agent with the name of the container (path -* relative to the root of container file system) as the argument. -* -* A container can only be deleted if both its 'count' of using tasks -* is zero, and its list of 'children' containers is empty. Since all -* tasks in the system use _some_ container, and since there is always at -* least one task in the system (init, pid == 1), therefore, top_container -* always has either children containers and/or using tasks. So we don't +* The fork and exit callbacks container_fork() and container_exit(), +* don't take either mutex, unless some subsystem has registered a +* fork/exit callback. +* +* A container can only be deleted if all three conditions below hold: +* +* - its 'count' of using container groups is zero +* - its list of 'children' containers is empty. +* - all of its subsystems' state records have a zero 'refcnt' +* +* Since all tasks in the system use _some_ container group, and since +* there is always at least one task in the system (init, pid == 1), +* therefore, the top_container in each hierarchy always has either +* children containers and/or using container groups. So we don't * need a special hack to ensure that top_container cannot be deleted. * * The above "Tale of Two Semaphores" would be complete, but for: * * The task_lock() exception * -* The need for this exception arises from the action of attach_task(), -* which overwrites one tasks container pointer with another. It does -* so using both mutexes, however there are several performance -* critical places that need to reference task->container without the -* expense of grabbing a system global mutex. Therefore except as -* noted below, when dereferencing or, as in attach_task(), modifying -* a tasks container pointer we use task_lock(), which acts on a spinlock +* The need for this exception arises from the action of +* attach_task(), which overwrites a task's container group pointer +* with a pointer to a different group. It does so using both +* mutexes, however there are several performance critical places that +* need to reference task->containers without the expense of grabbing +* a system global mutex. Therefore except as noted below, when +* dereferencing or, as in attach_task(), modifying a task's +* containers pointer we use task_lock(), which acts on a spinlock * (task->alloc_lock) already in the task_struct routinely used for * such matters. * * P.S. One more locking exception. RCU is used to guard the -* update of a tasks container pointer by attach_task() and the +* update of a task's containers pointer by attach_task() and the * access of task->container->mems_generation via that pointer in * the routine container_update_task_memory_state(). +* +* Some container subsystems and other external code also use these +* mutexes, exposed through the container_lock()/container_unlock() +* and container_manage_lock()/container_manage_unlock() functions. +* +* E.g. the out of memory (OOM) code needs to prevent containers from +* being changed while it scans the tasklist looking for a task in an +* overlapping container. The tasklist_lock is a spinlock, so must be +* taken inside callback_mutex. +* +* Some container subsystems (including cpusets) also use +* callback_mutex as a primary lock for synchronizing access to +* subsystem state. Deciding on best practices of when to use +* fine-grained locks vs container_lock()/container_unlock() is still +* a TODO. +* +* Note that manage_mutex and callback_mutex should both nest inside +* any inode->i_mutex, unless the inode isn't accessible to any code +* outside the current thread. */ static DEFINE_MUTEX (manage_mutex);static DEFINE_MUTEX (callback_mutex); +/** +* container_lock - lock out any changes to container structures +* +*/ + +void container_lock (void ) +{ +mutex_lock (&callback_mutex); +} + +/** +* container_unlock - release lock on container changes +* +* Undo the lock taken in a previous container_lock() call. +*/ + +void container_unlock (void ) +{ +mutex_unlock (&callback_mutex); +} + +/** +* container_manage_lock() - lock out anyone else considering making +* changes to container structures. This is a more heavy-weight lock +* than the callback_mutex taken by container_lock() */ + +void container_manage_lock (void ) +{ +mutex_lock (&manage_mutex); +} + +/** +* container_manage_unlock +* +* Undo the lock taken in a previous container_manage_lock() call. +*/ + +void container_manage_unlock (void ) +{ +mutex_unlock (&manage_mutex); +} + + +/* * A couple of forward declarations required, due to cyclic reference loop: * container_mkdir -> container_create -> container_populate_dir -> container_add_file @@ -202,15 +418,18 @@ static DEFINE_MUTEX(callback_mutex);static int container_mkdir (struct inode *dir,struct dentry *dentry,int mode);static int container_rmdir (struct inode *unused_dir,struct dentry *dentry); +static int container_populate_dir (struct container *cont); +static struct inode_operations container_dir_inode_operations; +struct file_operations proc_containerstats_operations;static struct backing_dev_info container_backing_dev_info = { .ra_pages =0 ,/* No readahead */ .capabilities =BDI_CAP_NO_ACCT_DIRTY |BDI_CAP_NO_WRITEBACK , }; -static struct inode *container_new_inode (mode_t mode) +static struct inode *container_new_inode (mode_t mode,struct super_block *sb) { -struct inode *inode =new_inode (container_sb); +struct inode *inode =new_inode (sb);if (inode) { inode->i_mode = mode;@@ -238,7 +457,8 @@ static struct dentry_operations containe .d_iput = container_diput, }; -static struct dentry *container_get_dentry (struct dentry *parent,const char *name) +static struct dentry *container_get_dentry (struct dentry *parent, +const char *name) {struct dentry *d =lookup_one_len (name, parent,strlen (name));if (!IS_ERR (d))@@ -255,19 +475,19 @@ static void remove_dir(struct dentry *d)dput (parent); } -/* -* NOTE : the dentry must have been dget()'ed -*/ -static void container_d_remove_dir (struct dentry *dentry) +static void container_clear_directory (struct dentry *dentry) {struct list_head *node; - +BUG_ON (!mutex_is_locked (&dentry->d_inode->i_mutex));spin_lock (&dcache_lock); node = dentry->d_subdirs.next;while (node != &dentry->d_subdirs) { struct dentry *d = list_entry(node, struct dentry, d_u.d_child);list_del_init (node);if (d->d_inode) { +/* This should never be called on a container +* directory with child containers */ +BUG_ON (d->d_inode->i_mode &S_IFDIR ); d =dget_locked (d);spin_unlock (&dcache_lock);d_delete (d);@@ -277,37 +497,222 @@ static void container_d_remove_dir(struc } node = dentry->d_subdirs.next; } +spin_unlock (&dcache_lock); +} + +/* +* NOTE : the dentry must have been dget()'ed +*/ +static void container_d_remove_dir (struct dentry *dentry) +{ +container_clear_directory (dentry); + +spin_lock (&dcache_lock);list_del_init (&dentry->d_u.d_child);spin_unlock (&dcache_lock);remove_dir (dentry); } +static int rebind_subsystems (struct containerfs_root *root, +unsigned long final_bits) +{ +unsigned long added_bits, removed_bits; +struct container *cont = &root->top_container; +int i; +int hierarchy = cont->hierarchy; + + removed_bits = root->subsys_bits & ~final_bits; + added_bits = final_bits & ~root->subsys_bits; +/* Check that any added subsystems are currently free */ +for (i =0 ; i < subsys_count; i++) { +unsigned long long bit =1ull << i; +struct container_subsys *ss = subsys[i]; +if (!(bit & added_bits)) +continue ; +if (ss->hierarchy !=0 ) { +/* Subsystem isn't free */ +return -EBUSY ; + } + } + +/* Currently we don't handle adding/removing subsystems when +* any subcontainers exist. This is theoretically supportable +* but involves complex erro r handling, so it's being left until +* later */ +if (!list_empty (&cont->children)) { +return -EBUSY ; + } + +mutex_lock (&callback_mutex); +/* Process each subsystem */ +for (i =0 ; i < subsys_count; i++) { +struct container_subsys *ss = subsys[i]; +unsigned long bit =1UL << i; +if (bit & added_bits) { +/* We're binding this subsystem to this hierarchy */ +BUG_ON (cont->subsys[i]); +BUG_ON (dummytop->subsys[i]->container != dummytop); + cont->subsys[i] = dummytop->subsys[i]; + cont->subsys[i]->container = cont; +list_add (&ss->sibling, &root->subsys_list); +rcu_assign_pointer (ss->hierarchy, hierarchy); +if (ss->bind) + ss->bind (ss, cont); + + }else if (bit & removed_bits) { +/* We're removing this subsystem */ +BUG_ON (cont->subsys[i] != dummytop->subsys[i]); +BUG_ON (cont->subsys[i]->container != cont); +if (ss->bind) + ss->bind (ss, dummytop); + dummytop->subsys[i]->container = dummytop; + cont->subsys[i] =NULL ; +rcu_assign_pointer (subsys[i]->hierarchy,0 ); +list_del (&ss->sibling); + }else if (bit & final_bits) { +/* Subsystem state should already exist */ +BUG_ON (!cont->subsys[i]); + }else { +/* Subsystem state shouldn't exist */ +BUG_ON (cont->subsys[i]); + } + } + root->subsys_bits = final_bits; +mutex_unlock (&callback_mutex); +synchronize_rcu (); + +return 0 ; +} + +/* +* Release the last use of a hierarchy. Will never be called when +* there are active subcontainers since each subcontainer bumps the +* value of sb->s_active. +*/ + +static void container_put_super (struct super_block *sb) { + +struct containerfs_root *root = sb->s_fs_info; +struct container *cont = &root->top_container; +int ret; + + root->sb =NULL ; + sb->s_fs_info =NULL ; + +mutex_lock (&manage_mutex); + +BUG_ON (root->number_of_containers !=1 ); +BUG_ON (!list_empty (&cont->children)); +BUG_ON (!list_empty (&cont->sibling)); +BUG_ON (!root->subsys_bits); + +/* Rebind all subsystems back to the default hierarchy */ + ret =rebind_subsystems (root,0 ); +BUG_ON (ret); + +mutex_unlock (&manage_mutex); +} + +static int container_show_options (struct seq_file *seq,struct vfsmount *vfs) +{ +struct containerfs_root *root = vfs->mnt_sb->s_fs_info; +struct container_subsys *ss; +for_each_subsys (root->top_container.hierarchy, ss) { +seq_printf (seq,",%s" , ss->name); + } +return 0 ; +} + +/* Convert a hierarchy specifier into a bitmask. LL=manage_mutex */ +static int parse_containerfs_options (char *opts,unsigned long *bits) +{ +char *token, *o = opts ?:"all" ; + + *bits =0 ; + +while ((token =strsep (&o,"," )) !=NULL ) { +if (!*token) +return -EINVAL ; +if (!strcmp (token,"all" )) { + *bits = (1 << subsys_count) -1 ; + }else { +struct container_subsys *ss; +int i; +for (i =0 ; i < subsys_count; i++) { + ss = subsys[i]; +if (!strcmp (token, ss->name)) { + *bits |=1 << i; +break ; + } + } +if (i == subsys_count) +return -ENOENT ; + } + } + +/* We can't have an empty hierarchy */ +if (!*bits) +return -EINVAL ; + +return 0 ; +} + +static int container_remount (struct super_block *sb,int *flags,char *data) +{ +int ret =0 ; +unsigned long subsys_bits; +struct containerfs_root *root = sb->s_fs_info; +struct container *cont = &root->top_container; + +mutex_lock (&cont->dentry->d_inode->i_mutex); +mutex_lock (&manage_mutex); + +/* See what subsystems are wanted */ + ret =parse_containerfs_options (data, &subsys_bits); +if (ret) + goto out_unlock; + + ret =rebind_subsystems (root, subsys_bits); + +/* (re)populate subsystem files */ +if (!ret) +container_populate_dir (cont); + + out_unlock: +mutex_unlock (&manage_mutex); +mutex_unlock (&cont->dentry->d_inode->i_mutex); +return ret; +} +static struct super_operations container_ops = { .statfs = simple_statfs, .drop_inode = generic_delete_inode, + .put_super = container_put_super, + .show_options = container_show_options, + .remount_fs = container_remount, }; -static int container_fill_super (struct super_block *sb,void *unused_data, -int unused_silent) +static int container_fill_super (struct super_block *sb,void *options, +int unused_silent) {struct inode *inode;struct dentry *root; +struct containerfs_root *hroot = options; sb->s_blocksize =PAGE_CACHE_SIZE ; sb->s_blocksize_bits =PAGE_CACHE_SHIFT ; sb->s_magic =CONTAINER_SUPER_MAGIC ; sb->s_op = &container_ops; - container_sb = sb; - inode =container_new_inode (S_IFDIR |S_IRUGO |S_IXUGO |S_IWUSR ); -if (inode) { - inode->i_op = &simple_dir_inode_operations; - inode->i_fop = &simple_dir_operations; -/* directories start off with i_nlink == 2 (for "." entry) */ - inode->i_nlink++; - }else { + inode =container_new_inode (S_IFDIR |S_IRUGO |S_IXUGO |S_IWUSR , sb); +if (!inode)return -ENOMEM ; - } + + inode->i_op = &simple_dir_inode_operations; + inode->i_fop = &simple_dir_operations; + inode->i_op = &container_dir_inode_operations; +/* directories start off with i_nlink == 2 (for "." entry) */ +inc_nlink (inode); root =d_alloc_root (inode);if (!root) {@@ -315,6 +720,12 @@ static int container_fill_super(struct sreturn -ENOMEM ; } sb->s_root = root; + root->d_fsdata = &hroot->top_container; + hroot->top_container.dentry = root; + + sb->s_fs_info = hroot; + hroot->sb = sb; +return 0 ; }@@ -322,7 +733,82 @@ static int container_get_sb(struct file_ int flags, const char *unused_dev_name, void *data, struct vfsmount *mnt) { - return get_sb_single(fs_type, flags, data, container_fill_super, mnt); + int i; +unsigned long subsys_bits =0 ; +int ret =0 ; +struct containerfs_root *root =NULL ; +int hierarchy; + +mutex_lock (&manage_mutex); + +/* First find the desired set of resource controllers */ + ret =parse_containerfs_options (data, &subsys_bits); +if (ret) + goto out_unlock; + +/* See if we already have a hierarchy containing this set */ + +for (i =1 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { + root = &rootnode[i]; +/* We match - use this hieracrchy */ +if (root->subsys_bits == subsys_bits)break ; +/* We clash - fail */ +if (root->subsys_bits & subsys_bits) { + ret = -EBUSY ; + goto out_unlock; + } + } + +if (i ==CONFIG_MAX_CONTAINER_HIERARCHIES ) { +/* No existing hierarchy matched this set - but we +* know that all the subsystems are free */ +for (i =1 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { + root = &rootnode[i]; +if (!root->sb && !root->subsys_bits)break ; + } + } + +if (i ==CONFIG_MAX_CONTAINER_HIERARCHIES ) { + ret = -ENOSPC ; + goto out_unlock; + } + + hierarchy = i; + +if (!root->sb) { +/* We need a new superblock for this container combination */ +struct container *cont = &root->top_container; + +BUG_ON (root->subsys_bits); + ret =get_sb_nodev (fs_type, flags, root, + container_fill_super, mnt); +if (ret) + goto out_unlock; + +BUG_ON (!list_empty (&cont->sibling)); +BUG_ON (!list_empty (&cont->children)); +BUG_ON (root->number_of_containers !=1 ); + + ret =rebind_subsystems (root, subsys_bits); + +/* It's safe to nest i_mutex inside manage_mutex in +* this case, since no-one else can be accessing this +* directory yet */ +mutex_lock (&cont->dentry->d_inode->i_mutex); +container_populate_dir (cont); +mutex_unlock (&cont->dentry->d_inode->i_mutex); +BUG_ON (ret); + + }else { +/* Reuse the existing superblock */ + ret =simple_set_mnt (mnt, root->sb); +if (!ret) +atomic_inc (&root->sb->s_active); + } + + out_unlock: +mutex_unlock (&manage_mutex); +return ret; } static struct file_system_type container_fs_type = {@@ -372,135 +858,79 @@ int container_path(const struct containe }/* -* Notify userspace when a container is released, by running -* /sbin/container_release_agent with the name of the container (path -* relative to the root of container file system) as the argument. -* -* Most likely, this user command will try to rmdir this container. -* -* This races with the possibility that some other task will be -* attached to this container before it is removed, or that some other -* user task will 'mkdir' a child container of this container. That's ok. -* The presumed 'rmdir' will fail quietly if this container is no longer -* unused, and this container will be reprieved from its death sentence, -* to continue to serve a useful existence. Next time it's released, -* we will get notified again, if it still has 'notify_on_release' set. -* -* The final arg to call_usermodehelper() is 0, which means don't -* wait. The separate /sbin/container_release_agent task is forked by -* call_usermodehelper(), then control in this thread returns here, -* without waiting for the release agent task. We don't bother to -* wait because the caller of this routine has no use for the exit -* status of the /sbin/container_release_agent task, so no sense holding -* our caller up for that. -* -* When we had only one container mutex, we had to call this -* without holding it, to avoid deadlock when call_usermodehelper() -* allocated memory. With two locks, we could now call this while -* holding manage_mutex, but we still don't, so as to minimize -* the time manage_mutex is held. +* Attach task 'tsk' to container 'cont' +* +* Call holding manage_mutex. May take callback_mutex and task_lock of +* the task 'pid' during call. */ -static void container_release_agent (const char *pathbuf) +static int attach_task (struct container *cont,struct task_struct *tsk) { -char *argv[3 ], *envp[3 ]; -int i; - -if (!pathbuf) -return ; - - i =0 ; - argv[i++] = release_agent_path; - argv[i++] = (char *)pathbuf; - argv[i] =NULL ; - - i =0 ; -/* minimal command environment */ - envp[i++] ="HOME=/" ; - envp[i++] ="PATH=/sbin:/bin:/usr/sbin:/usr/bin" ; - envp[i] =NULL ; +int retval =0 ; +struct container_subsys *ss; +struct container_group *oldcg, *newcg; +struct container *oldcont; +int h = cont->hierarchy; -call_usermodehelper (argv[0 ], argv, envp,0 ); -kfree (pathbuf); -} +/* Nothing to do if the task is already in that container */ +if (tsk->containers->container[h] == cont) +return 0 ; -/* -* Either cont->count of using tasks transitioned to zero, or the -* cont->children list of child containers just became empty. If this -* cont is notify_on_release() and now both the user count is zero and -* the list of children is empty, prepare container path in a kmalloc'd -* buffer, to be returned via ppathbuf, so that the caller can invoke -* container_release_agent() with it later on, once manage_mutex is dropped. -* Call here with manage_mutex held. -* -* This check_for_release() routine is responsible for kmalloc'ing -* pathbuf. The above container_release_agent() is responsible for -* kfree'ing pathbuf. The caller of these routines is responsible -* for providing a pathbuf pointer, initialized to NULL, then -* calling check_for_release() with manage_mutex held and the address -* of the pathbuf pointer, then dropping manage_mutex, then calling -* container_release_agent() with pathbuf, as set by check_for_release(). -*/ - -static void check_for_release (struct container *cont,char **ppathbuf) -{ -if (notify_on_release (cont) &&atomic_read (&cont->count) ==0 && -list_empty (&cont->children)) { -char *buf; - - buf =kmalloc (PAGE_SIZE ,GFP_KERNEL ); -if (!buf) -return ; - -if (container_path (cont, buf,PAGE_SIZE ) <0 ) -kfree (buf); -else - *ppathbuf = buf; +for_each_subsys (h, ss) { +if (ss->can_attach) { + retval = ss->can_attach (ss, cont, tsk); +if (retval) { +put_task_struct (tsk); +return retval; + } + } } -} +/* Locate or allocate a new container_group for this task, +* based on its final set of containers */ + oldcg = tsk->containers; + newcg =find_container_group (oldcg, cont); +if (!newcg) { +put_task_struct (tsk); +return -ENOMEM ; + } -/* -* update_flag - read a 0 or a 1 in a file and update associated flag -* bit: the bit to update (CONT_NOTIFY_ON_RELEASE) -* cont: the container to update -* buf: the buffer where we read the 0 or 1 -* -* Call with manage_mutex held. -*/ - -static int update_flag (container_flagbits_t bit,struct container *cont,char *buf) -{ -int turning_on; +mutex_lock (&callback_mutex); +task_lock (tsk); +rcu_assign_pointer (tsk->containers, newcg); +task_unlock (tsk); - turning_on = (simple_strtoul (buf,NULL ,10 ) !=0 ); + oldcont = oldcg->container[h]; +for_each_subsys (h, ss) { +if (ss->attach) { + ss->attach (ss, cont, oldcont, tsk); + } + } -mutex_lock (&callback_mutex); -if (turning_on) -set_bit (bit, &cont->flags); -else -clear_bit (bit, &cont->flags);mutex_unlock (&callback_mutex); +for_each_subsys (h, ss) { +if (ss->post_attach) { + ss->post_attach (ss, cont, oldcont, tsk); + } + } + +synchronize_rcu (); +put_container_group (oldcg);return 0 ; } -/* -* Attack task specified by pid in 'pidbuf' to container 'cont', possibly -* writing the path of the old container in 'ppathbuf' if it needs to be -* notified on release. +* Attach task with pid 'pid' to container 'cont'. Call with +* manage_mutex, may take callback_mutex and task_lock of task * -* Call holding manage_mutex. May take callback_mutex and task_lock of -* the task 'pid' during call. */ -static int attach_task (struct container *cont,char *pidbuf,char **ppathbuf) +static int attach_task_by_pid (struct container *cont,char *pidbuf) {pid_t pid;struct task_struct *tsk; -struct container *oldcont; -int retval =0 ; +int ret;if (sscanf (pidbuf,"%d" , &pid) !=1 )return -EIO ;@@ -527,43 +957,9 @@ static int attach_task(struct containerget_task_struct (tsk); } -#ifdef CONFIG_CPUSETS - retval =cpuset_can_attach_task (cont, tsk); -#endif -if (retval) { -put_task_struct (tsk); -return retval; - } - -mutex_lock (&callback_mutex); - -task_lock (tsk); - oldcont = tsk->container; -if (!oldcont) { -task_unlock (tsk); -mutex_unlock (&callback_mutex); -put_task_struct (tsk); -return -ESRCH ; - } -atomic_inc (&cont->count); -rcu_assign_pointer (tsk->container, cont); -task_unlock (tsk); - -#ifdef CONFIG_CPUSETS -cpuset_attach_task (cont, tsk); -#endif - -mutex_unlock (&callback_mutex); - -#ifdef CONFIG_CPUSETS -cpuset_post_attach_task (cont, oldcont, tsk); -#endif - + ret =attach_task (cont, tsk);put_task_struct (tsk); -synchronize_rcu (); -if (atomic_dec_and_test (&oldcont->count)) -check_for_release (oldcont, ppathbuf); -return 0 ; +return ret; }/* The various types of files and directories in a container file system */ @@ -571,9 +967,7 @@ static int attach_task(struct containertypedef enum {FILE_ROOT ,FILE_DIR , -FILE_NOTIFY_ON_RELEASE ,FILE_TASKLIST , -FILE_RELEASE_AGENT , }container_filetype_t ; static ssize_t container_common_file_write(struct container *cont,@@ -584,7 +978,6 @@ static ssize_t container_common_file_wri {container_filetype_t type = cft->private;char *buffer; -char *pathbuf =NULL ;int retval =0 ;if (nbytes >=PATH_MAX )@@ -608,26 +1001,9 @@ static ssize_t container_common_file_wri }switch (type) { -case FILE_NOTIFY_ON_RELEASE : - retval =update_flag (CONT_NOTIFY_ON_RELEASE , cont, buffer); -break ;case FILE_TASKLIST : - retval =attach_task (cont, buffer, &pathbuf); -break ; -case FILE_RELEASE_AGENT : - { -if (nbytes <sizeof (release_agent_path)) { -/* We never write anything other than '\0' -* into the last char of release_agent_path, -* so it always remains a NUL-terminated -* string */ -strncpy (release_agent_path, buffer, nbytes); - release_agent_path[nbytes] =0 ; - }else { - retval = -ENOSPC ; - } + retval =attach_task_by_pid (cont, buffer);break ; - }default : retval = -EINVAL ; goto out2;@@ -637,7 +1013,6 @@ static ssize_t container_common_file_wri retval = nbytes; out2:mutex_unlock (&manage_mutex); -container_release_agent (pathbuf); out1:kfree (buffer);return retval;@@ -646,80 +1021,27 @@ out1:static ssize_t container_file_write (struct file *file,const char __user *buf,size_t nbytes,loff_t *ppos) { -ssize_t retval =0 ;struct cftype *cft =__d_cft (file->f_dentry);struct container *cont =__d_cont (file->f_dentry->d_parent);if (!cft)return -ENODEV ; +if (!cft->write) +return -EINVAL ; -/* special function ? */ -if (cft->write) - retval = cft->write (cont, cft, file, buf, nbytes, ppos); -else - retval = -EINVAL ; - -return retval; +return cft->write (cont, cft, file, buf, nbytes, ppos); } -static ssize_t container_common_file_read (struct container *cont, -struct cftype *cft, -struct file *file, -char __user *buf, -size_t nbytes,loff_t *ppos) +static ssize_t container_file_read (struct file *file,char __user *buf, +size_t nbytes,loff_t *ppos) { -container_filetype_t type = cft->private; -char *page; -ssize_t retval =0 ; -char *s; - -if (!(page = (char *)__get_free_page (GFP_KERNEL ))) -return -ENOMEM ; - - s = page; - -switch (type) { -case FILE_NOTIFY_ON_RELEASE : - *s++ =notify_on_release (cont) ?'1' :'0' ; -break ; -case FILE_RELEASE_AGENT : - { -size_t n; -container_manage_lock (); - n =strnlen (release_agent_path,sizeof (release_agent_path)); - n =min (n, (size_t )PAGE_SIZE ); -strncpy (s, release_agent_path, n); -container_manage_unlock (); - s += n; -break ; - } -default : - retval = -EINVAL ; - goto out; - } - *s++ ='\n' ; - - retval =simple_read_from_buffer (buf, nbytes, ppos, page, s - page); -out: -free_page ((unsigned long )page); -return retval; -} - -static ssize_t container_file_read (struct file *file,char __user *buf,size_t nbytes, -loff_t *ppos) -{ -ssize_t retval =0 ;struct cftype *cft =__d_cft (file->f_dentry);struct container *cont =__d_cont (file->f_dentry->d_parent);if (!cft)return -ENODEV ; +if (!cft->read) +return -EINVAL ; -/* special function ? */ -if (cft->read) - retval = cft->read (cont, cft, file, buf, nbytes, ppos); -else - retval = -EINVAL ; - -return retval; +return cft->read (cont, cft, file, buf, nbytes, ppos); }static int container_file_open (struct inode *inode,struct file *file)@@ -780,7 +1102,7 @@ static struct inode_operations container .rename = container_rename, }; -static int container_create_file (struct dentry *dentry,int mode) +static int container_create_file (struct dentry *dentry,int mode,struct super_block *sb) {struct inode *inode;@@ -789,7 +1111,7 @@ static int container_create_file(structif (dentry->d_inode)return -EEXIST ; - inode =container_new_inode (mode); + inode =container_new_inode (mode, sb);if (!inode)return -ENOMEM ;@@ -798,7 +1120,11 @@ static int container_create_file(struct inode->i_fop = &simple_dir_operations;/* start off with i_nlink == 2 (for "." entry) */ - inode->i_nlink++; +inc_nlink (inode); + +/* start with the directory inode held, so that we can +* populate it without racing with another mkdir */ +mutex_lock (&inode->i_mutex); }else if (S_ISREG (mode)) { inode->i_size =0 ; inode->i_fop = &container_file_operations;@@ -818,20 +1144,19 @@ static int container_create_file(struct* mode: mode to set on new directory. */ -static int container_create_dir (struct container *cont,const char *name,int mode) +static int container_create_dir (struct container *cont,struct dentry *dentry, +int mode) { -struct dentry *dentry =NULL ;struct dentry *parent;int error =0 ; parent = cont->parent->dentry; - dentry =container_get_dentry (parent, name);if (IS_ERR (dentry))return PTR_ERR (dentry); - error =container_create_file (dentry,S_IFDIR | mode); + error =container_create_file (dentry,S_IFDIR | mode, cont->root->sb);if (!error) { dentry->d_fsdata = cont; - parent->d_inode->i_nlink++; +inc_nlink (parent->d_inode); cont->dentry = dentry; }dput (dentry);@@ -845,19 +1170,40 @@ int container_add_file(struct containerstruct dentry *dentry;int error; -mutex_lock (&dir->d_inode->i_mutex); +BUG_ON (!mutex_is_locked (&dir->d_inode->i_mutex)); dentry =container_get_dentry (dir, cft->name);if (!IS_ERR (dentry)) { - error =container_create_file (dentry,0644 |S_IFREG ); + error =container_create_file (dentry,0644 |S_IFREG , cont->root->sb);if (!error) dentry->d_fsdata = (void *)cft;dput (dentry); }else error =PTR_ERR (dentry); -mutex_unlock (&dir->d_inode->i_mutex);return error; } +/* Count the number of tasks in a container. Could be made more +* time-efficient but less space-efficient with more linked lists +* running through each container and the container_group structures +* that referenced it. */ + +int container_task_count (const struct container *cont) { +int count =0 ; +int hierarchy = cont->hierarchy; +struct list_head *l; +spin_lock (&container_group_lock); + l = &init_container_group.list; +do { +struct container_group *cg = + list_entry(l, struct container_group, list); +if (cg->container[hierarchy] == cont) + count +=atomic_read (&cg->ref.refcount); + l = l->next; + }while (l != &init_container_group.list); +spin_unlock (&container_group_lock); +return count; +} +/* * Stuff for reading the 'tasks' file. * @@ -881,20 +1227,23 @@ struct ctr_struct { };/* -* Load into 'pidarray' up to 'npids' of the tasks using container 'cont'. -* Return actual number of pids loaded. No need to task_lock(p) -* when reading out p->container, as we don't really care if it changes -* on the next cycle, and we are not going to try to dereference it. +* Load into 'pidarray' up to 'npids' of the tasks using container +* 'cont'. Return actual number of pids loaded. No need to +* task_lock(p) when reading out p->container, since we're in an RCU +* read section, so the container_group can't go away, and is +* immutable after creation. */ static int pid_array_load (pid_t *pidarray,int npids,struct container *cont) {int n =0 ;struct task_struct *g, *p; +int h = cont->hierarchy; +rcu_read_lock ();read_lock (&tasklist_lock); do_each_thread(g, p) { -if (p->container == cont) { +if (p->containers->container[h] == cont) { pidarray[n++] =pid_nr (task_pid (p));if (unlikely (n == npids)) goto array_full;@@ -903,6 +1252,7 @@ static int pid_array_load(pid_t *pidarra array_full:read_unlock (&tasklist_lock); +rcu_read_unlock ();return n; }@@ -953,7 +1303,7 @@ static int container_tasks_open(struct i* caller from the case that the additional container users didn't * show up until sometime later on. */ - npids =atomic_read (&cont->count); + npids =container_task_count (cont); pidarray =kmalloc (npids *sizeof (pid_t),GFP_KERNEL );if (!pidarray) goto err1;@@ -1020,38 +1370,34 @@ static struct cftype cft_tasks = { .private = FILE_TASKLIST, }; -static struct cftype cft_notify_on_release = { - .name ="notify_on_release" , - .read = container_common_file_read, - .write = container_common_file_write, - .private =FILE_NOTIFY_ON_RELEASE , -}; - -static struct cftype cft_release_agent = { - .name ="release_agent" , - .read = container_common_file_read, - .write = container_common_file_write, - .private =FILE_RELEASE_AGENT , -}; -static int container_populate_dir (struct container *cont) {int err; +struct container_subsys *ss; + +/* First clear out any existing files */ +container_clear_directory (cont->dentry); -if ((err =container_add_file (cont, &cft_notify_on_release)) <0 ) -return err;if ((err =container_add_file (cont, &cft_tasks)) <0 )return err; -if ((cont == &top_container) && - (err =container_add_file (cont, &cft_release_agent)) <0 ) -return err; -#ifdef CONFIG_CPUSETS -if ((err =cpuset_populate_dir (cont)) <0 ) -return err; -#endif + +for_each_subsys (cont->hierarchy, ss) { +if (ss->populate && (err = ss->populate (ss, cont)) <0 ) +return err; + } +return 0 ; } +static void init_container_css (struct container_subsys *ss, +struct container *cont) +{ +struct container_subsys_state *css = cont->subsys[ss->subsys_id]; + css->container = cont; +spin_lock_init (&css->refcnt_lock); +atomic_set (&css->refcnt,0 ); +} +/* * container_create - create a container * parent: container that will be parent of the new container. @@ -1061,61 +1407,83 @@ static int container_populate_dir(struct* Must be called with the mutex on the parent inode held */ -static long container_create (struct container *parent,const char *name,int mode) +static long container_create (struct container *parent,struct dentry *dentry, +int mode) {struct container *cont; -int err; +struct containerfs_root *root = parent->root; +int err =0 ; +struct container_subsys *ss; +struct super_block *sb = root->sb; - cont =kmalloc (sizeof (*cont),GFP_KERNEL ); + cont =kzalloc (sizeof (*cont),GFP_KERNEL );if (!cont)return -ENOMEM ; +/* Grab a reference on the superblock so the hierarchy doesn't +* get deleted on unmount if there are child containers. This +* can be done outside manage_mutex, since the sb can't +* disappear while someone has an open control file on the +* fs */ +atomic_inc (&sb->s_active); +mutex_lock (&manage_mutex); + cont->flags =0 ; -if (notify_on_release (parent)) -set_bit (CONT_NOTIFY_ON_RELEASE , &cont->flags);atomic_set (&cont->count,0 );INIT_LIST_HEAD (&cont->sibling);INIT_LIST_HEAD (&cont->children); cont->parent = parent; - -#ifdef CONFIG_CPUSETS - err =cpuset_create (cont); -if (err) - goto err_unlock_free; -#endif + cont->root = parent->root; + cont->hierarchy = parent->hierarchy; + cont->top_container = parent->top_container; + +for_each_subsys (cont->hierarchy, ss) { + err = ss->create (ss, cont); +if (err) goto err_destroy; +init_container_css (ss, cont); + }mutex_lock (&callback_mutex);list_add (&cont->sibling, &cont->parent->children); - number_of_containers++; + root->number_of_containers++;mutex_unlock (&callback_mutex); - err =container_create_dir (cont, name, mode); + err =container_create_dir (cont, dentry, mode);if (err <0 ) goto err_remove; -/* -* Release manage_mutex before container_populate_dir() because it -* will down() this new directory's i_mutex and if we race with -* another mkdir, we might deadlock. -*/ -mutex_unlock (&manage_mutex); +/* The container directory was pre-locked for us */ +BUG_ON (!mutex_is_locked (&cont->dentry->d_inode->i_mutex)); err =container_populate_dir (cont);/* If err < 0, we have a half-filled directory - oh well ;) */ + +mutex_unlock (&manage_mutex); +mutex_unlock (&cont->dentry->d_inode->i_mutex); +return 0 ; err_remove: -#ifdef CONFIG_CPUSETS -cpuset_destroy (cont ); -#endif +mutex_lock (&callback_mutex);list_del (&cont->sibling); - number_of_containers--; + root->number_of_containers--;mutex_unlock (&callback_mutex); - err_unlock_free: + + err_destroy: + +for_each_subsys (cont->hierarchy, ss) { +if (cont->subsys[ss->subsys_id]) + ss->destroy (ss, cont); + } +mutex_unlock (&manage_mutex); + +/* Release the reference count that we took on the superblock */ +deactivate_super (sb); +kfree (cont);return err; }@@ -1125,26 +1493,20 @@ static int container_mkdir(struct inodestruct container *c_parent = dentry->d_parent->d_fsdata;/* the vfs holds inode->i_mutex already */ -return container_create (c_parent, dentry->d_name.name, mode |S_IFDIR ); +return container_create (c_parent, dentry, mode |S_IFDIR ); } -/* -* Locking note on the strange update_flag() call below: -* -* If the container being removed is marked cpu_exclusive, then simulate -* turning cpu_exclusive off, which will call update_cpu_domains(). -* The lock_cpu_hotplug() call in update_cpu_domains() must not be -* made while holding callback_mutex. Elsewhere the kernel nests -* callback_mutex inside lock_cpu_hotplug() calls. So the reverse -* nesting would risk an ABBA deadlock. -*/ -static int container_rmdir (struct inode *unused_dir,struct dentry *dentry) {struct container *cont = dentry->d_fsdata;struct dentry *d;struct container *parent; -char *pathbuf =NULL ; +struct container_subsys *ss; +struct super_block *sb; +struct containerfs_root *root; +unsigned long flags; +int css_busy =0 ; +int hierarchy;/* the vfs holds both inode->i_mutex already */ @@ -1157,82 +1519,331 @@ static int container_rmdir(struct inodemutex_unlock (&manage_mutex);return -EBUSY ; } + + hierarchy = cont->hierarchy; parent = cont->parent; + root = cont->root; + sb = root->sb; + +local_irq_save (flags); +/* Check each container, locking the refcnt lock and testing +* the refcnt. This will lock out any calls to css_get() */ +for_each_subsys (hierarchy, ss) { +struct container_subsys_state *css; + css = cont->subsys[ss->subsys_id]; +spin_lock (&css->refcnt_lock); + css_busy +=atomic_read (&css->refcnt); + } +/* Go through and release all the locks; if we weren't busy, +* then set the refcount to -1 to prevent css_get() from adding +* a refcount */ +for_each_subsys (hierarchy, ss) { +struct container_subsys_state *css; + css = cont->subsys[ss->subsys_id]; +if (!css_busy)atomic_dec (&css->refcnt); +spin_unlock (&css->refcnt_lock); + } +local_irq_restore (flags); +if (css_busy) { +mutex_unlock (&manage_mutex); +return -EBUSY ; + } + +for_each_subsys (hierarchy, ss) { +if (cont->subsys[ss->subsys_id]) + ss->destroy (ss, cont); + } +mutex_lock (&callback_mutex);set_bit (CONT_REMOVED , &cont->flags); -list_del (&cont->sibling);/* delete my sibling from parent->children */ +/* delete my sibling from parent->children */ +list_del (&cont->sibling);spin_lock (&cont->dentry->d_lock); d =dget (cont->dentry); cont->dentry =NULL ;spin_unlock (&d->d_lock); +container_d_remove_dir (d);dput (d); - number_of_containers--; + root->number_of_containers--;mutex_unlock (&callback_mutex); -#ifdef CONFIG_CPUSETS -cpuset_destroy (cont); -#endif -if (list_empty (&parent->children)) -check_for_release (parent, &pathbuf); +mutex_unlock (&manage_mutex); -container_release_agent (pathbuf); +/* Drop the active superblock reference that we took when we +* created the container */ +deactivate_super (sb);return 0 ; } -/* -* container_init_early - probably not needed yet, but will be needed -* once cpusets are hooked into this code +static atomic_t namecnt; +static void get_unused_name (char *buf) { +sprintf (buf,"node%d" ,atomic_inc_return (&namecnt)); +} + +/** +* container_clone - duplicate the current container and move this +* task into the new child */ +int container_clone (struct task_struct *tsk,struct container_subsys *subsys) +{ +struct dentry *dentry; +int ret =0 ; +char nodename[32 ]; +struct container *parent, *child; +struct inode *inode; +int h; + +/* We shouldn't be called by an unregistered subsystem */ +BUG_ON (subsys->subsys_id <0 ); + +/* First figure out what hierarchy and container we're dealing +* with, and pin them so we can drop manage_mutex */ +mutex_lock (&manage_mutex); + again: + h = subsys->hierarchy; +if (h ==0 ) { +printk (KERN_INFO +"Not cloning container for unused subsystem %s\n" , + subsys->name); +mutex_unlock (&manage_mutex); +return 0 ; + } + parent = tsk->containers->container[h]; +/* Pin the hierarchy */ +atomic_inc (&parent->root->sb->s_active); +/* Keep the container alive */ +atomic_inc (&parent->count); +mutex_unlock (&manage_mutex); + +/* Now do the VFS work to create a container */ +get_unused_name (nodename); + inode = parent->dentry->d_inode; + +/* Hold the parent directory mutex across this operation to +* stop anyone else deleting the new container */ +mutex_lock (&inode->i_mutex); + dentry =container_get_dentry (parent->dentry, nodename); +if (IS_ERR (dentry)) { +printk (KERN_INFO +"Couldn't allocate dentry for %s: %ld\n" , nodename, +PTR_ERR (dentry)); + ret =PTR_ERR (dentry); + goto out_release; + } + +/* Create the container directory, which also creates the container */ + ret =vfs_mkdir (inode, dentry,S_IFDIR |0755 ); + child =__d_cont (dentry); +dput (dentry); +if (ret) { +printk (KERN_INFO +"Failed to create container %s: %d\n" , nodename, + ret); + goto out_release; + } + +if (!child) { +printk (KERN_INFO +"Couldn't find new container %s\n" , nodename); + ret = -ENOMEM ; + goto out_release; + } + +/* The container now exists. Retake manage_mutex and check +* that we're still in the same state that we thought we +* were. */ +mutex_lock (&manage_mutex); +if ((h != subsys->hierarchy) || + (parent != tsk->containers->container[h])) { +/* Aargh, we raced ... */ +mutex_unlock (&inode->i_mutex); +atomic_dec (&parent->count); +deactivate_super (parent->root->sb); +printk (KERN_INFO +"Race in container_clone() - leaking container %s\n" , + nodename); + goto again; + } + +/* All seems fine. Finish by moving the task into the new container */ + ret =attach_task (child, tsk); +mutex_unlock (&manage_mutex); + + out_release: +mutex_unlock (&inode->i_mutex); +atomic_dec (&parent->count); +deactivate_super (parent->root->sb); +return ret; +} + +int container_is_descendant (const struct container *cont) { +int ret; +struct container *target; +container_lock (); + target = current->containers->container[cont->hierarchy]; +while (cont != target && cont!= target->top_container) { + cont = cont->parent; + } + ret = (cont == target); +container_unlock (); +return ret; +} + +/** +* container_init_early - initialize containers at system boot +* +* Description: Initialize the container housekeeping structures +**/ int __init container_init_early (void ) { -struct task_struct *tsk = current; +int i; + +kref_init (&init_container_group.ref); +get_container_group (&init_container_group); +INIT_LIST_HEAD (&init_container_group.list); + container_group_count =1 ; + +for (i =0 ; i <CONFIG_MAX_CONTAINER_HIERARCHIES ; i++) { +struct containerfs_root *root = &rootnode[i]; +struct container *cont = &root->top_container; +INIT_LIST_HEAD (&root->subsys_list); + root->number_of_containers =1 ; + + cont->root = root; + cont->hierarchy = i; +INIT_LIST_HEAD (&cont->sibling); +INIT_LIST_HEAD (&cont->children); + cont->top_container = cont; +atomic_set (&cont->count,1 ); + + init_container_group.container[i] = cont; + } + init_task.containers = &init_container_group; - tsk->container = &top_container;return 0 ; }/** -* container_init - initialize containers at system boot -* -* Description: Initialize top_container and the container internal file system, +* container_init - register container filesystem and /proc file **/ int __init container_init (void ) { -struct dentry *root;int err; - - init_task.container = &top_container; +struct proc_dir_entry *entry; err =register_filesystem (&container_fs_type);if (err <0 ) goto out; - container_mount =kern_mount (&container_fs_type); -if (IS_ERR (container_mount)) { -printk (KERN_ERR "container: could not mount!\n" ); - err =PTR_ERR (container_mount); - container_mount =NULL ; - goto out; - } - root = container_mount->mnt_sb->s_root; - root->d_fsdata = &top_container; - root->d_inode->i_nlink++; - top_container.dentry = root; - root->d_inode->i_op = &container_dir_inode_operations; - number_of_containers =1 ; - err =container_populate_dir (&top_container);
kernel/container.c static int container_rmdir(struct inode · line 1748
Quoted 10 lines
kernel/container.c +6
-0+ + entry =create_proc_entry ("containers" ,0 ,NULL ); +if (entry) + entry->proc_fops = &proc_containerstats_operations; + out:return err; } +#include <asm/proto.h>
I did have a problem with this include. On s390 it didn't exist so I've just been running without it (with no problems). A quick 'find' suggests it only exists on x86_64, so I'd expect failures on all other arches.
-serge
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Quoted 6 lines
Quoted 1 line
+#include <
asm/proto.h>I did have a problem with this include. On s390 it didn't exist so I've just been running without it (with no problems). A quick 'find' suggests it only exists on
x86_64, so I'd expect failures on all other arches.
confirmed on x86 also.
C.
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
Quoted 9 lines · Cedric Le Goater
Quoted 6 lines
Quoted 1 line
+#include <
asm/proto.h>I did have a problem with this include. On s390 it didn't exist so I've just been running without it (with no problems). A quick 'find' suggests it only exists on
x86_64, so I'd expect failures on all other arches.confirmed on x86 also.
Sorry, this turned out to be left over from when I was debugging an initialization problem that required early_printk(). I should have removed it.
You can safely remove it from the patch.
Paul
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
In reply to a message not in the archiveQuoted 6 lines · [email protected] wrote:
This patch removes all cpuset-specific knowlege from the container system, replacing it with a generic API that can be used by multiple subsystems. Cpusets is adapted to be a container subsystem.
Signed-off-by Paul Menage <[email protected]>
Hi, Paul,
This patch fails to apply for me
balbir@balbir-laptop:~/ocbalbir/images/kernels/containers/linux-2.6.20$ pushpatch patching file include/linux/container.h patching file include/linux/cpuset.h patching file kernel/container.c patch: **** malformed patch at line 640: @@ -202,15 +418,18 @@ static DEFINE_MUTEX(callback_mutex);
multiuser_container does not apply
Is anybody else seeing this problem?
Signature · 3 lines
Balbir Singh Linux Technology Center IBM, ISTL
Re: [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
In reply to a message not in the archiveReview
Quoted 12 lines · On Mon, Feb 12, 2007 at 12:15:24AM -0800, [email protected] wrote:
(quoted diff) +12
-0+static struct container_group *find_container_group( + struct container_group *oldcg, struct container *cont) +{ + struct container_group *res; + struct container_subsys *ss; + int h = cont->hierarchy; + int i; + + BUG_ON(oldcg->container[h] == cont); + /* First see if we already have a container group that matches + * the desired set */ + spin_lock(&container_group_lock);
I think this should be spin_lock_irq().
I have been bitten by lockdep warning in the rcfs patchs. But I think they should apply here as well.
Essentially, container_task_count() is called from cpuset.c after doing write_lock_irq(tasklist_lock). This creates the dependency:
tasklist_lock (irqs disabled) -> container_group_lock
However in above function, find_container_group (and several other functions) we take container_group_lock w/o blocking interrupts. An interrupt handler now trying to take tasklist_lock will cause a deadlock.
Signature · 2 lines
Regards, vatsa
Re: [ckrm-tech] [PATCH 3/7] containers (V7): Add generic multi-subsystem API to containers
In reply to a message not in the archiveQuoted 2 lines · On Mon, Feb 12, 2007 at 12:15:24AM -0800, [email protected] wrote:
+static int
attach_task(struct container *cont, struct task_struct *tsk)
{
[snip]
Review
Quoted 5 lines
(quoted diff) +5
-0+ for_each_subsys(h, ss) { + if (ss->can_attach) { + retval = ss->can_attach(ss, cont, tsk); + if (retval) { + put_task_struct(tsk);
We don't need this put_task_struct(), since our caller attach_task_by_pid() would do it for us.
Review
Quoted 12 lines
(quoted diff) +9
-1+ return retval; + } + } } -} + /* Locate or allocate a new container_group for this task, + * based on its final set of containers */ + oldcg = tsk->containers; + newcg = find_container_group(oldcg, cont); + if (!newcg) { + put_task_struct(tsk);
Ditto
Quoted 2 lines
(quoted diff) +2
-0+ return -ENOMEM; + }
Signature · 2 lines
Regards, vatsa
Quoted 12 lines · [email protected] wrote:
Generic Process Containers
--------------------------There have recently been various proposals floating around for resource management/accounting and other task grouping subsystems in the kernel, including ResGroups, User BeanCounters, NSProxy containers, and others. These all need the basic abstraction of being able to group together multiple processes in an aggregate, in order to track/limit the resources permitted to those processes, or control other behaviour of the processes, and all implement this grouping in different ways.
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container. We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
This would suggesting re-write this patchset, part 2 as a "CPUSet namespace", part 4 as a "CPU scheduling namespace", parts 5 and 6 as "Resource Limits Namespace" (drop this "BeanCounter" brand), and of course part 7 falls away.
Sam.
Quoted 14 lines · Sam Vilain
Quoted 12 lines · [email protected] wrote:
Generic Process Containers
--------------------------There have recently been various proposals floating around for resource management/accounting and other task grouping subsystems in the kernel, including ResGroups, User BeanCounters, NSProxy containers, and others. These all need the basic abstraction of being able to group together multiple processes in an aggregate, in order to track/limit the resources permitted to those processes, or control other behaviour of the processes, and all implement this grouping in different ways.
Quoted 1 line
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container.
The nsproxy is an implementation detail, actually, and might go away at any time. But yes we had been talking for quite awhile about sets of namespaces as containers, so that a vserver would be a container, as would a lightweight checkpoint/restart job.
Quoted 4 lines
We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
This would suggesting re-write this patchset, part 2 as a "CPUSet
Well it's an unfortunate conflict, but I don't see where we have any standing to make Paul change his terminology :)
Lately I've been talking about "Paul's containers", our "namespaces", and the "namespace container subsystem" in patch 7.
Quoted 3 lines
namespace", part 4 as a "CPU scheduling namespace", parts 5 and 6 as "Resource Limits Namespace" (drop this "BeanCounter" brand), and of course part 7 falls away.
Part 7 does not fall away, it uses Paul's patchset to start to provide a management interface for namespaces.
-serge
Quoted 3 lines · Serge E. Hallyn
Well it's an unfortunate conflict, but I don't see where we have any standing to make Paul change his terminology :)
I have no huge problem with changing my terminology in the interest of wider adoption. "Container" seems like an appropriate name for the abstraction, and possibly more appropriate than for a virtual server, but if it was decreed that the only thing stopping the container patch being merged was that it should be called, say, "Process Sets", I would happily s/container/pset/g across the entire patchset.
Paul
Quoted 4 lines · Sam Vilain
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container. We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
You may have done that, but the CKRM/ResGroups independently decided a long time ago that the fundamental unit was the resource class, and the OpenVZ folks decided that the fundamental unit was the BeanCounter, and the CPUSet folks decided that the fundamental unit was the CPUSet, etc ... :-)
But there's a lot of common ground between these different approaches, and potential for synergy, so the point of this patch set is to provide a unification point for all of them, and a stepping stone for other new resource controllers and process control modules.
Paul
Quoted 10 lines · Paul Menage
Quoted 4 lines
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container. We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
You may have done that, but the CKRM/ResGroups independently decided a long time ago that the fundamental unit was the resource class, and the OpenVZ folks decided that the fundamental unit was the BeanCounter, and the CPUSet folks decided that the fundamental unit was the CPUSet, etc ... :-)
There was a consensus that emerged that resulted in the nsproxy being included in the kernel. That is why I used "we". What was included varies greatly from the patch I put forward, as I mentioned.
You missed the point of what I was trying to say. Let me try again.
Originally I too thought that in order to begin on the path of virtualisation merging we would just make a simple container/vserver/whatever structure and hang everything off that.
I'll now say the same thing that was said before of my patch, that I don't think that adding the containers structure inside the kernel adds anything interesting or useful. In fact, I'd go further to say that very thing you think is a useful abstraction is locking yourself into inflexibility. This is what Eric recognised early on and eventually brought me around to the idea of too.
So, we have the current implementation - individual subsystems are virtualised, and it is the subsystems that you can see and work with. You can group together your virtualisation decisions and call that a container, great.
Ask yourself this - what do you need the container structure for so badly, that virtualising the individual resources does not provide for? You don't need it to copy the namespaces of another process ("enter") and you don't need it for checkpoint/migration. What does it mean to make a new container? You're making a "nothing" namespace for some as-yet-unspecified grouping of tasks. That's a great idea for a set of tightly integrated userland utilities to simplify the presentation to the admin, but I don't see why you need to enshrine this in the kernel. Certainly not for any of the other patches in your set as far as I can see.
Quoted 5 lines
But there's a lot of common ground between these different approaches, and potential for synergy, so the point of this patch set is to provide a unification point for all of them, and a stepping stone for other new resource controllers and process control modules.
That precisely echos my sentiment on my submissions of some 12 months ago.
Sam.
Quoted 2 lines · Sam Vilain
Ask yourself this - what do you need the container structure for so badly, that virtualising the individual resources does not provide for?
Primarily, that otherwise every module that wants to affect/monitor behaviour of a group of associated processes has to implement its own process grouping abstraction.
As an example, the CPU accounting patch that in included in my patch set as an illustration of a simple resource monitoring module is just 250 lines, almost entirely in one file; if it also had to handle associating tasks together into groups and presenting a filesystem interface to the user it would be far larger and would have a much bigger footprint on the kernel.
Quoted 1 line
From the point of view of the virtual server containers, the advantage
is that you're integrated with a standard filesystem interface for determining group membership. It does become simpler to combine virtual servers and resource controllers, although I grant you that you could juggle that from userspace without the additional kernel support.
Paul
Quoted 7 lines · Paul Menage
Quoted 3 lines
Ask yourself this - what do you need the container structure for so badly, that virtualising the individual resources does not provide for?
Primarily, that otherwise every module that wants to affect/monitor behaviour of a group of associated processes has to implement its own process grouping abstraction.
Not every module, you just make them on sensible, planned groupings. The danger is that the "container" group becomes a fallback grouping for things when people can't be bothered thinking about it properly, and everything including the kitchen sink gets thrown in. Then later you find a real use case where you don't want them together, but it's too late because it's already a part of the official API.
Quoted 7 lines
As an example, the CPU accounting patch that in included in my patch set as an illustration of a simple resource monitoring module is just 250 lines, almost entirely in one file; if it also had to handle associating tasks together into groups and presenting a filesystem interface to the user it would be far larger and would have a much bigger footprint on the kernel.
It's also less flexible. What if I want to do CPU accounting on some other boundaries than the "virtual server" a process is a part of?
Quoted 7 lines
From the point of view of the virtual server containers, the advantage is that you're integrated with a standard filesystem interface for determining group membership. It does become simpler to combine virtual servers and resource controllers, although I grant you that you could juggle that from userspace without the additional kernel support.
I'm not disagreeing it's a pragmatic shortcut that has been successful for a number of projects including vserver which I use every day. But it reduces "synergy" by excluding the people working with virtualisation in ways that don't fit its model.
Yes, there should be a similarity in the way that you manage namespaces and it should be easy to develop new namespaces without constantly re-inventing the wheel. But why does that imply making binding decisions about the nature of how you can virtualise? IMHO those decisions should be made on a per-subsystem basis.
Sam.
Quoted 7 lines · Sam Vilain
Not every module, you just make them on sensible, planned groupings. The danger is that the "container" group becomes a fallback grouping for things when people can't be bothered thinking about it properly, and everything including the kitchen sink gets thrown in. Then later you find a real use case where you don't want them together, but it's too late because it's already a part of the official API.
This paragraph seems to have a contradiction - first you say that modules should use "planned groupings", then you complain that modules using generic containers will find that they can't separate from other modules "because they're part of an official API".
On the contrary - a module that's written over generic containers can be combined with other such modules in the same groupings, or used in different groupings, depending on what the user desires. This is discussed in some detail in Documentation/containers.txt in patch 3 of my patchset.
My patchset doesn't try to make all modules use the same *grouping*, it tries to make them use the same *abstraction* (and hence share code), but allows them to easily use different groupings just by binding modules to different hierarchies.
Quoted 11 lines
Quoted 8 lines
As an example, the CPU accounting patch that in included in my patch set as an illustration of a simple resource monitoring module is just 250 lines, almost entirely in one file; if it also had to handle associating tasks together into groups and presenting a filesystem interface to the user it would be far larger and would have a much bigger footprint on the kernel.
It's also less flexible. What if I want to do CPU accounting on some other boundaries than the "virtual server" a process is a part of?
No problem - that's why there are multiple hierarchies available. You can mount the CPU accounting on one hierarchy, the virtual servers on a different hierarchy, and they have completely independent mappings from processes to containers - essentially parallel trees of containers.
Paul
Quoted 11 lines · Paul Menage
Quoted 4 lines · Sam Vilain
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container. We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
You may have done that, but the CKRM/ResGroups independently decided a long time ago that the fundamental unit was the resource class, and the OpenVZ folks decided that the fundamental unit was the BeanCounter, and the CPUSet folks decided that the fundamental unit was the CPUSet, etc ... :-)
Using the container name is bad and it led to this stupid argument.
The fundamental unit of what we have merged into the kernel is the namespace. The aggregate of all namespaces and everything is the container.
Please, please pick a different name so people don't take one look at your stuff and get confused like Sam did.
Quoted 4 lines
But there's a lot of common ground between these different approaches, and potential for synergy, so the point of this patch set is to provide a unification point for all of them, and a stepping stone for other new resource controllers and process control modules.
For the case of namespaces I don't see how your code makes things better. I do not see a real problem that you are solving.
I do agree that a common interface to the code and a common set of infrastructure may help.
Personally until the pid namespace (which includes within it process groupings) is complete I'm not certain the problem will be concrete enough to solve. Hopefully we can have a working version of that shortly.
Eric
Quoted 14 lines · Eric W. Biederman
Quoted 11 lines · Paul Menage
Quoted 4 lines · Sam Vilain
I know I'm a bit out of touch, but AIUI the NSProxy *is* the container. We decided a long time ago that a container was basically just a set of namespaces, which includes all of the subsystems you mention.
You may have done that, but the CKRM/ResGroups independently decided a long time ago that the fundamental unit was the resource class, and the OpenVZ folks decided that the fundamental unit was the BeanCounter, and the CPUSet folks decided that the fundamental unit was the CPUSet, etc ... :-)
Quoted 5 lines
Using the container name is bad and it led to this stupid argument.
The fundamental unit of what we have merged into the kernel is the namespace. The aggregate of all namespaces and everything is the container.
What are you defining here as "everything"? If you mean "all things that could be applied to a segregated group of processes such as a virtual server", then "container" seems like a good name for my patches, since it allows you to aggregate namespaces, resource control, other virtualization, etc.
Sam said "the NSProxy *is* the container". You appear to be planning to have some namespaces, possibly not aggregated within the nsproxy (pid namespace?) but are you planning to have some higher-level "container" object that aggregates the nsproxy and the other namespaces? If so, what is it? Does it track process membership, etc? What's the userspace API? In what fundamental ways would it be different from my generic containers patches with Serge Hallyn's namespace subsystem. (If these questions are answered already by designs or code, I'd be happy to be pointed at them).
I guess what it comes down to, is why is an aggregation of namespaces suitable for the name "container", when an aggregation of namespaces and other resource controllers isn't?
What do you think might be a better name for the generic process groups that I'm pushing? As I said, I'm happy to do a simple search/replace on my code to give a different name if that turned out to be the gating factor to getting it merged. But I'd be inclined to leave that decision up to Andrew/Linus.
Quoted 7 lines
Quoted 4 lines
But there's a lot of common ground between these different approaches, and potential for synergy, so the point of this patch set is to provide a unification point for all of them, and a stepping stone for other new resource controllers and process control modules.
For the case of namespaces I don't see how your code makes things better. I do not see a real problem that you are solving.
I'm trying to solve the problem that lots of different folks (including us) are trying to do things that aggregate multiple process into some kind of constrained group, and are all trying to use different and incompatible ways of grouping/tracking those processes.
I agree that namespaces fit slightly less well into this model than some other subsystems like resource management. But by integrating with it you'd get automatic access to all the various different resource controller work that's being done.
Paul
Quoted 5 lines · Paul Menage
What are you defining here as "everything"? If you mean "all things that could be applied to a segregated group of processes such as a virtual server", then "container" seems like a good name for my patches, since it allows you to aggregate namespaces, resource control, other virtualization, etc.
The aggregation happens at the task struct when you apply different pieces to a task we don't need a structure to allow aggregation.
Quoted 5 lines
Sam said "the NSProxy *is* the container". You appear to be planning to have some namespaces, possibly not aggregated within the nsproxy (pid namespace?) but are you planning to have some higher-level "container" object that aggregates the nsproxy and the other namespaces?
No. A reverse mapping is not needed and is not interesting. As long as I can walk all processes and ask what namespace are you in I don't care.
Quoted 10 lines
If so, what is it? Does it track process membership, etc? What's the userspace API? In what fundamental ways would it be different from my generic containers patches with Serge Hallyn's namespace subsystem. (If these questions are answered already by designs or code, I'd be happy to be pointed at them).
I guess what it comes down to, is why is an aggregation of namespaces suitable for the name "container", when an aggregation of namespaces and other resource controllers isn't?
To me at least the interesting part of your work is not the aggregation portion. But the infrastructure for building the different process groups.
You seem to be calling your groups of processes lumped together for one purpose or another a container.
Quoted 5 lines
What do you think might be a better name for the generic process groups that I'm pushing? As I said, I'm happy to do a simple search/replace on my code to give a different name if that turned out to be the gating factor to getting it merged. But I'd be inclined to leave that decision up to Andrew/Linus.
As we have previously been using the term container is like application. The end product that the user sees built out of individual kernel provided components, but not something that exists in the kernel as such.
We need a term for the non-aggregated case for the individual process groups we build this out of in your infrastructure. Because you clearly have more than one kind of process group a set of processes can belong to.
Quoted 13 lines
Quoted 7 lines
Quoted 4 lines
But there's a lot of common ground between these different approaches, and potential for synergy, so the point of this patch set is to provide a unification point for all of them, and a stepping stone for other new resource controllers and process control modules.
For the case of namespaces I don't see how your code makes things better. I do not see a real problem that you are solving.
I'm trying to solve the problem that lots of different folks (including us) are trying to do things that aggregate multiple process into some kind of constrained group, and are all trying to use different and incompatible ways of grouping/tracking those processes.
Quoted 4 lines
I agree that namespaces fit slightly less well into this model than some other subsystems like resource management. But by integrating with it you'd get automatic access to all the various different resource controller work that's being done.
I don't need to integrate with it to get access to the resource controller work. The resource controller work applies to a set of processes. I have a set of processes. I just need to apply the resource controllers to my set of processes.
At least when we don't talk about the names for these things it is that simple. Now global names I have issues with because almost always the global names fall into their own category, but that is a different story.
Now I have some half backed ideas that might be useful for providing a better abstraction. But it requires setting down and looking at the problem in detail, and understanding what people are trying to accomplish with these things they are building. What subset of process groups do you find interesting.
All that is necessary to have a group of processes do something in an unnamed fashion is to hang a pointer off of the task_struct. That's easy. You are adding a lot more to that, and there is enough abstraction I haven't been able to easily work my way through the infrastructure to what is really going on. Admittedly I haven't looked closely yet.
Eric
Quoted 7 lines · Eric W. Biederman
Quoted 5 lines
Sam said "the NSProxy *is* the container". You appear to be planning to have some namespaces, possibly not aggregated within the nsproxy (pid namespace?) but are you planning to have some higher-level "container" object that aggregates the nsproxy and the other namespaces?
No. A reverse mapping is not needed and is not interesting.
... to you.
Quoted 2 lines
As long as I can walk all processes and ask what namespace are you in I don't care.
How do you currently do that?
Quoted 4 lines
To me at least the interesting part of your work is not the aggregation portion. But the infrastructure for building the different process groups.
In that case you can easily use it to just assign one namespace type to each tree of process groups. The aggregation is something that other groups find useful, but isn't required for the user to actually make use of.
Quoted 3 lines
You seem to be calling your groups of processes lumped together for one purpose or another a container.
Correct.
Quoted 5 lines
We need a term for the non-aggregated case for the individual process groups we build this out of in your infrastructure. Because you clearly have more than one kind of process group a set of processes can belong to.
Yes. That's why my system supports multiple unrelated hierarchies of groups.
Quoted 7 lines
Quoted 4 lines
I agree that namespaces fit slightly less well into this model than some other subsystems like resource management. But by integrating with it you'd get automatic access to all the various different resource controller work that's being done.
I don't need to integrate with it to get access to the resource controller work.
Right, you could certainly do the extra work of tying your virtual servers together with resource controllers in userspace. But you'd still need an API to allow those resource controllers to be associated with groups of processes and configured with limits/guarantees, which is one of the main aims of my containers patch.
Quoted 5 lines
Now I have some half backed ideas that might be useful for providing a better abstraction. But it requires setting down and looking at the problem in detail, and understanding what people are trying to accomplish with these things they are building. What subset of process groups do you find interesting.
I'm primarily interested in getting something in the kernel that can be used as a base for interesting subsystems that apply behavioural or QoS changes to defined groups of processes. Resource controllers and namespaces seem to be good examples of this, but I can think of useful subsystems for monitoring, permission control, etc.
Quoted 4 lines
All that is necessary to have a group of processes do something in an unnamed fashion is to hang a pointer off of the
task_struct. That's easy.
Right, adding a pointer to task_struct is easy. Configuring how/when to not directly inherit it from the parent, or to change it for a running task, or configuring state associated with the thing that the pointer is pointing to, naming that group, and determining which group a given process is assocaited with, is something that's effectively repeated boiler plate for each different subsystem, and which can be accomplished more generically via an abstraction like my containers patch.
Quoted 1 line
You are adding a lot more to that, and there is
The main thing that I'm adding on top of the operations mentioned in the previous paragraph, which pretty much essentially have to be in the kernel, is the ability to group multiple different subsystems together so that they share the same process->container mappings. Yes, that is something that could potentially be done from userspace instead, and just provide an independent tree for each subsystem or namespace. But there seems to be interest from other parties (including me) in having kernel support for it.
Paul
Quoted 4 lines · Paul Menage
Quoted 2 lines
No. A reverse mapping is not needed and is not interesting.
... to you.
You're missing the point of Eric's next sentence. If you can achieve everything you need to achieve and get all the information you are after without it, then it is uninteresting.
Quoted 6 lines
Quoted 3 lines
As long as I can walk all processes and ask what namespace are you in I don't care.
How do you currently do that?
Take a look at /proc/PID/mounts for example.
Quoted 4 lines
Quoted 4 lines
All that is necessary to have a group of processes do something in an unnamed fashion is to hang a pointer off of the
task_struct. That's easy.
Quoted 9 lines
Right, adding a pointer to
task_structis easy. Configuring how/when to not directly inherit it from the parent, or to change it for a running task, or configuring state associated with the thing that the pointer is pointing to, naming that group, and determining which group a given process is assocaited with, is something that's effectively repeated boiler plate for each different subsystem, and which can be accomplished more generically via an abstraction like my containers patch.
So make helpers. Macros. Anything, just don't introduce model limitations like the container structure, because we've already got the structure; the nsproxy.
Sam.
Quoted 9 lines · Sam Vilain
Quoted 4 lines · Paul Menage
Quoted 2 lines
No. A reverse mapping is not needed and is not interesting.
... to you.
You're missing the point of Eric's next sentence. If you can achieve everything you need to achieve and get all the information you are after without it, then it is uninteresting.
Yes, you can do it with an exhaustive trawl of /proc. That can be very expensive on busy machines.
Quoted 9 lines
Quoted 7 lines
Quoted 4 lines
As long as I can walk all processes and ask what namespace are you in I don't care.
How do you currently do that?
Take a look at
/proc/PID/mountsfor example.
That doesn't tell you what mounts namespace a process is in - it tells you what the process can *view* in the namespace.
Quoted 5 lines
So make helpers. Macros. Anything, just don't introduce model limitations like the container structure, because we've already got the structure; the nsproxy.
As I mentioned in another email, nsproxy is fine for things that don't need explicit configuration or reporting, or which already have configuration methods (such as fork(), mount(), etc) available, and which don't need to support movement of processes between different "namespaces". If it was extended to support the naming/config/movement, then it would be fine to use it as the equivalent of a container.
I'd actually be interested in trying to combine my container object and the nsproxy object into a single concept.
Paul
Quoted 5 lines · Paul Menage
Quoted 3 lines · Eric W. Biederman
All that is necessary to have a group of processes do something in an unnamed fashion is to hang a pointer off of the
task_struct. That's easy.
Quoted 8 lines
Right, adding a pointer to
task_structis easy. Configuring how/when to not directly inherit it from the parent, or to change it for a running task, or configuring state associated with the thing that the pointer is pointing to, naming that group, and determining which group a given process is assocaited with, is something that's effectively repeated boiler plate for each different subsystem, and which can be accomplished more generically via an abstraction like my containers patch.
Eric,
what you gain with this patchset is, one very simple container subsystem can tie a container to a cpu, another can limit it's RSS, and suddenly you can
mount -t container -o ns,rss,cpuwhatever ns /container
And each virtual server you create by unsharing can get automatic cpu and rss controls.
That is worthwhile imo.
-serge
Quoted 9 lines · Paul Menage
Quoted 6 lines
Using the container name is bad and it led to this stupid argument.
The fundamental unit of what we have merged into the kernel is the namespace. The aggregate of all namespaces and everything is the container.
What are you defining here as "everything"? If you mean "all things that could be applied to a segregated group of processes such as a virtual server",
The term "segregated group of processes" is too vague. Segregated for what? What is the kernel supposed to do with this information?
Quoted 4 lines
I guess what it comes down to, is why is an aggregation of namespaces suitable for the name "container", when an aggregation of namespaces and other resource controllers isn't?
This argument goes away if you just rename these resource groups to resource namespaces.
Quoted 6 lines
What do you think might be a better name for the generic process groups that I'm pushing? As I said, I'm happy to do a simple search/replace on my code to give a different name if that turned out to be the gating factor to getting it merged. But I'd be inclined to leave that decision up to Andrew/Linus.
Did you like the names I came up with in my original reply?
- CPUset namespace for CPU partitioning
- Resource namespaces:
- cpusched namespace for CPU
- ulimit namespace for memory
- quota namespace for disk space
- io namespace for disk activity
- etc
Quoted 8 lines
Quoted 3 lines
For the case of namespaces I don't see how your code makes things better. I do not see a real problem that you are solving.
I'm trying to solve the problem that lots of different folks (including us) are trying to do things that aggregate multiple process into some kind of constrained group, and are all trying to use different and incompatible ways of grouping/tracking those processes.
Maybe what's missing is a set of helper macros/functions that assist with writing new namespaces. Perhaps you can give some more examples and we can consider these on a case by case basis.
Sam.
Quoted 3 lines · Sam Vilain
The term "segregated group of processes" is too vague. Segregated for what? What is the kernel supposed to do with this information?
The generic part of the kernel just keeps track of the fact that they're segregated (and their children, etc).
It's the clients of this subsystem (virtual servers, resource controllers) that can decide to give different per-process behaviour based on the membership of various groups.
Quoted 9 lines
Did you like the names I came up with in my original reply?
- CPUset namespace for CPU partitioning
- Resource namespaces:
- cpusched namespace for CPU
- ulimit namespace for memory
- quota namespace for disk space
- io namespace for disk activity
- etc
This is a strange abuse of the term "namespace".
http://en.wikipedia.org/wiki/Namespace_%28computer_science%29
For the virtual server work that you're doing, namespace is a good term:
pids name processes, hence a pid namespace lets you have multiple distinct mappings from pids to processes
filenames name files, so a filename (or mount) namespace lets you have multiple distinct mappings from filenames to files.
For resource QoS control, it doesn't really make sense to talk about namespaces. We're not virtualizing resources to rename them for different virtual servers, we're limiting the quality of access to the resources.
But the semantics of the term "namespace" notwithstanding, you're equating a virtual server namespace (pid, ipc, uts, mounts, etc) with a resource controller (memory, I/O, etc) in terms of their place in a hierarchy, which is something I agree with. All of these subsystems can be considered to be units that can be associated with groups of processes; the ultimate grouping of processes is something that we're both ultimately referring to as a container.
Quoted 3 lines
Maybe what's missing is a set of helper macros/functions that assist with writing new namespaces.
That's pretty much what my containers patch does - provides a generic way to associate state (what you're referring to as a namespace) with groups of processes and a consistent user-space API for manipulating them.
Paul
Quoted 10 lines · Paul Menage
Quoted 3 lines
The term "segregated group of processes" is too vague. Segregated for what? What is the kernel supposed to do with this information?
The generic part of the kernel just keeps track of the fact that they're segregated (and their children, etc).
It's the clients of this subsystem (virtual servers, resource controllers) that can decide to give different per-process behaviour based on the membership of various groups.
So those clients can use helper functions to use their own namespaces. If they happen to group things in the same way they they'll all end up using the same nsproxy.
Quoted 26 lines
Quoted 10 lines
Did you like the names I came up with in my original reply?
- CPUset namespace for CPU partitioning
- Resource namespaces:
- cpusched namespace for CPU
- ulimit namespace for memory
- quota namespace for disk space
- io namespace for disk activity
- etcThis is a strange abuse of the term "namespace".
http://en.wikipedia.org/wiki/Namespace_%28computer_science%29For the virtual server work that you're doing, namespace is a good term:
pids name processes, hence a pid namespace lets you have multiple distinct mappings from pids to processes
filenames name files, so a filename (or mount) namespace lets you have multiple distinct mappings from filenames to files.For resource QoS control, it doesn't really make sense to talk about namespaces. We're not virtualizing resources to rename them for different virtual servers, we're limiting the quality of access to the resources.
Quoted 8 lines
But the semantics of the term "namespace" notwithstanding, you're equating a virtual server namespace (pid, ipc, uts, mounts, etc) with a resource controller (memory, I/O, etc) in terms of their place in a hierarchy, which is something I agree with. All of these subsystems can be considered to be units that can be associated with groups of processes; the ultimate grouping of processes is something that we're both ultimately referring to as a container.
I don't necessarily agree with the 'heirarchy' bit. It doesn't have to be so segregated. But I think we already covered that in this thread.
I agree with the comment on the abuse of the term "namespace", though consider that it has already been abused with the term IPC namespaces. We have for some time been using it to refer to groupable entities within the kernel that are associated with tasks, even if they don't involve named entities that clash within a particular domain. But there is always an entity and a domain, and that is the key point I'm trying to make - the features you are putting forward are no different to the examples that we made specifically for the purpose of setting the standard for further features to follow.
We talked about naming a bit before, see
http://lkml.org/lkml/2006/3/21/403 and possibly other threads. So, anyway, feel free to flog this old dead horse and suggest different terms. We've all had long enough to think about it since so maybe it's worth it, but with any new term it should be really darned clear that they're essentially the same thing as namespaces, or otherwise be really likable.
Sam.
Quoted 3 lines · Sam Vilain
I don't necessarily agree with the 'heirarchy' bit. It doesn't have to be so segregated. But I think we already covered that in this thread.
OK, but it's much easier to use a hierarchical system as a flat system (just don't create children) than a flat system as a hierarchical system. And others do seem to want a hierarchy for their process groupings.
Quoted 3 lines
I agree with the comment on the abuse of the term "namespace", though consider that it has already been abused with the term IPC namespaces.
That doesn't seem like an abuse to me at all - you're controlling what IPC object a given name (shm_id, sem_id or msg_id) refers to for any given group of processes.
Quoted 7 lines
We have for some time been using it to refer to groupable entities within the kernel that are associated with tasks, even if they don't involve named entities that clash within a particular domain. But there is always an entity and a domain, and that is the key point I'm trying to make - the features you are putting forward are no different to the examples that we made specifically for the purpose of setting the standard for further features to follow.
They're very similar, I agree. An important difference is that things like pid/mount namespaces are simply ways of controlling the visibility of existing unix concepts, such as processes or filesystems. You don't need additional configuration to make them useful, as unix already has standard ways of manipulating them.
Things like resource controllers typically require additional configuration to control how much resources each group of processes can consume, etc. Also, it appears to be much more common to want to move tasks between different resource controllers than to move them between different namespaces. (And in order to configure and move, you need to be able to name)
The configuration, naming and movement are the features that my container patch provides on top of the features that nsproxy provides for namespaces.
Quoted 5 lines
anyway, feel free to flog this old dead horse and suggest different terms. We've all had long enough to think about it since so maybe it's worth it, but with any new term it should be really darned clear that they're essentially the same thing as namespaces, or otherwise be really likable.
What you're calling a "namespace", I'm calling a "subsystem state". Essentially they're the same thing. The important thing that generic containers provide is a standard way to manipulate "subsystem states" or "namespaces".
Paul
Quoted 6 lines · Sam Vilain
Paul Menage wrote:
We talked about naming a bit before, see
http://lkml.org/lkml/2006/3/21/403 and possibly other threads. So, anyway, feel free to flog this old dead horse and suggest different terms. We've all had long enough to think about it since so maybe it's
Bah,
Paul, how about s/container/rug/, as in resource usage group?
At least you can be pretty sure that (a) rug won't conflict with any other computer science therm, and (b) for that exact reason it should be pretty memorable.
-serge