user/sven/linux.git/kernel/kthread.c, branch v4.9.51

cgroup, kthread: close race window where new kthreads can be migrated to non-root cgroups

2017-04-21T07:31:18Z

commit 77f88796cee819b9c4562b0b6b44691b3b7755b1 upstream. Creation of a kthread goes through a couple interlocked stages between the kthread itself and its creator. Once the new kthread starts running, it initializes itself and wakes up the creator. The creator then can further configure the kthread and then let it start doing its job by waking it up. In this configuration-by-creator stage, the creator is the only one that can wake it up but the kthread is visible to userland. When altering the kthread's attributes from userland is allowed, this is fine; however, for cases where CPU affinity is critical, kthread_bind() is used to first disable affinity changes from userland and then set the affinity. This also prevents the kthread from being migrated into non-root cgroups as that can affect the CPU affinity and many other things. Unfortunately, the cgroup side of protection is racy. While the PF_NO_SETAFFINITY flag prevents further migrations, userland can win the race before the creator sets the flag with kthread_bind() and put the kthread in a non-root cgroup, which can lead to all sorts of problems including incorrect CPU affinity and starvation. This bug got triggered by userland which periodically tries to migrate all processes in the root cpuset cgroup to a non-root one. Per-cpu workqueue workers got caught while being created and ended up with incorrected CPU affinity breaking concurrency management and sometimes stalling workqueue execution. This patch adds task->no_cgroup_migration which disallows the task to be migrated by userland. kthreadd starts with the flag set making every child kthread start in the root cgroup with migration disallowed. The flag is cleared after the kthread finishes initialization by which time PF_NO_SETAFFINITY is set if the kthread should stay in the root cgroup. It'd be better to wait for the initialization instead of failing but I couldn't think of a way of implementing that without adding either a new PF flag, or sleeping and retrying from waiting side. Even if userland depends on changing cgroup membership of a kthread, it either has to be synchronized with kthread_create() or periodically repeat, so it's unlikely that this would break anything. v2: Switch to a simpler implementation using a new task_struct bit field suggested by Oleg. Signed-off-by: Tejun Heo Suggested-by: Oleg Nesterov Cc: Linus Torvalds Cc: Andrew Morton Cc: Peter Zijlstra (Intel) Cc: Thomas Gleixner Reported-and-debugged-by: Chris Mason Signed-off-by: Tejun Heo Signed-off-by: Greg Kroah-Hartman

kthread: better support freezable kthread workers

2016-10-11T22:06:33Z

This patch allows to make kthread worker freezable via a new @flags parameter. It will allow to avoid an init work in some kthreads. It currently does not affect the function of kthread_worker_fn() but it might help to do some optimization or fixes eventually. I currently do not know about any other use for the @flags parameter but I believe that we will want more flags in the future. Finally, I hope that it will not cause confusion with @flags member in struct kthread. Well, I guess that we will want to rework the basic kthreads implementation once all kthreads are converted into kthread workers or workqueues. It is possible that we will merge the two structures. Link: http://lkml.kernel.org/r/1470754545-17632-12-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: allow to modify delayed kthread work

2016-10-11T22:06:33Z

There are situations when we need to modify the delay of a delayed kthread work. For example, when the work depends on an event and the initial delay means a timeout. Then we want to queue the work immediately when the event happens. This patch implements kthread_mod_delayed_work() as inspired workqueues. It cancels the timer, removes the work from any worker list and queues it again with the given timeout. A very special case is when the work is being canceled at the same time. It might happen because of the regular kthread_cancel_delayed_work_sync() or by another kthread_mod_delayed_work(). In this case, we do nothing and let the other operation win. This should not normally happen as the caller is supposed to synchronize these operations a reasonable way. Link: http://lkml.kernel.org/r/1470754545-17632-11-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: allow to cancel kthread work

2016-10-11T22:06:33Z

We are going to use kthread workers more widely and sometimes we will need to make sure that the work is neither pending nor running. This patch implements cancel_*_sync() operations as inspired by workqueues. Well, we are synchronized against the other operations via the worker lock, we use del_timer_sync() and a counter to count parallel cancel operations. Therefore the implementation might be easier. First, we check if a worker is assigned. If not, the work has newer been queued after it was initialized. Second, we take the worker lock. It must be the right one. The work must not be assigned to another worker unless it is initialized in between. Third, we try to cancel the timer when it exists. The timer is deleted synchronously to make sure that the timer call back is not running. We need to temporary release the worker->lock to avoid a possible deadlock with the callback. In the meantime, we set work->canceling counter to avoid any queuing. Fourth, we try to remove the work from a worker list. It might be the list of either normal or delayed works. Fifth, if the work is running, we call kthread_flush_work(). It might take an arbitrary time. We need to release the worker-lock again. In the meantime, we again block any queuing by the canceling counter. As already mentioned, the check for a pending kthread work is done under a lock. In compare with workqueues, we do not need to fight for a single PENDING bit to block other operations. Therefore we do not suffer from the thundering storm problem and all parallel canceling jobs might use kthread_flush_work(). Any queuing is blocked until the counter gets zero. Link: http://lkml.kernel.org/r/1470754545-17632-10-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: initial support for delayed kthread work

2016-10-11T22:06:33Z

We are going to use kthread_worker more widely and delayed works will be pretty useful. The implementation is inspired by workqueues. It uses a timer to queue the work after the requested delay. If the delay is zero, the work is queued immediately. In compare with workqueues, each work is associated with a single worker (kthread). Therefore the implementation could be much easier. In particular, we use the worker->lock to synchronize all the operations with the work. We do not need any atomic operation with a flags variable. In fact, we do not need any state variable at all. Instead, we add a list of delayed works into the worker. Then the pending work is listed either in the list of queued or delayed works. And the existing check of pending works is the same even for the delayed ones. A work must not be assigned to another worker unless reinitialized. Therefore the timer handler might expect that dwork->work->worker is valid and it could simply take the lock. We just add some sanity checks to help with debugging a potential misuse. Link: http://lkml.kernel.org/r/1470754545-17632-9-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: detect when a kthread work is used by more workers

2016-10-11T22:06:33Z

Nothing currently prevents a work from queuing for a kthread worker when it is already running on another one. This means that the work might run in parallel on more than one worker. Also some operations are not reliable, e.g. flush. This problem will be even more visible after we add kthread_cancel_work() function. It will only have "work" as the parameter and will use worker->lock to synchronize with others. Well, normally this is not a problem because the API users are sane. But bugs might happen and users also might be crazy. This patch adds a warning when we try to insert the work for another worker. It does not fully prevent the misuse because it would make the code much more complicated without a big benefit. It adds the same warning also into kthread_flush_work() instead of the repeated attempts to get the right lock. A side effect is that one needs to explicitly reinitialize the work if it must be queued into another worker. This is needed, for example, when the worker is stopped and started again. It is a bit inconvenient. But it looks like a good compromise between the stability and complexity. I have double checked all existing users of the kthread worker API and they all seems to initialize the work after the worker gets started. Just for completeness, the patch adds a check that the work is not already in a queue. The patch also puts all the checks into a separate function. It will be reused when implementing delayed works. Link: http://lkml.kernel.org/r/1470754545-17632-8-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Cc: Oleg Nesterov Cc: Tejun Heo Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: add kthread_destroy_worker()

2016-10-11T22:06:33Z

The current kthread worker users call flush() and stop() explicitly. This function does the same plus it frees the kthread_worker struct in one call. It is supposed to be used together with kthread_create_worker*() that allocates struct kthread_worker. Link: http://lkml.kernel.org/r/1470754545-17632-7-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Cc: Oleg Nesterov Cc: Tejun Heo Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: add kthread_create_worker*()

2016-10-11T22:06:33Z

Kthread workers are currently created using the classic kthread API, namely kthread_run(). kthread_worker_fn() is passed as the @threadfn parameter. This patch defines kthread_create_worker() and kthread_create_worker_on_cpu() functions that hide implementation details. They enforce using kthread_worker_fn() for the main thread. But I doubt that there are any plans to create any alternative. In fact, I think that we do not want any alternative main thread because it would be hard to support consistency with the rest of the kthread worker API. The naming and function of kthread_create_worker() is inspired by the workqueues API like the rest of the kthread worker API. The kthread_create_worker_on_cpu() variant is motivated by the original kthread_create_on_cpu(). Note that we need to bind per-CPU kthread workers already when they are created. It makes the life easier. kthread_bind() could not be used later for an already running worker. This patch does _not_ convert existing kthread workers. The kthread worker API need more improvements first, e.g. a function to destroy the worker. IMPORTANT: kthread_create_worker_on_cpu() allows to use any format of the worker name, in compare with kthread_create_on_cpu(). The good thing is that it is more generic. The bad thing is that most users will need to pass the cpu number in two parameters, e.g. kthread_create_worker_on_cpu(cpu, "helper/%d", cpu). To be honest, the main motivation was to avoid the need for an empty va_list. The only legal way was to create a helper function that would be called with an empty list. Other attempts caused compilation warnings or even errors on different architectures. There were also other alternatives, for example, using #define or splitting __kthread_create_worker(). The used solution looked like the least ugly. Link: http://lkml.kernel.org/r/1470754545-17632-6-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread: allow to call __kthread_create_on_node() with va_list args

2016-10-11T22:06:33Z

kthread_create_on_node() implements a bunch of logic to create the kthread. It is already called by kthread_create_on_cpu(). We are going to extend the kthread worker API and will need to call kthread_create_on_node() with va_list args there. This patch does only a refactoring and does not modify the existing behavior. Link: http://lkml.kernel.org/r/1470754545-17632-5-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Acked-by: Tejun Heo Cc: Oleg Nesterov Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Thomas Gleixner Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

kthread/smpboot: do not park in kthread_create_on_cpu()

2016-10-11T22:06:33Z

kthread_create_on_cpu() was added by the commit 2a1d446019f9a5983e ("kthread: Implement park/unpark facility"). It is currently used only when enabling new CPU. For this purpose, the newly created kthread has to be parked. The CPU binding is a bit tricky. The kthread is parked when the CPU has not been allowed yet. And the CPU is bound when the kthread is unparked. The function would be useful for more per-CPU kthreads, e.g. bnx2fc_thread, fcoethread. For this purpose, the newly created kthread should stay in the uninterruptible state. This patch moves the parking into smpboot. It binds the thread already when created. Then the function might be used universally. Also the behavior is consistent with kthread_create() and kthread_create_on_node(). Link: http://lkml.kernel.org/r/1470754545-17632-4-git-send-email-pmladek@suse.com Signed-off-by: Petr Mladek Reviewed-by: Thomas Gleixner Cc: Oleg Nesterov Cc: Tejun Heo Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Steven Rostedt Cc: "Paul E. McKenney" Cc: Josh Triplett Cc: Jiri Kosina Cc: Borislav Petkov Cc: Michal Hocko Cc: Vlastimil Babka Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds