user/sven/linux.git/kernel/exit.c, branch v5.14.9

Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace

2021-06-29T03:39:26Z

Pull user namespace rlimit handling update from Eric Biederman: "This is the work mainly by Alexey Gladkov to limit rlimits to the rlimits of the user that created a user namespace, and to allow users to have stricter limits on the resources created within a user namespace." * 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace: cred: add missing return error code when set_cred_ucounts() failed ucounts: Silence warning in dec_rlimit_ucounts ucounts: Set ucount_max to the largest positive value the type can hold kselftests: Add test to check for rlimit changes in different user namespaces Reimplement RLIMIT_MEMLOCK on top of ucounts Reimplement RLIMIT_SIGPENDING on top of ucounts Reimplement RLIMIT_MSGQUEUE on top of ucounts Reimplement RLIMIT_NPROC on top of ucounts Use atomic_t for ucounts reference counting Add a reference to ucounts for each cred Increase size of ucounts to atomic_long_t

Revert "signal: Allow tasks to cache one sigqueue struct"

2021-06-27T20:32:54Z

This reverts commits 4bad58ebc8bc4f20d89cff95417c9b4674769709 (and 399f8dd9a866e107639eabd3c1979cd526ca3a98, which tried to fix it). I do not believe these are correct, and I'm about to release 5.13, so am reverting them out of an abundance of caution. The locking is odd, and appears broken. On the allocation side (in __sigqueue_alloc()), the locking is somewhat straightforward: it depends on sighand->siglock. Since one caller doesn't hold that lock, it further then tests 'sigqueue_flags' to avoid the case with no locks held. On the freeing side (in sigqueue_cache_or_free()), there is no locking at all, and the logic instead depends on 'current' being a single thread, and not able to race with itself. To make things more exciting, there's also the data race between freeing a signal and allocating one, which is handled by using WRITE_ONCE() and READ_ONCE(), and being mutually exclusive wrt the initial state (ie freeing will only free if the old state was NULL, while allocating will obviously only use the value if it was non-NULL, so only one or the other will actually act on the value). However, while the free->alloc paths do seem mutually exclusive thanks to just the data value dependency, it's not clear what the memory ordering constraints are on it. Could writes from the previous allocation possibly be delayed and seen by the new allocation later, causing logical inconsistencies? So it's all very exciting and unusual. And in particular, it seems that the freeing side is incorrect in depending on "current" being single-threaded. Yes, 'current' is a single thread, but in the presense of asynchronous events even a single thread can have data races. And such asynchronous events can and do happen, with interrupts causing signals to be flushed and thus free'd (for example - sending a SIGCONT/SIGSTOP can happen from interrupt context, and can flush previously queued process control signals). So regardless of all the other questions about the memory ordering and locking for this new cached allocation, the sigqueue_cache_or_free() assumptions seem to be fundamentally incorrect. It may be that people will show me the errors of my ways, and tell me why this is all safe after all. We can reinstate it if so. But my current belief is that the WRITE_ONCE() that sets the cached entry needs to be a smp_store_release(), and the READ_ONCE() that finds a cached entry needs to be a smp_load_acquire() to handle memory ordering correctly. And the sequence in sigqueue_cache_or_free() would need to either use a lock or at least be interrupt-safe some way (perhaps by using something like the percpu 'cmpxchg': it doesn't need to be SMP-safe, but like the percpu operations it needs to be interrupt-safe). Fixes: 399f8dd9a866 ("signal: Prevent sigqueue caching after task got released") Fixes: 4bad58ebc8bc ("signal: Allow tasks to cache one sigqueue struct") Cc: Thomas Gleixner Cc: Peter Zijlstra Cc: Oleg Nesterov Cc: Christian Brauner Signed-off-by: Linus Torvalds

do_wait: make PIDTYPE_PID case O(1) instead of O(n)

2021-05-07T02:24:13Z

Add a special-case when waiting on a pid (via waitpid, waitid, wait4, etc) to avoid doing an O(n) scan of children and tracees, and instead do an O(1) lookup. This improves performance when waiting on a pid from a thread group with many children and/or tracees. Time to fork and then call waitpid on the child, from a task that already has N children [1]: N | Before | After -----|---------|------ 1 | 74 us | 74 us 20 | 72 us | 75 us 100 | 83 us | 77 us 500 | 99 us | 74 us 1000 | 179 us | 75 us 5000 | 804 us | 79 us 8000 | 1268 us | 78 us [1]: https://lkml.org/lkml/2021/3/12/1567 This can make a substantial performance improvement for applications with a thread that has many children or tracees and frequently needs to wait on them. Tools that use ptrace to intercept syscalls for a large number of processes are likely to fall into this category. In particular this patch was developed while building a ptrace-based second generation of the Shadow emulator [2], for which it allows us to avoid quadratic scaling (without having to use a workaround that introduces a ~40% performance penalty) [3]. Other examples of tools that fall into this category which this patch may help include User Mode Linux [4] and DetTrace [5]. [2]: https://shadow.github.io/ [3]: https://github.com/shadow/shadow/issues/1134#issuecomment-798992292 [4]: https://en.wikipedia.org/wiki/User-mode_Linux [5]: https://github.com/dettrace/dettrace Link: https://lkml.kernel.org/r/20210314231544.9379-1-jnewsome@torproject.org Signed-off-by: James Newsome Reviewed-by: Oleg Nesterov Cc: "Eric W . Biederman" Cc: Christian Brauner Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds

Reimplement RLIMIT_NPROC on top of ucounts

2021-04-30T19:14:01Z

The rlimit counter is tied to uid in the user_namespace. This allows rlimit values to be specified in userns even if they are already globally exceeded by the user. However, the value of the previous user_namespaces cannot be exceeded. To illustrate the impact of rlimits, let's say there is a program that does not fork. Some service-A wants to run this program as user X in multiple containers. Since the program never fork the service wants to set RLIMIT_NPROC=1. service-A \- program (uid=1000, container1, rlimit_nproc=1) \- program (uid=1000, container2, rlimit_nproc=1) The service-A sets RLIMIT_NPROC=1 and runs the program in container1. When the service-A tries to run a program with RLIMIT_NPROC=1 in container2 it fails since user X already has one running process. We cannot use existing inc_ucounts / dec_ucounts because they do not allow us to exceed the maximum for the counter. Some rlimits can be overlimited by root or if the user has the appropriate capability. Changelog v11: * Change inc_rlimit_ucounts() which now returns top value of ucounts. * Drop inc_rlimit_ucounts_and_test() because the return code of inc_rlimit_ucounts() can be checked. Signed-off-by: Alexey Gladkov Link: https://lkml.kernel.org/r/c5286a8aa16d2d698c222f7532f3d735c82bc6bc.1619094428.git.legion@kernel.org Signed-off-by: Eric W. Biederman

signal: Allow tasks to cache one sigqueue struct

2021-04-14T16:04:08Z

The idea for this originates from the real time tree to make signal delivery for realtime applications more efficient. In quite some of these application scenarios a control tasks signals workers to start their computations. There is usually only one signal per worker on flight. This works nicely as long as the kmem cache allocations do not hit the slow path and cause latencies. To cure this an optimistic caching was introduced (limited to RT tasks) which allows a task to cache a single sigqueue in a pointer in task_struct instead of handing it back to the kmem cache after consuming a signal. When the next signal is sent to the task then the cached sigqueue is used instead of allocating a new one. This solved the problem for this set of application scenarios nicely. The task cache is not preallocated so the first signal sent to a task goes always to the cache allocator. The cached sigqueue stays around until the task exits and is freed when task::sighand is dropped. After posting this solution for mainline the discussion came up whether this would be useful in general and should not be limited to realtime tasks: https://lore.kernel.org/r/m11rcu7nbr.fsf@fess.ebiederm.org One concern leading to the original limitation was to avoid a large amount of pointlessly cached sigqueues in alive tasks. The other concern was vs. RLIMIT_SIGPENDING as these cached sigqueues are not accounted for. The accounting problem is real, but on the other hand slightly academic. After gathering some statistics it turned out that after boot of a regular distro install there are less than 10 sigqueues cached in ~1500 tasks. In case of a 'mass fork and fire signal to child' scenario the extra 80 bytes of memory per task are well in the noise of the overall memory consumption of the fork bomb. If this should be limited then this would need an extra counter in struct user, more atomic instructions and a seperate rlimit. Yet another tunable which is mostly unused. The caching is actually used. After boot and a full kernel compile on a 64CPU machine with make -j128 the number of 'allocations' looks like this: From slab: 23996 From task cache: 52223 I.e. it reduces the number of slab cache operations by ~68%. A typical pattern there is: <...>-58490 __sigqueue_alloc: for 58488 from slab ffff8881132df460 <...>-58488 __sigqueue_free: cache ffff8881132df460 <...>-58488 __sigqueue_alloc: for 1149 from cache ffff8881103dc550 bash-1149 exit_task_sighand: free ffff8881132df460 bash-1149 __sigqueue_free: cache ffff8881103dc550 The interesting sequence is that the exiting task 58488 grabs the sigqueue from bash's task cache to signal exit and bash sticks it back into it's own cache. Lather, rinse and repeat. The caching is probably not noticable for the general use case, but the benefit for latency sensitive applications is clear. While kmem caches are usually just serving from the fast path the slab merging (default) can depending on the usage pattern of the merged slabs cause occasional slow path allocations. The time spared per cached entry is a few micro seconds per signal which is not relevant for e.g. a kernel build, but for signal heavy workloads it's measurable. As there is no real downside of this caching mechanism making it unconditionally available is preferred over more conditional code or new magic tunables. Signed-off-by: Thomas Gleixner Signed-off-by: Peter Zijlstra (Intel) Reviewed-by: Oleg Nesterov Link: https://lkml.kernel.org/r/87sg4lbmxo.fsf@nanos.tec.linutronix.de

kernel/io_uring: cancel io_uring before task works

2020-12-31T02:36:54Z

For cancelling io_uring requests it needs either to be able to run currently enqueued task_works or having it shut down by that moment. Otherwise io_uring_cancel_files() may be waiting for requests that won't ever complete. Go with the first way and do cancellations before setting PF_EXITING and so before putting the task_work infrastructure into a transition state where task_work_run() would better not be called. Cc: stable@vger.kernel.org # 5.5+ Signed-off-by: Pavel Begunkov Signed-off-by: Jens Axboe

Merge branch 'linus' into sched/core, to resolve semantic conflict

2020-11-27T10:10:50Z

Signed-off-by: Ingo Molnar

sched: fix exit_mm vs membarrier (v4)

2020-10-29T10:00:30Z

exit_mm should issue memory barriers after user-space memory accesses, before clearing current->mm, to order user-space memory accesses performed prior to exit_mm before clearing tsk->mm, which has the effect of skipping the membarrier private expedited IPIs. exit_mm should also update the runqueue's membarrier_state so membarrier global expedited IPIs are not sent when they are not needed. The membarrier system call can be issued concurrently with do_exit if we have thread groups created with CLONE_VM but not CLONE_THREAD. Here is the scenario I have in mind: Two thread groups are created, A and B. Thread group B is created by issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD. Let's assume we have a single thread within each thread group (Thread A and Thread B). The AFAIU we can have: Userspace variables: int x = 0, y = 0; CPU 0 CPU 1 Thread A Thread B (in thread group A) (in thread group B) x = 1 barrier() y = 1 exit() exit_mm() current->mm = NULL; r1 = load y membarrier() skips CPU 0 (no IPI) because its current mm is NULL r2 = load x BUG_ON(r1 == 1 && r2 == 0) Signed-off-by: Mathieu Desnoyers Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20201020134715.13909-2-mathieu.desnoyers@efficios.com

don't dump the threads that had been already exiting when zapped.

2020-10-28T20:39:49Z

Coredump logics needs to report not only the registers of the dumping thread, but (since 2.5.43) those of other threads getting killed. Doing that might require extra state saved on the stack in asm glue at kernel entry; signal delivery logics does that (we need to be able to save sigcontext there, at the very least) and so does seccomp. That covers all callers of do_coredump(). Secondary threads get hit with SIGKILL and caught as soon as they reach exit_mm(), which normally happens in signal delivery, so those are also fine most of the time. Unfortunately, it is possible to end up with secondary zapped when it has already entered exit(2) (or, worse yet, is oopsing). In those cases we reach exit_mm() when mm->core_state is already set, but the stack contents is not what we would have in signal delivery. At least on two architectures (alpha and m68k) it leads to infoleaks - we end up with a chunk of kernel stack written into coredump, with the contents consisting of normal C stack frames of the call chain leading to exit_mm() instead of the expected copy of userland registers. In case of alpha we leak 312 bytes of stack. Other architectures (including the regset-using ones) might have similar problems - the normal user of regsets is ptrace and the state of tracee at the time of such calls is special in the same way signal delivery is. Note that had the zapper gotten to the exiting thread slightly later, it wouldn't have been included into coredump anyway - we skip the threads that have already cleared their ->mm. So let's pretend that zapper always loses the race. IOW, have exit_mm() only insert into the dumper list if we'd gotten there from handling a fatal signal[*] As the result, the callers of do_exit() that have *not* gone through get_signal() are not seen by coredump logics as secondary threads. Which excludes voluntary exit()/oopsen/traps/etc. The dumper thread itself is unaffected by that, so seccomp is fine. [*] originally I intended to add a new flag in tsk->flags, but ebiederman pointed out that PF_SIGNALED is already doing just what we need. Cc: stable@vger.kernel.org Fixes: d89f3847def4 ("[PATCH] thread-aware coredumps, 2.5.43-C3") History-tree: https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git Acked-by: "Eric W. Biederman" Signed-off-by: Al Viro

pid: move pidfd_get_pid() to pid.c

2020-10-18T16:27:10Z

process_madvise syscall needs pidfd_get_pid function to translate pidfd to pid so this patch move the function to kernel/pid.c. Suggested-by: Alexander Duyck Signed-off-by: Minchan Kim Signed-off-by: Andrew Morton Reviewed-by: Suren Baghdasaryan Reviewed-by: Alexander Duyck Reviewed-by: Vlastimil Babka Acked-by: Christian Brauner Acked-by: David Rientjes Cc: Jens Axboe Cc: Jann Horn Cc: Brian Geffon Cc: Daniel Colascione Cc: Joel Fernandes Cc: Johannes Weiner Cc: John Dias Cc: Kirill Tkhai Cc: Michal Hocko Cc: Oleksandr Natalenko Cc: Sandeep Patil Cc: SeongJae Park Cc: SeongJae Park Cc: Shakeel Butt Cc: Sonny Rao Cc: Tim Murray Cc: Christian Brauner Cc: Florian Weimer Cc: Link: http://lkml.kernel.org/r/20200302193630.68771-5-minchan@kernel.org Link: http://lkml.kernel.org/r/20200622192900.22757-3-minchan@kernel.org Link: https://lkml.kernel.org/r/20200901000633.1920247-3-minchan@kernel.org Signed-off-by: Linus Torvalds