Thorvald reported a WARNING [1]. And the root cause is below race:
CPU 1 CPU 2
fork hugetlbfs_fallocate
dup_mmap hugetlbfs_punch_hole
i_mmap_lock_write(mapping);
vma_interval_tree_insert_after -- Child vma is visible through i_mmap tree.
i_mmap_unlock_write(mapping);
hugetlb_dup_vma_private -- Clear vma_lock outside i_mmap_rwsem!
i_mmap_lock_write(mapping);
hugetlb_vmdelete_list
vma_interval_tree_foreach
hugetlb_vma_trylock_write -- Vma_lock is cleared.
tmp->vm_ops->open -- Alloc new vma_lock outside i_mmap_rwsem!
hugetlb_vma_unlock_write -- Vma_lock is assigned!!!
i_mmap_unlock_write(mapping);
hugetlb_dup_vma_private() and hugetlb_vm_op_open() are called outside
i_mmap_rwsem lock while vma lock can be used in the same time. Fix this
by deferring linking file vma until vma is fully initialized. Those vmas
should be initialized first before they can be used.
Link: https://lkml.kernel.org/r/20240410091441.3539905-1-linmiaohe@huawei.com
Fixes: 8d9bfb260814 ("hugetlb: add vma based lock for pmd sharing")
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Reported-by: Thorvald Natvig <thorvald@google.com>
Closes: https://lore.kernel.org/linux-mm/20240129161735.6gmjsswx62o4pbja@revolver/T/ [1]
Reviewed-by: Jane Chu <jane.chu@oracle.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peng Zhang <zhangpeng.00@bytedance.com>
Cc: Tycho Andersen <tandersen@netflix.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Pull RCU updates from Boqun Feng:
- Eliminate deadlocks involving do_exit() and RCU tasks, by Paul:
Instead of SRCU read side critical sections, now a percpu list is
used in do_exit() for scaning yet-to-exit tasks
- Fix a deadlock due to the dependency between workqueue and RCU
expedited grace period, reported by Anna-Maria Behnsen and Thomas
Gleixner and fixed by Frederic: Now RCU expedited always uses its own
kthread worker instead of a workqueue
- RCU NOCB updates, code cleanups, unnecessary barrier removals and
minor bug fixes
- Maintain real-time response in rcu_tasks_postscan() and a minor fix
for tasks trace quiescence check
- Misc updates, comments and readibility improvement, boot time
parameter for lazy RCU and rcutorture improvement
- Documentation updates
* tag 'rcu.next.v6.9' of git://git.kernel.org/pub/scm/linux/kernel/git/boqun/linux: (34 commits)
rcu-tasks: Maintain real-time response in rcu_tasks_postscan()
rcu-tasks: Eliminate deadlocks involving do_exit() and RCU tasks
rcu-tasks: Maintain lists to eliminate RCU-tasks/do_exit() deadlocks
rcu-tasks: Initialize data to eliminate RCU-tasks/do_exit() deadlocks
rcu-tasks: Initialize callback lists at rcu_init() time
rcu-tasks: Add data to eliminate RCU-tasks/do_exit() deadlocks
rcu-tasks: Repair RCU Tasks Trace quiescence check
rcu/sync: remove un-used rcu_sync_enter_start function
rcutorture: Suppress rtort_pipe_count warnings until after stalls
srcu: Improve comments about acceleration leak
rcu: Provide a boot time parameter to control lazy RCU
rcu: Rename jiffies_till_flush to jiffies_lazy_flush
doc: Update checklist.rst discussion of callback execution
doc: Clarify use of slab constructors and SLAB_TYPESAFE_BY_RCU
context_tracking: Fix kerneldoc headers for __ct_user_{enter,exit}()
doc: Add EARLY flag to early-parsed kernel boot parameters
doc: Add CONFIG_RCU_STRICT_GRACE_PERIOD to checklist.rst
doc: Make checklist.rst note that spinlocks are implied RCU readers
doc: Make whatisRCU.rst note that spinlocks are RCU readers
doc: Spinlocks are implied RCU readers
...
Pull pdfd updates from Christian Brauner:
- Until now pidfds could only be created for thread-group leaders but
not for threads. There was no technical reason for this. We simply
had no users that needed support for this. Now we do have users that
need support for this.
This introduces a new PIDFD_THREAD flag for pidfd_open(). If that
flag is set pidfd_open() creates a pidfd that refers to a specific
thread.
In addition, we now allow clone() and clone3() to be called with
CLONE_PIDFD | CLONE_THREAD which wasn't possible before.
A pidfd that refers to an individual thread differs from a pidfd that
refers to a thread-group leader:
(1) Pidfds are pollable. A task may poll a pidfd and get notified
when the task has exited.
For thread-group leader pidfds the polling task is woken if the
thread-group is empty. In other words, if the thread-group
leader task exits when there are still threads alive in its
thread-group the polling task will not be woken when the
thread-group leader exits but rather when the last thread in the
thread-group exits.
For thread-specific pidfds the polling task is woken if the
thread exits.
(2) Passing a thread-group leader pidfd to pidfd_send_signal() will
generate thread-group directed signals like kill(2) does.
Passing a thread-specific pidfd to pidfd_send_signal() will
generate thread-specific signals like tgkill(2) does.
The default scope of the signal is thus determined by the type
of the pidfd.
Since use-cases exist where the default scope of the provided
pidfd needs to be overriden the following flags are added to
pidfd_send_signal():
- PIDFD_SIGNAL_THREAD
Send a thread-specific signal.
- PIDFD_SIGNAL_THREAD_GROUP
Send a thread-group directed signal.
- PIDFD_SIGNAL_PROCESS_GROUP
Send a process-group directed signal.
The scope change will only work if the struct pid is actually
used for this scope.
For example, in order to send a thread-group directed signal the
provided pidfd must be used as a thread-group leader and
similarly for PIDFD_SIGNAL_PROCESS_GROUP the struct pid must be
used as a process group leader.
- Move pidfds from the anonymous inode infrastructure to a tiny pseudo
filesystem. This will unblock further work that we weren't able to do
simply because of the very justified limitations of anonymous inodes.
Moving pidfds to a tiny pseudo filesystem allows for statx on pidfds
to become useful for the first time. They can now be compared by
inode number which are unique for the system lifetime.
Instead of stashing struct pid in file->private_data we can now stash
it in inode->i_private. This makes it possible to introduce concepts
that operate on a process once all file descriptors have been closed.
A concrete example is kill-on-last-close. Another side-effect is that
file->private_data is now freed up for per-file options for pidfds.
Now, each struct pid will refer to a different inode but the same
struct pid will refer to the same inode if it's opened multiple
times. In contrast to now where each struct pid refers to the same
inode.
The tiny pseudo filesystem is not visible anywhere in userspace
exactly like e.g., pipefs and sockfs. There's no lookup, there's no
complex inode operations, nothing. Dentries and inodes are always
deleted when the last pidfd is closed.
We allocate a new inode and dentry for each struct pid and we reuse
that inode and dentry for all pidfds that refer to the same struct
pid. The code is entirely optional and fairly small. If it's not
selected we fallback to anonymous inodes. Heavily inspired by nsfs.
The dentry and inode allocation mechanism is moved into generic
infrastructure that is now shared between nsfs and pidfs. The
path_from_stashed() helper must be provided with a stashing location,
an inode number, a mount, and the private data that is supposed to be
used and it will provide a path that can be passed to dentry_open().
The helper will try retrieve an existing dentry from the provided
stashing location. If a valid dentry is found it is reused. If not a
new one is allocated and we try to stash it in the provided location.
If this fails we retry until we either find an existing dentry or the
newly allocated dentry could be stashed. Subsequent openers of the
same namespace or task are then able to reuse it.
- Currently it is only possible to get notified when a task has exited,
i.e., become a zombie and userspace gets notified with EPOLLIN. We
now also support waiting until the task has been reaped, notifying
userspace with EPOLLHUP.
- Ensure that ESRCH is reported for getfd if a task is exiting instead
of the confusing EBADF.
- Various smaller cleanups to pidfd functions.
* tag 'vfs-6.9.pidfd' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (23 commits)
libfs: improve path_from_stashed()
libfs: add stashed_dentry_prune()
libfs: improve path_from_stashed() helper
pidfs: convert to path_from_stashed() helper
nsfs: convert to path_from_stashed() helper
libfs: add path_from_stashed()
pidfd: add pidfs
pidfd: move struct pidfd_fops
pidfd: allow to override signal scope in pidfd_send_signal()
pidfd: change pidfd_send_signal() to respect PIDFD_THREAD
signal: fill in si_code in prepare_kill_siginfo()
selftests: add ESRCH tests for pidfd_getfd()
pidfd: getfd should always report ESRCH if a task is exiting
pidfd: clone: allow CLONE_THREAD | CLONE_PIDFD together
pidfd: exit: kill the no longer used thread_group_exited()
pidfd: change do_notify_pidfd() to use __wake_up(poll_to_key(EPOLLIN))
pid: kill the obsolete PIDTYPE_PID code in transfer_pid()
pidfd: kill the no longer needed do_notify_pidfd() in de_thread()
pidfd_poll: report POLLHUP when pid_task() == NULL
pidfd: implement PIDFD_THREAD flag for pidfd_open()
...
This moves pidfds from the anonymous inode infrastructure to a tiny pseudo filesystem. This has been on my todo for quite a while as it will unblock further work that we weren't able to do simply because of the very justified limitations of anonymous inodes. Moving pidfds to a tiny pseudo filesystem allows: * statx() on pidfds becomes useful for the first time. * pidfds can be compared simply via statx() and then comparing inode numbers. * pidfds have unique inode numbers for the system lifetime. * struct pid is now stashed in inode->i_private instead of file->private_data. This means it is now possible to introduce concepts that operate on a process once all file descriptors have been closed. A concrete example is kill-on-last-close. * file->private_data is freed up for per-file options for pidfds. * Each struct pid will refer to a different inode but the same struct pid will refer to the same inode if it's opened multiple times. In contrast to now where each struct pid refers to the same inode. Even if we were to move to anon_inode_create_getfile() which creates new inodes we'd still be associating the same struct pid with multiple different inodes. The tiny pseudo filesystem is not visible anywhere in userspace exactly like e.g., pipefs and sockfs. There's no lookup, there's no complex inode operations, nothing. Dentries and inodes are always deleted when the last pidfd is closed. We allocate a new inode for each struct pid and we reuse that inode for all pidfds. We use iget_locked() to find that inode again based on the inode number which isn't recycled. We allocate a new dentry for each pidfd that uses the same inode. That is similar to anonymous inodes which reuse the same inode for thousands of dentries. For pidfds we're talking way less than that. There usually won't be a lot of concurrent openers of the same struct pid. They can probably often be counted on two hands. I know that systemd does use separate pidfd for the same struct pid for various complex process tracking issues. So I think with that things actually become way simpler. Especially because we don't have to care about lookup. Dentries and inodes continue to be always deleted. The code is entirely optional and fairly small. If it's not selected we fallback to anonymous inodes. Heavily inspired by nsfs which uses a similar stashing mechanism just for namespaces. Link: https://lore.kernel.org/r/20240213-vfs-pidfd_fs-v1-2-f863f58cfce1@kernel.org Signed-off-by: Christian Brauner <brauner@kernel.org>
Move the pidfd file operations over to their own file in preparation of implementing pidfs and to isolate them from other mostly unrelated functionality in other files. Link: https://lore.kernel.org/r/20240213-vfs-pidfd_fs-v1-1-f863f58cfce1@kernel.org Signed-off-by: Christian Brauner <brauner@kernel.org>
Holding a mutex across synchronize_rcu_tasks() and acquiring that same mutex in code called from do_exit() after its call to exit_tasks_rcu_start() but before its call to exit_tasks_rcu_stop() results in deadlock. This is by design, because tasks that are far enough into do_exit() are no longer present on the tasks list, making it a bit difficult for RCU Tasks to find them, let alone wait on them to do a voluntary context switch. However, such deadlocks are becoming more frequent. In addition, lockdep currently does not detect such deadlocks and they can be difficult to reproduce. In addition, if a task voluntarily context switches during that time (for example, if it blocks acquiring a mutex), then this task is in an RCU Tasks quiescent state. And with some adjustments, RCU Tasks could just as well take advantage of that fact. This commit therefore initializes the data structures that will be needed to rely on these quiescent states and to eliminate these deadlocks. Link: https://lore.kernel.org/all/20240118021842.290665-1-chenzhongjin@huawei.com/ Reported-by: Chen Zhongjin <chenzhongjin@huawei.com> Reported-by: Yang Jihong <yangjihong1@huawei.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Tested-by: Yang Jihong <yangjihong1@huawei.com> Tested-by: Chen Zhongjin <chenzhongjin@huawei.com> Reviewed-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Turn kill_pid_info() into kill_pid_info_type(), this allows to pass any pid_type to group_send_sig_info(), despite its name it should work fine even if type = PIDTYPE_PID. Change pidfd_send_signal() to use PIDTYPE_PID or PIDTYPE_TGID depending on PIDFD_THREAD. While at it kill another TODO comment in pidfd_show_fdinfo(). As Christian expains fdinfo reports f_flags, userspace can already detect PIDFD_THREAD. Reviewed-by: Tycho Andersen <tandersen@netflix.com> Signed-off-by: Oleg Nesterov <oleg@redhat.com> Link: https://lore.kernel.org/r/20240209130650.GA8048@redhat.com Signed-off-by: Christian Brauner <brauner@kernel.org>
copy_process() just needs to pass PIDFD_THREAD to __pidfd_prepare() if clone_flags & CLONE_THREAD. We can also add another CLONE_ flag (or perhaps reuse CLONE_DETACHED) to enforce PIDFD_THREAD without CLONE_THREAD. Originally-from: Tycho Andersen <tycho@tycho.pizza> Signed-off-by: Oleg Nesterov <oleg@redhat.com> Link: https://lore.kernel.org/r/20240205145532.GA28823@redhat.com Reviewed-by: Tycho Andersen <tandersen@netflix.com> Signed-off-by: Christian Brauner <brauner@kernel.org>
Add another wake_up_all(wait_pidfd) into __change_pid() and change pidfd_poll() to include EPOLLHUP if task == NULL. This allows to wait until the target process/thread is reaped. TODO: change do_notify_pidfd() to use the keyed wakeups. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Link: https://lore.kernel.org/r/20240202131226.GA26018@redhat.com Signed-off-by: Christian Brauner <brauner@kernel.org>
With this flag: - pidfd_open() doesn't require that the target task must be a thread-group leader - pidfd_poll() succeeds when the task exits and becomes a zombie (iow, passes exit_notify()), even if it is a leader and thread-group is not empty. This means that the behaviour of pidfd_poll(PIDFD_THREAD, pid-of-group-leader) is not well defined if it races with exec() from its sub-thread; pidfd_poll() can succeed or not depending on whether pidfd_task_exited() is called before or after exchange_tids(). Perhaps we can improve this behaviour later, pidfd_poll() can probably take sig->group_exec_task into account. But this doesn't really differ from the case when the leader exits before other threads (so pidfd_poll() succeeds) and then another thread execs and pidfd_poll() will block again. thread_group_exited() is no longer used, perhaps it can die. Co-developed-by: Tycho Andersen <tycho@tycho.pizza> Signed-off-by: Oleg Nesterov <oleg@redhat.com> Link: https://lore.kernel.org/r/20240131132602.GA23641@redhat.com Tested-by: Tycho Andersen <tandersen@netflix.com> Reviewed-by: Tycho Andersen <tandersen@netflix.com> Signed-off-by: Christian Brauner <brauner@kernel.org>