Re: [PATCH v4] proc: Fix a dentry lock race between release_task and lookup

[Zhihao Cheng <chengzhihao1@huawei.com>]

在 2022/6/1 14:23, Zhihao Cheng 写道:
friendly ping
> Commit 7bc3e6e55acf06 ("proc: Use a list of inodes to flush from proc")
> moved proc_flush_task() behind __exit_signal(). Then, process systemd
> can take long period high cpu usage during releasing task in following
> concurrent processes:
>
>    systemd                                 ps
> kernel_waitid                 stat(/proc/tgid)
>    do_wait                       filename_lookup
>      wait_consider_task            lookup_fast
>        release_task
>          __exit_signal
>            __unhash_process
>              detach_pid
>                __change_pid // remove task->pid_links
>                                       d_revalidate -> pid_revalidate  // 0
>                                       d_invalidate(/proc/tgid)
>                                         shrink_dcache_parent(/proc/tgid)
>                                           d_walk(/proc/tgid)
>                                             spin_lock_nested(/proc/tgid/fd)
>                                             // iterating opened fd
>          proc_flush_pid                                    |
>             d_invalidate (/proc/tgid/fd)                   |
>                shrink_dcache_parent(/proc/tgid/fd)         |
>                  shrink_dentry_list(subdirs)               ↓
>                    shrink_lock_dentry(/proc/tgid/fd) --> race on dentry lock
>
> Function d_invalidate() will remove dentry from hash firstly, but why does
> proc_flush_pid() process dentry '/proc/tgid/fd' before dentry '/proc/tgid'?
> That's because proc_pid_make_inode() adds proc inode in reverse order by
> invoking hlist_add_head_rcu(). But proc should not add any inodes under
> '/proc/tgid' except '/proc/tgid/task/pid', fix it by adding inode into
> 'pid->inodes' only if the inode is /proc/tgid or /proc/tgid/task/pid.
>
> Performance regression:
> Create 200 tasks, each task open one file for 50,000 times. Kill all
> tasks when opened files exceed 10,000,000 (cat /proc/sys/fs/file-nr).
>