On 2026/2/1 7:13, Waiman Long wrote:
> 
> On 1/30/26 9:53 PM, Chen Ridong wrote:
>>
>> On 2026/1/30 23:42, Waiman Long wrote:
>>> The current cpuset partition code is able to dynamically update
>>> the sched domains of a running system and the corresponding
>>> HK_TYPE_DOMAIN housekeeping cpumask to perform what is essentally the
>>> "isolcpus=domain,..." boot command line feature at run time.
>>>
>>> The housekeeping cpumask update requires flushing a number of different
>>> workqueues which may not be safe with cpus_read_lock() held as the
>>> workqueue flushing code may acquire cpus_read_lock() or acquiring locks
>>> which have locking dependency with cpus_read_lock() down the chain. Below
>>> is an example of such circular locking problem.
>>>
>>>    ======================================================
>>>    WARNING: possible circular locking dependency detected
>>>    6.18.0-test+ #2 Tainted: G S
>>>    ------------------------------------------------------
>>>    test_cpuset_prs/10971 is trying to acquire lock:
>>>    ffff888112ba4958 ((wq_completion)sync_wq){+.+.}-{0:0}, at:
>>> touch_wq_lockdep_map+0x7a/0x180
>>>
>>>    but task is already holding lock:
>>>    ffffffffae47f450 (cpuset_mutex){+.+.}-{4:4}, at:
>>> cpuset_partition_write+0x85/0x130
>>>
>>>    which lock already depends on the new lock.
>>>
>>>    the existing dependency chain (in reverse order) is:
>>>    -> #4 (cpuset_mutex){+.+.}-{4:4}:
>>>    -> #3 (cpu_hotplug_lock){++++}-{0:0}:
>>>    -> #2 (rtnl_mutex){+.+.}-{4:4}:
>>>    -> #1 ((work_completion)(&arg.work)){+.+.}-{0:0}:
>>>    -> #0 ((wq_completion)sync_wq){+.+.}-{0:0}:
>>>
>>>    Chain exists of:
>>>      (wq_completion)sync_wq --> cpu_hotplug_lock --> cpuset_mutex
>>>
>>>    5 locks held by test_cpuset_prs/10971:
>>>     #0: ffff88816810e440 (sb_writers#7){.+.+}-{0:0}, at: 
>>> ksys_write+0xf9/0x1d0
>>>     #1: ffff8891ab620890 (&of->mutex#2){+.+.}-{4:4}, at:
>>> kernfs_fop_write_iter+0x260/0x5f0
>>>     #2: ffff8890a78b83e8 (kn->active#187){.+.+}-{0:0}, at:
>>> kernfs_fop_write_iter+0x2b6/0x5f0
>>>     #3: ffffffffadf32900 (cpu_hotplug_lock){++++}-{0:0}, at:
>>> cpuset_partition_write+0x77/0x130
>>>     #4: ffffffffae47f450 (cpuset_mutex){+.+.}-{4:4}, at:
>>> cpuset_partition_write+0x85/0x130
>>>
>>>    Call Trace:
>>>     <TASK>
>>>       :
>>>     touch_wq_lockdep_map+0x93/0x180
>>>     __flush_workqueue+0x111/0x10b0
>>>     housekeeping_update+0x12d/0x2d0
>>>     update_parent_effective_cpumask+0x595/0x2440
>>>     update_prstate+0x89d/0xce0
>>>     cpuset_partition_write+0xc5/0x130
>>>     cgroup_file_write+0x1a5/0x680
>>>     kernfs_fop_write_iter+0x3df/0x5f0
>>>     vfs_write+0x525/0xfd0
>>>     ksys_write+0xf9/0x1d0
>>>     do_syscall_64+0x95/0x520
>>>     entry_SYSCALL_64_after_hwframe+0x76/0x7e
>>>
>>> To avoid such a circular locking dependency problem, we have to
>>> call housekeeping_update() without holding the cpus_read_lock() and
>>> cpuset_mutex. The current set of wq's flushed by housekeeping_update()
>>> may not have work functions that call cpus_read_lock() directly,
>>> but we are likely to extend the list of wq's that are flushed in the
>>> future. Moreover, the current set of work functions may hold locks that
>>> may have cpu_hotplug_lock down the dependency chain.
>>>
>>> One way to do that is to introduce a new top level cpuset_top_mutex
>>> which will be acquired first.  This new cpuset_top_mutex will provide
>>> the need mutual exclusion without the need to hold cpus_read_lock().
>>>
>> Introducing a new global lock warrants careful consideration. I wonder if we
>> could make all updates to isolated_cpus asynchronous. If that is feasible, we
>> could avoid adding a global lock altogether. If not, we need to clarify which
>> updates must remain synchronous and which ones can be handled asynchronously.
> 
> Almost all the cpuset code are run with cpuset_mutex held with either
> cpus_read_lock or cpus_write_lock. So there is no concurrent access/update to
> any of the cpuset internal data. The new cpuset_top_mutex is aded to resolve 
> the
> possible deadlock scenarios with the new housekeeping_update() call without
> breaking this model. Allow parallel concurrent access/update to cpuset data 
> will
> greatly complicate the code and we will likely missed some corner cases that 
> we

I agree with that point. However, we already have paths where isolated_cpus is
updated asynchronously, meaning parallel concurrent access/update is already
happening. Therefore, we cannot entirely avoid such scenarios, so why not keep
the locking simple(make all updates to isolated_cpus asynchronous)?

This is just a thought in my mind.

> have to fix in the future. We will only do that if cpuset is in a critical
> performance path, but it is not. It is not just isolated_cpus that we are
> protecting, all the other cpuset data may be at risk if we don't have another
> top level mutex to protect them.
> 
> Cheers,
> Longman
> 

-- 
Best regards,
Ridong


Reply via email to