Re: [RFC][PATCH 2/2] add a counter for writers spinning on a rwlock

[Mandeep Baines <msb@google.com>]

Hi Frédéric,

Unfortunately, this can't be done for hung_task. It writes to the
task_struct here:

static void check_hung_task(struct task_struct *t, unsigned long now,
                            unsigned long timeout)
{
        unsigned long switch_count = t->nvcsw + t->nivcsw;

        if (t->flags & PF_FROZEN)
                return;

        if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
                t->last_switch_count = switch_count;
                t->last_switch_timestamp = now;
                return;
        }

It is able to get away with using only a read_lock because no one else
reads or writes to these fields.

Regards,
Mandeep

On Mon, Jan 26, 2009 at 8:04 AM, Frédéric Weisbecker <fweisbec@gmail.com> wrote:
> 2009/1/26 Peter Zijlstra <peterz@infradead.org>:
>> On Mon, 2009-01-26 at 16:25 +0100, Frédéric Weisbecker wrote:
>>> 2009/1/26 Peter Zijlstra <peterz@infradead.org>:
>>> > On Sun, 2009-01-25 at 12:50 -0800, Frederic Weisbecker wrote:
>>> >> This patch adds a counter for writers that enter a rwlock slow path.
>>> >> For example it can be useful for slow background tasks which perform some jobs
>>> >> on the tasklist, such as the hung_task detector (kernel/hung_task.c).
>>> >>
>>> >> It adds a inc/dec pair on the slow path and 4 bytes for each rwlocks, so the overhead
>>> >> is not null.
>>> >>
>>> >> Only x86 is supported for now, writers_spinning_lock() will return 0 on other archs (which
>>> >> is perhaps not a good idea).
>>> >>
>>> >> Comments?
>>> >
>>> > _why_ ?
>>>
>>> The hung task detector runs a periodic loop through the task_list, and
>>> currently it doesn't run
>>> over an arbitrary threshold of tasks to not hold the task_list lock
>>> for too long.
>>>
>>> So we thought about a way to detect if there are some writers waiting
>>> for the lock, anf if so, release
>>> the lock, schedule and retry.
>>
>> Ah, if it can do that, then it can also use RCU, no? Only users who
>> really have to hold off new tasks need the read-task_lock. The rest can
>> use RCU.
>
>
> Really?
> That sounds a good news. Mandeep, what do you think?
>