Re: write_lock_irq(&tasklist_lock)

[Peter Zijlstra]

On Thu, May 24, 2018 at 10:37:25AM -0700, Sodagudi Prasad wrote:
>  Kernel version is locked for couple of products and same issue observed on
> both 4.14.41
> and 4.9.96 kernels. We can only accept the stable updates from upstream for
> these products.

> If QUEUED_RWLOCKS works on above listed kernel versions without any issues,
> we can enabled QUEUED_RWLOCKS.

You want:

e0d02285f16e ("locking/qrwlock: Use 'struct qrwlock' instead of 'struct __qrwlock'")
4df714be4dcf ("locking/atomic: Add atomic_cond_read_acquire()")
b519b56e378e ("locking/qrwlock: Use atomic_cond_read_acquire() when spinning in qrwlock")
087133ac9076 ("locking/qrwlock, arm64: Move rwlock implementation over to qrwlocks")
d13316614633 ("locking/qrwlock: Prevent slowpath writers getting held up by fastpath")

IIRC, enabling QUEUED_RWLOCKS will 'work' but esp. that
atomic_cond_read_acquire() one is goodness for ARM64.

> Can we go ahead with Linus suggestion for these kernel version?
> So that IRQ wont be disabled for quite a long time.
> 
>  static void tasklist_write_lock(void)
>    {
>          unsigned long flags;
>          local_irq_save(flags);
>          while (!write_trylock(&tasklist_lock)) {
>                  local_irq_restore(flags);
>                  do { cpu_relax(); } while (write_islocked(&tasklist_lock));
>                  local_irq_disable();
>          }
>    }

First you say you can only take stable updates, then you ask for a
gruesome hack that will seriously regress architectures that do use
qrwlock which will hence never make it into stable.

Just take the arm64 qrwlock patches and pretend they're from stable.