Re: [net-next v3 05/15] net: cavium/liquidio: Convert tasklet API to new bottom half workqueue mechanism

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 8 Aug 2024 19:31:57 -0700 Allen wrote:
> > > In the context of of the driver, the conversion from tasklet_enable()
> > > to enable_and_queue_work() is correct because the callback function
> > > associated with the work item is designed to be safe even if there
> > > is no immediate work to process. The callback function can handle
> > > being invoked in such situations without causing errors or undesirable
> > > behavior. This makes the workqueue approach a suitable and safe
> > > replacement for the current tasklet mechanism, as it provides the
> > > necessary flexibility and ensures that the work item is properly
> > > scheduled and executed.  
> >
> > Fewer words, clearer indication that you read the code would be better
> > for the reviewer. Like actually call out what in the code makes it safe.
> >  
> Okay.
> > Just to be clear -- conversions to enable_and_queue_work() will require
> > manual inspection in every case.  
> 
> Attempting again.
> 
> The enable_and_queue_work() only schedules work if it is not already
> enabled, similar to how tasklet_enable() would only allow a tasklet to run
> if it had been previously scheduled.
> 
> In the current driver, where we are attempting conversion, enable_work()
> checks whether the work is already enabled and only enables it if it
> was disabled. If no new work is queued, queue_work() won't be called.
> Hence, the callback is safe even if there's no work.

Hm. Let me give you an example of what I was hoping to see for this
patch (in addition to your explanation of the API difference):

 The conversion for oct_priv->droq_bh_work should be safe. While
 the work is per adapter, the callback (octeon_droq_bh()) walks all
 queues, and for each queue checks whether the oct->io_qmask.oq mask
 has a bit set. In case of spurious scheduling of the work - none of
 the bits should be set, making the callback a noop.