Re: [patch V2 46/46] irqchip: Add IMS (Interrupt Message Storm) driver - NOT FOR MERGING
From: Jason Gunthorpe
Date: Mon Aug 31 2020 - 10:45:44 EST
On Wed, Aug 26, 2020 at 01:17:14PM +0200, Thomas Gleixner wrote:
> + * ims_queue_info - Information to create an IMS queue domain
> + * @queue_lock: Callback which informs the device driver that
> + * an interrupt management operation starts.
> + * @queue_sync_unlock: Callback which informs the device driver that an
> + * interrupt management operation ends.
> +
> + * @queue_get_shadow: Callback to retrieve te shadow storage for a MSI
> + * entry associated to a queue. The queue is
> + * identified by the device struct which is used for
> + * allocating interrupts and the msi entry index.
> + *
> + * @queue_lock() and @queue_sync_unlock() are only called for management
> + * operations on a particular interrupt: request, free, enable, disable,
> + * affinity setting. These functions are never called from atomic context,
> + * like low level interrupt handling code. The purpose of these functions
> + * is to signal the device driver the start and end of an operation which
> + * affects the IMS queue shadow state. @queue_lock() allows the driver to
> + * do preperatory work, e.g. locking. Note, that @queue_lock() has to
> + * preserve the sleepable state on return. That means the driver cannot
> + * disable preemption and (soft)interrupts in @queue_lock and then undo
> + * that operation in @queue_sync_unlock() which restricts the lock types
> + * for eventual serialization of these operations to sleepable locks. Of
> + * course the driver can disable preemption and (soft)interrupts
> + * temporarily for internal work.
> + *
> + * On @queue_sync_unlock() the driver has to check whether the shadow state
> + * changed and issue a command to update the hardware state and wait for
> + * the command to complete. If the command fails or times out then the
> + * driver has to take care of the resulting mess as this is called from
> + * functions which have no return value and none of the callers can deal
> + * with the failure. The lock which is used by the driver to protect a
> + * operation sequence must obviously not be released before the command
> + * completes or fails. Otherwise new operations on the same interrupt line
> + * could take place and change the shadow state before the driver was able
> + * to compose the command.
I haven't looked through everything in detail, but this does look like
it is good for the mlx5 devices. Looked like it was only one small
update to the set_affinity, so not very disruptive?
Thanks,
Jason