[PATCH v3 1/2] cxl: Add mechanism for delivering AFU driver specific events

Matt Ochs mrochs at us.ibm.com
Thu Mar 10 03:41:04 AEDT 2016


> On Mar 9, 2016, at 8:37 AM, Vaibhav Jain <vaibhav at linux.vnet.ibm.com> wrote:
>> +/*
>> + * AFU driver ops allows an AFU driver to create their own events to pass to
>> + * userspace through the file descriptor as a simpler alternative to overriding
>> + * the read() and poll() calls that works with the generic cxl events. These
>> + * events are given priority over the generic cxl events, so they will be
>> + * delivered first if multiple types of events are pending.
>> + *
>> + * event_pending() will be called by the cxl driver to check if an event is
>> + * pending (e.g. in select/poll/read calls).
>> + *
>> + * deliver_event() will be called to fill out a cxl_event structure with the
>> + * driver specific event. The header will already have the type and
>> + * process_element fields filled in, and header.size will be set to
>> + * sizeof(struct cxl_event_header). The AFU driver can extend that size up to
>> + * max_size (if an afu driver requires more space, they should submit a patch
>> + * increasing the size in the struct cxl_event_afu_driver_reserved definition).
>> + *
>> + * Both of these calls are made with a spin lock held, so they must not sleep.
>> + */
>> +struct cxl_afu_driver_ops {
>> +	bool (*event_pending) (struct cxl_context *ctx);
>> +	void (*deliver_event) (struct cxl_context *ctx,
>> +			struct cxl_event *event, size_t max_size);
>> +};
>> +
> 
> I would propose these two apis.
> 
> /*
> *  fetches an event from the driver event queue. NULL means that queue
> *  is empty. Can sleep if needed. The memory for cxl_event is allocated
> *  by module being called. Hence it can be potentially be larger then
> *  sizeof(struct cxl_event). Multiple calls to this should return same
> *  pointer untill ack_event is called.
> */
> struct cxl_event * fetch_event(struct cxl_context * ctx);
> 
> /*
> * Returns and acknowledge the struct cxl_event * back to the driver
> * which can then free it or maybe put it back in a kmem_cache. This
> * should be called once we have completely returned the current
> * struct cxl_event from the readcall
> */
> void ack_event(struct cxl_context * ctx, struct cxl_event *);
> 
> I think above apis would give us more flexbility in the future when
> drivers would want to send larger events without breaking the abi.

From a cxlflash perspective, I think we'd be fine with this model as
long as the driver events are still prioritized. I do like the removal of
the no-sleep requirement and this would allow us to simply hand off
an already populated event reference.





More information about the Linuxppc-dev mailing list