Re: [RFC PATCH v2 7/9] vfio-ccw: Wire up the CRW irq and CRW region

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu,  6 Feb 2020 22:38:23 +0100
Eric Farman <farman@xxxxxxxxxxxxx> wrote:

> From: Farhan Ali <alifm@xxxxxxxxxxxxx>
> 
> Use an IRQ to notify userspace that there is a CRW
> pending in the region, related to path-availability
> changes on the passthrough subchannel.
> 
> Signed-off-by: Farhan Ali <alifm@xxxxxxxxxxxxx>
> Signed-off-by: Eric Farman <farman@xxxxxxxxxxxxx>
> ---
> 
> Notes:
>     v1->v2:
>      - Remove extraneous 0x0 in crw.rsid assignment [CH]
>      - Refactor the building/queueing of a crw into its own routine [EF]
>     
>     v0->v1: [EF]
>      - Place the non-refactoring changes from the previous patch here
>      - Clean up checkpatch (whitespace) errors
>      - s/chp_crw/crw/
>      - Move acquire/release of io_mutex in vfio_ccw_crw_region_read()
>        into patch that introduces that region
>      - Remove duplicate include from vfio_ccw_drv.c
>      - Reorder include in vfio_ccw_private.h
> 
>  drivers/s390/cio/vfio_ccw_chp.c     |  5 ++
>  drivers/s390/cio/vfio_ccw_drv.c     | 73 +++++++++++++++++++++++++++++
>  drivers/s390/cio/vfio_ccw_ops.c     |  4 ++
>  drivers/s390/cio/vfio_ccw_private.h |  9 ++++
>  include/uapi/linux/vfio.h           |  1 +
>  5 files changed, 92 insertions(+)

[I may have gotten all muddled up from staring at this, but please bear
with me...]

> diff --git a/drivers/s390/cio/vfio_ccw_chp.c b/drivers/s390/cio/vfio_ccw_chp.c
> index 8fde94552149..328b4e1d1972 100644
> --- a/drivers/s390/cio/vfio_ccw_chp.c
> +++ b/drivers/s390/cio/vfio_ccw_chp.c
> @@ -98,6 +98,11 @@ static ssize_t vfio_ccw_crw_region_read(struct vfio_ccw_private *private,
>  		ret = count;
>  
>  	mutex_unlock(&private->io_mutex);
> +
> +	/* Notify the guest if more CRWs are on our queue */
> +	if (!list_empty(&private->crw) && private->crw_trigger)
> +		eventfd_signal(private->crw_trigger, 1);

Here we possibly arm the eventfd again, but don't do anything regarding
queued crws and the region.

> +
>  	return ret;
>  }
>  
> diff --git a/drivers/s390/cio/vfio_ccw_drv.c b/drivers/s390/cio/vfio_ccw_drv.c
> index 1e1360af1b34..c48c260a129d 100644
> --- a/drivers/s390/cio/vfio_ccw_drv.c
> +++ b/drivers/s390/cio/vfio_ccw_drv.c
> @@ -108,6 +108,31 @@ static void vfio_ccw_sch_io_todo(struct work_struct *work)
>  		eventfd_signal(private->io_trigger, 1);
>  }
>  
> +static void vfio_ccw_crw_todo(struct work_struct *work)
> +{
> +	struct vfio_ccw_private *private;
> +	struct vfio_ccw_crw *crw;
> +
> +	private = container_of(work, struct vfio_ccw_private, crw_work);
> +
> +	/* FIXME Ugh, need better control of this list */
> +	crw = list_first_entry_or_null(&private->crw,
> +				       struct vfio_ccw_crw, next);
> +
> +	if (crw) {
> +		list_del(&crw->next);
> +
> +		mutex_lock(&private->io_mutex);
> +		memcpy(&private->crw_region->crw0, crw->crw, sizeof(*crw->crw));
> +		mutex_unlock(&private->io_mutex);
> +
> +		kfree(crw);
> +
> +		if (private->crw_trigger)
> +			eventfd_signal(private->crw_trigger, 1);
> +	}
> +}

This function copies one outstanding crw and arms the eventfd.

> +
>  /*
>   * Css driver callbacks
>   */

(...)

> @@ -276,6 +309,44 @@ static int vfio_ccw_sch_event(struct subchannel *sch, int process)
>  	return rc;
>  }
>  
> +static void vfio_ccw_alloc_crw(struct vfio_ccw_private *private,
> +			       struct chp_link *link,
> +			       unsigned int erc)
> +{
> +	struct vfio_ccw_crw *vc_crw;
> +	struct crw *crw;
> +
> +	/*
> +	 * If unable to allocate a CRW, just drop the event and
> +	 * carry on.  The guest will either see a later one or
> +	 * learn when it issues its own store subchannel.
> +	 */
> +	vc_crw = kzalloc(sizeof(*vc_crw), GFP_ATOMIC);
> +	if (!vc_crw)
> +		return;
> +
> +	/*
> +	 * Build in the first CRW space, but don't chain anything
> +	 * into the second one even though the space exists.
> +	 */
> +	crw = &vc_crw->crw[0];
> +
> +	/*
> +	 * Presume every CRW we handle is reported by a channel-path.
> +	 * Maybe not future-proof, but good for what we're doing now.
> +	 *
> +	 * FIXME Sort of a lie, since we're converting a CRW
> +	 * reported by a channel-path into one issued to each
> +	 * subchannel, but still saying it's coming from the path.
> +	 */
> +	crw->rsc = CRW_RSC_CPATH;
> +	crw->rsid = (link->chpid.cssid << 8) | link->chpid.id;
> +	crw->erc = erc;
> +
> +	list_add_tail(&vc_crw->next, &private->crw);
> +	queue_work(vfio_ccw_work_q, &private->crw_work);

This function allocates a new crw and queues it. After that, it
triggers the function doing the copy-to-region-and-notify stuff.

> +}
> +
>  static int vfio_ccw_chp_event(struct subchannel *sch,
>  			      struct chp_link *link, int event)
>  {
> @@ -303,6 +374,7 @@ static int vfio_ccw_chp_event(struct subchannel *sch,
>  	case CHP_OFFLINE:
>  		/* Path is gone */
>  		cio_cancel_halt_clear(sch, &retry);
> +		vfio_ccw_alloc_crw(private, link, CRW_ERC_PERRN);
>  		break;
>  	case CHP_VARY_ON:
>  		/* Path logically turned on */
> @@ -312,6 +384,7 @@ static int vfio_ccw_chp_event(struct subchannel *sch,
>  	case CHP_ONLINE:
>  		/* Path became available */
>  		sch->lpm |= mask & sch->opm;
> +		vfio_ccw_alloc_crw(private, link, CRW_ERC_INIT);
>  		break;
>  	}
>  

These two (path online/offline handling) are the only code paths
triggering an update to the queued crws.

Aren't we missing copying in a new queued crw after userspace had done
a read?




[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux