Message ID | 1550152269-6317-5-git-send-email-pmorel@linux.ibm.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | vfio: ap: ioctl definitions for AP Queue Interrupt Control | expand |
On Thu, 14 Feb 2019 14:51:04 +0100 Pierre Morel <pmorel@linux.ibm.com> wrote: > We need to find the queue with a specific APQN during the > handling of the interception of the PQAP/AQIC instruction. > > To handle the AP associated device reference count we keep > track of it in the vfio_ap_queue until we put the device. So, the relationship is (struct ap_device)--(driver_data)-->(struct vfio_ap_queue)--(pointer)-->(struct ap_device) ? IOW, a backlink? If so, can't you already set that up during probe? Or am I confused by the various similar devices again? Maybe a diagram would help... > > Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> > --- > drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ > drivers/s390/crypto/vfio_ap_private.h | 1 + > 2 files changed, 55 insertions(+) > > diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c > index 900b9cf..2a52c9b 100644 > --- a/drivers/s390/crypto/vfio_ap_ops.c > +++ b/drivers/s390/crypto/vfio_ap_ops.c > @@ -24,6 +24,60 @@ > #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" > #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" > > +/** > + * vfio_ap_check_apqn: check if a ap_queue is of a given APQN > + * > + * Returns 1 if we have a match. > + * Otherwise returns 0. > + */ > +static int vfio_ap_check_apqn(struct device *dev, void *data) > +{ > + struct vfio_ap_queue *q = dev_get_drvdata(dev); > + > + return (q->apqn == *(int *)data); > +} > + > +/** > + * vfio_ap_get_queue: Retrieve a queue with a specific APQN > + * @apqn: The queue APQN > + * > + * Retrieve a queue with a specific APQN from the list of the > + * devices associated to the vfio_ap_driver. > + * > + * The vfio_ap_queue has been already associated with the device > + * during the probe. > + * Store the associated device for reference counting > + * > + * Returns the pointer to the associated vfio_ap_queue > + */ > +static __attribute__((unused)) Eww. Can you get rid of that by reordering or squashing patches? > + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) > +{ > + struct device *dev; > + struct vfio_ap_queue *q; > + > + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, > + vfio_ap_check_apqn); > + if (!dev) > + return NULL; > + q = dev_get_drvdata(dev); > + q->dev = dev; > + return q; > +} > + > +/** > + * vfio_ap_put_queue: lower device reference count for a queue > + * @q: The queue > + * > + * put the associated device > + * > + */ > +static __attribute__((unused)) void vfio_ap_put_queue(struct vfio_ap_queue *q) > +{ > + put_device(q->dev); > + q->dev = NULL; > +} > + > static void vfio_ap_matrix_init(struct ap_config_info *info, > struct ap_matrix *matrix) > { > diff --git a/drivers/s390/crypto/vfio_ap_private.h b/drivers/s390/crypto/vfio_ap_private.h > index 8836f01..081f0d7 100644 > --- a/drivers/s390/crypto/vfio_ap_private.h > +++ b/drivers/s390/crypto/vfio_ap_private.h > @@ -87,6 +87,7 @@ extern int vfio_ap_mdev_register(void); > extern void vfio_ap_mdev_unregister(void); > > struct vfio_ap_queue { > + struct device *dev; > int apqn; > }; > #endif /* _VFIO_AP_PRIVATE_H_ */
On 15/02/2019 10:49, Cornelia Huck wrote: > On Thu, 14 Feb 2019 14:51:04 +0100 > Pierre Morel <pmorel@linux.ibm.com> wrote: > >> We need to find the queue with a specific APQN during the >> handling of the interception of the PQAP/AQIC instruction. >> >> To handle the AP associated device reference count we keep >> track of it in the vfio_ap_queue until we put the device. > > So, the relationship is > (struct ap_device)--(driver_data)-->(struct vfio_ap_queue)--(pointer)-->(struct ap_device) > ? IOW, a backlink? > > If so, can't you already set that up during probe? Will do. > > Or am I confused by the various similar devices again? Maybe a diagram > would help... No you are right. > >> >> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> >> --- >> drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ >> drivers/s390/crypto/vfio_ap_private.h | 1 + >> 2 files changed, 55 insertions(+) >> >> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c >> index 900b9cf..2a52c9b 100644 >> --- a/drivers/s390/crypto/vfio_ap_ops.c >> +++ b/drivers/s390/crypto/vfio_ap_ops.c >> @@ -24,6 +24,60 @@ >> #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" >> #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" >> >> +/** >> + * vfio_ap_check_apqn: check if a ap_queue is of a given APQN >> + * >> + * Returns 1 if we have a match. >> + * Otherwise returns 0. >> + */ >> +static int vfio_ap_check_apqn(struct device *dev, void *data) >> +{ >> + struct vfio_ap_queue *q = dev_get_drvdata(dev); >> + >> + return (q->apqn == *(int *)data); >> +} >> + >> +/** >> + * vfio_ap_get_queue: Retrieve a queue with a specific APQN >> + * @apqn: The queue APQN >> + * >> + * Retrieve a queue with a specific APQN from the list of the >> + * devices associated to the vfio_ap_driver. >> + * >> + * The vfio_ap_queue has been already associated with the device >> + * during the probe. >> + * Store the associated device for reference counting >> + * >> + * Returns the pointer to the associated vfio_ap_queue >> + */ >> +static __attribute__((unused)) > > Eww. Can you get rid of that by reordering or squashing patches? I did this to avoid posting a very big patch. I will of course squash 4 and 5 with patch 6, when the two patches 4 and 5 are reviewed. If you think it brings more clarity to squash all for the next iteration I will do. Regards, Pierre
On Fri, 15 Feb 2019 11:10:43 +0100 Pierre Morel <pmorel@linux.ibm.com> wrote: > On 15/02/2019 10:49, Cornelia Huck wrote: > > On Thu, 14 Feb 2019 14:51:04 +0100 > > Pierre Morel <pmorel@linux.ibm.com> wrote: > > > >> We need to find the queue with a specific APQN during the > >> handling of the interception of the PQAP/AQIC instruction. > >> > >> To handle the AP associated device reference count we keep > >> track of it in the vfio_ap_queue until we put the device. > > > > So, the relationship is > > (struct ap_device)--(driver_data)-->(struct vfio_ap_queue)--(pointer)-->(struct ap_device) > > ? IOW, a backlink? > > > > If so, can't you already set that up during probe? > > Will do. > > > > > Or am I confused by the various similar devices again? Maybe a diagram > > would help... > > No you are right. Good, I was fearing that I was more confused than normal for Fridays ;) > > > > > >> > >> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> > >> --- > >> drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ > >> drivers/s390/crypto/vfio_ap_private.h | 1 + > >> 2 files changed, 55 insertions(+) > >> > >> diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c > >> index 900b9cf..2a52c9b 100644 > >> --- a/drivers/s390/crypto/vfio_ap_ops.c > >> +++ b/drivers/s390/crypto/vfio_ap_ops.c > >> @@ -24,6 +24,60 @@ > >> #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" > >> #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" > >> > >> +/** > >> + * vfio_ap_check_apqn: check if a ap_queue is of a given APQN > >> + * > >> + * Returns 1 if we have a match. > >> + * Otherwise returns 0. > >> + */ > >> +static int vfio_ap_check_apqn(struct device *dev, void *data) > >> +{ > >> + struct vfio_ap_queue *q = dev_get_drvdata(dev); > >> + > >> + return (q->apqn == *(int *)data); > >> +} > >> + > >> +/** > >> + * vfio_ap_get_queue: Retrieve a queue with a specific APQN > >> + * @apqn: The queue APQN > >> + * > >> + * Retrieve a queue with a specific APQN from the list of the > >> + * devices associated to the vfio_ap_driver. > >> + * > >> + * The vfio_ap_queue has been already associated with the device > >> + * during the probe. > >> + * Store the associated device for reference counting > >> + * > >> + * Returns the pointer to the associated vfio_ap_queue > >> + */ > >> +static __attribute__((unused)) > > > > Eww. Can you get rid of that by reordering or squashing patches? > > I did this to avoid posting a very big patch. > I will of course squash 4 and 5 with patch 6, when the two patches 4 and > 5 are reviewed. > > If you think it brings more clarity to squash all for the next iteration > I will do. Let's just see what the patches look like in the end. If a squashed patch is not too unwieldy, I'd prefer that over those unused annotations, though. Hoping for review from others as well ;)
On 2/14/19 8:51 AM, Pierre Morel wrote: > We need to find the queue with a specific APQN during the > handling of the interception of the PQAP/AQIC instruction. > > To handle the AP associated device reference count we keep > track of it in the vfio_ap_queue until we put the device. > > Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> > --- > drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ > drivers/s390/crypto/vfio_ap_private.h | 1 + > 2 files changed, 55 insertions(+) > > diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c > index 900b9cf..2a52c9b 100644 > --- a/drivers/s390/crypto/vfio_ap_ops.c > +++ b/drivers/s390/crypto/vfio_ap_ops.c > @@ -24,6 +24,60 @@ > #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" > #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" > > +/** > + * vfio_ap_check_apqn: check if a ap_queue is of a given APQN > + * > + * Returns 1 if we have a match. > + * Otherwise returns 0. > + */ > +static int vfio_ap_check_apqn(struct device *dev, void *data) > +{ > + struct vfio_ap_queue *q = dev_get_drvdata(dev); > + > + return (q->apqn == *(int *)data); > +} > + > +/** > + * vfio_ap_get_queue: Retrieve a queue with a specific APQN > + * @apqn: The queue APQN > + * > + * Retrieve a queue with a specific APQN from the list of the > + * devices associated to the vfio_ap_driver. > + * > + * The vfio_ap_queue has been already associated with the device > + * during the probe. > + * Store the associated device for reference counting > + * > + * Returns the pointer to the associated vfio_ap_queue > + */ > +static __attribute__((unused)) > + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) I think you should change this signature for the reasons I stated below: struct device *vfio_ap_get_queue_dev(int apqn) > +{ > + struct device *dev; > + struct vfio_ap_queue *q; > + > + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, > + vfio_ap_check_apqn); > + if (!dev) > + return NULL; > + q = dev_get_drvdata(dev); > + q->dev = dev; Why store the device with the vfio_ap_queue object? Why not just return the device. The caller can get the vfio_ap_queue from the device's driver data. It seems the only reason for the 'dev' field is to temporarily hold a ref to the device so it can be put later. Why not just put the device. > + return q; > +} > + > +/** > + * vfio_ap_put_queue: lower device reference count for a queue > + * @q: The queue > + * > + * put the associated device > + * > + */ > +static __attribute__((unused)) void vfio_ap_put_queue(struct vfio_ap_queue *q) > +{ > + put_device(q->dev); > + q->dev = NULL; > +} I would get rid of this function. If you take my suggestion above, you can just call the put_device() directly. There will be no need for this superfluous 'dev' field. > + > static void vfio_ap_matrix_init(struct ap_config_info *info, > struct ap_matrix *matrix) > { > diff --git a/drivers/s390/crypto/vfio_ap_private.h b/drivers/s390/crypto/vfio_ap_private.h > index 8836f01..081f0d7 100644 > --- a/drivers/s390/crypto/vfio_ap_private.h > +++ b/drivers/s390/crypto/vfio_ap_private.h > @@ -87,6 +87,7 @@ extern int vfio_ap_mdev_register(void); > extern void vfio_ap_mdev_unregister(void); > > struct vfio_ap_queue { > + struct device *dev; No need for this as per my comments above. > int apqn; > }; > #endif /* _VFIO_AP_PRIVATE_H_ */ >
On Fri, 15 Feb 2019 17:13:21 -0500 Tony Krowiak <akrowiak@linux.ibm.com> wrote: > On 2/14/19 8:51 AM, Pierre Morel wrote: > > We need to find the queue with a specific APQN during the > > handling of the interception of the PQAP/AQIC instruction. > > > > To handle the AP associated device reference count we keep > > track of it in the vfio_ap_queue until we put the device. > > > > Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> > > --- > > drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ > > drivers/s390/crypto/vfio_ap_private.h | 1 + > > 2 files changed, 55 insertions(+) > > +/** > > + * vfio_ap_get_queue: Retrieve a queue with a specific APQN > > + * @apqn: The queue APQN > > + * > > + * Retrieve a queue with a specific APQN from the list of the > > + * devices associated to the vfio_ap_driver. > > + * > > + * The vfio_ap_queue has been already associated with the device > > + * during the probe. > > + * Store the associated device for reference counting > > + * > > + * Returns the pointer to the associated vfio_ap_queue > > + */ > > +static __attribute__((unused)) > > + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) > > I think you should change this signature for the reasons I stated > below: > > struct device *vfio_ap_get_queue_dev(int apqn) > > > +{ > > + struct device *dev; > > + struct vfio_ap_queue *q; > > + > > + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, > > + vfio_ap_check_apqn); > > + if (!dev) > > + return NULL; > > + q = dev_get_drvdata(dev); > > + q->dev = dev; > > Why store the device with the vfio_ap_queue object? Why not just return > the device. The caller can get the vfio_ap_queue from the device's > driver data. It seems the only reason for the 'dev' field is to > temporarily hold a ref to the device so it can be put later. Why not > just put the device. Having looked at the remainder of the patches, I tend to agree that we don't really need the backlink; we walk the driver's list of devices in any case IIUC. We *might* want a mechanism to grab the queue quickly (i.e. without walking the list) if there's anything performance sensitive in there; but from the patch descriptions, I don't think anything is done in a hot path, so it should be fine.
On 18/02/2019 13:21, Cornelia Huck wrote: > On Fri, 15 Feb 2019 17:13:21 -0500 > Tony Krowiak <akrowiak@linux.ibm.com> wrote: > >> On 2/14/19 8:51 AM, Pierre Morel wrote: >>> We need to find the queue with a specific APQN during the >>> handling of the interception of the PQAP/AQIC instruction. >>> >>> To handle the AP associated device reference count we keep >>> track of it in the vfio_ap_queue until we put the device. >>> >>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> >>> --- >>> drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ >>> drivers/s390/crypto/vfio_ap_private.h | 1 + >>> 2 files changed, 55 insertions(+) > >>> +/** >>> + * vfio_ap_get_queue: Retrieve a queue with a specific APQN >>> + * @apqn: The queue APQN >>> + * >>> + * Retrieve a queue with a specific APQN from the list of the >>> + * devices associated to the vfio_ap_driver. >>> + * >>> + * The vfio_ap_queue has been already associated with the device >>> + * during the probe. >>> + * Store the associated device for reference counting >>> + * >>> + * Returns the pointer to the associated vfio_ap_queue >>> + */ >>> +static __attribute__((unused)) >>> + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) >> >> I think you should change this signature for the reasons I stated >> below: >> >> struct device *vfio_ap_get_queue_dev(int apqn) >> >>> +{ >>> + struct device *dev; >>> + struct vfio_ap_queue *q; >>> + >>> + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, >>> + vfio_ap_check_apqn); >>> + if (!dev) >>> + return NULL; >>> + q = dev_get_drvdata(dev); >>> + q->dev = dev; >> >> Why store the device with the vfio_ap_queue object? Why not just return >> the device. The caller can get the vfio_ap_queue from the device's >> driver data. It seems the only reason for the 'dev' field is to >> temporarily hold a ref to the device so it can be put later. Why not >> just put the device. > > Having looked at the remainder of the patches, I tend to agree that we > don't really need the backlink; we walk the driver's list of devices in > any case IIUC. > > We *might* want a mechanism to grab the queue quickly (i.e. without > walking the list) if there's anything performance sensitive in there; > but from the patch descriptions, I don't think anything is done in a > hot path, so it should be fine. > OK you are right, I ll drop it
On 2/18/19 7:21 AM, Cornelia Huck wrote: > On Fri, 15 Feb 2019 17:13:21 -0500 > Tony Krowiak <akrowiak@linux.ibm.com> wrote: > >> On 2/14/19 8:51 AM, Pierre Morel wrote: >>> We need to find the queue with a specific APQN during the >>> handling of the interception of the PQAP/AQIC instruction. >>> >>> To handle the AP associated device reference count we keep >>> track of it in the vfio_ap_queue until we put the device. >>> >>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> >>> --- >>> drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ >>> drivers/s390/crypto/vfio_ap_private.h | 1 + >>> 2 files changed, 55 insertions(+) > >>> +/** >>> + * vfio_ap_get_queue: Retrieve a queue with a specific APQN >>> + * @apqn: The queue APQN >>> + * >>> + * Retrieve a queue with a specific APQN from the list of the >>> + * devices associated to the vfio_ap_driver. >>> + * >>> + * The vfio_ap_queue has been already associated with the device >>> + * during the probe. >>> + * Store the associated device for reference counting >>> + * >>> + * Returns the pointer to the associated vfio_ap_queue >>> + */ >>> +static __attribute__((unused)) >>> + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) >> >> I think you should change this signature for the reasons I stated >> below: >> >> struct device *vfio_ap_get_queue_dev(int apqn) >> >>> +{ >>> + struct device *dev; >>> + struct vfio_ap_queue *q; >>> + >>> + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, >>> + vfio_ap_check_apqn); >>> + if (!dev) >>> + return NULL; >>> + q = dev_get_drvdata(dev); >>> + q->dev = dev; >> >> Why store the device with the vfio_ap_queue object? Why not just return >> the device. The caller can get the vfio_ap_queue from the device's >> driver data. It seems the only reason for the 'dev' field is to >> temporarily hold a ref to the device so it can be put later. Why not >> just put the device. After thinking about this further, I question whether we even need this function if it is going to return 'struct device *'. In that case, why not just call driver_find_device() when the device is needed? If you want to keep the function, then the function needs only one statement: return driver_find_device(...). > > Having looked at the remainder of the patches, I tend to agree that we > don't really need the backlink; we walk the driver's list of devices in > any case IIUC. > > We *might* want a mechanism to grab the queue quickly (i.e. without > walking the list) if there's anything performance sensitive in there; > but from the patch descriptions, I don't think anything is done in a > hot path, so it should be fine. >
diff --git a/drivers/s390/crypto/vfio_ap_ops.c b/drivers/s390/crypto/vfio_ap_ops.c index 900b9cf..2a52c9b 100644 --- a/drivers/s390/crypto/vfio_ap_ops.c +++ b/drivers/s390/crypto/vfio_ap_ops.c @@ -24,6 +24,60 @@ #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" +/** + * vfio_ap_check_apqn: check if a ap_queue is of a given APQN + * + * Returns 1 if we have a match. + * Otherwise returns 0. + */ +static int vfio_ap_check_apqn(struct device *dev, void *data) +{ + struct vfio_ap_queue *q = dev_get_drvdata(dev); + + return (q->apqn == *(int *)data); +} + +/** + * vfio_ap_get_queue: Retrieve a queue with a specific APQN + * @apqn: The queue APQN + * + * Retrieve a queue with a specific APQN from the list of the + * devices associated to the vfio_ap_driver. + * + * The vfio_ap_queue has been already associated with the device + * during the probe. + * Store the associated device for reference counting + * + * Returns the pointer to the associated vfio_ap_queue + */ +static __attribute__((unused)) + struct vfio_ap_queue *vfio_ap_get_queue(int apqn) +{ + struct device *dev; + struct vfio_ap_queue *q; + + dev = driver_find_device(&matrix_dev->vfio_ap_drv->driver, NULL, &apqn, + vfio_ap_check_apqn); + if (!dev) + return NULL; + q = dev_get_drvdata(dev); + q->dev = dev; + return q; +} + +/** + * vfio_ap_put_queue: lower device reference count for a queue + * @q: The queue + * + * put the associated device + * + */ +static __attribute__((unused)) void vfio_ap_put_queue(struct vfio_ap_queue *q) +{ + put_device(q->dev); + q->dev = NULL; +} + static void vfio_ap_matrix_init(struct ap_config_info *info, struct ap_matrix *matrix) { diff --git a/drivers/s390/crypto/vfio_ap_private.h b/drivers/s390/crypto/vfio_ap_private.h index 8836f01..081f0d7 100644 --- a/drivers/s390/crypto/vfio_ap_private.h +++ b/drivers/s390/crypto/vfio_ap_private.h @@ -87,6 +87,7 @@ extern int vfio_ap_mdev_register(void); extern void vfio_ap_mdev_unregister(void); struct vfio_ap_queue { + struct device *dev; int apqn; }; #endif /* _VFIO_AP_PRIVATE_H_ */
We need to find the queue with a specific APQN during the handling of the interception of the PQAP/AQIC instruction. To handle the AP associated device reference count we keep track of it in the vfio_ap_queue until we put the device. Signed-off-by: Pierre Morel <pmorel@linux.ibm.com> --- drivers/s390/crypto/vfio_ap_ops.c | 54 +++++++++++++++++++++++++++++++++++ drivers/s390/crypto/vfio_ap_private.h | 1 + 2 files changed, 55 insertions(+)