| Message ID | 980897e53f7cc2ec9bbbf58d9d451ee56a249309.1307875512.git.hans.verkuil@cisco.com (mailing list archive) |
|---|---|
| State | RFC |
| Headers | show |
Em 12-06-2011 07:59, Hans Verkuil escreveu: > From: Hans Verkuil <hans.verkuil@cisco.com> > > The check_mode function checks whether a mode is supported. So calling it > supported_mode is more appropriate. In addition it returned either 0 or > -EINVAL which suggests that the -EINVAL error should be passed on. However, > that's not the case so change the return type to bool. I prefer to keep returning -EINVAL. This is the proper thing to do, and to return the result to the caller. A fixme should be added though, so, after someone add a subdev call that would properly handle the -EINVAL code for multiple tuners, the functions should return the error code instead of 0. > > Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com> > --- > drivers/media/video/tuner-core.c | 19 ++++++++----------- > 1 files changed, 8 insertions(+), 11 deletions(-) > > diff --git a/drivers/media/video/tuner-core.c b/drivers/media/video/tuner-core.c > index 5748d04..083b9f1 100644 > --- a/drivers/media/video/tuner-core.c > +++ b/drivers/media/video/tuner-core.c > @@ -723,22 +723,19 @@ static int tuner_remove(struct i2c_client *client) > */ > > /** > - * check_mode - Verify if tuner supports the requested mode > + * supported_mode - Verify if tuner supports the requested mode > * @t: a pointer to the module's internal struct_tuner > * > * This function checks if the tuner is capable of tuning analog TV, > * digital TV or radio, depending on what the caller wants. If the > - * tuner can't support that mode, it returns -EINVAL. Otherwise, it > - * returns 0. > + * tuner can't support that mode, it returns false. Otherwise, it > + * returns true. > * This function is needed for boards that have a separate tuner for > * radio (like devices with tea5767). > */ > -static inline int check_mode(struct tuner *t, enum v4l2_tuner_type mode) > +static bool supported_mode(struct tuner *t, enum v4l2_tuner_type mode) > { > - if ((1 << mode & t->mode_mask) == 0) > - return -EINVAL; > - > - return 0; > + return 1 << mode & t->mode_mask; > } > > /** > @@ -759,7 +756,7 @@ static int set_mode_freq(struct i2c_client *client, struct tuner *t, > struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; > > if (mode != t->mode) { > - if (check_mode(t, mode) == -EINVAL) { > + if (!supported_mode(t, mode)) { > tuner_dbg("Tuner doesn't support mode %d. " > "Putting tuner to sleep\n", mode); > t->standby = true; > @@ -1138,7 +1135,7 @@ static int tuner_g_frequency(struct v4l2_subdev *sd, struct v4l2_frequency *f) > struct tuner *t = to_tuner(sd); > struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; > > - if (check_mode(t, f->type) == -EINVAL) > + if (!supported_mode(t, f->type)) > return 0; > f->type = t->mode; > if (fe_tuner_ops->get_frequency && !t->standby) { > @@ -1161,7 +1158,7 @@ static int tuner_g_tuner(struct v4l2_subdev *sd, struct v4l2_tuner *vt) > struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; > struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; > > - if (check_mode(t, vt->type) == -EINVAL) > + if (!supported_mode(t, vt->type)) > return 0; > vt->type = t->mode; > if (analog_ops->get_afc) -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: > Em 12-06-2011 07:59, Hans Verkuil escreveu: > > From: Hans Verkuil <hans.verkuil@cisco.com> > > > > The check_mode function checks whether a mode is supported. So calling it > > supported_mode is more appropriate. In addition it returned either 0 or > > -EINVAL which suggests that the -EINVAL error should be passed on. However, > > that's not the case so change the return type to bool. > > I prefer to keep returning -EINVAL. This is the proper thing to do, and > to return the result to the caller. A fixme should be added though, so, > after someone add a subdev call that would properly handle the -EINVAL > code for multiple tuners, the functions should return the error code > instead of 0. No, you can't return -EINVAL here. It is the responsibility of the bridge driver to determine whether there is e.g. a radio tuner. So if one of these tuner subdevs is called with mode radio while it is a tv tuner, then that is not an error, but instead it is a request that can safely be ignored as not relevant for that tuner. It is up to the bridge driver to ensure that a tuner is loaded that is actually valid for the radio mode. Subdev ops should only return errors when there is a real problem (e.g. i2c errors) and should just return 0 if a request is not for them. That's why I posted these first two patches: these functions suggest that you can return an error if the mode doesn't match when you really cannot. If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error that is returned should match a real error (e.g. an i2c error), not that one of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, otherwise there wouldn't be a /dev/radio node. I firmly believe that the RFCv4 series is correct and just needs an additional patch adding some documentation. That said, it would make sense to move the first three patches to the end instead if you prefer. Since these are cleanups, not bug fixes like the others. Regards, Hans > > > > > Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com> > > --- > > drivers/media/video/tuner-core.c | 19 ++++++++----------- > > 1 files changed, 8 insertions(+), 11 deletions(-) > > > > diff --git a/drivers/media/video/tuner-core.c b/drivers/media/video/tuner-core.c > > index 5748d04..083b9f1 100644 > > --- a/drivers/media/video/tuner-core.c > > +++ b/drivers/media/video/tuner-core.c > > @@ -723,22 +723,19 @@ static int tuner_remove(struct i2c_client *client) > > */ > > > > /** > > - * check_mode - Verify if tuner supports the requested mode > > + * supported_mode - Verify if tuner supports the requested mode > > * @t: a pointer to the module's internal struct_tuner > > * > > * This function checks if the tuner is capable of tuning analog TV, > > * digital TV or radio, depending on what the caller wants. If the > > - * tuner can't support that mode, it returns -EINVAL. Otherwise, it > > - * returns 0. > > + * tuner can't support that mode, it returns false. Otherwise, it > > + * returns true. > > * This function is needed for boards that have a separate tuner for > > * radio (like devices with tea5767). > > */ > > -static inline int check_mode(struct tuner *t, enum v4l2_tuner_type mode) > > +static bool supported_mode(struct tuner *t, enum v4l2_tuner_type mode) > > { > > - if ((1 << mode & t->mode_mask) == 0) > > - return -EINVAL; > > - > > - return 0; > > + return 1 << mode & t->mode_mask; > > } > > > > /** > > @@ -759,7 +756,7 @@ static int set_mode_freq(struct i2c_client *client, struct tuner *t, > > struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; > > > > if (mode != t->mode) { > > - if (check_mode(t, mode) == -EINVAL) { > > + if (!supported_mode(t, mode)) { > > tuner_dbg("Tuner doesn't support mode %d. " > > "Putting tuner to sleep\n", mode); > > t->standby = true; > > @@ -1138,7 +1135,7 @@ static int tuner_g_frequency(struct v4l2_subdev *sd, struct v4l2_frequency *f) > > struct tuner *t = to_tuner(sd); > > struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; > > > > - if (check_mode(t, f->type) == -EINVAL) > > + if (!supported_mode(t, f->type)) > > return 0; > > f->type = t->mode; > > if (fe_tuner_ops->get_frequency && !t->standby) { > > @@ -1161,7 +1158,7 @@ static int tuner_g_tuner(struct v4l2_subdev *sd, struct v4l2_tuner *vt) > > struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; > > struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; > > > > - if (check_mode(t, vt->type) == -EINVAL) > > + if (!supported_mode(t, vt->type)) > > return 0; > > vt->type = t->mode; > > if (analog_ops->get_afc) > > -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Em 12-06-2011 13:07, Hans Verkuil escreveu: > On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: >> Em 12-06-2011 07:59, Hans Verkuil escreveu: >>> From: Hans Verkuil <hans.verkuil@cisco.com> >>> >>> The check_mode function checks whether a mode is supported. So calling it >>> supported_mode is more appropriate. In addition it returned either 0 or >>> -EINVAL which suggests that the -EINVAL error should be passed on. However, >>> that's not the case so change the return type to bool. >> >> I prefer to keep returning -EINVAL. This is the proper thing to do, and >> to return the result to the caller. A fixme should be added though, so, >> after someone add a subdev call that would properly handle the -EINVAL >> code for multiple tuners, the functions should return the error code >> instead of 0. > > No, you can't return -EINVAL here. It is the responsibility of the bridge > driver to determine whether there is e.g. a radio tuner. So if one of these > tuner subdevs is called with mode radio while it is a tv tuner, then that > is not an error, but instead it is a request that can safely be ignored > as not relevant for that tuner. It is up to the bridge driver to ensure > that a tuner is loaded that is actually valid for the radio mode. > > Subdev ops should only return errors when there is a real problem (e.g. i2c > errors) and should just return 0 if a request is not for them. > > That's why I posted these first two patches: these functions suggest that you > can return an error if the mode doesn't match when you really cannot. > > If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error > that is returned should match a real error (e.g. an i2c error), not that one > of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, > otherwise there wouldn't be a /dev/radio node. > > I firmly believe that the RFCv4 series is correct and just needs an additional > patch adding some documentation. > > That said, it would make sense to move the first three patches to the end > instead if you prefer. Since these are cleanups, not bug fixes like the others. The errors at tuner should be propagated. If there's just one tuner, the error code should just be returned by the ioctl. But, if there are two tuners, if the bridge driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the other tuner returns 0, then it should return 0. So, it is about the opposite behaviour implemented at v4l2_device_call_until_err(). In order to implement the correct behaviour, the tuner driver should return -EINVAL if check_mode/set_mode fails. However, this breaks any bridge that may be using v4l2_device_call_until_err(). That's why the current code returns 0. The proper fix for it is: 1) create a call_all function that returns 0 if one of the subdevs returned 0, or returns an error code otherwise; 2) change all bridge calls to tuner stuff to the new call_all function; 3) return the check_mode/set_mode error to the bridge. One alternative for (1) would be to simply change the v4l2_device_call_all() to return 0 if one of the subdrivers returned 0. Something like (not tested): #define __v4l2_device_call_subdevs_p(v4l2_dev, sd, cond, o, f, args..$ ({ \ long __rc = 0, __err = 0; \ \ list_for_each_entry((sd), &(v4l2_dev)->subdevs, list) { \ if ((cond) && (sd)->ops->o && (sd)->ops->o->f) { \ __err = (sd)->ops->o->f((sd) , ##args); \ if (_err) \ __rc = __err; \ } \ } \ __rc; \ }) #define v4l2_device_call_all(v4l2_dev, grpid, o, f, args...) \ do { \ struct v4l2_subdev *__sd; \ \ __v4l2_device_call_subdevs_p(v4l2_dev, __sd, \ !(grpid) || __sd->grp_id == (grpid), o, f , \ ##args); \ } while (0) As it currently doesn't return any error, this shouldn't break any driver (as nobody expects an error code there). We'll need to review the bridge drivers anyway, so that they'll return the error code from v4l_device_call(). Cheers, Mauro. -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Em 12-06-2011 14:27, Mauro Carvalho Chehab escreveu: > Em 12-06-2011 13:07, Hans Verkuil escreveu: >> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: >>> Em 12-06-2011 07:59, Hans Verkuil escreveu: >>>> From: Hans Verkuil <hans.verkuil@cisco.com> >>>> >>>> The check_mode function checks whether a mode is supported. So calling it >>>> supported_mode is more appropriate. In addition it returned either 0 or >>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, >>>> that's not the case so change the return type to bool. >>> >>> I prefer to keep returning -EINVAL. This is the proper thing to do, and >>> to return the result to the caller. A fixme should be added though, so, >>> after someone add a subdev call that would properly handle the -EINVAL >>> code for multiple tuners, the functions should return the error code >>> instead of 0. >> >> No, you can't return -EINVAL here. It is the responsibility of the bridge >> driver to determine whether there is e.g. a radio tuner. So if one of these >> tuner subdevs is called with mode radio while it is a tv tuner, then that >> is not an error, but instead it is a request that can safely be ignored >> as not relevant for that tuner. It is up to the bridge driver to ensure >> that a tuner is loaded that is actually valid for the radio mode. >> >> Subdev ops should only return errors when there is a real problem (e.g. i2c >> errors) and should just return 0 if a request is not for them. >> >> That's why I posted these first two patches: these functions suggest that you >> can return an error if the mode doesn't match when you really cannot. >> >> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error >> that is returned should match a real error (e.g. an i2c error), not that one >> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, >> otherwise there wouldn't be a /dev/radio node. >> >> I firmly believe that the RFCv4 series is correct and just needs an additional >> patch adding some documentation. >> >> That said, it would make sense to move the first three patches to the end >> instead if you prefer. Since these are cleanups, not bug fixes like the others. > > > The errors at tuner should be propagated. If there's just one tuner, the error > code should just be returned by the ioctl. But, if there are two tuners, if the bridge > driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, > then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the > other tuner returns 0, then it should return 0. So, it is about the opposite behaviour > implemented at v4l2_device_call_until_err(). > > In order to implement the correct behaviour, the tuner driver should return -EINVAL if > check_mode/set_mode fails. However, this breaks any bridge that may be using > v4l2_device_call_until_err(). That's why the current code returns 0. > > The proper fix for it is: > > 1) create a call_all function that returns 0 if one of the subdevs returned 0, > or returns an error code otherwise; > 2) change all bridge calls to tuner stuff to the new call_all function; > 3) return the check_mode/set_mode error to the bridge. > > One alternative for (1) would be to simply change the v4l2_device_call_all() to return 0 if > one of the subdrivers returned 0. Something like (not tested): > Sorry, wrong logic. It should be, instead: #define __v4l2_device_call_subdevs_p(v4l2_dev, sd, cond, o, f, args..$ ({ \ long __rc = -ENOIOCTLCMD, __err = _rc; \ \ list_for_each_entry((sd), &(v4l2_dev)->subdevs, list) { \ if ((cond) && (sd)->ops->o && (sd)->ops->o->f) { \ __err = (sd)->ops->o->f((sd) , ##args); \ if (!_err) \ __rc = 0; \ } \ } \ (__rc == 0) ? 0 : __err; \ }) -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: > Em 12-06-2011 13:07, Hans Verkuil escreveu: > > On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: > >> Em 12-06-2011 07:59, Hans Verkuil escreveu: > >>> From: Hans Verkuil <hans.verkuil@cisco.com> > >>> > >>> The check_mode function checks whether a mode is supported. So calling it > >>> supported_mode is more appropriate. In addition it returned either 0 or > >>> -EINVAL which suggests that the -EINVAL error should be passed on. However, > >>> that's not the case so change the return type to bool. > >> > >> I prefer to keep returning -EINVAL. This is the proper thing to do, and > >> to return the result to the caller. A fixme should be added though, so, > >> after someone add a subdev call that would properly handle the -EINVAL > >> code for multiple tuners, the functions should return the error code > >> instead of 0. > > > > No, you can't return -EINVAL here. It is the responsibility of the bridge > > driver to determine whether there is e.g. a radio tuner. So if one of these > > tuner subdevs is called with mode radio while it is a tv tuner, then that > > is not an error, but instead it is a request that can safely be ignored > > as not relevant for that tuner. It is up to the bridge driver to ensure > > that a tuner is loaded that is actually valid for the radio mode. > > > > Subdev ops should only return errors when there is a real problem (e.g. i2c > > errors) and should just return 0 if a request is not for them. > > > > That's why I posted these first two patches: these functions suggest that you > > can return an error if the mode doesn't match when you really cannot. > > > > If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error > > that is returned should match a real error (e.g. an i2c error), not that one > > of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, > > otherwise there wouldn't be a /dev/radio node. > > > > I firmly believe that the RFCv4 series is correct and just needs an additional > > patch adding some documentation. > > > > That said, it would make sense to move the first three patches to the end > > instead if you prefer. Since these are cleanups, not bug fixes like the others. > > > The errors at tuner should be propagated. If there's just one tuner, the error > code should just be returned by the ioctl. But, if there are two tuners, if the bridge > driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, > then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the > other tuner returns 0, then it should return 0. So, it is about the opposite behaviour > implemented at v4l2_device_call_until_err(). Sorry, but no, that's not true. You are trying to use the error codes from tuner subdevs to determine whether none of the tuner subdevs support a certain tuner mode. E.g., you want to change something for a radio tuner and there are no radio tuner subdevs. But that's the job of the bridge driver to check. That has the overview, the lowly subdevs do not. The subdevs just filter the ops and check the mode to see if they should handle it and ignore it otherwise. Only if they have to handle it will they return a possible error. The big problem with trying to use subdev errors codes for this is that you don't see the difference between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev that returns -EINVAL just because it didn't understand the tuner mode. So the bridge may return -EINVAL to the application instead of the real error, which is -EIO. That's the whole principle behind the sub-device model: sub-devices do not know about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no radio tuner, then the bridge driver is the one that should detect that and return -EINVAL. Actually, as mentioned before, it can also be done in video_ioctl2.c by checking the tuner mode against the device node it's called on. But that requires tightening of the V4L2 spec first. Regards, Hans > In order to implement the correct behaviour, the tuner driver should return -EINVAL if > check_mode/set_mode fails. However, this breaks any bridge that may be using > v4l2_device_call_until_err(). That's why the current code returns 0. > > The proper fix for it is: > > 1) create a call_all function that returns 0 if one of the subdevs returned 0, > or returns an error code otherwise; > 2) change all bridge calls to tuner stuff to the new call_all function; > 3) return the check_mode/set_mode error to the bridge. > > One alternative for (1) would be to simply change the v4l2_device_call_all() to return 0 if > one of the subdrivers returned 0. Something like (not tested): > > > #define __v4l2_device_call_subdevs_p(v4l2_dev, sd, cond, o, f, args..$ > ({ \ > long __rc = 0, __err = 0; \ > \ > list_for_each_entry((sd), &(v4l2_dev)->subdevs, list) { \ > if ((cond) && (sd)->ops->o && (sd)->ops->o->f) { \ > __err = (sd)->ops->o->f((sd) , ##args); \ > if (_err) \ > __rc = __err; \ > } \ > } \ > __rc; \ > }) > > > #define v4l2_device_call_all(v4l2_dev, grpid, o, f, args...) \ > do { \ > struct v4l2_subdev *__sd; \ > \ > __v4l2_device_call_subdevs_p(v4l2_dev, __sd, \ > !(grpid) || __sd->grp_id == (grpid), o, f , \ > ##args); \ > } while (0) > > > As it currently doesn't return any error, this shouldn't break any driver > (as nobody expects an error code there). We'll need to review the bridge > drivers anyway, so that they'll return the error code from v4l_device_call(). > > Cheers, > Mauro. > -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Em 12-06-2011 15:09, Hans Verkuil escreveu: > On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: >> Em 12-06-2011 13:07, Hans Verkuil escreveu: >>> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: >>>> Em 12-06-2011 07:59, Hans Verkuil escreveu: >>>>> From: Hans Verkuil <hans.verkuil@cisco.com> >>>>> >>>>> The check_mode function checks whether a mode is supported. So calling it >>>>> supported_mode is more appropriate. In addition it returned either 0 or >>>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, >>>>> that's not the case so change the return type to bool. >>>> >>>> I prefer to keep returning -EINVAL. This is the proper thing to do, and >>>> to return the result to the caller. A fixme should be added though, so, >>>> after someone add a subdev call that would properly handle the -EINVAL >>>> code for multiple tuners, the functions should return the error code >>>> instead of 0. >>> >>> No, you can't return -EINVAL here. It is the responsibility of the bridge >>> driver to determine whether there is e.g. a radio tuner. So if one of these >>> tuner subdevs is called with mode radio while it is a tv tuner, then that >>> is not an error, but instead it is a request that can safely be ignored >>> as not relevant for that tuner. It is up to the bridge driver to ensure >>> that a tuner is loaded that is actually valid for the radio mode. >>> >>> Subdev ops should only return errors when there is a real problem (e.g. i2c >>> errors) and should just return 0 if a request is not for them. >>> >>> That's why I posted these first two patches: these functions suggest that you >>> can return an error if the mode doesn't match when you really cannot. >>> >>> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error >>> that is returned should match a real error (e.g. an i2c error), not that one >>> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, >>> otherwise there wouldn't be a /dev/radio node. >>> >>> I firmly believe that the RFCv4 series is correct and just needs an additional >>> patch adding some documentation. >>> >>> That said, it would make sense to move the first three patches to the end >>> instead if you prefer. Since these are cleanups, not bug fixes like the others. >> >> >> The errors at tuner should be propagated. If there's just one tuner, the error >> code should just be returned by the ioctl. But, if there are two tuners, if the bridge >> driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, >> then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the >> other tuner returns 0, then it should return 0. So, it is about the opposite behaviour >> implemented at v4l2_device_call_until_err(). > > Sorry, but no, that's not true. You are trying to use the error codes from tuner > subdevs to determine whether none of the tuner subdevs support a certain tuner mode. Not only that. There are some cases where the tuner driver may not bind for some reason. So, even if the bridge driver does support a certain mode, a call to G_TUNER may fail (for example, if tea5767 probe failed). Only with a proper return code, the bridge driver can detect if the driver found some issue. > E.g., you want to change something for a radio tuner and there are no radio tuner > subdevs. But that's the job of the bridge driver to check. That has the overview, > the lowly subdevs do not. The subdevs just filter the ops and check the mode to see > if they should handle it and ignore it otherwise. > > Only if they have to handle it will they return a possible error. The big problem > with trying to use subdev errors codes for this is that you don't see the difference > between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev > that returns -EINVAL just because it didn't understand the tuner mode. > > So the bridge may return -EINVAL to the application instead of the real error, which > is -EIO. An -EIO would also be discarded, as errors at v4l2_device_call_all() calls don't return anything. So, currently, the bridge has to assume that no errors happened and return 0. > That's the whole principle behind the sub-device model: sub-devices do not know > about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no > radio tuner, then the bridge driver is the one that should detect that and return > -EINVAL. > > Actually, as mentioned before, it can also be done in video_ioctl2.c by checking > the tuner mode against the device node it's called on. But that requires tightening > of the V4L2 spec first. Yes, video_ioctl2 (or, currently, the bridge driver) shouldn't allow an invalid operation. But if the call returns an error, this error should be propagated. Also, as I've explained before, even adding the invalid mode check inside video_ioctl, you may still have errors if the registered tuners don't support the mode, because one of the tuners didn't registered properly. Cheers, Mauro -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Monday, June 13, 2011 00:06:19 Mauro Carvalho Chehab wrote: > Em 12-06-2011 15:09, Hans Verkuil escreveu: > > On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: > >> Em 12-06-2011 13:07, Hans Verkuil escreveu: > >>> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: > >>>> Em 12-06-2011 07:59, Hans Verkuil escreveu: > >>>>> From: Hans Verkuil <hans.verkuil@cisco.com> > >>>>> > >>>>> The check_mode function checks whether a mode is supported. So calling it > >>>>> supported_mode is more appropriate. In addition it returned either 0 or > >>>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, > >>>>> that's not the case so change the return type to bool. > >>>> > >>>> I prefer to keep returning -EINVAL. This is the proper thing to do, and > >>>> to return the result to the caller. A fixme should be added though, so, > >>>> after someone add a subdev call that would properly handle the -EINVAL > >>>> code for multiple tuners, the functions should return the error code > >>>> instead of 0. > >>> > >>> No, you can't return -EINVAL here. It is the responsibility of the bridge > >>> driver to determine whether there is e.g. a radio tuner. So if one of these > >>> tuner subdevs is called with mode radio while it is a tv tuner, then that > >>> is not an error, but instead it is a request that can safely be ignored > >>> as not relevant for that tuner. It is up to the bridge driver to ensure > >>> that a tuner is loaded that is actually valid for the radio mode. > >>> > >>> Subdev ops should only return errors when there is a real problem (e.g. i2c > >>> errors) and should just return 0 if a request is not for them. > >>> > >>> That's why I posted these first two patches: these functions suggest that you > >>> can return an error if the mode doesn't match when you really cannot. > >>> > >>> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error > >>> that is returned should match a real error (e.g. an i2c error), not that one > >>> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, > >>> otherwise there wouldn't be a /dev/radio node. > >>> > >>> I firmly believe that the RFCv4 series is correct and just needs an additional > >>> patch adding some documentation. > >>> > >>> That said, it would make sense to move the first three patches to the end > >>> instead if you prefer. Since these are cleanups, not bug fixes like the others. > >> > >> > >> The errors at tuner should be propagated. If there's just one tuner, the error > >> code should just be returned by the ioctl. But, if there are two tuners, if the bridge > >> driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, > >> then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the > >> other tuner returns 0, then it should return 0. So, it is about the opposite behaviour > >> implemented at v4l2_device_call_until_err(). > > > > Sorry, but no, that's not true. You are trying to use the error codes from tuner > > subdevs to determine whether none of the tuner subdevs support a certain tuner mode. > > Not only that. There are some cases where the tuner driver may not bind for some reason. > So, even if the bridge driver does support a certain mode, a call to G_TUNER may fail > (for example, if tea5767 probe failed). Only with a proper return code, the bridge driver > can detect if the driver found some issue. Surely, that's an error reported by tuner_probe, isn't it? That's supposed to ensure that the tuner driver was loaded and initialized correctly. I'm not sure if it does, but that's definitely where any errors of that kind should be reported. Looking at it some more, it seems to me that s_type_addr should also return an error if there are problems. Ditto for tuner_s_config. An alternative solution is to keep a 'initialized' boolean that is set to true once the tuner is fully configured. If g_tuner et al are called when the tuner is not fully configured, then you can return -ENODEV or -EIO or something like that. But that's a separate status check and has nothing to do with mode checking. > > E.g., you want to change something for a radio tuner and there are no radio tuner > > subdevs. But that's the job of the bridge driver to check. That has the overview, > > the lowly subdevs do not. The subdevs just filter the ops and check the mode to see > > if they should handle it and ignore it otherwise. > > > > Only if they have to handle it will they return a possible error. The big problem > > with trying to use subdev errors codes for this is that you don't see the difference > > between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev > > that returns -EINVAL just because it didn't understand the tuner mode. > > > > So the bridge may return -EINVAL to the application instead of the real error, which > > is -EIO. > > An -EIO would also be discarded, as errors at v4l2_device_call_all() calls don't return > anything. So, currently, the bridge has to assume that no errors happened and return 0. Obviously, v4l2_device_call_all calls should be replaced with v4l2_device_call_until_err. I've no problem with that. > > > That's the whole principle behind the sub-device model: sub-devices do not know > > about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no > > radio tuner, then the bridge driver is the one that should detect that and return > > -EINVAL. > > > > Actually, as mentioned before, it can also be done in video_ioctl2.c by checking > > the tuner mode against the device node it's called on. But that requires tightening > > of the V4L2 spec first. > > Yes, video_ioctl2 (or, currently, the bridge driver) shouldn't allow an invalid operation. > But if the call returns an error, this error should be propagated. > > Also, as I've explained before, even adding the invalid mode check inside video_ioctl, > you may still have errors if the registered tuners don't support the mode, because one > of the tuners didn't registered properly. And that's something that tuner_probe/s_type_addr/s_config should have detected. Or, should that be impossible (I would have to spend more time to analyze that) we might have to add a 'validate_tuner' op that can be called to verify all tuners are configured correctly. Regards, Hans -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Em 13-06-2011 07:23, Hans Verkuil escreveu: > On Monday, June 13, 2011 00:06:19 Mauro Carvalho Chehab wrote: >> Em 12-06-2011 15:09, Hans Verkuil escreveu: >>> On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: >>>> Em 12-06-2011 13:07, Hans Verkuil escreveu: >>>>> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: >>>>>> Em 12-06-2011 07:59, Hans Verkuil escreveu: >>>>>>> From: Hans Verkuil <hans.verkuil@cisco.com> >>>>>>> >>>>>>> The check_mode function checks whether a mode is supported. So calling it >>>>>>> supported_mode is more appropriate. In addition it returned either 0 or >>>>>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, >>>>>>> that's not the case so change the return type to bool. >>>>>> >>>>>> I prefer to keep returning -EINVAL. This is the proper thing to do, and >>>>>> to return the result to the caller. A fixme should be added though, so, >>>>>> after someone add a subdev call that would properly handle the -EINVAL >>>>>> code for multiple tuners, the functions should return the error code >>>>>> instead of 0. >>>>> >>>>> No, you can't return -EINVAL here. It is the responsibility of the bridge >>>>> driver to determine whether there is e.g. a radio tuner. So if one of these >>>>> tuner subdevs is called with mode radio while it is a tv tuner, then that >>>>> is not an error, but instead it is a request that can safely be ignored >>>>> as not relevant for that tuner. It is up to the bridge driver to ensure >>>>> that a tuner is loaded that is actually valid for the radio mode. >>>>> >>>>> Subdev ops should only return errors when there is a real problem (e.g. i2c >>>>> errors) and should just return 0 if a request is not for them. >>>>> >>>>> That's why I posted these first two patches: these functions suggest that you >>>>> can return an error if the mode doesn't match when you really cannot. >>>>> >>>>> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error >>>>> that is returned should match a real error (e.g. an i2c error), not that one >>>>> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, >>>>> otherwise there wouldn't be a /dev/radio node. >>>>> >>>>> I firmly believe that the RFCv4 series is correct and just needs an additional >>>>> patch adding some documentation. >>>>> >>>>> That said, it would make sense to move the first three patches to the end >>>>> instead if you prefer. Since these are cleanups, not bug fixes like the others. >>>> >>>> >>>> The errors at tuner should be propagated. If there's just one tuner, the error >>>> code should just be returned by the ioctl. But, if there are two tuners, if the bridge >>>> driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, >>>> then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the >>>> other tuner returns 0, then it should return 0. So, it is about the opposite behaviour >>>> implemented at v4l2_device_call_until_err(). >>> >>> Sorry, but no, that's not true. You are trying to use the error codes from tuner >>> subdevs to determine whether none of the tuner subdevs support a certain tuner mode. >> >> Not only that. There are some cases where the tuner driver may not bind for some reason. >> So, even if the bridge driver does support a certain mode, a call to G_TUNER may fail >> (for example, if tea5767 probe failed). Only with a proper return code, the bridge driver >> can detect if the driver found some issue. > > Surely, that's an error reported by tuner_probe, isn't it? That's supposed to ensure > that the tuner driver was loaded and initialized correctly. I'm not sure if it does, > but that's definitely where any errors of that kind should be reported. > > Looking at it some more, it seems to me that s_type_addr should also return an > error if there are problems. Ditto for tuner_s_config. > > An alternative solution is to keep a 'initialized' boolean that is set to true > once the tuner is fully configured. If g_tuner et al are called when the tuner > is not fully configured, then you can return -ENODEV or -EIO or something like that. NACK. This would be just an ugly workaround. > But that's a separate status check and has nothing to do with mode checking. > >>> E.g., you want to change something for a radio tuner and there are no radio tuner >>> subdevs. But that's the job of the bridge driver to check. That has the overview, >>> the lowly subdevs do not. The subdevs just filter the ops and check the mode to see >>> if they should handle it and ignore it otherwise. >>> >>> Only if they have to handle it will they return a possible error. The big problem >>> with trying to use subdev errors codes for this is that you don't see the difference >>> between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev >>> that returns -EINVAL just because it didn't understand the tuner mode. >>> >>> So the bridge may return -EINVAL to the application instead of the real error, which >>> is -EIO. >> >> An -EIO would also be discarded, as errors at v4l2_device_call_all() calls don't return >> anything. So, currently, the bridge has to assume that no errors happened and return 0. > > Obviously, v4l2_device_call_all calls should be replaced with v4l2_device_call_until_err. > I've no problem with that. See bellow. >> >>> That's the whole principle behind the sub-device model: sub-devices do not know >>> about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no >>> radio tuner, then the bridge driver is the one that should detect that and return >>> -EINVAL. >>> >>> Actually, as mentioned before, it can also be done in video_ioctl2.c by checking >>> the tuner mode against the device node it's called on. But that requires tightening >>> of the V4L2 spec first. >> >> Yes, video_ioctl2 (or, currently, the bridge driver) shouldn't allow an invalid operation. >> But if the call returns an error, this error should be propagated. >> >> Also, as I've explained before, even adding the invalid mode check inside video_ioctl, >> you may still have errors if the registered tuners don't support the mode, because one >> of the tuners didn't registered properly. > > And that's something that tuner_probe/s_type_addr/s_config should have detected. I'm almost sure that they don't do it, currently: s_type_addr/s_config also calls v4l2_device_call_all(). No errors are returned back. The tuner_probe call also can't do much, as it doesn't know in advance if the device has one or two tuners. > Or, should that be impossible (I would have to spend more time to analyze that) > we might have to add a 'validate_tuner' op that can be called to verify all tuners > are configured correctly. You're wanting to create a very complex patchset just to justify that replacing from -EINVAL to a bool is the right thing to do. It isn't. The point is that: if, for any reason, an ioctl fails, it should return an _error_, and _not_ a boolean. After fixing v4l2_device_call_all() to allow it to return errors, the next step is to review all calls to it, and add a proper handler for the errors. s_type_addr, s_config, g_tuner, s_tuner, etc should be handling errors. In other words, the original v4l2_device_call_all() that were just replicating the previous I2C behaviour is a mistake, as it doesn't provide any feedback about errors. This needs to be replaced by something that it is aware of the errors. If you take a look at v4l2-subdev, there's just one operation that doesn't return an error (v4l2_subdev_internal_ops.unregistered, never called from drivers). All the others returns an error. However, the default usage is to simply discard errors. This is wrong. Errors should be propagated. AFAIK, there are only a two types error propagation that are currently needed: 1) Call all subdevices. If one returns 0, assumes that the operation succeeded. This is used when there are multiple subdevs, but they're mutually exclusive: only one of them will handle such call. It is needed by tuners and by controls, on devices that have several subdevs providing different sets of controls. 2) Call all subdevices until an error. Used when the same operation needs to be set on multiple subdevices. The subdevices that don't implement such operation should return -ENOIOCTLCMD. Btw, v4l2_device_call_subdevs_until_err() has currently a bug: if all sub-devices return -ENOIOCTLCMD, it returns 0. It should, instead, return -ENOIOCTLCMD, in order to allow the bridge drivers to return an error code to the userspace, to indicate that the IOCTL was not handled. Eventually, we may just use (2) for everything, if we patch all subdev drivers to return -ENOIOCTLCMD if they are discarding a subdev call, but, in this case, the bridge drivers will need to replace the -ENOIOCTLCMD to an error code defined at the V4L2 spec (or we can have a macro for that). A side note: the only error codes defined at the media API DocBook are: -EACCES, -EAGAIN, -EBADF, -EBUSY, -EFAULT, -EIO, -EINTR, -EINVAL, -ENFILE, -ENOMEM, -ENOSPC, -ENOTTY, -ENXIO, -EMFILE, -EPERM, -ERANGE and EPIPE. On most places, the error codes are defined per ioctl. We need to review the DocBook and the drivers to be sure that they match the API specs, in terms of returned codes. It probably makes sense to create a section with the valid error codes, remove most of error codes comments from each ioctl, and add a link to the global error code section. Ah, -ENODEV is not currently defined, but -ENXIO is defined on a few places. -ENXIO means "No such device or address". So, it may make sense to replace all -ENODEV to -ENXIO at the drivers. Cheers, Mauro -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Monday, June 13, 2011 13:45:14 Mauro Carvalho Chehab wrote: > Em 13-06-2011 07:23, Hans Verkuil escreveu: > > On Monday, June 13, 2011 00:06:19 Mauro Carvalho Chehab wrote: > >> Em 12-06-2011 15:09, Hans Verkuil escreveu: > >>> On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: > >>>> Em 12-06-2011 13:07, Hans Verkuil escreveu: > >>>>> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: > >>>>>> Em 12-06-2011 07:59, Hans Verkuil escreveu: > >>>>>>> From: Hans Verkuil <hans.verkuil@cisco.com> > >>>>>>> > >>>>>>> The check_mode function checks whether a mode is supported. So calling it > >>>>>>> supported_mode is more appropriate. In addition it returned either 0 or > >>>>>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, > >>>>>>> that's not the case so change the return type to bool. > >>>>>> > >>>>>> I prefer to keep returning -EINVAL. This is the proper thing to do, and > >>>>>> to return the result to the caller. A fixme should be added though, so, > >>>>>> after someone add a subdev call that would properly handle the -EINVAL > >>>>>> code for multiple tuners, the functions should return the error code > >>>>>> instead of 0. > >>>>> > >>>>> No, you can't return -EINVAL here. It is the responsibility of the bridge > >>>>> driver to determine whether there is e.g. a radio tuner. So if one of these > >>>>> tuner subdevs is called with mode radio while it is a tv tuner, then that > >>>>> is not an error, but instead it is a request that can safely be ignored > >>>>> as not relevant for that tuner. It is up to the bridge driver to ensure > >>>>> that a tuner is loaded that is actually valid for the radio mode. > >>>>> > >>>>> Subdev ops should only return errors when there is a real problem (e.g. i2c > >>>>> errors) and should just return 0 if a request is not for them. > >>>>> > >>>>> That's why I posted these first two patches: these functions suggest that you > >>>>> can return an error if the mode doesn't match when you really cannot. > >>>>> > >>>>> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error > >>>>> that is returned should match a real error (e.g. an i2c error), not that one > >>>>> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, > >>>>> otherwise there wouldn't be a /dev/radio node. > >>>>> > >>>>> I firmly believe that the RFCv4 series is correct and just needs an additional > >>>>> patch adding some documentation. > >>>>> > >>>>> That said, it would make sense to move the first three patches to the end > >>>>> instead if you prefer. Since these are cleanups, not bug fixes like the others. > >>>> > >>>> > >>>> The errors at tuner should be propagated. If there's just one tuner, the error > >>>> code should just be returned by the ioctl. But, if there are two tuners, if the bridge > >>>> driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, > >>>> then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the > >>>> other tuner returns 0, then it should return 0. So, it is about the opposite behaviour > >>>> implemented at v4l2_device_call_until_err(). > >>> > >>> Sorry, but no, that's not true. You are trying to use the error codes from tuner > >>> subdevs to determine whether none of the tuner subdevs support a certain tuner mode. > >> > >> Not only that. There are some cases where the tuner driver may not bind for some reason. > >> So, even if the bridge driver does support a certain mode, a call to G_TUNER may fail > >> (for example, if tea5767 probe failed). Only with a proper return code, the bridge driver > >> can detect if the driver found some issue. > > > > Surely, that's an error reported by tuner_probe, isn't it? That's supposed to ensure > > that the tuner driver was loaded and initialized correctly. I'm not sure if it does, > > but that's definitely where any errors of that kind should be reported. > > > > Looking at it some more, it seems to me that s_type_addr should also return an > > error if there are problems. Ditto for tuner_s_config. > > > > An alternative solution is to keep a 'initialized' boolean that is set to true > > once the tuner is fully configured. If g_tuner et al are called when the tuner > > is not fully configured, then you can return -ENODEV or -EIO or something like that. > > NACK. This would be just an ugly workaround. Agreed :-) > > > But that's a separate status check and has nothing to do with mode checking. > > > >>> E.g., you want to change something for a radio tuner and there are no radio tuner > >>> subdevs. But that's the job of the bridge driver to check. That has the overview, > >>> the lowly subdevs do not. The subdevs just filter the ops and check the mode to see > >>> if they should handle it and ignore it otherwise. > >>> > >>> Only if they have to handle it will they return a possible error. The big problem > >>> with trying to use subdev errors codes for this is that you don't see the difference > >>> between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev > >>> that returns -EINVAL just because it didn't understand the tuner mode. > >>> > >>> So the bridge may return -EINVAL to the application instead of the real error, which > >>> is -EIO. > >> > >> An -EIO would also be discarded, as errors at v4l2_device_call_all() calls don't return > >> anything. So, currently, the bridge has to assume that no errors happened and return 0. > > > > Obviously, v4l2_device_call_all calls should be replaced with v4l2_device_call_until_err. > > I've no problem with that. > > See bellow. > > >> > >>> That's the whole principle behind the sub-device model: sub-devices do not know > >>> about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no > >>> radio tuner, then the bridge driver is the one that should detect that and return > >>> -EINVAL. > >>> > >>> Actually, as mentioned before, it can also be done in video_ioctl2.c by checking > >>> the tuner mode against the device node it's called on. But that requires tightening > >>> of the V4L2 spec first. > >> > >> Yes, video_ioctl2 (or, currently, the bridge driver) shouldn't allow an invalid operation. > >> But if the call returns an error, this error should be propagated. > >> > >> Also, as I've explained before, even adding the invalid mode check inside video_ioctl, > >> you may still have errors if the registered tuners don't support the mode, because one > >> of the tuners didn't registered properly. > > > > And that's something that tuner_probe/s_type_addr/s_config should have detected. > > I'm almost sure that they don't do it, currently: s_type_addr/s_config also calls > v4l2_device_call_all(). No errors are returned back. The tuner_probe call also can't > do much, as it doesn't know in advance if the device has one or two tuners. I know they don't. But that's what should happen. > > Or, should that be impossible (I would have to spend more time to analyze that) > > we might have to add a 'validate_tuner' op that can be called to verify all tuners > > are configured correctly. > > You're wanting to create a very complex patchset just to justify that replacing > from -EINVAL to a bool is the right thing to do. It isn't. The point is that: > if, for any reason, an ioctl fails, it should return an _error_, and _not_ a boolean. True for an ioctl, not necessarily true for a subdev op. I think one thing that confuses the issue here is that we have no clear error code for subdev ops that return a 'not handled' code. To some extent ENOIOCTLCMD is used for that, but it's not consistently used. Only if all subdevs called from a v4l2_device_call_all type macro return 'not handled' should it actually return some error. > After fixing v4l2_device_call_all() to allow it to return errors, the next step > is to review all calls to it, and add a proper handler for the errors. s_type_addr, > s_config, g_tuner, s_tuner, etc should be handling errors. > > In other words, the original v4l2_device_call_all() that were just replicating the > previous I2C behaviour is a mistake, as it doesn't provide any feedback about errors. > This needs to be replaced by something that it is aware of the errors. If you take > a look at v4l2-subdev, there's just one operation that doesn't return an error > (v4l2_subdev_internal_ops.unregistered, never called from drivers). All the others > returns an error. However, the default usage is to simply discard errors. This is wrong. > Errors should be propagated. It's not a mistake, it's just that nobody had the time to sort out the mess. The current behavior is basically bug-compatible with the pre-subdev days. It's only since all bridge drivers were converted to use subdevs that we can even think about cleaning up error handling. > > AFAIK, there are only a two types error propagation that are currently needed: > > 1) Call all subdevices. If one returns 0, assumes that the operation succeeded. This is > used when there are multiple subdevs, but they're mutually exclusive: only one of them > will handle such call. It is needed by tuners and by controls, on devices that have > several subdevs providing different sets of controls. > > 2) Call all subdevices until an error. Used when the same operation needs to be set > on multiple subdevices. The subdevices that don't implement such operation should > return -ENOIOCTLCMD. > > Btw, v4l2_device_call_subdevs_until_err() has currently a bug: if all sub-devices return > -ENOIOCTLCMD, it returns 0. It should, instead, return -ENOIOCTLCMD, in order to allow > the bridge drivers to return an error code to the userspace, to indicate that the > IOCTL was not handled. > > Eventually, we may just use (2) for everything, if we patch all subdev drivers to return > -ENOIOCTLCMD if they are discarding a subdev call, but, in this case, the bridge drivers > will need to replace the -ENOIOCTLCMD to an error code defined at the V4L2 spec (or we > can have a macro for that). Option 2 is the correct approach. If all subdev drivers return -ENOIOCTLCMD, then return an error. If one driver returns a non-0 and non-ENOIOCTLCMD error, then return that, otherwise return 0. > A side note: the only error codes defined at the media API DocBook are: -EACCES, -EAGAIN, > -EBADF, -EBUSY, -EFAULT, -EIO, -EINTR, -EINVAL, -ENFILE, -ENOMEM, -ENOSPC, -ENOTTY, -ENXIO, > -EMFILE, -EPERM, -ERANGE and EPIPE. On most places, the error codes are defined per ioctl. > We need to review the DocBook and the drivers to be sure that they match the API specs, > in terms of returned codes. It probably makes sense to create a section with the valid error > codes, remove most of error codes comments from each ioctl, and add a link to the global > error code section. > > Ah, -ENODEV is not currently defined, but -ENXIO is defined on a few places. -ENXIO means > "No such device or address". So, it may make sense to replace all -ENODEV to -ENXIO at > the drivers. Right, all very lovely, but I just want to fix the broken tuner code. We all know error handling is a big mess and could keep a small army of janitors busy for a year. For now I'll remove those two offending patches and redo the patch series without them. Regards, Hans -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Em 13-06-2011 09:07, Hans Verkuil escreveu: > On Monday, June 13, 2011 13:45:14 Mauro Carvalho Chehab wrote: >> Em 13-06-2011 07:23, Hans Verkuil escreveu: >>> On Monday, June 13, 2011 00:06:19 Mauro Carvalho Chehab wrote: >>>> Em 12-06-2011 15:09, Hans Verkuil escreveu: >>>>> On Sunday, June 12, 2011 19:27:21 Mauro Carvalho Chehab wrote: >>>>>> Em 12-06-2011 13:07, Hans Verkuil escreveu: >>>>>>> On Sunday, June 12, 2011 16:37:55 Mauro Carvalho Chehab wrote: >>>>>>>> Em 12-06-2011 07:59, Hans Verkuil escreveu: >>>>>>>>> From: Hans Verkuil <hans.verkuil@cisco.com> >>>>>>>>> >>>>>>>>> The check_mode function checks whether a mode is supported. So calling it >>>>>>>>> supported_mode is more appropriate. In addition it returned either 0 or >>>>>>>>> -EINVAL which suggests that the -EINVAL error should be passed on. However, >>>>>>>>> that's not the case so change the return type to bool. >>>>>>>> >>>>>>>> I prefer to keep returning -EINVAL. This is the proper thing to do, and >>>>>>>> to return the result to the caller. A fixme should be added though, so, >>>>>>>> after someone add a subdev call that would properly handle the -EINVAL >>>>>>>> code for multiple tuners, the functions should return the error code >>>>>>>> instead of 0. >>>>>>> >>>>>>> No, you can't return -EINVAL here. It is the responsibility of the bridge >>>>>>> driver to determine whether there is e.g. a radio tuner. So if one of these >>>>>>> tuner subdevs is called with mode radio while it is a tv tuner, then that >>>>>>> is not an error, but instead it is a request that can safely be ignored >>>>>>> as not relevant for that tuner. It is up to the bridge driver to ensure >>>>>>> that a tuner is loaded that is actually valid for the radio mode. >>>>>>> >>>>>>> Subdev ops should only return errors when there is a real problem (e.g. i2c >>>>>>> errors) and should just return 0 if a request is not for them. >>>>>>> >>>>>>> That's why I posted these first two patches: these functions suggest that you >>>>>>> can return an error if the mode doesn't match when you really cannot. >>>>>>> >>>>>>> If I call v4l2_device_call_until_err() for e.g. s_frequency, then the error >>>>>>> that is returned should match a real error (e.g. an i2c error), not that one >>>>>>> of the tv tuners refused the radio mode. I know there is a radio tuner somewhere, >>>>>>> otherwise there wouldn't be a /dev/radio node. >>>>>>> >>>>>>> I firmly believe that the RFCv4 series is correct and just needs an additional >>>>>>> patch adding some documentation. >>>>>>> >>>>>>> That said, it would make sense to move the first three patches to the end >>>>>>> instead if you prefer. Since these are cleanups, not bug fixes like the others. >>>>>> >>>>>> >>>>>> The errors at tuner should be propagated. If there's just one tuner, the error >>>>>> code should just be returned by the ioctl. But, if there are two tuners, if the bridge >>>>>> driver calls G_TUNER (or any other tuner subdev call) and both tuners return -EINVAL, >>>>>> then it needs to return -EINVAL to userspace. If just one returns -EINVAL, and the >>>>>> other tuner returns 0, then it should return 0. So, it is about the opposite behaviour >>>>>> implemented at v4l2_device_call_until_err(). >>>>> >>>>> Sorry, but no, that's not true. You are trying to use the error codes from tuner >>>>> subdevs to determine whether none of the tuner subdevs support a certain tuner mode. >>>> >>>> Not only that. There are some cases where the tuner driver may not bind for some reason. >>>> So, even if the bridge driver does support a certain mode, a call to G_TUNER may fail >>>> (for example, if tea5767 probe failed). Only with a proper return code, the bridge driver >>>> can detect if the driver found some issue. >>> >>> Surely, that's an error reported by tuner_probe, isn't it? That's supposed to ensure >>> that the tuner driver was loaded and initialized correctly. I'm not sure if it does, >>> but that's definitely where any errors of that kind should be reported. >>> >>> Looking at it some more, it seems to me that s_type_addr should also return an >>> error if there are problems. Ditto for tuner_s_config. >>> >>> An alternative solution is to keep a 'initialized' boolean that is set to true >>> once the tuner is fully configured. If g_tuner et al are called when the tuner >>> is not fully configured, then you can return -ENODEV or -EIO or something like that. >> >> NACK. This would be just an ugly workaround. > > Agreed :-) > >> >>> But that's a separate status check and has nothing to do with mode checking. >>> >>>>> E.g., you want to change something for a radio tuner and there are no radio tuner >>>>> subdevs. But that's the job of the bridge driver to check. That has the overview, >>>>> the lowly subdevs do not. The subdevs just filter the ops and check the mode to see >>>>> if they should handle it and ignore it otherwise. >>>>> >>>>> Only if they have to handle it will they return a possible error. The big problem >>>>> with trying to use subdev errors codes for this is that you don't see the difference >>>>> between a real error of a subdev (e.g. -EIO when an i2c access fails) and a subdev >>>>> that returns -EINVAL just because it didn't understand the tuner mode. >>>>> >>>>> So the bridge may return -EINVAL to the application instead of the real error, which >>>>> is -EIO. >>>> >>>> An -EIO would also be discarded, as errors at v4l2_device_call_all() calls don't return >>>> anything. So, currently, the bridge has to assume that no errors happened and return 0. >>> >>> Obviously, v4l2_device_call_all calls should be replaced with v4l2_device_call_until_err. >>> I've no problem with that. >> >> See bellow. >> >>>> >>>>> That's the whole principle behind the sub-device model: sub-devices do not know >>>>> about 'the world outside'. So if you pass RADIO mode to S_FREQUENCY and there is no >>>>> radio tuner, then the bridge driver is the one that should detect that and return >>>>> -EINVAL. >>>>> >>>>> Actually, as mentioned before, it can also be done in video_ioctl2.c by checking >>>>> the tuner mode against the device node it's called on. But that requires tightening >>>>> of the V4L2 spec first. >>>> >>>> Yes, video_ioctl2 (or, currently, the bridge driver) shouldn't allow an invalid operation. >>>> But if the call returns an error, this error should be propagated. >>>> >>>> Also, as I've explained before, even adding the invalid mode check inside video_ioctl, >>>> you may still have errors if the registered tuners don't support the mode, because one >>>> of the tuners didn't registered properly. >>> >>> And that's something that tuner_probe/s_type_addr/s_config should have detected. >> >> I'm almost sure that they don't do it, currently: s_type_addr/s_config also calls >> v4l2_device_call_all(). No errors are returned back. The tuner_probe call also can't >> do much, as it doesn't know in advance if the device has one or two tuners. > > I know they don't. But that's what should happen. > >>> Or, should that be impossible (I would have to spend more time to analyze that) >>> we might have to add a 'validate_tuner' op that can be called to verify all tuners >>> are configured correctly. >> >> You're wanting to create a very complex patchset just to justify that replacing >> from -EINVAL to a bool is the right thing to do. It isn't. The point is that: >> if, for any reason, an ioctl fails, it should return an _error_, and _not_ a boolean. > > True for an ioctl, not necessarily true for a subdev op. True also for subdev op. If the bridge driver will do something else or not if a subdev has an error is up to the bridge, but the errors should be reported there. > I think one thing that > confuses the issue here is that we have no clear error code for subdev ops that > return a 'not handled' code. To some extent ENOIOCTLCMD is used for that, but > it's not consistently used. Only if all subdevs called from a v4l2_device_call_all > type macro return 'not handled' should it actually return some error. Agreed. >> After fixing v4l2_device_call_all() to allow it to return errors, the next step >> is to review all calls to it, and add a proper handler for the errors. s_type_addr, >> s_config, g_tuner, s_tuner, etc should be handling errors. >> >> In other words, the original v4l2_device_call_all() that were just replicating the >> previous I2C behaviour is a mistake, as it doesn't provide any feedback about errors. >> This needs to be replaced by something that it is aware of the errors. If you take >> a look at v4l2-subdev, there's just one operation that doesn't return an error >> (v4l2_subdev_internal_ops.unregistered, never called from drivers). All the others >> returns an error. However, the default usage is to simply discard errors. This is wrong. >> Errors should be propagated. > > It's not a mistake, it's just that nobody had the time to sort out the mess. > The current behavior is basically bug-compatible with the pre-subdev days. > It's only since all bridge drivers were converted to use subdevs that we can > even think about cleaning up error handling. > >> >> AFAIK, there are only a two types error propagation that are currently needed: >> >> 1) Call all subdevices. If one returns 0, assumes that the operation succeeded. This is >> used when there are multiple subdevs, but they're mutually exclusive: only one of them >> will handle such call. It is needed by tuners and by controls, on devices that have >> several subdevs providing different sets of controls. >> >> 2) Call all subdevices until an error. Used when the same operation needs to be set >> on multiple subdevices. The subdevices that don't implement such operation should >> return -ENOIOCTLCMD. >> >> Btw, v4l2_device_call_subdevs_until_err() has currently a bug: if all sub-devices return >> -ENOIOCTLCMD, it returns 0. It should, instead, return -ENOIOCTLCMD, in order to allow >> the bridge drivers to return an error code to the userspace, to indicate that the >> IOCTL was not handled. >> >> Eventually, we may just use (2) for everything, if we patch all subdev drivers to return >> -ENOIOCTLCMD if they are discarding a subdev call, but, in this case, the bridge drivers >> will need to replace the -ENOIOCTLCMD to an error code defined at the V4L2 spec (or we >> can have a macro for that). > > Option 2 is the correct approach. If all subdev drivers return -ENOIOCTLCMD, > then return an error. If one driver returns a non-0 and non-ENOIOCTLCMD error, > then return that, otherwise return 0. I'm ok with that. >> A side note: the only error codes defined at the media API DocBook are: -EACCES, -EAGAIN, >> -EBADF, -EBUSY, -EFAULT, -EIO, -EINTR, -EINVAL, -ENFILE, -ENOMEM, -ENOSPC, -ENOTTY, -ENXIO, >> -EMFILE, -EPERM, -ERANGE and EPIPE. On most places, the error codes are defined per ioctl. >> We need to review the DocBook and the drivers to be sure that they match the API specs, >> in terms of returned codes. It probably makes sense to create a section with the valid error >> codes, remove most of error codes comments from each ioctl, and add a link to the global >> error code section. >> >> Ah, -ENODEV is not currently defined, but -ENXIO is defined on a few places. -ENXIO means >> "No such device or address". So, it may make sense to replace all -ENODEV to -ENXIO at >> the drivers. > > Right, all very lovely, but I just want to fix the broken tuner code. We all know > error handling is a big mess and could keep a small army of janitors busy for a year. > > For now I'll remove those two offending patches and redo the patch series without > them. OK. > > Regards, > > Hans -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/media/video/tuner-core.c b/drivers/media/video/tuner-core.c index 5748d04..083b9f1 100644 --- a/drivers/media/video/tuner-core.c +++ b/drivers/media/video/tuner-core.c @@ -723,22 +723,19 @@ static int tuner_remove(struct i2c_client *client) */ /** - * check_mode - Verify if tuner supports the requested mode + * supported_mode - Verify if tuner supports the requested mode * @t: a pointer to the module's internal struct_tuner * * This function checks if the tuner is capable of tuning analog TV, * digital TV or radio, depending on what the caller wants. If the - * tuner can't support that mode, it returns -EINVAL. Otherwise, it - * returns 0. + * tuner can't support that mode, it returns false. Otherwise, it + * returns true. * This function is needed for boards that have a separate tuner for * radio (like devices with tea5767). */ -static inline int check_mode(struct tuner *t, enum v4l2_tuner_type mode) +static bool supported_mode(struct tuner *t, enum v4l2_tuner_type mode) { - if ((1 << mode & t->mode_mask) == 0) - return -EINVAL; - - return 0; + return 1 << mode & t->mode_mask; } /** @@ -759,7 +756,7 @@ static int set_mode_freq(struct i2c_client *client, struct tuner *t, struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; if (mode != t->mode) { - if (check_mode(t, mode) == -EINVAL) { + if (!supported_mode(t, mode)) { tuner_dbg("Tuner doesn't support mode %d. " "Putting tuner to sleep\n", mode); t->standby = true; @@ -1138,7 +1135,7 @@ static int tuner_g_frequency(struct v4l2_subdev *sd, struct v4l2_frequency *f) struct tuner *t = to_tuner(sd); struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; - if (check_mode(t, f->type) == -EINVAL) + if (!supported_mode(t, f->type)) return 0; f->type = t->mode; if (fe_tuner_ops->get_frequency && !t->standby) { @@ -1161,7 +1158,7 @@ static int tuner_g_tuner(struct v4l2_subdev *sd, struct v4l2_tuner *vt) struct analog_demod_ops *analog_ops = &t->fe.ops.analog_ops; struct dvb_tuner_ops *fe_tuner_ops = &t->fe.ops.tuner_ops; - if (check_mode(t, vt->type) == -EINVAL) + if (!supported_mode(t, vt->type)) return 0; vt->type = t->mode; if (analog_ops->get_afc)