media: stm32: dcmi: Switch to __v4l2_subdev_state_alloc()
Subsystems (3)
PATCH v2 1/1 media: stm32: dcmi: Switch to __v4l2_subdev_state_alloc()
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
---
V2: Add FIXME comment above __v4l2_subdev_state_alloc() calls
--- drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)
drivers/media/platform/st/stm32/stm32-dcmi.c +37 -22
@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;int ret; +/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); + sd_fmt =find_format_by_fourcc (dcmi, pix->pixelformat);if (!sd_fmt) { -if (!dcmi->num_of_sd_formats) -return -ENODATA ; +if (!dcmi->num_of_sd_formats) { + ret = -ENODATA ; + goto done; + } sd_fmt = dcmi->sd_formats[dcmi->num_of_sd_formats -1 ]; pix->pixelformat = sd_fmt->fourcc;
@@ -1036,10 +1044,9 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); + ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, sd_state, &format);if (ret <0 ) -return ret; + goto done;/* Update pix regarding to what sensor can do */ v4l2_fill_pix_format (pix, &format.format);
@@ -1079,7 +1086,9 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f,if (sd_framesize) *sd_framesize = sd_fsize; -return 0 ; +done: +__v4l2_subdev_state_free (sd_state); +return ret; }static int dcmi_set_fmt (struct stm32_dcmi *dcmi,struct v4l2_format *f)
@@ -1183,31 +1192,37 @@ static int dcmi_set_sensor_format(struct stm32_dcmi *dcmi, struct v4l2_pix_format *pix) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , }; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - };int ret; +/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); + sd_fmt =find_format_by_fourcc (dcmi, pix->pixelformat);if (!sd_fmt) { -if (!dcmi->num_of_sd_formats) -return -ENODATA ; +if (!dcmi->num_of_sd_formats) { + ret = -ENODATA ; + goto done; + } sd_fmt = dcmi->sd_formats[dcmi->num_of_sd_formats -1 ]; pix->pixelformat = sd_fmt->fourcc; }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); -if (ret <0 ) -return ret; - -return 0 ; + ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, sd_state, &format); +done: +__v4l2_subdev_state_free (sd_state); +return ret; } static int dcmi_get_sensor_bounds(struct stm32_dcmi *dcmi,
Signature · 1 line
2.35.1
Hi Marek, Tomi, Laurent,
Quoted 41 lines · Marek Vasut
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
Signed-off-by Marek Vasut <[email protected]>Cc Alain Volmat <[email protected]>Cc Alexandre Torgue <[email protected]>Cc Amelie DELAUNAY <[email protected]>Cc Hugues FRUCHET <[email protected]>Cc Laurent Pinchart <[email protected]>Cc Philippe CORNU <[email protected]>---
V2: Add FIXME comment above__v4l2_subdev_state_alloc()calls--- drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;
drivers/media/platform/st/stm32/stm32-dcmi.c static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, · line 1016
Quoted 10 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +8
-0int ret; +/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); +
I've been reading the discussion for the v1 patch, and I seriously do not like this.
My comments are not specifically for this patch, but for all cases where __v4l2_subdev_state_alloc is called.
It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.
My suggestion would be to document that in the kerneldoc for this function in media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).
And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Regards,
Hans
Quoted 85 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +27
-18sd_fmt =find_format_by_fourcc (dcmi, pix->pixelformat);if (!sd_fmt) { -if (!dcmi->num_of_sd_formats) -return -ENODATA ; +if (!dcmi->num_of_sd_formats) { + ret = -ENODATA ; + goto done; + } sd_fmt = dcmi->sd_formats[dcmi->num_of_sd_formats -1 ]; pix->pixelformat = sd_fmt->fourcc;@@ -1036,10 +1044,9 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); + ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, sd_state, &format);if (ret <0 ) -return ret; + goto done;/* Update pix regarding to what sensor can do */ v4l2_fill_pix_format (pix, &format.format);@@ -1079,7 +1086,9 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f,if (sd_framesize) *sd_framesize = sd_fsize; -return 0 ; +done: +__v4l2_subdev_state_free (sd_state); +return ret; }static int dcmi_set_fmt (struct stm32_dcmi *dcmi,struct v4l2_format *f)@@ -1183,31 +1192,37 @@ static int dcmi_set_sensor_format(struct stm32_dcmi *dcmi, struct v4l2_pix_format *pix) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , }; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - };int ret; +/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); + sd_fmt =find_format_by_fourcc (dcmi, pix->pixelformat);if (!sd_fmt) { -if (!dcmi->num_of_sd_formats) -return -ENODATA ; +if (!dcmi->num_of_sd_formats) { + ret = -ENODATA ; + goto done; + } sd_fmt = dcmi->sd_formats[dcmi->num_of_sd_formats -1 ]; pix->pixelformat = sd_fmt->fourcc; }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); -if (ret <0 ) -return ret; - -return 0 ; + ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, sd_state, &format); +done: +__v4l2_subdev_state_free (sd_state); +return ret; } static int dcmi_get_sensor_bounds(struct stm32_dcmi *dcmi,
Quoted 1 line · Hans Verkuil
Hi Marek, Tomi, Laurent,
Hi,
[...]
Quoted 42 lines
Quoted 23 lines
drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;drivers/media/platform/st/stm32/stm32-dcmi.c static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, · line 1016
Quoted 10 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +8
-0int ret; +/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); +I've been reading the discussion for the v1 patch, and I seriously do not like this.
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
Quoted 7 lines
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).
Would it be OK if I left the core rework/documentation to Tomi as a subsequent patch to this one ?
Quoted 2 lines
And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Probably, I only ran into it with the DCMI so far.
Quoted 52 lines · Marek Vasut
Quoted 1 line · Hans Verkuil
Hi Marek, Tomi, Laurent,
Hi,
[...]
Quoted 45 lines
Quoted 24 lines
drivers/media/platform/st/stm32/stm32-dcmi.c| 59 ++++++++++++--------1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) { const struct dcmi_format *sd_fmt; + static struct lock_class_key key; struct dcmi_framesize sd_fsize; struct v4l2_pix_format *pix = &f->fmt.pix; - struct v4l2_subdev_pad_config pad_cfg; - struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; + struct v4l2_subdev_state *sd_state; struct v4l2_subdev_format format = { .which = V4L2_SUBDEV_FORMAT_TRY, }; bool do_crop; int ret;Quoted 8 lines
+ /*
+ * FIXME: Drop this call, drivers are not supposed to use
+ *__v4l2_subdev_state_alloc().
+ */
+sd_state=__v4l2_subdev_state_alloc(dcmi->source, "dcmi:state->lock", &key); + if (IS_ERR(sd_state))
+ returnPTR_ERR(sd_state);
+I've been reading the discussion for the v1 patch, and I seriously do not like this.
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.
Quoted 5 lines
Quoted 3 lines
My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).Would it be OK if I left the core rework/documentation to Tomi as a subsequent patch to this one ?
Yes. It would be nice if Tomi can make a patch fixing the comments as suggested above, and then your patch can go on top. Adding a helper function can be done later, it's not my main concern.
Quoted 5 lines
Quoted 3 lines
And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Probably, I only ran into it with the DCMI so far.
Tomi, do you know?
Regards,
Hans
Quoted 46 lines · Hans Verkuil
Hi Marek, Tomi, Laurent,
Quoted 43 lines · Marek Vasut
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
Signed-off-by Marek Vasut <[email protected]>Cc Alain Volmat <[email protected]>Cc Alexandre Torgue <[email protected]>Cc Amelie DELAUNAY <[email protected]>Cc Hugues FRUCHET <[email protected]>Cc Laurent Pinchart <[email protected]>Cc Philippe CORNU <[email protected]>---
V2: Add FIXME comment above__v4l2_subdev_state_alloc()calls--- drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;int ret;
Quoted 10 lines
drivers/media/platform/st/stm32/stm32-dcmi.c static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, · line 1016
Quoted 8 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +8
-0+/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); +I've been reading the discussion for the v1 patch, and I seriously do not like this.
I don't like it either.
Quoted 2 lines
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.
Just to emphasize it: afaics this is not an issue with the subdev state. This driver was already broken. Before the subdev state change the fix would have been to call source subdev's init_cfg. Now __v4l2_subdev_state_alloc handles calling init_cfg (along with a few other inits).
Quoted 2 lines
It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
Counting this one? I found 3 cases in v5.18-rc4:
1) drivers/media/platform/renesas/rcar-vin/rcar-v4l2.c:
Allocates the state for v4l2_subdev_call, set_fmt, TRY.
Here, a helper macro which does the alloc would be a solution.
2) drivers/media/platform/renesas/vsp1/vsp1_entity.c:
Allocates the state for storing internal active state.
Here, I think, the easiest fix is for the driver to use the subdev state properly.
3) drivers/staging/media/tegra-video/vi.c:
Allocates the state for v4l2_subdev_call, enum_frame_size and set_fmt, TRY. Interestingly the code also calls get_selection but without passing the state...
This is a more interesting case as the source's subdev state is actually modified by the driver. The driver calls enum_frame_size, then modifies the state, then calls set_fmt. I'm not sure if that's really legal... The driver directly modifies the state, instead of calling set_selection?
Quoted 3 lines
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.
That is true. There's no single answer to that. I think instead of trying to document that in the v4l2-subdev doc, we can enhance the comments in those three call sites to explain how it needs to be fixed.
Quoted 6 lines
My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).And another question: are more drivers affected by this? Is it possible to find those and fix them all?
I think any driver that calls a source subdev's subdev ops which a subdev state as a parameter (the ones that used to take v4l2_subdev_pad_config), and does not call init_cfg is broken in the same way. With some grepping, I couldn't find anyone calling init_cfg. Finding those drivers which do those calls is a bit more difficult, but can be done with some efforts.
atmel-isc-base.c is one I found, and looks like it's doing something a bit similar to the 3) case above.
Perhaps the best way to solve this is just to remove the underscores from __v4l2_subdev_state_alloc, and change all the drivers which create temporary v4l2_subdev_states to use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).
It would've been nice to keep __v4l2_subdev_state_alloc internal to the v4l2-subdev, but maybe the v4l2 drivers are not there yet. The non-MC drivers seem to be doing all kinds of interesting things.
Tomi
Quoted 113 lines · Tomi Valkeinen
Quoted 46 lines · Hans Verkuil
Hi Marek, Tomi, Laurent,
Quoted 43 lines · Marek Vasut
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
Signed-off-by Marek Vasut <[email protected]>Cc Alain Volmat <[email protected]>Cc Alexandre Torgue <[email protected]>Cc Amelie DELAUNAY <[email protected]>Cc Hugues FRUCHET <[email protected]>Cc Laurent Pinchart <[email protected]>Cc Philippe CORNU <[email protected]>---
V2: Add FIXME comment above__v4l2_subdev_state_alloc()calls--- drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;int ret;Quoted 10 lines
drivers/media/platform/st/stm32/stm32-dcmi.c static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, · line 1016
Quoted 8 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +8
-0+/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); +I've been reading the discussion for the v1 patch, and I seriously do not like this.
I don't like it either.
Quoted 2 lines
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.Just to emphasize it: afaics this is not an issue with the subdev state. This driver was already broken. Before the subdev state change the fix would have been to call source subdev's
init_cfg. Now__v4l2_subdev_state_allochandles callinginit_cfg(along with a few other inits).Quoted 2 lines
It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
Counting this one? I found 3 cases in v5.18-rc4:
1)
drivers/media/platform/renesas/rcar-vin/rcar-v4l2.c:Allocates the state for
v4l2_subdev_call,set_fmt, TRY.Here, a helper macro which does the alloc would be a solution.
2)
drivers/media/platform/renesas/vsp1/vsp1_entity.c:Allocates the state for storing internal active state.
Here, I think, the easiest fix is for the driver to use the subdev state properly.
3)
drivers/staging/media/tegra-video/vi.c:Allocates the state for
v4l2_subdev_call,enum_frame_sizeandset_fmt, TRY. Interestingly the code also callsget_selectionbut without passing the state...This is a more interesting case as the source's subdev state is actually modified by the driver. The driver calls
enum_frame_size, then modifies the state, then callsset_fmt. I'm not sure if that's really legal... The driver directly modifies the state, instead of callingset_selection?Quoted 3 lines
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.That is true. There's no single answer to that. I think instead of trying to document that in the v4l2-subdev doc, we can enhance the comments in those three call sites to explain how it needs to be fixed.
Quoted 6 lines
My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Quoted 6 lines
I think any driver that calls a source subdev's subdev ops which a subdev state as a parameter (the ones that used to take
v4l2_subdev_pad_config), and does not callinit_cfgis broken in the same way. With some grepping, I couldn't find anyone callinginit_cfg. Finding those drivers which do those calls is a bit more difficult, but can be done with some efforts.
A quick check identified the following files:
atmel-isc-base.catmel-isi.ccxusb-analog.cfimc-capture.cmcam-core.cpxa_camera.crenesas-ceu.csaa7134-empress.cvia-camera.c
A few drivers with more complex call patterns may be missing.
Quoted 8 lines
atmel-isc-base.cis one I found, and looks like it's doing something a bit similar to the 3) case above.Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).
As long as we prevent any new driver from using that API, that's fine with me.
Quoted 3 lines
It would've been nice to keep
__v4l2_subdev_state_allocinternal to the v4l2-subdev, but maybe the v4l2 drivers are not there yet. The non-MC drivers seem to be doing all kinds of interesting things.
Signature · 3 lines
Regards, Laurent Pinchart
Quoted 138 lines · Laurent Pinchart
Quoted 113 lines · Tomi Valkeinen
Quoted 46 lines · Hans Verkuil
Hi Marek, Tomi, Laurent,
Quoted 43 lines · Marek Vasut
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
Signed-off-by Marek Vasut <[email protected]>Cc Alain Volmat <[email protected]>Cc Alexandre Torgue <[email protected]>Cc Amelie DELAUNAY <[email protected]>Cc Hugues FRUCHET <[email protected]>Cc Laurent Pinchart <[email protected]>Cc Philippe CORNU <[email protected]>---
V2: Add FIXME comment above__v4l2_subdev_state_alloc()calls--- drivers/media/platform/st/stm32/stm32-dcmi.c | 59 ++++++++++++-------- 1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) {const struct dcmi_format *sd_fmt; +static struct lock_class_key key;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; +struct v4l2_subdev_state *sd_state;struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };bool do_crop;int ret;Quoted 10 lines
drivers/media/platform/st/stm32/stm32-dcmi.c static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, · line 1016
Quoted 8 lines
drivers/media/platform/st/stm32/stm32-dcmi.c +8
-0+/* +* FIXME: Drop this call, drivers are not supposed to use +* __v4l2_subdev_state_alloc(). +*/ + sd_state =__v4l2_subdev_state_alloc (dcmi->source,"dcmi:state->lock" , &key); +if (IS_ERR (sd_state)) +return PTR_ERR (sd_state); +I've been reading the discussion for the v1 patch, and I seriously do not like this.
I don't like it either.
Quoted 2 lines
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.Just to emphasize it: afaics this is not an issue with the subdev state. This driver was already broken. Before the subdev state change the fix would have been to call source subdev's
init_cfg. Now__v4l2_subdev_state_allochandles callinginit_cfg(along with a few other inits).Quoted 2 lines
It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
Counting this one? I found 3 cases in v5.18-rc4:
1)
drivers/media/platform/renesas/rcar-vin/rcar-v4l2.c:Allocates the state for
v4l2_subdev_call,set_fmt, TRY.Here, a helper macro which does the alloc would be a solution.
2)
drivers/media/platform/renesas/vsp1/vsp1_entity.c:Allocates the state for storing internal active state.
Here, I think, the easiest fix is for the driver to use the subdev state properly.
3)
drivers/staging/media/tegra-video/vi.c:Allocates the state for
v4l2_subdev_call,enum_frame_sizeandset_fmt, TRY. Interestingly the code also callsget_selectionbut without passing the state...This is a more interesting case as the source's subdev state is actually modified by the driver. The driver calls
enum_frame_size, then modifies the state, then callsset_fmt. I'm not sure if that's really legal... The driver directly modifies the state, instead of callingset_selection?Quoted 3 lines
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.That is true. There's no single answer to that. I think instead of trying to document that in the v4l2-subdev doc, we can enhance the comments in those three call sites to explain how it needs to be fixed.
Quoted 6 lines
My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Quoted 6 lines
I think any driver that calls a source subdev's subdev ops which a subdev state as a parameter (the ones that used to take
v4l2_subdev_pad_config), and does not callinit_cfgis broken in the same way. With some grepping, I couldn't find anyone callinginit_cfg. Finding those drivers which do those calls is a bit more difficult, but can be done with some efforts.A quick check identified the following files:
atmel-isc-base.catmel-isi.ccxusb-analog.cfimc-capture.cmcam-core.cpxa_camera.crenesas-ceu.csaa7134-empress.cvia-camera.cA few drivers with more complex call patterns may be missing.
Quoted 3 lines
atmel-isc-base.cis one I found, and looks like it's doing something a bit similar to the 3) case above.
Quoted 8 lines
Quoted 5 lines
Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).As long as we prevent any new driver from using that API, that's fine with me.
An alternative would be to keep the v4l2_subdev_state as a local variable (allocated in the stack), and add a new function, v4l2_subdev_state_local_init() or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).
This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Tomi
On 6/29/22 14:26, Tomi Valkeinen wrote:
[...]
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).
Quoted 8 lines
Quoted 2 lines
As long as we prevent any new driver from using that API, that's fine with me.
An alternative would be to keep the
v4l2_subdev_stateas a local variable (allocated in the stack), and add a new function,v4l2_subdev_state_local_init()or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Don't you think a driver fix shouldn't involve "rework the subsystem" requirement to be applicable ?
Quoted 13 lines · Marek Vasut
On 6/29/22 14:26, Tomi Valkeinen wrote:
[...]
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).As long as we prevent any new driver from using that API, that's fine with me.
Quoted 7 lines
Quoted 5 lines
An alternative would be to keep the
v4l2_subdev_stateas a local variable (allocated in the stack), and add a new function,v4l2_subdev_state_local_init()or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Don't you think a driver fix shouldn't involve "rework the subsystem" requirement to be applicable ?
No, but we should think what's the best way to do the fix, if the fix is controversial. Otherwise we might just break things even worse. Adding the macro seems like a much better way, and far from "rework the subsystem". Granted, this was just a quick edit without testing so it may fail miserably...
Can you try this out?
Also, a bit unrelated thing: dcmi_get_sensor_format gets the active format from the source subdev, but dcmi_set_sensor_format sets the try format. That sounds like a bug. Is it on purpose?
drivers/media/platform/st/stm32/stm32-dcmi.c +2 -12
@@ -999,10 +999,6 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f,const struct dcmi_format *sd_fmt;struct dcmi_framesize sd_fsize;struct v4l2_pix_format *pix = &f->fmt.pix; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - };struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , };
@@ -1037,8 +1033,7 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); + ret =v4l2_subdev_call_state_try (dcmi->source, pad, set_fmt, &format);if (ret <0 )return ret;
@@ -1187,10 +1182,6 @@ static int dcmi_set_sensor_format(struct stm32_dcmi *dcmi,struct v4l2_subdev_format format = { .which =V4L2_SUBDEV_FORMAT_TRY , }; -struct v4l2_subdev_pad_config pad_cfg; -struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - };int ret; sd_fmt =find_format_by_fourcc (dcmi, pix->pixelformat);
@@ -1203,8 +1194,7 @@ static int dcmi_set_sensor_format(struct stm32_dcmi *dcmi, }v4l2_fill_mbus_format (&format.format, pix, sd_fmt->mbus_code); - ret =v4l2_subdev_call (dcmi->source, pad, set_fmt, - &pad_state, &format); + ret =v4l2_subdev_call_state_try (dcmi->source, pad, set_fmt, &format);if (ret <0 )return ret;
include/media/v4l2-subdev.h +13 -0
@@ -1433,6 +1433,19 @@ extern const struct v4l2_subdev_ops v4l2_subdev_call_wrappers; __result; \ }) +#define v4l2_subdev_call_state_try(sd, o, f, args...) \ + ({ \ + int __result; \ + static struct lock_class_key __key; \ + const char *name = KBUILD_BASENAME \ + ":" __stringify(__LINE__) ":state->lock"; \ + struct v4l2_subdev_state *state = \ + __v4l2_subdev_state_alloc(sd, name, &__key); \ + __result = v4l2_subdev_call(sd, o, f, state, ##args); \ + __v4l2_subdev_state_free(state); \ + __result; \ + }) +/** * v4l2_subdev_has_op - Checks if a subdev defines a certain operation. *
Tomi
Quoted 22 lines · Tomi Valkeinen
Quoted 13 lines · Marek Vasut
On 6/29/22 14:26, Tomi Valkeinen wrote:
[...]
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).As long as we prevent any new driver from using that API, that's fine with me.
Quoted 7 lines
Quoted 5 lines
An alternative would be to keep the
v4l2_subdev_stateas a local variable (allocated in the stack), and add a new function,v4l2_subdev_state_local_init()or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Don't you think a driver fix shouldn't involve "rework the subsystem" requirement to be applicable ?
Quoted 7 lines
No, but we should think what's the best way to do the fix, if the fix is controversial. Otherwise we might just break things even worse. Adding the macro seems like a much better way, and far from "rework the subsystem". Granted, this was just a quick edit without testing so it may fail miserably...
Can you try this out?
It seems to work as well. How shall we proceed ?
Quoted 29 lines · Marek Vasut
Quoted 28 lines · Tomi Valkeinen
Quoted 13 lines · Marek Vasut
On 6/29/22 14:26, Tomi Valkeinen wrote:
[...]
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines
Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).As long as we prevent any new driver from using that API, that's fine with me.
Quoted 7 lines
Quoted 5 lines
An alternative would be to keep the
v4l2_subdev_stateas a local variable (allocated in the stack), and add a new function,v4l2_subdev_state_local_init()or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Don't you think a driver fix shouldn't involve "rework the subsystem" requirement to be applicable ?
No, but we should think what's the best way to do the fix, if the fix is controversial. Otherwise we might just break things even worse. Adding the macro seems like a much better way, and far from "rework the subsystem". Granted, this was just a quick edit without testing so it may fail miserably...
Quoted 3 lines
Quoted 1 line
Can you try this out?
It seems to work as well. How shall we proceed ?
I haven't had to look at this more closely, but I made proper patches of this and sent them for review. Note that they're not exactly the same as the diff here: the macro was missing lock and unlock for the subdev state, which I've added. Again, only compile tested.
Tomi
Quoted 144 lines · Tomi Valkeinen
Quoted 143 lines · Laurent Pinchart
Quoted 112 lines · Tomi Valkeinen
Quoted 47 lines · Hans Verkuil
Hi Marek, Tomi, Laurent,
Quoted 44 lines · Marek Vasut
Any local subdev state should be allocated and free'd using __v4l2_subdev_state_alloc()/__v4l2_subdev_state_free(), which takes care of calling .init_cfg() subdev op. Without this, subdev internal state might be uninitialized by the time any other subdev op is called.
Signed-off-by Marek Vasut <[email protected]>Cc Alain Volmat <[email protected]>Cc Alexandre Torgue <[email protected]>Cc Amelie DELAUNAY <[email protected]>Cc Hugues FRUCHET <[email protected]>Cc Laurent Pinchart <[email protected]>Cc Philippe CORNU <[email protected]>---
V2: Add FIXME comment above__v4l2_subdev_state_alloc()calls
---
drivers/media/platform/st/stm32/stm32-dcmi.c| 59 ++++++++++++--------1 file changed, 37 insertions(+), 22 deletions(-)drivers/media/platform/st/stm32/stm32-dcmi.c +2
-4@@ -996,22 +996,30 @@ static int dcmi_try_fmt(struct stm32_dcmi *dcmi, struct v4l2_format *f, struct dcmi_framesize *sd_framesize) { const struct dcmi_format *sd_fmt; + static struct lock_class_key key; struct dcmi_framesize sd_fsize; struct v4l2_pix_format *pix = &f->fmt.pix; - struct v4l2_subdev_pad_config pad_cfg; - struct v4l2_subdev_state pad_state = { - .pads = &pad_cfg - }; + struct v4l2_subdev_state *sd_state; struct v4l2_subdev_format format = { .which = V4L2_SUBDEV_FORMAT_TRY, }; bool do_crop; int ret;+ /*
+ * FIXME: Drop this call, drivers are not supposed to useQuoted 8 lines
Quoted 6 lines
+ *
__v4l2_subdev_state_alloc().
+ */
+sd_state=__v4l2_subdev_state_alloc(dcmi->source, "dcmi:state->lock", &key); + if (IS_ERR(sd_state))
+ returnPTR_ERR(sd_state);
+I've been reading the discussion for the v1 patch, and I seriously do not like this.
I don't like it either.
Quoted 2 lines
My comments are not specifically for this patch, but for all cases where
__v4l2_subdev_state_allocis called.Just to emphasize it: afaics this is not an issue with the subdev state. This driver was already broken. Before the subdev state change the fix would have been to call source subdev's
init_cfg. Now__v4l2_subdev_state_allochandles callinginit_cfg(along with a few other inits).Quoted 2 lines
It is now used in 4 drivers, so that's no longer a rare case, and the code isn't exactly trivial either.
Counting this one? I found 3 cases in v5.18-rc4:
1)
drivers/media/platform/renesas/rcar-vin/rcar-v4l2.c:Allocates the state for
v4l2_subdev_call,set_fmt, TRY.Here, a helper macro which does the alloc would be a solution.
2)
drivers/media/platform/renesas/vsp1/vsp1_entity.c:Allocates the state for storing internal active state.
Here, I think, the easiest fix is for the driver to use the subdev state properly.
3)
drivers/staging/media/tegra-video/vi.c:Allocates the state for
v4l2_subdev_call,enum_frame_sizeandset_fmt, TRY. Interestingly the code also callsget_selectionbut without passing the state...This is a more interesting case as the source's subdev state is actually modified by the driver. The driver calls
enum_frame_size, then modifies the state, then callsset_fmt. I'm not sure if that's really legal... The driver directly modifies the state, instead of callingset_selection?Quoted 3 lines
I think a helper function might be beneficial, but the real problem is with the
comment: it does not explain why you shouldn't use it and what needs to be done to fix it.That is true. There's no single answer to that. I think instead of trying to document that in the v4l2-subdev doc, we can enhance the comments in those three call sites to explain how it needs to be fixed.
Quoted 6 lines
My suggestion would be to document that in the kerneldoc for this function in
media/v4l2-subdev.h, and then refer to that from this comment (and similar comments in the other drivers that use this).And another question: are more drivers affected by this? Is it possible to find those and fix them all?
Quoted 6 lines
I think any driver that calls a source subdev's subdev ops which a subdev state as a parameter (the ones that used to take
v4l2_subdev_pad_config), and does not callinit_cfgis broken in the same way. With some grepping, I couldn't find anyone callinginit_cfg. Finding those drivers which do those calls is a bit more difficult, but can be done with some efforts.A quick check identified the following files:
atmel-isc-base.catmel-isi.ccxusb-analog.cfimc-capture.cmcam-core.cpxa_camera.crenesas-ceu.csaa7134-empress.cvia-camera.cA few drivers with more complex call patterns may be missing.
Quoted 8 lines
atmel-isc-base.cis one I found, and looks like it's doing something a bit similar to the 3) case above.Perhaps the best way to solve this is just to remove the underscores from
__v4l2_subdev_state_alloc, and change all the drivers which create temporaryv4l2_subdev_statesto use that (and the free) functions. And also create the helper macro which can be used in those cases where the call is simple (the state is not modified or accessed by the caller).
Quoted 5 lines
Quoted 2 lines
As long as we prevent any new driver from using that API, that's fine with me.
An alternative would be to keep the
v4l2_subdev_stateas a local variable (allocated in the stack), and add a new function,v4l2_subdev_state_local_init()or such. The function would initialize the given state, expecting the allocatable fields to be already allocated (state->pads, which in the above cases points to another local variable, i.e. stack).
Certainly in the case of this stm32 driver, that seems much closer to the original code so also easier to understand.
The reason I am blocking this patch now is that I want to prevent seeing these terrible FIXMEs all over the place. It was in hindsight a mistake to accept it in the first place, so we need something better in place first before we fix this (and other) drivers.
I do apologize to Marek for that, but this issue highlighted a real problem that needs addressing first.
Regards,
Hans
Quoted 6 lines
This would prevent the need of a free call, which, while not complex as such, might cause a bigger amount of changes in some cases to handle the error paths correctly.
Of course, if the above-mentioned macro works, then that's the easiest solution. But that won't work for all drivers.
Tomi