Em 17-10-2011 15:03, Hans Verkuil escreveu:
> On Monday, October 17, 2011 17:36:10 Mauro Carvalho Chehab wrote:
>> Em 17-10-2011 06:32, Hans Verkuil escreveu:
>>> RFC: Improved handling of presets
>>> =================================
>>>
>>> This RFC attempts to resolve the issues raised by this thread a few months
>>> ago:
>>>
>>> http://www.mail-archive.com/linux-media@vger.kernel.org/msg33981.html
>>>
>>> The last post in this thread did a good job of summarizing the discussion:
>>>
>>> http://www.mail-archive.com/linux-media@vger.kernel.org/msg34190.html
>>>
>>> I think it is time to revisit this problem. It basically boils down to the
>>> fact
>>> that the current preset API is too limited and somewhat awkward and I agree
>>> with that.
>>>
>>> Now, I really do not like the idea of creating a preset2 API. Instead I
>>> think
>>> it is better to extend the current dv_timings API with new ioctls:
>>>
>>> VIDIOC_QUERY_DV_TIMINGS
>>> VIDIOC_ENUM_DV_TIMINGS
>>> VIDIOC_DV_TIMINGS_CAP
>>>
>>> These form a superset of the preset API and once this is in place we can
>>> deprecate some or all of the preset ioctls and eventually remove them (say
>>> in one or two years).
>>
>> Those 3 new ioctl's are, in practice, a preset2 API. I hate needing to
>> deprecate
>> an API, but, if this is needed, better to do earlier than later. Let's not
>> spend
>> our time investing on a dead horse.
>>
>>> Here is the current definition of the v4l2_dv_timings struct from
>>> videodev2.h:
>>>
>>> /* BT.656/BT.1120 timing data */
>>> struct v4l2_bt_timings {
>>> __u32 width; /* width in pixels */
>>> __u32 height; /* height in lines */
>>> __u32 interlaced; /* Interlaced or progressive */
>>> __u32 polarities; /* Positive or negative polarity */
>>> __u64 pixelclock; /* Pixel clock in HZ. Ex. 74.25MHz->74250000 */
>>> __u32 hfrontporch; /* Horizpontal front porch in pixels */
>>> __u32 hsync; /* Horizontal Sync length in pixels */
>>> __u32 hbackporch; /* Horizontal back porch in pixels */
>>> __u32 vfrontporch; /* Vertical front porch in pixels */
>>> __u32 vsync; /* Vertical Sync length in lines */
>>> __u32 vbackporch; /* Vertical back porch in lines */
>>> __u32 il_vfrontporch; /* Vertical front porch for bottom field of
>>> * interlaced field formats
>>> */
>>> __u32 il_vsync; /* Vertical sync length for bottom field of
>>> * interlaced field formats
>>> */
>>> __u32 il_vbackporch; /* Vertical back porch for bottom field of
>>> * interlaced field formats
>>> */
>>> __u32 reserved[16];
>>> } __attribute__ ((packed));
>>>
>>> /* Interlaced or progressive format */
>>> #define V4L2_DV_PROGRESSIVE 0
>>> #define V4L2_DV_INTERLACED 1
>>>
>>> /* Polarities. If bit is not set, it is assumed to be negative polarity */
>>> #define V4L2_DV_VSYNC_POS_POL 0x00000001
>>> #define V4L2_DV_HSYNC_POS_POL 0x00000002
>>>
>>> /* DV timings */
>>> struct v4l2_dv_timings {
>>> __u32 type;
>>> union {
>>> struct v4l2_bt_timings bt;
>>> __u32 reserved[32];
>>> };
>>> } __attribute__ ((packed));
>>>
>>> /* Values for the type field */
>>> #define V4L2_DV_BT_656_1120 0 /* BT.656/1120 timing type */
>>>
>>> This API allows you to set and get all the timings details. Its current
>>> use is to allow userspace to set non-standard timings and it is used only in
>>> the dm646x davinci drivers at the moment.
>>>
>>> I propose the following additions:
>>>
>>> 1) Add a standards field to v4l2_bt_timings:
>>>
>>> __u32 standards;
>>
>> Seems fine for get/query/enum operations. It probably doesn't make sense for
>> set.
>
> No, it doesn't. For set it will just be ignored by the driver.
>
>>>
>>> Currently I have four standards:
>>>
>>> #define V4L2_DV_BT_STD_CEA861 (1 << 0)
>>> #define V4L2_DV_BT_STD_DMT (1 << 1) /* VESA Discrete Monitor
>>> Timings */
>>> #define V4L2_DV_BT_STD_CVT (1 << 2) /* VESA Coordinated Video
>>> Timings */
>>> #define V4L2_DV_BT_STD_GTF (1 << 3) /* VESA Generalized Timings
>>> Formula */
>>>
>>> A particular timing can be part of 0 or more standards.
>>> Both CVT and GTF timings have a so-called 'reduced blanking' mode. It
>>> would be
>>> nice to represent this with a flag somewhere. I guess we need a flags
>>> field
>>> for that. The 'polarities' field really should have been called a
>>> 'flags' field.
>>> Oh well...
>>
>> instead of flags, we may consider using something like:
>> _u32 polarity:1
>>
>> We need to double check if this is portable enough, however.
>
> Bitfields aren't portable, unfortunately. But I wonder if we could do
> something like this:
>
> union {
> __u32 polarities; /* Positive or negative polarity */
> __u32 flags;
> };
>
> and deprecate 'polarities' and remove it in a year.
If the idea is to replace the existing ioctl's, just name it as flags for the
structs
used by the new ones. The old "polarities" will be removed together with the
removal
of the old ioctl's.
> I suspect that Cisco might be the only user of this API anyway with the
> davinci drivers.
Likely.
>>> 2) Create a VIDIOC_ENUM_DV_TIMINGS ioctl:
>>>
>>> struct v4l2_enum_dv_timings {
>>> __u32 index;
>>> char name[32];
>>
>> Not sure about the "name" field. An u32 working as an enum could work better.
>
> Huh? You need a name field for the human-readable description of the
> timings, just like all other enum ioctls.
The naming here could be something completely arbitrary, especially for custom
timings. I'm ok if you insist on keeping it, but I don't think that this would
bring any value. If userspace needs naming, it can give whatever name it wants,
as all data for the timings are there: standards, resolutions, fps, etc.
So, userspace could for example do something like:
sprintf ("%s%dx%d_%.02ffps",
dv_standard_name(dv),
dv->width,
dv->height,
dv->pixelclock/(dv->width * dv->height));
>>> struct v4l2_dv_timings timings;
>>> __u32 reserved[];
>>
>> Adding a reserved here is probably an overkill, as there are already reserved
>> fields at timings struct.
>
> Probably true.
>
>>> };
>>>
>>> #define VIDIOC_ENUM_DV_TIMINGS _IOWR('V', XX, struct
>>> v4l2_enum_dv_timings)
>>>
>>> This ioctl enumerates over all discrete supported timings and returns
>>> their
>>> name and timings.
>>>
>>> The timings field can be used as an input to S_DV_TIMINGS. The timings in
>>> this enumeration are guaranteed to be supported by the hardware.
>>> However, other custom timings may be supported as well (see my proposal
>>> for VIDIOC_DV_TIMINGS_CAP).
>>
>> If I understood well, enum will show all timings officially supported by the
>> hardware, right? The ones that are custom (e. g. doesn't belong to any
>> standard,
>> but, for some reason, the chipset vendor decided to add there) will have the
>> standards field equal to 0, right?
>
> Correct.
>
> I have seen roughly two types of receivers/transmitters: those that only have
> a
> limited list of timings, and those that can handle an almost unlimited range
> of
> timings, typically only restricted by the maximum pixel clock and sometimes
> blanking requirements.
>
> The enum ioctl will return either the fixed list, or a representative list of
> formats. In the latter case I am thinking of the various VGA-derived
> resolutions
> and the 720p/1080p variants. That's typically what a user wants to see.
As proposed, this ioctl will only fine for those with a limited list of timings,
but it doesn't cover well the "unlimited range" case.
For nowadays needs, the discrete list is probably OK, but adding a "type" field
like what I've proposed below would allow adding other ways to better cover the
"unlimited range of timings".
>
> Vendor-specific formats may be added as well if there is some good reason for
> it.
>
>> If so, it seems to be doing its job.
>>
>>> 3) Create a VIDIOC_QUERY_DV_TIMINGS ioctl:
>>>
>>> struct v4l2_query_dv_timings {
>>> __u32 state;
>>> __u32 index;
>>> struct v4l2_dv_timings timings;
>>> __u32 reserved[];
>>> };
>>>
>>> #define V4L2_QUERY_STATE_NO_TIMINGS 0
>>> #define V4L2_QUERY_STATE_UNSUPP_TIMINGS 1
>>> #define V4L2_QUERY_STATE_SUPP_TIMINGS 2
>>>
>>> #define VIDIOC_QUERY_DV_TIMINGS _IOR('V', XX, struct
>>> v4l2_query_dv_timings)
>>
>> Hmm... are you meant to using it to detect the supported DV's from a certain
>> input?
>
> It is the DV equivalent of QUERYSTD: so detecting what video format is found
> on the input.
>
>> I'm not a TV/monitor set expert. I'm not sure if this is enough, as some
>> devices might
>> accept a continuous range for some timings parameter. Maybe it makes sense
>> to have a
>> "type" field here in order to allow future expansion.
>>
>> Something like:
>>
>> struct v4l2_query_dv_timings {
>> __u32 state; /* In fact, I think this should be
>> removed. See bellow */
>
> bellow -> below. I'm going to cure you of this typo one of these days :-)
:)
>
>> __u32 index;
>> #define V4L2_DV_QUERY_TYPE_DISCRETE
>> __u32 type;
>
> I don't really see what this adds. The BT656 et al doesn't have a range.
> Should a new interface appear that needs a range, then that also means a new
> timings type (and we already have a field for that).
The non-discrete types would cover the cases that you called before as
"unlimited range of timings".
>> union {
>> struct v4l2_dv_timings timings;
>> __u32 reserved[128];
>> };
>> };
>>
>>
>>> There are three states:
>>>
>>> 1 - no timings could be detected. Call ENUM_INPUT to find out why.
>>
>> If the input doesn't accept DV timings, just return some error code.
>> It makes sense to define different error codes for each possible condition:
>> - input(or output) is analog (so, no DV is supported);
>> - TV set/monitor doesn't support querying DV;
>> - chipset doesn't support it (ENOTTY);
>> - index out of range (EINVAL).
>
> An error could indeed be returned in this case.
>
>>> 2 - timings could be detected, and the timings struct is filled in, but
>>> they
>>> are not supported by other parts of the hardware. Call DV_TIMINGS_CAP
>>> and check the timings against the capabilities to find out why.
>>
>> If the timing is known to not be supported by part of the hardware, any
>> trial to use it
>> should be denied (EINVAL).
>>
>> As such, it doesn't make any sense to return a DV timing that is known to
>> not be
>> supported by the hardware.
>
> It does. Where this becomes important is when you have to tell the user *why*
> it
> isn't working. It is quite common that the receiver i2c device can handle the
> format
> it receives, but other parts in the pipeline can't. Examples: the receiver can
> handle interlaced formats, but the FPGA only does progressive, or the
> receiver can
> handle higher clock frequencies than other parts of the pipeline can (these
> are
> real-life examples).
>
> By returning this information you can check against the capabilities and
> actualy produce a sensible user message.
There's no way to tell userspace where the problem is, e. g. in each
part of the pipeline the timing is not supported. Also, it doesn't seem to be
the
usual usecase to show invalid entries, as it helps only when there's no match.
IMO, if this feature is really needed, I would add a "filter" bitmask parameter
that would
allow showing those invalid timings. if filter is zero, it will show only the
timings
that are valid. Otherwise, it will show the valid timings, plus the invalid
ones that
match the filter. At the return, those valid parameters will fill the filter
with 0, and
the invalid will fill it with the parts of the pipeline where it is not
supported.
Another option would be to implement the DV's at pad level. So, userspace could
query it
at pad level, when no timing were found. This is probably the better solution.
>> It should be noticed that the "standards" field equal to 0 already indicates
>> that such
>> timing is not part of any known standard. So, if the idea here was to let
>> userspace know
>> what are the "preferred" timings, it can just use the standards field.
>
> Hmm, I think you misinterpret this ioctl. It's not about preferred timings,
> it's
> about discovering what timings the video on the input uses.
No. You miss-interpreted my comment. What I meant to say is that userspace may
prefer to set
a DV timing that is supported by some standard (as it is better to get
something known than
to just randomly choosing a non-official timing).
A DV query will return all supported timings. By having the standards field
properly filled,
the userspace detection algorithm that chooses the DV timing to be used could
benefit of the
"standards" information, in order to select the DV timing to be used.
>> Maybe it makes sense to add a "standard" flag to indicate that the timing
>> matches an
>> existing timing at the drivers timing table, as it may have some de-facto
>> timings supported
>> by both the monitor/tv set and the hardware, but not officially (yet) part
>> of any standard.
>>
>>> 3 - timings are detected and are supported.
>>>
>>> The index can be used with ENUM_DV_TIMINGS to get the name. If the
>>> timings
>>> are not part of the enumerated timings list, then index is set to
>>> 0xffffffff
>>> (or at least some value that will cause ENUM_DV_TIMINGS to return
>>> EINVAL).
>>> That value would be represented by a macro such as
>>> V4L2_QUERY_UNKNOWN_INDEX.
>>
>> The index should be something between 0 to n, where n is the last supported
>> DV timing.
>> using it for anything else than that will make this different than the other
>> VIDIOC_ENUM
>> ioctl's, making the API messy.
>>
>> If the device needs to return EINVAL, just return it, instead of adding any
>> tricks.
>
> I'd better add a new field then, such as: 'is_custom', if the incoming video
> has
> valid timings that are not in the enum list.
You don't need: standards = 0 already says that. That's what I said before, and
that you've
miss-interpreted ;)
> For example: 1024x768 is a standard XGA resolution which comes in a number of
> fps
> values. So XGA@60 would be a resolution that's part of the enum list. But it's
> quite possible to send it 1028x768@60. This is non-standard, but
> QUERY_DV_TIMINGS
> should still be able to detect it (if the hardware can, of course).
>
> In that case the 'index' field can't be used, but the timings struct must be
> instead.
index != enum
The index here should be a plain sequential number, just like any other enum
ioctl. Anything
different than that is messy and will prevent userspace to properly query the
supported
standards.
The typical query loop on userspace should be something like:
int rc, index = 0;
for (index = 0; 1; index++) {
rc = ioctl(fd, VIDIOC_QUERY_DV_TIMINGS, &data);
if (rc == -1 && rc == -EINVAL)
break;
if (rc == 0) {
/* Handle the DV timing */
} else {
/* Handle an error condition */
}
}
>>> 4) Create a VIDIOC_DV_TIMINGS_CAP ioctl:
>>>
>>> /* BT.656/BT.1120 timing capabilities */
>>> struct v4l2_bt_timings_cap {
>>> __u32 min_width; /* width in pixels */
>>> __u32 max_width; /* width in pixels */
>>> __u32 min_height; /* height in lines */
>>> __u32 max_height; /* height in lines */
>>> __u64 min_pixelclock; /* Pixel clock in HZ. Ex.
>>> 74.25MHz->74250000 */
>>> __u64 max_pixelclock; /* Pixel clock in HZ. Ex.
>>> 74.25MHz->74250000 */
>>> __u32 standards; /* Supported standards */
>>> __u32 flags; /* Interlaced or progressive */
>>> __u32 reserved[16];
>>> } __attribute__ ((packed));
>>>
>>> /* Supports interlaced formats */
>>> #define V4L2_DV_BT_CAP_INTERLACED (1 << 0)
>>> /* Supports progressive formats */
>>> #define V4L2_DV_BT_CAP_PROGRESSIVE (1 << 1)
>>> /* Supports reduced blanking formats */
>>> #define V4L2_DV_BT_CAP_REDUCED_BLANKING (1 << 2)
>>> /* Supports custom formats */
>>> #define V4L2_DV_BT_CAP_CUSTOM (1 << 3)
>>>
>>> /* DV timings capabilities */
>>> struct v4l2_dv_timings_cap {
>>> __u32 type;
>>> union {
>>> struct v4l2_bt_timings_cap bt;
>>> __u32 reserved[32];
>>> };
>>> } __attribute__ ((packed));
>>>
>>> #define VIDIOC_DV_TIMINGS_CAP _IOWR('V', XX, struct
>>> v4l2_dv_timings_cap)
>>>
>>> This ioctl can be used to query the driver for the supported
>>> capabilities.
>>> Most speak for themselves. If V4L2_DV_BT_CAP_CUSTOM is not set, then only
>>> the timings from ENUM_DV_TIMINGS can be used, if it is set, then the
>>> hardware
>>> also supports timings that are not in that list.
>>
>> What information would this ioctl return? The DV times supported by the
>> chipset?
>> The ones supported by the TV set/monitor? The subset that it is supported by
>> the
>> entire pipeline?
>
> The latter. It's about finding the limitations of the hardware. It's not about
> the limitations of whatever you hook up to the hardware, that's handled by
> EDID negotiations, etc.
>
>>
>> What happens if the user changes the TV set and/or monitor? A practical case
>> to consider:
>> how this would handle things like a KVM (Keyboard, Video Monitor) switch?
>>
>> I think it may need a field to indicate if the TV set/monitor were detected,
>> and maybe
>> the screen size, in cm or pols (in order to help sizing the letters for OSD).
>>
>> Side note: I have this problem here on my current environment: if I start X
>> here with my
>> KVM switched to another monitor, X miss-detects the screen size (in pols),
>> and uses bigger fonts.
>> This happens because, when X starts, it tries to read the EDID information
>> from my monitor, but
>> it won't answer, as is connected to another computer. I'm not sure what my
>> KVM
>> device answers to it, but it is clear to me that it is not providing the
>> complete data to X.
>
> This is faulty EDID handling, nothing to do with this API.
>
> The better KVM switches should be able to deal with this intelligently, but
> intelligent
> EDID handling is a rather hit and miss in my experience.
I agree, but reality is different. Not sure if we can actually do something, but
an unconnected monitor (or a bad KVM) would result on an empty return from such
query at device level.
>>> 5) It can be dangerous to allow userspace to set up random timings for an
>>> output
>>> as this can damage monitors. It depends on various factors whether or
>>> not this
>>> should be protected by requiring root access. This has to be signalled
>>> somehow.
>>> I think that we should add a flag to v4l2_enum_dv_timings that tells
>>> whether
>>> selecting that specific timing requires root permissions. And we also
>>> need a
>>> V4L2_DV_BT_CAP_CUSTOM_ROOT flag in struct v4l2_bt_timings_cap to signal
>>> whether setting up custom timings requires root permissions.
>>
>> I think that custom settings should always require root, as they're not
>> officially
>> supported by the hardware, so, they could potentially cause damages.
>
> 'custom' != 'not officially supported'. 'Custom' means: not part of the
> enumerated
> timings list. Since many receivers and transmitters support an almost
> infinite number
> of timings you cannot say in general that custom settings require root. This
> has
> more to do with how the input or output is hooked up.
True, but if the timing were not enumerated, it means that there were no test
for that
specific timing (as it is very doubt that the vendors will test the entire
timings range,
as there are literally millions or billions of different combinations).
Look: if both the TV/monitor set and the pipeline says that they support an
specific DV timing
combination, such timing should be safe, as those timings are officially
supported. So, root
is not needed.
All the other possible timings are "likely" safe, at best. So, root access
should be
required.
> Any transmitter that is hooked up to a DVI/HDMI output probably needs root,
Why? if the TV set connected into it supports EDID, and the selected timing
were enumerated
by both, root shouldn't be needed, as both chipset and pipeline explicitly says
that such
timing is supported.
> but
> a transmitter that is hooked up to a known device (e.g. a monitor panel) and
> where
> the driver can guarantee that the timings are always safe, than root
> shouldn't be
> necessary. This is in the end a board-specific decision.
On a system where the monitor is physically part of the device and there's no
way to
replace the monitor with something else, then all hardware is well known.
However, on
such system, it doesn't make sense for the driver to offer support for custom
timings.
With great powers, comes great responsibility.
> What I want to avoid is that applications are forced to run as root, even if
> the designer of the board knows that there is no need for it.
The entire application doesn't need to be root. Just the code that would set a
custom DV.
Xawtv has a v4l-conf application specifically designed for those ioctl's that
require
root. Similar approaches could be used when DV custom timings are needed.
>> The ones returned by the query ioctl probably won't need, as the query
>> should be
>> checking at the entire pipeline if those timings are supported, so they're
>> potentially safe.
>>
>>
>>> 6) Parsing the timings structures can be difficult, and for many
>>> applications
>>> you are just interested in finding certain industry standard timings.
>>>
>>> This was the idea behind the original preset API, and I still think that
>>> that is very useful in practice.
>>>
>>> My general view is that the preset API should be deprecated and
>>> eventually
>>> removed. After extending the timings API I see no more need for the
>>> preset
>>> API.
>>
>> That's my impression too. Considering that the end result is that the old
>> API will
>> be removed, I don't think we should be bind to keep the same data structure.
>> If/where
>> it doesn't fit well, just replace it by something that fits (like the idea
>> of having
>> a "flags" field).
>>
>>> So that leaves the question how to incorporate the preset functionality
>>> in
>>> the timings API.
>>>
>>> I see two options:
>>>
>>> a) Add a preset or alias field to v4l2_enum_dv_timings. This just makes
>>> it
>>> easy for the application to check for specific formats and to store
>>> the timings.
>>>
>>> b) As a), but also add a new v4l2_dv_timings type: V4L2_DV_PRESET (or
>>> DV_ALIAS).
>>> This also has a struct v4l2_dv_preset (or dv_alias) that is used to
>>> store
>>> the preset value. This way you can set a preset directly using
>>> VIDIOC_S_DV_TIMINGS.
>>
>> c) discard struct v4l2_dv_preset, and add an optional
>> v4l2_enum_dv_timings_v2.
>>
>> At v4l2_enum_dv_timings_v2, zero would mean to use the timers. Any other
>> value means that
>> the timings should be discarded, an it should just use the timings for that
>> enum.
>> A query or enum iocl should fill it with a non-zero value for those enum
>> standards that
>> are supported by the driver.
>
> You've lost me here. I've no idea what you mean here, sorry.
What I'm proposing is something like:
#define V4L2_DV2_CUSTOM 0
#define V4L2_DV2_480P59_94 1 /* BT.1362 */
#define V4L2_DV2_576P50 2 /* BT.1362 */
...
(Disclaimer note: I just renamed the enums to _DV2_ to avoid re-using an
existing name
for the sake of my example. I'm not proposing to use the above names)
If a DV set operation fills the enum field with V4L2_DV2_CUSTOM, then the
timings will
be specified at the DV timings structure. If any other value is used, the
timings will
come from the preset.
>> In any case, we should create an enum namespace that won't have duplicate
>> names for
>> different settings, as we've discussed before.
>
> Agreed.
>
>> I suggest to implement some core functions to easy driver to handle the
>> presets that are
>> part of the existing standards.
>
> That was the plan.
>
>
> Regards,
>
> Hans
>
>>
>>>
>>>
>>> Comments?
>>>
>>> Hans Verkuil
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-media" in
>>> the body of a message to majord...@vger.kernel.org
>>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>
>> Regards,
>> Mauro
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-media" in
>> the body of a message to majord...@vger.kernel.org
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-media" in
> the body of a message to majord...@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
