[Linux-kernel-mentees] [RFC, WIP, v4 09/11] media: vidtv: implement a PES packetizer

Mauro Carvalho Chehab mchehab+huawei at kernel.org
Wed May 6 08:59:25 UTC 2020


Em Wed, 6 May 2020 03:55:48 -0300
"Daniel W. S. Almeida" <dwlsalmeida at gmail.com> escreveu:

> Hi Mauro,
> 
> 
> > As commented, don't use WARN_ON(). At most, you could use WARN_ON_ONCE(),
> > as otherwise, you may end by causing serious performance issues if
> > the code starts to produce a flood of warnings at the dmesg.
> > 
> > I would use pr_warn_ratelimit() on all such cases.
> >   
> 
> OK.
> 
> 
> 
> 
> > I don't like the idea of changing the "from" buffer endiannes, copy
> > and then restore it back to the original state. Is this really needed?
> > 
> > I would, instead, define:
> > 
> > 	struct pes_header {
> > 	...
> > 		__be32 bitfield;
> > 		__be16 length;
> > 	...
> > 	};
> > 
> > Then wherever you would touch them:
> > 
> > 	u32 bitfield;
> > 	u16 len;
> > 
> > 	/* Write into BE fields */
> > 	pes_header.bitfield = cpu_to_be32(bitfield);
> > 	pes_header.length = cpu_to_be16(len);
> > 
> > 	/* Read from BE fields */
> > 	bitfield = be32_to_cpu(pes_header.bitfield);
> > 	len = be16_to_cpu(pes_header.length);
> > 
> > 
> > As a side effect, when you use "__be16" and "__be32" types, gcc
> > and smatch/sparse will warn you if you mess with endiannes.
> > 
> > Same applies to similar code elsewhere.
> >   
> 
> I don't like it either, it is error prone. I did not know about this
> other possibility. Does this work for _bitfields_ though?

See my comment below.

> I think the authors for libdvbv5 used unions precisely so bswap() could
> be called on a 'bitfield' member and from then on the bitfields could be
> accessed directly, e.g.:
> 
> 	union {
> 		u16 bitfield; <-- call bswap() on this
> 		struct {
>                         --> then use these directly:  
> 			u8  syntax:1;
> 			u8  zero:1;
> 			u8  one:2;
> 			u16 section_length:12;
> 		} __packed;
> 	} __packed
> 
> At least that's what I understood.

You should double-check the structs from the specs. If I'm not mistaken,
bytes were swapped on some places. As I commented for patch 08/11,
the focus there were to make life simpler for userspace, and not to
store a precise copy of the byte order.

> 
> I found this: 
> https://lwn.net/Articles/741762/
> 
> Maybe *_get_bits, *_replace_bits is the equivalent that I should use for bitfields?

I never used them, but, based on their definition:

static __always_inline base type##_get_bits(__##type v, base field)	\
{									\
	return (from(v) & field)/field_multiplier(field);		\
}

Calling be16_get_bits should do the right cast to the type.

I don't know what the "from()" and "to()" macros would do.

I guess you will need to do some tests to see if this works as
expected.

> 
> Because I'd rather not do this:
> 
> > 	u32 bitfield;
> > 	/* Write into BE fields */
> > 	pes_header.bitfield = cpu_to_be32(bitfield);  
> 
> Since I'd have to write the (many!) bitwise operations myself and I'm
> sure I will mess this up at _some_ point.

If you mess up, gcc (and/or smatch) will complain. I mean,

if bitfield is declared as __be32, if you do:

	u32 bitfield;
	pes_header.bitfield = bitfield;

this will produce warnings.

Thanks,
Mauro


More information about the Linux-kernel-mentees mailing list