[FFmpeg-devel] [PATCH 1/2] avutil/random_seed: add av_random()

[Anton Khirnov]

Quoting James Almer (2023-07-04 22:18:40)
> On 7/4/2023 5:14 PM, Anton Khirnov wrote:
> > Quoting James Almer (2023-07-04 22:08:40)
> >> On 7/4/2023 4:59 PM, Anton Khirnov wrote:
> >>>>
> >>>> Not if any of the functions above were called but failed to fill the buffer.
> >>>>
> >>>> I could add return AVERROR(ENOSYS) to the HAVE_UNISTD_H block, and
> >>>> return AVERROR_INVALIDDATA outside.
> >>>
> >>> AVERROR_INVALIDDATA is defined as 'Invalid data found when processing
> >>> input'.
> >>> This function does not process any input, so IMO that code never makes
> >>> sense for it.
> >>>
> >>> I'd say make it ENOSYS, AVERROR_UNKNOWN, or keep around one of
> >>> individual method errors.
> >>
> >> For the cases read() is used for /dev/random/, i can return
> >> AVERROR(errno), given the doxy states read() returns -1 on error and
> >> sets errno to some value. Although if it succeeds and returns a value
> >> smaller than len i would need to return AVERROR_UNKNOWN like you suggested.
> > 
> > I wonder if we should be trying /dev/random at all. I cannot think of
> > any remotely sane configuration where /dev/urandom fails, but
> > /dev/random works. So it only makes sense to use /dev/urandom.
> 
> I think the problem is that one may not generate the requested amount of 
> bytes.

It may happen that /dev/random blocks, but /dev/urandom should not. On
Linux, the manpage says:
       Since Linux 3.16, a read(2) from /dev/urandom will return at most
       32 MB.  A read(2) from /dev/random will return at most 512 bytes
       (340  bytes  be‐ fore Linux 2.6.12).
so we may want to read in a loop if we want to handle callers requesting
very large amounts of data (though I can't think of a valid use case for
that).

But however it may be, it should never happen that /dev/urandom is not
usable, but /dev/random is.

-- 
Anton Khirnov