======================================================

Kernel module (device driver) for dt3155 frame grabber

======================================================

This driver is completely rewritten (almost from scratch).

Initially I used the linux version found on Internet and
it worked well for kernels 2.4.xx. But in the times I had
to switch to 2.6.17 because got a new motherboard with Promise
PDC20378 disk controller that was not supported in 2.4.xx.
As far as I had 2GB of RAM the driver was unstable due to
probable interference between the allocator an HIGHMEM,
thus I decided to rewrite and adapt it to act as a true
PCI compliant character device driver

The first thing to do was to design a new allocator based
on allocating a configurable number of 4MB chunks of memory,
that latter are broken into frame buffers of 768x576
bytes kept in different FIFOs. As far as the driver autoloads
as a kernel module during kernel boot, the allocation of 4MB
chunks succeeds.

Since then the driver is used in the lab and adapted to the
changes in the driver interface up to the current kernel
(2.6.39).

The driver keeps two FIFOs: one for free buffers and one for
buffers containing image data. When one start the acquisition
(via dedicated ioctl) a buffer is taken from the free fifo
and the DMA machine is programed to load image data to it and
started. Under completion (during the interrupt), another free
buffer is programed into the DMA and the filled in buffer is
pushed onto the image data FIFO. The user land is reading
image data via regular read system calls from the image data
FIFO (in the order of the acquisition). When finishing another
ioctl stops the DMA machine and moves all not read image data
buffers to the free FIFO. This double buffering permit us to
acquire data send them to the application window and write them
on HDD in real time (without lost frames on the disk).

Sure, there are problems:

1. The device can be opened only once.

2. Reading from the device is not POSIX compliant. If a
small user land buffer is used copy_to_user() fills it and
the rest of the kernel buffer data is disregarded (and lost).

3. How to mmap the buffer (without faulting for each page).
For now all buffer pages are marked as reserved to make
remap_pfn_range() happy. I know this is not good.

4. Not tested on a BIG ENDIAN architecture.

5. Many others.... :-)