when trying to port "kmouse" to Linux 2.1.5, I noticed that mouse
movements would suddenly report bogus Y coordinates. Upon examining
the source code I noticed that the code fragment
long kmouse_read_data(struct inode *inode, struct file *filp,
char *buf, unsigned long count)
{
signed char localbuf[5];
.
.
.
copy_to_user(buf,localbuf,5);
.
.
.
}
did not have the desired effect. The "copy_to_user" instruction would
only copy four instead of five bytes! Replacing the code with "memcpy"
or avoiding the use of a constant length parameter did fix the
problem:
long kmouse_read_data(struct inode *inode, struct file *filp,
char *buf, unsigned long count)
{
static int constant5 = 5;
signed char localbuf[5];
.
.
.
copy_to_user(buf,localbuf,constant5);
.
.
.
}
I assume, there is something wrong with the operand constraints, but I
cannot spot any obvious solution.
Markus
P.S.: I also noticed that passing complex expressions as the source
and/or destination address for copy_to_user() would result in
GCC (2.7.2) complaining about a reserved register being
clobbered.