We found now the problem. Basically it's a LINUX kernel bug. It seems
to appear at least on kernels between 2.0.0 and 2.0.22. Older kernels
do work fine. In case I don't bore anybody, here what's really
happening. If you do a read() system call in a non-blocking device,
and it get's interrupted in the system call (which kind of should not
happen at all), then a posix compilant system (and LINUX for that
matter) should return an error of EAGAIN. However the kernels
mentioned above randomly return either EINTR (which is for blocking
devices) or ERESTARTSYS, which the docs say "Should never be seen by
user programs".
Umm... that sounds interesting. What evidence do you have to
demonstrate this? (You realize that just because strace shows
ERESTARTSYS doesn't mean that the user program ever gets the
ERESTARTSYS; the system call just gets restarted.)
What device were you finding this before? I've just desk-checked the
tty N_TTY code, and it sure doesn't look like it has this problem. Each
device implements its own reader code, though, so you really need to
specify which device you were using.
- Ted