If user spaces calls read(2), the vfs_read() in fs/read_write.c
does not create an iov array any more since v5.19-10288-g3e20a751aff0.

That commit does as follows:

   -       struct iovec iov = { .iov_base = buf, .iov_len = len };
   -       iov_iter_init(&iter, READ, &iov, 1, len);
   +       iov_iter_ubuf(&iter, READ, buf, len);

as a resuly, our code that checks iter->iov->nr_segs sees 0 segments,
understands we have no buffer, and returns 0. Same for write(2).

This is a hack, as I don'w know what happens if the user space uses readv/writev.
Actually, I'd better have a simple fops->read() and fops->write() in the driver,
but I'm sure I lack a lot of background knowledge here.

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Sebastian Andrzej Siewior authored 1 year ago and Stefan Rauch committed 1 year ago

After handle_write_cfg(, WBA_ERR, ) the function wakes a worker which
will talk to the hardware and process the MSI. This worker can be
delayed if the system is busy leading to latency spikes during the
communication.

The detour via the worker can be avoided and the request can be
completed within the write request. Doing it within handle_write_cfg()
could be possible because wishbone_process_msi() is using its own buffer
for communication and only the locking is in the way.
It looks safer and simpler to simply wait until all locks are released
and process the MSI afterwards if needed.
This also mimics the previous workflow except that there is less delay
now.

Don't schedule the worker from handle_write_cfg() but process the MSI
request directly from the write callback if needed.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Kurt Kanzenbach <kurt@linutronix.de>

Sebastian Andrzej Siewior authored 1 year ago and Stefan Rauch committed 1 year ago

The interrupt service routine schedules a worker in order to perform the
main task including signaling userspace. This can lead to increased
latency since the worke is not sychronized with the interrupt handler
and be invoked "later" depending on the load on the system.

This can be avoided by using a threaded interrupt and invoking the work
directly in the threaded interrupt. There are still other places which
invoke wishbone_slave_ready() (and use the worker) in order to judge if
this is sane to do or not.

Use threaded interrupts and perform the work in the threaded interrupt.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Kurt Kanzenbach <kurt@linutronix.de>

sdbfs: relicense lib to LGPL v2.1

See merge request !1

The userspace and kernel binaries are kept as GPL v2 but the library has
been relicensed as LGPL v2.1 in order to allow it to be linked to
proprietary software.

The permission to relicense the files has been approved by CERN in 2018
and then in 2021 (and also by its main author).

This avoids sleeping in interrupt handlers.