The following functions now reports to the user that an overrun has
occured by returning errno EIO:
* grtc_read, on overrun and no more data in the DMA buffer
* ioctl(GRTC_IOC_RECV), on overrun and no more frames in the DMA buffer
* ioctl(GRTC_IOC_ISSTARTED), EIO is returned on overrun
This driver was initially intended for APBUART access on GRLIB PCI
peripherals (AMBA-over-PCI). There are already many APBUART drivers
for LEON. Instead of using this one the termios APBUART should work
on AMBA-over-PCI systems with recent changes on LEON3, however that
has not been tested yet.
The common interrupt layer for GRLIB PCI perihperals is prepared for SMP
support by this patch. The existing locking (interrupt disabling) is
replaced by a new requirement on the user to implement locking before
calling the genirq API. This approach avoids taking more locks than
necessary.
The split up of the locks also introduces that the user must allocate
memory to describe ISR handlers, to avoid calling malloc()/free() while
possessing a spin-lock and interrupts are globally disabled.
It is also required to use semaphore release instead of flush when stopping
or on BUSOFF/AHBERR condition. Otherwise a task just about to wait
(taking the semaphore) could end up locked because the semaphore count is
still the same.
There was previously a scenario where the semaphore flush would not always make
semaphore obtain to return in case of BUSOFF, AHBERROR or grcan_stop. It has to
do with the rtems_semaphore_flush() not releasing the semaphore but just aborts
any _current_ waiter.
When bus-off condition is detected by the ISR, it sets the started flag to
STATE_BUSOFF. This is monitored by the user functions grcan_read() and
grcan_write() each time they want to enable DMA or update interrupt mask. If
they detect that ISR has detected bus-off then they will return either with an
error or with the number of CAN messages processed. Future calls to functions
which require STARTED mode will be rejected and grcan_isstarted() will return
0. The next call to grcan_stop() will do the transition from BUSOFF->STOPPED
and the device can then be started again with grcan_start().
Similar to a bus-off condition, the AHB error condition detected by the ISR
will trigger the same shut-down logic of the driver. The difference is that
the state entered is STATE_AHBERR and the routines will return a different
value to indicate AHB error state.
This commit also fixes an issue where ISR was not always unregistered on close.
User functions can cause these transitions:
STATE_STOPPED -> STATE_STARTED (grcan_start)
STATE_STARTED -> STATE_STOPPED (grcan_stop)
STATE_BUSOFF -> STATE_STOPPED (grcan_stop, grcan_close)
STATE_AHBERR -> STATE_STOPPED (grcan_stop, grcan_close)
ISR can cause these transition
STATE_STARTED -> STATE_BUSOFF (grcan_interrupt)
STATE_STARTED -> STATE_AHBERR (grcan_interrupt)
STATE_BUSOFF/AHBERR is entered from ISR on bus-off condition. At transition
the ISR disables DMA, masks all interrupts and flushes semaphores.
Other related updates:
* Statistics are updated from the ISR. Update is now spin-locked to ensure a
consistent user view.
* The debug output has been updated to include state changes.
* For read/write/flush, return error (-4) if driver aborted the operation
due to bus-off. Likewise if abourted due to AHB error -5 is returned.
* Collect bus-off statistics
Related to the new BUSOFF and AHBERR states the API has been updated to
reflect the current SW driver state. The isstarted() function has been
replaced with get_state().
This commit preserves all driver services, using a function based user
interface instead of the I/O interface.
The messages count parameter is now number of CAN messages instead of
number of bytes.
The EDCL and Duplex-Detection are now disabled during soft-reset
of the GRETH core.
The speed settings are preserved on boot and greth_stop() also,
this is required to keep EDCL operational when DD is set.
Probing of separate interrupts was done by storing the GPTIMER_CFG_SI bit. But
it was never actually stored since it is bit 8 and the datatype is 8-bit. Now
store the AND result as boolean value instead.
This commit updates the OCCAN driver locking mechanism:
1. Convert interrupt disable/enable to interrupt locks.
2. Make sure interrupt service routines use proper locking to deal with threads
running in parallel.
Do not disable transmitter until shift register is empty when closing device.
It is correct to look at the sending software state when closing the device.
However, the sending state must be cleared only when everything is sent.
