The standard FreeBSD MII support causes severe problems on the LPC3200
chip family. If an Ethernet module register is accessed while there is
no clock from the PHY, the chip completely locks up and only an external
watchdog can recover from this state. The legacy driver had a custom
PHY management code which helped to avoid such issues. The if_lpe.c
driver is no longer maintained by FreeBSD.
Under unknown conditions the receive path ended up in a frozen state.
In this state, the DMA and driver descriptor head were equal and all
receive descriptors had the used bit set. So, the DMA was unable to
store received frames. However, the receive daemon was never woken up
to refill the receive buffers. It seems that the RXUBR interrupt can be
used to recover from this state.
Update #4652.
Do not use the interface mutex in the receive loop. Avoid multiple
reads of DMA descriptor words. Use a compile-time constant for the
receive DMA descriptor count to simplify calculations.
Update #4652.
Use the transmit interface handler to avoid a transmit task/interrupt.
Use a compile-time constant for the transmit DMA descriptor count to
simplify calculations.
Update #4652.
The pppstart expected that a driver write would somehow magically
process all data passed to the write function. Because ppp disables all
buffering that originally has been in termios, that assumption is not
true for all but polled drivers.
With this patch, the pppstart now gets and processes the feedback that
is returned from the driver via rtems_termios_dequeue_characters.
Update #4493
Possible data sources for SD driver:
- mmc_sd_switch():
- length: 64 byte;
- buffer on stack
- mmc_test_bus_width():
- length: 4 or 8 byte
- buffer in program memory or on stack
- mmc_app_send_scr():
- length: 8 byte
- buffer from device ivar structure
- mmc_app_sd_status():
- length: 64 byte
- buffer from device ivar structure
- mmc_send_ext_csd():
- length: MMC_EXTCSD_SIZE = 512
- buffer from device ivar structure
- rtems_bsd_mmcsd_disk_read_write():
- length: depends on read
- buffer from rtems_blkdev buffer -> already aligned
- mmcsd_ioctl_cmd():
- length: depends on call
- buffer malloced, not aligned -> patched in RTEMS
So the problematic buffers are only the ones up to 512 bytes. Copy these data
into a buffer to avoid that problem.
