tree 560b0e17466a4b633f51f9ab402339e5ac23e5ea
parent 8c66d6d2a1a572768616ddca2c3863384b14d846
author Shinya Kuribayashi <shinya.kuribayashi.px@renesas.com> 1353030855 +0900
committer Greg Kroah-Hartman <gregkh@linuxfoundation.org> 1353031371 -0800

serial: sh-sci: add locking to console write function to avoid SMP lockup

Symptom:

When entering the suspend with Android logcat running, printk() call
gets stuck and never returns.  The issue can be observed at printk()s
on nonboot CPUs when going to offline with their interrupts disabled,
and never seen at boot CPU (core0 in our case).

Details:

serial_console_write() lacks of appropriate spinlock handling.

In SMP systems, as long as sci_transmit_chars() is being processed
at one CPU core, serial_console_write() can stuck at the other CPU
core(s), when it tries to access to the same serial port _without_
a proper locking.  serial_console_write() waits for the transmit FIFO
getting empty, while sci_transmit_chars() writes data to the FIFO.

In general, peripheral interrupts are routed to boot CPU (core0) by
Linux ARM standard affinity settings.  SCI(F) interrupts are handled
by core0, so sci_transmit_chars() is processed on core0 as well.

When logcat is running, it writes enormous log data to the kernel at
every moment, forever.  So core0 can repeatedly continue to process
sci_transmit_chars() in its interrupt handler, which eventually makes
the other CPU core(s) stuck at serial_console_write().

Looking at serial/8250.c, this is a known console write lockup issue
with SMP kernels.  Fix the sh-sci driver in the same way 8250.c does.

Signed-off-by: Shinya Kuribayashi <shinya.kuribayashi.px@renesas.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>