Messages in this thread | | | From | Peter Hurley <> | Subject | [PATCH v3 0/6] N_TTY input path fixes | Date | Fri, 16 Jan 2015 15:05:33 -0500 |
| |
Greg,
Here's the unordered access fix blown up to a series :(
The first patch removes receive_room() from the non-producer code paths; this avoids the complication of read_head access from the consumer-side. Exclusive-use paths (where the termios_rwsem is write-locked, such as n_tty_set_termios(), n_tty_ioctl() and n_tty_flush_buffer()) can access any of the buffer indexes without racing, since both the input worker and reader are excluded by the write lock.
The second patch fixes a failure to make forward progress if an unterminated line longer than 3967 chars is received in canonical mode; the tty will mistakenly be throttled and thus potentially never receive the line termination.
The third patch simplifies the throttle threshold calculation, _and_ removes the last-but-one receive_room() call site.
The fourth patch fixes the unordered read buffer accesses with weakly- ordered arches.
The fifth patch fixes the space calculation issue noted by Christian; while this is low priority, the subsequent patch requires it.
The sixth patch fixes input handling when a line termination has not been received and the read buffer is full.
I'll send a separate email for which stables these can be directly applied; then I'll get to work on backporting these fixes to the other longterm stable trees.
Regards,
Peter Hurley (6): n_tty: Eliminate receive_room() from consumer/exclusive paths n_tty: Fix throttle for canon lines > 3967 chars n_tty: Simplify throttle threshold calculation n_tty: Fix unordered accesses to lockless read buffer n_tty: Fix PARMRK over-throttling n_tty: Fix read buffer overwrite when no newline
drivers/tty/n_tty.c | 209 ++++++++++++++++++++++++++++++---------------------- 1 file changed, 119 insertions(+), 90 deletions(-)
-- 2.2.2
| |