tty/serial eating \n regression in 3.13-rc1

[Jason Gunthorpe]

Hello Peter,

I have been testing 3.13-rc1 and I noticed a change in behavior in the
tty/serial layer. Specifically I have a login running on serial that
presents the usual login:/password: prompt.

3.10.12 does this:

login: admin
password: 
prompt>

While 3.13-rc1 does this:

login: admin
password: prompt>

That is to say, a \n got lost. I have also noticed while using the
serial console that '\n's are occasionally lost, and the lossage is
highly repeatable and seemingly based on number of chars being output
in a burst, eg entering 'foo bar' at a shell prompt might loose the \n,
while entering 'foo bar ' will not.

I ran a full git bisect search and determined that the attached patch
introduced the problem.

This seems like a bug - at least the commit comment seems like it
should not cause any change in behaviour. I can reproduce it and help
debug, if you can give some guidance, I am not a tty layer expert :)

I have seen this bug on ppc with an out-of-tree serial driver, and on
ARM kirkwood with the serial8250 driver. I did the bisection search on
PPC.

good 8bb495e3f02401ee6f76d1b1d77f3ac9f079e376 Linux 3.10
bad 6ce4eac1f600b34f2f7f58f9cd8f0503d79e42ae Linux 3.13-rc1
good 6e4664525b1db28f8c4e1130957f70a94c19213e Linux 3.11
bad 5e01dc7b26d9f24f39abace5da98ccbd6a5ceb52 Linux 3.12

bad cc998ff8811530be521f6b316f37ab7676a07938
bad 816434ec4a674fcdb3c2221a6dffdc8f34020550
good 751144271f4b63d5de9005ea4e5e6e5c7c6fd629
bad 40031da445fb4d269af9c7c445b2adf674f171e7
bad f66c83d059d1ed90968caa81d401f160912b063a

bad 1ef39808ca27837017a433f94aa7055cb8490e80 serial: sirf: add DT-binding document to describle detailed properties
bad 10d8b34a421716d55a4ff7c2d427c35e88b8fd60 serial: max310x: Driver rework
good 0f56bd2f6a97d8b0eb5c8f9bc04b83a6c16d1d48 tty: Use non-atomic state to signal flip buffer flush pending
bad a1dd30e9b4c38a4a177106741b03ac6a55777286 n_tty: Special case EXTPROC receive_buf() as raw mode
bad 29c7c5ca36d9c132cf9c37a09bc43626e790fb4c n_tty: Eliminate counter in __process_echoes
good addaebccf63a8c9f7c7a62ce1ceb9da4307c5a1c n_tty: Use separate head and tail indices for echo_buf
good 019ebdf9f26fd2e43b9e1af576835183e95dc82e n_tty: Eliminate echo_commit memory barrier
bad bc5b1ec5860a9bf52842be4a3a9d96e19f06c11d (*) n_tty: Only flush echo output if actually output
bad cbfd0340ae1993378fd47179db949e050e16e697 n_tty: Process echoes in blocks

==> cbfd0340ae1993378fd47179db949e050e16e697

Thanks,
Jason