stable inclusion
from stable-5.10.21
commit 0c78bf9c55f1b7e5301cc79a6cc0514bc0d60c99
bugzilla: 50609

--------------------------------

commit 15ea8ae8 upstream.

With the conversion to do the tty ldisc read operations in small chunks,
the n_tty line discipline became noticeably slower for throughput
oriented loads, because rather than read things in up to 2kB chunks, it
would return at most 64 bytes per read() system call.

The cost is mainly all in the "do system calls over and over", not
really in the new "copy to an extra kernel buffer".

This can be fixed by teaching the n_tty line discipline about the
"cookie continuation" model, which the chunking code supports because
things like hdlc need to be able to handle packets up to 64kB in size.

Doing that doesn't just get us back to the old performace, but to much
better performance: my stupid "copy 10MB of data over a pty" test
program is now almost twice as fast as it used to be (going down from
0.1s to 0.054s).

This is entirely because it now creates maximal chunks (which happens to
be "one byte less than one page" due to how we do the circular tty
buffers).

NOTE! This case only handles the simpler non-icanon case, which is the
one where people may care about throughput.  I'm going to do the icanon
case later too, because while performance isn't a major issue for that,
there may be programs that think they'll always get a full line and
don't like the 64-byte chunking for that reason.

Such programs are arguably buggy (signals etc can cause random partial
results from tty reads anyway), and good programs will handle such
partial reads, but expecting everybody to write "good programs" has
never been a winning policy for the kernel..
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: NChen Jun <chenjun102@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>