The flush_to_ldisc() work entry has special logic to notice when it has
seen the original tail of the data queue, and it avoids continuing the
flush if it sees that _original_ tail rather than the current tail.
This logic can trigger in case somebody is constantly adding new data to
the tty while the flushing is active - and the intent is to avoid
excessive CPU usage while flushing the tty, especially as we used to do
this from a softirq context which made it non-preemptible.
However, since we no longer re-arm the work-queue from within itself
(because that causes other trouble: see commit a5660b41af "tty: fix
endless work loop when the buffer fills up"), this just leads to
possible hung tty's (most easily seen in SMP and with a test-program
that floods a pty with data - nobody seems to have reported this for any
real-life situation yet).
And since the workqueue isn't done from timers and softirq's any more,
it's doubtful whether the CPU useage issue is really relevant any more.
So just remove the logic entirely, and see if anybody ever notices.
Alternatively, we might want to re-introduce the "re-arm the work" for
just this case, but then we'd have to re-introduce the delayed work
model or some explicit timer, which really doesn't seem worth it for
this.
Reported-and-tested-by: Guillaume Chazarain <guichaz@gmail.com>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Felipe Balbi <balbi@ti.com>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit b1c43f82c5.
It was broken in so many ways, and results in random odd pty issues.
It re-introduced the buggy schedule_work() in flush_to_ldisc() that can
cause endless work-loops (see commit a5660b41af: "tty: fix endless
work loop when the buffer fills up").
It also used an "unsigned int" return value fo the ->receive_buf()
function, but then made multiple functions return a negative error code,
and didn't actually check for the error in the caller.
And it didn't actually work at all. BenH bisected down odd tty behavior
to it:
"It looks like the patch is causing some major malfunctions of the X
server for me, possibly related to PTYs. For example, cat'ing a
large file in a gnome terminal hangs the kernel for -minutes- in a
loop of what looks like flush_to_ldisc/workqueue code, (some ftrace
data in the quoted bits further down).
...
Some more data: It -looks- like what happens is that the
flush_to_ldisc work queue entry constantly re-queues itself (because
the PTY is full ?) and the workqueue thread will basically loop
forver calling it without ever scheduling, thus starving the consumer
process that could have emptied the PTY."
which is pretty much exactly the problem we fixed in a5660b41af.
Milton Miller pointed out the 'unsigned int' issue.
Reported-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reported-by: Milton Miller <miltonm@bga.com>
Cc: Stefan Bigler <stefan.bigler@keymile.com>
Cc: Toby Gray <toby.gray@realvnc.com>
Cc: Felipe Balbi <balbi@ti.com>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
it makes it simpler to keep track of the amount of
bytes received and simplifies how flush_to_ldisc counts
the remaining bytes. It also fixes a bug of lost bytes
on n_tty when flushing too many bytes via the USB
serial gadget driver.
Tested-by: Stefan Bigler <stefan.bigler@keymile.com>
Tested-by: Toby Gray <toby.gray@realvnc.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Commit f23eb2b2b2 ('tty: stop using "delayed_work" in the tty layer')
ended up causing hung machines on UP with no preemption, because the
work routine to flip the buffer data to the ldisc would endlessly re-arm
itself if the destination buffer had filled up.
With the delayed work, that only caused a timer-driving polling of the
tty state every timer tick, but without the delay we just ended up with
basically a busy loop instead.
Stop the insane polling, and instead make the code that opens up the
receive room re-schedule the buffer flip work. That's what we should
have been doing anyway.
This same "poll for tty room" issue is almost certainly also the cause
of excessive kworker activity when idle reported by Dave Jones, who also
reported "flush_to_ldisc executing 2500 times a second" back in Nov 2010:
http://lkml.org/lkml/2010/11/30/592
which is that silly flushing done every timer tick. Wasting both power
and CPU for no good reason.
Reported-and-tested-by: Alexander Beregalov <a.beregalov@gmail.com>
Reported-and-tested-by: Sitsofe Wheeler <sitsofe@yahoo.com>
Cc: Greg KH <gregkh@suse.de>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Dave Jones <davej@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Using delayed-work for tty flip buffers ends up causing us to wait for
the next tick to complete some actions. That's usually not all that
noticeable, but for certain latency-critical workloads it ends up being
totally unacceptable.
As an extreme case of this, passing a token back-and-forth over a pty
will take two ticks per iteration, so even just a thousand iterations
will take 8 seconds assuming a common 250Hz configuration.
Avoiding the whole delayed work issue brings that ping-pong test-case
down to 0.009s on my machine.
In more practical terms, this latency has been a performance problem for
things like dive computer simulators (simulating the serial interface
using the ptys) and for other environments (Alan mentions a CP/M emulator).
Reported-by: Jef Driesen <jefdriesen@telenet.be>
Acked-by: Greg KH <gregkh@suse.de>
Acked-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There's a small window inside the flush_to_ldisc function,
where the tty is unlocked and calling ldisc's receive_buf
function. If in this window new buffer is added to the tty,
the processing might never leave the flush_to_ldisc function.
This scenario will hog the cpu, causing other tty processing
starving, and making it impossible to interface the computer
via tty.
I was able to exploit this via pty interface by sending only
control characters to the master input, causing the flush_to_ldisc
to be scheduled, but never actually generate any output.
To reproduce, please run multiple instances of following code.
- SNIP
#define _XOPEN_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
int main(int argc, char **argv)
{
int i, slave, master = getpt();
char buf[8192];
sprintf(buf, "%s", ptsname(master));
grantpt(master);
unlockpt(master);
slave = open(buf, O_RDWR);
if (slave < 0) {
perror("open slave failed");
return 1;
}
for(i = 0; i < sizeof(buf); i++)
buf[i] = rand() % 32;
while(1) {
write(master, buf, sizeof(buf));
}
return 0;
}
- SNIP
The attached patch (based on -next tree) fixes this by checking on the
tty buffer tail. Once it's reached, the current work is rescheduled
and another could run.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Cc: stable <stable@kernel.org>
Acked-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
The tty code should be in its own subdirectory and not in the char
driver with all of the cruft that is currently there.
Based on work done by Arnd Bergmann <arnd@arndb.de>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
