Skip to contentSkip to conversation

tty: Add driver unthrottle in ioctl(...,TCFLSH,..).

PATCH Superseded Ilya Zykov 3 participants 10 messages Nov 21, 2012 – Nov 22, 2012 TTY layer and serial drivers

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Ilya Zykov

author
#0 lore raw

PATCH 1/1 tty: Add driver unthrottle in ioctl(...,TCFLSH,..).

Revert 'tty: fix "IRQ45: nobody cared"'

This revert commit 7b292b4bf9a9d6098440d85616d6ca4c608b8304

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have
empty buffers and throttled driver and there is no place to unthrottle driver. It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do writers wake up.

About 'tty: fix "IRQ45: nobody cared"':
We don't call tty_unthrottle() if release last filp - ('tty->count == 0') In other case it must be safely. Maybe we have bug in other place.

---

drivers/tty/n_tty.c +0 -0
index 26f0d0e..a783d0e 100644
drivers/tty/n_tty.c +1 -1
@@ -184,6 +184,7 @@ static void reset_buffer_flags(struct tty_struct *tty)
        tty->canon_head = tty->canon_data = tty->erasing = 0;
        memset(&tty->read_flags, 0, sizeof tty->read_flags);
        n_tty_set_room(tty);
+       check_unthrottle(tty);
 }
 
 /**
@@ -1585,7 +1586,6 @@ static int n_tty_open(struct tty_struct *tty)
                        return -ENOMEM;
        }
        reset_buffer_flags(tty);
-       tty_unthrottle(tty);
        tty->column = 0;
        n_tty_set_termios(tty, NULL);
        tty->minimum_to_wake = 1;
#1 lore raw
Quoted 7 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have
empty buffers and throttled driver and there is no place to unthrottle driver. It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do writers wake up.

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Alan

author
#2 lore raw
Quoted 9 lines · Alan Cox
Quoted 7 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have
empty buffers and throttled driver and there is no place to unthrottle driver. It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do writers wake up.

Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

#3 lore raw
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 12 lines · <[email protected]> wrote:
Quoted 5 lines · Alan Cox
Quoted 5 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 2 lines
Quoted 2 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 2 lines
Quoted 7 lines
Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual observed bug, where a driver would lock up. So the behaviour needs preserved.

Andrew

author
#4 lore raw
Quoted 19 lines · andrew mcgregor
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 15 lines · <[email protected]> wrote:
Quoted 5 lines · Alan Cox
Quoted 5 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 2 lines
Quoted 2 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Quoted 9 lines
Quoted 4 lines
Quoted 2 lines

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual observed bug, where a driver would lock up. So the behaviour needs preserved.

Andrew

Maybe it was wrong driver, unfortunately, I didn't find full information about this bug. As an example, if driver indirectly call reset_buffer_flags() in driver's close() function it will be before decrement last (tty->count).

author
#5 lore raw
Quoted 32 lines · Ilya Zykov
Quoted 31 lines · andrew mcgregor
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 13 lines · <[email protected]> wrote:
Quoted 5 lines · Alan Cox
Quoted 5 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 3 lines
Quoted 3 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait,
and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 2 lines
Quoted 8 lines
Quoted 6 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual observed bug, where a driver would lock up. So the behaviour needs preserved.

Quoted 9 lines
Quoted 2 lines

Andrew

Maybe it was wrong driver, unfortunately, I didn't find full information about this bug. As an example, if driver indirectly call reset_buffer_flags() in driver's close() function it will be before decrement last (tty->count).

Particularly, many drivers and 'serial_core.c' use tty_ldisc_flush() in
own close() function. tty_ldisc_flush() call reset_buffer_flags() indirectly.
I think is wrong way use tty_ldisc_flush() in driver's close() function, because tty layer 'tty_release()' call tty_ldisc_release() after decremented (tty->count), and clear all buffers. We don't care about this in driver. And call ldisc's function.

author
#6 lore raw
Quoted 45 lines · Ilya Zykov
Quoted 34 lines · Ilya Zykov
Quoted 33 lines · andrew mcgregor
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 19 lines · <[email protected]> wrote:
Quoted 6 lines · Alan Cox
Quoted 6 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 3 lines
Quoted 3 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait,
and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 7 lines

So instead of revertng it why not just fix it ? Just add an argument to
the reset_buffer_flags function to indicate if unthrottling is permitted.

Quoted 4 lines
Quoted 2 lines

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual observed bug, where a driver would lock up. So the behaviour needs preserved.

Quoted 9 lines
Quoted 2 lines

Andrew

Maybe it was wrong driver, unfortunately, I didn't find full information about this bug. As an example, if driver indirectly call reset_buffer_flags() in driver's close() function it will be before decrement last (tty->count).

Quoted 9 lines

Particularly, many drivers and 'serial_core.c' use tty_ldisc_flush() in own close() function. tty_ldisc_flush() call reset_buffer_flags() indirectly.
I think is wrong way use tty_ldisc_flush() in driver's close() function, because tty layer 'tty_release()' call tty_ldisc_release() after decremented (tty->count), and clear all buffers.
We don't care about this in driver. And call ldisc's function.

Sorry, last phrase not correct.
We don't NEED care about this in driver. And call ldisc's function.

#7 lore raw
In reply to Ilya Zykov
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 05:29 PM, in message <[email protected]>, Ilya Zykov

Quoted 29 lines · <[email protected]> wrote:
Quoted 19 lines · andrew mcgregor
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 15 lines · <[email protected]> wrote:
Quoted 5 lines · Alan Cox
Quoted 5 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 2 lines
Quoted 2 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual

observed bug, where a driver would lock up. So the behaviour needs preserved.

Quoted 1 line
Quoted 7 lines
Quoted 2 lines

Andrew

Maybe it was wrong driver, unfortunately, I didn't find full information about this bug. As an example, if driver indirectly call reset_buffer_flags() in driver's close() function it will be before decrement last (tty->count).

Well, the driver in question was just 8250.c, so you should be able to see that the original condition can exist.

Here's the commit message again:

tty: fix "IRQ45: nobody cared"

Unthrottling the TTY during close ends up enabling interrupts on a device not on the active list, which will never have the interrupts cleared. Doctor, it hurts when I do this.

Quoted 9 lines
Quoted 1 line
Quoted 1 line · Alan Cox
Quoted 5 lines · On Wed, 01 Jun 2011 10:34:07 +1200 "andrew mcgregor"…

The LKML message
http://kerneltrap.org/mailarchive/linux-kernel/2010/2/25/4541847from
February doesn't seem to have been resolved since. We struck the issue, and the patch below (against 2.6.32) fixes it. Should I supply a patch against 3.0.0rc?

Quoted 8 lines

I think that would be sensible. I don't actually see how you hit it as the IRQ ought to be masked by then but it's certainly wrong for n_tty to be calling into check_unthrottle at that point.

So yes please send a patch with a suitable Signed-off-by: line to linux-serial and cc GregKH <[email protected]> as well.

Alan

What part of this no longer applies? I'm happy enough if you can prove that this can't happen any more, but otherwise the fix should remain.

Andrew

author
#8 lore raw
Quoted 67 lines · andrew mcgregor
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 05:29 PM, in message <[email protected]>, Ilya Zykov

Quoted 29 lines · <[email protected]> wrote:
Quoted 19 lines · andrew mcgregor
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

On 11/22/2012 at 10:39 AM, in message <[email protected]>, Ilya Zykov

Quoted 15 lines · <[email protected]> wrote:
Quoted 5 lines · Alan Cox
Quoted 5 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have empty buffers and throttled driver and there is no place to unthrottle

driver.

Quoted 2 lines
Quoted 2 lines

It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do

writers wake up.

Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

Alan

Because in my opinion, unthrottling permitted always, except release last filp (tty->count == 0)

Maybe so, but the patch was there in the first place to resolve an actual

observed bug, where a driver would lock up. So the behaviour needs preserved.

Quoted 1 line
Quoted 7 lines
Quoted 2 lines

Andrew

Maybe it was wrong driver, unfortunately, I didn't find full information about this bug. As an example, if driver indirectly call reset_buffer_flags() in driver's close() function it will be before decrement last (tty->count).

Well, the driver in question was just 8250.c, so you should be able to see that the original condition can exist.

Here's the commit message again:

tty: fix "IRQ45: nobody cared"

Unthrottling the TTY during close ends up enabling interrupts on a device not on the active list, which will never have the interrupts cleared. Doctor, it hurts when I do this.

Quoted 16 lines
Quoted 1 line
Quoted 1 line · Alan Cox
Quoted 5 lines · On Wed, 01 Jun 2011 10:34:07 +1200 "andrew mcgregor"…

The LKML message
http://kerneltrap.org/mailarchive/linux-kernel/2010/2/25/4541847from
February doesn't seem to have been resolved since. We struck the issue, and the patch below (against 2.6.32) fixes it. Should I supply a patch against 3.0.0rc?

I think that would be sensible. I don't actually see how you hit it as the IRQ ought to be masked by then but it's certainly wrong for n_tty to be calling into check_unthrottle at that point.

So yes please send a patch with a suitable Signed-off-by: line to linux-serial and cc GregKH <[email protected]> as well.

Quoted 8 lines
Quoted 1 line

Alan

What part of this no longer applies? I'm happy enough if you can prove that this can't happen any more, but otherwise the fix should remain.

Andrew

This patch must help for 8250.c

drivers/tty/serial/serial_core.c +0 -2
index a21dc8e..2e197c3 100644
@@ -1288,8 +1288,6 @@ static void uart_close(struct tty_struct *tty, struct file *filp)
        uart_shutdown(tty, state);
        uart_flush_buffer(tty);
 
-       tty_ldisc_flush(tty);
-
        tty_port_tty_set(port, NULL);
        spin_lock_irqsave(&port->lock, flags);
        tty->closing = 0;
author
#9 lore raw
In reply to Alan Cox
Quoted 9 lines · Alan Cox
Quoted 7 lines

Function reset_buffer_flags() also invoked during the ioctl(...,TCFLSH,..). At the time of request we can have full buffers and throttled driver too. If we don't unthrottle driver, we can get forever throttled driver, because after request, we will have
empty buffers and throttled driver and there is no place to unthrottle driver. It simple reproduce with "pty" pair then one side sleep on tty->write_wait, and other side do ioctl(...,TCFLSH,..). Then there is no place to do writers wake up.

Quoted 5 lines

So instead of revertng it why not just fix it ? Just add an argument to the reset_buffer_flags function to indicate if unthrottling is permitted.

Alan

Maybe we don't need call reset_buffer_flags() then execute tty_release() and (tty->count != 0)? Why in this case we reset buffers?