PROBLEM: Race condition in tty buffer's function flush_to_ldisc().
Function flush_to_ldisc() call disc->ops->receive_buf(), without tty->buf.lock and with TTY_FLUSHING bit set. If we have deferred TTY_FLUSHPENDING request, another thread can grab tty->buf.lock, and flush tty's buffer when receive_buf() use its.
---
drivers/tty/tty_buffer.c +8 -11
@@ -427,11 +427,15 @@ static void flush_to_ldisc(struct work_stty_buffer_free (tty, head);continue ; } -/* Ldisc or user is trying to flush the buffers -we are feeding to the ldisc, stop feeding the -line discipline as we want to empty the queue */ -if (test_bit (TTY_FLUSHPENDING , &tty->flags)) +/* Ldisc or user is trying to flush the buffers. +We may have a deferred request to flush the input buffer, +if so pull the chain under the lock and empty the queue */ +if (test_bit (TTY_FLUSHPENDING , &tty->flags)) { +__tty_buffer_flush (tty); +clear_bit (TTY_FLUSHPENDING , &tty->flags); +wake_up (&tty->read_wait);break ; + }if (!tty->receive_room)break ;if (count > tty->receive_room)
@@ -447,13 +451,6 @@ static void flush_to_ldisc(struct work_sclear_bit (TTY_FLUSHING , &tty->flags); } -/* We may have a deferred request to flush the input buffer, -if so pull the chain under the lock and empty the queue */ -if (test_bit (TTY_FLUSHPENDING , &tty->flags)) { -__tty_buffer_flush (tty); -clear_bit (TTY_FLUSHPENDING , &tty->flags); -wake_up (&tty->read_wait); - }spin_unlock_irqrestore (&tty->buf.lock, flags);tty_ldisc_deref (disc);
Quoted 2 lines · On Mon, 07 Nov 2011 15:14:34 +0400 Ilya Zykov <[email protected]> wrote:
Function
flush_to_ldisc()calldisc->ops->receive_buf(), withouttty->buf.lockand withTTY_FLUSHINGbit set.
flush_to_ldisc is single threaded for a given tty. If you fail to ensure that is the case everything breaks.
What cases does this occur ?
Also for the other patches do you have benchmarks yet ?
Alan
Quoted 8 lines · Alan Cox
Quoted 2 lines · On Mon, 07 Nov 2011 15:14:34 +0400 Ilya Zykov <[email protected]> wrote:
Function
flush_to_ldisc()calldisc->ops->receive_buf(), withouttty->buf.lockand withTTY_FLUSHINGbit set.
flush_to_ldiscis single threaded for a given tty. If you fail to ensure that is the case everything breaks.
I think so. But if add in flush_to_ldisc():
if (!test_and_set_bit(TTY_FLUSHING, &tty->flags)) {
.....
} else
prink...In my syslog appear this prink on pty devices.
Quoted 7 lines
What cases does this occur ?
Also for the other patches do you have benchmarks yet ?
Alan
Quoted 8 lines · Alan Cox
Quoted 2 lines · On Mon, 07 Nov 2011 15:14:34 +0400 Ilya Zykov <[email protected]> wrote:
Function
flush_to_ldisc()calldisc->ops->receive_buf(), withouttty->buf.lockand withTTY_FLUSHINGbit set.
flush_to_ldiscis single threaded for a given tty. If you fail to ensure that is the case everything breaks.
Why we need: "if (!test_and_set_bit(TTY_FLUSHING, &tty->flags)) {" if flush_to_ldisc is single threaded?
we can: set_bit(TTY_FLUSHING, &tty->flags)
without if() at all.
Quoted 7 lines
What cases does this occur ?
Also for the other patches do you have benchmarks yet ?
Alan
Quoted 5 lines
if (!test_and_set_bit(TTY_FLUSHING, &tty->flags)) { ..... } else prink...In my syslog appear this prink on pty devices.
Then we need to understand the call paths that did this. I suspect it may be a bug in the hacks Linus made to n_tty. They've always bothered me as they never looked correct.
Can you get traces to see if it is actually calling flush_to_ldisc multithreaded on a given tty
Quoted 4 lines
Why we need: "if (!test_and_set_bit(TTY_FLUSHING,
&tty->flags)) {" ifflush_to_ldiscis single threaded?
we can:set_bit(TTY_FLUSHING, &tty->flags)
withoutif()at all.
It is single threaded with respect to itself (you can't have two flush_to_ldisc on the same tty at once) but you can have a parallel call to tty_buffer_flush. The tty_buffer_flush path needs to pick the right approach reliably.
So the theory is
If TTY_FLUSHING is set then tty_buffer_flush lets the flush_to_ldisc do the work. During this time we won't fluish a buffer under the ldisc.
If TTY_FLUSHING is not set then we will do the flush directly. As we hold tty->buf.lock at that point the two cannot race as far as I can see.
We are actually now probably at the point we could take a per tty mutex on the flush_to_ldisc and use it to tidy up ldisc change, flush and other things. So this code could welll be something worth simplifying once the rest of the tty_lock() is cleaned out.
Alan
Quoted 11 lines · Alan Cox
Quoted 5 lines
Why we need: "if (!test_and_set_bit(TTY_FLUSHING,
&tty->flags)) {" ifflush_to_ldiscis single threaded?
we can:set_bit(TTY_FLUSHING, &tty->flags)
withoutif()at all.It is single threaded with respect to itself (you can't have two
flush_to_ldiscon the same tty at once) but you can have a parallel call totty_buffer_flush. Thetty_buffer_flushpath needs to pick the right approach reliably.
Of course I know about tty_buffer_flush(), it only read TTY_FLUSHING, it can't change TTY_FLUSHING, if flush_to_ldisc() single threaded, we can change TTY_FLUSHING only in one place in one time(in flush_to_ldisc()), therefor we can use only "set_bit(TTY_FLUSHING, &tty->flags)" without test.
Quoted 5 lines
Of course I know about
tty_buffer_flush(), it only readTTY_FLUSHING, it can't changeTTY_FLUSHING, ifflush_to_ldisc()single threaded, we can changeTTY_FLUSHINGonly in one place in onetime(in flush_to_ldisc()), therefor we can use only "set_bit(TTY_FLUSHING, &tty->flags)" without test.
Yes.. if you can pin down why in your testing you see the other case sometimes being true.
Alan
Quoted 6 lines · Alan Cox
Quoted 5 lines
Of course I know about
tty_buffer_flush(), it only readTTY_FLUSHING, it can't changeTTY_FLUSHING, ifflush_to_ldisc()single threaded, we can changeTTY_FLUSHINGonly in one place in onetime(in flush_to_ldisc()), therefor we can use only "set_bit(TTY_FLUSHING, &tty->flags)" without test.
Quoted 7 lines
Yes.. if you can pin down why in your testing you see the other case sometimes being true.
Alan
OK, I try.
Ilya.
Quoted 11 lines · Alan Cox
Quoted 5 lines
Of course I know about
tty_buffer_flush(), it only readTTY_FLUSHING, it can't changeTTY_FLUSHING, ifflush_to_ldisc()single threaded, we can changeTTY_FLUSHINGonly in one place in onetime(in flush_to_ldisc()), therefor we can use only "set_bit(TTY_FLUSHING, &tty->flags)" without test.Yes.. if you can pin down why in your testing you see the other case sometimes being true.
Alan
Nested call flush_to_ldisc() happened on different CPU only. It happens because one side pty call "schedule_work(&tty->buf.work)" on one CPU from "tty_flip_buffer_push()", the other side call "schedule_work(&tty->buf.work)" from n_tty's layer from "n_tty_set_room()" on different CPU. It also can happen in interrupt if interrupt handle on different CPU. Schedule_work() schedule work on CPU which was called(IMHO).
For testing I use in xterminal: cat big_file
and flush_to_ldisc() with this patch:
drivers/tty/tty_buffer.c +8 -1
@@ -405,13 +405,14 @@ static void flush_to_ldisc(struct work_s container_of(work, struct tty_struct, buf.work);unsigned long flags;struct tty_ldisc *disc; +static int mthread; disc =tty_ldisc_ref (tty);if (disc ==NULL )/* !TTY_LDISC */ return ;spin_lock_irqsave (&tty->buf.lock, flags); - + mthread =0 ;if (!test_and_set_bit (TTY_FLUSHING , &tty->flags)) {struct tty_buffer *head;while ((head = tty->buf.head) !=NULL ) {
@@ -445,6 +446,12 @@ static void flush_to_ldisc(struct work_sspin_lock_irqsave (&tty->buf.lock, flags); }clear_bit (TTY_FLUSHING , &tty->flags); + }else { + mthread =1 ; +printk (KERN_WARNING "Tty %s FLUSHING CPU %d.\n" , tty->name,percpu_read (cpu_number)); + } +if (mthread) { +printk (KERN_WARNING "Tty %s FLUSHING multithreaded CPU %d.\n" , tty->name,percpu_read (cpu_number)); }/* We may have a deferred request to flush the input buffer,
And get in my syslog:
Nov 7 21:34:13 serh kernel: [ 76.323848] Tty ptm0 FLUSHING CPU 0. Nov 7 21:34:13 serh kernel: [ 76.323850] Tty ptm0 FLUSHING multithreaded CPU 0. Nov 7 21:34:13 serh kernel: [ 76.323856] Tty ptm0 FLUSHING CPU 0. Nov 7 21:34:13 serh kernel: [ 76.323857] Tty ptm0 FLUSHING multithreaded CPU 0. Nov 7 21:34:13 serh kernel: [ 76.323861] Tty ptm0 FLUSHING multithreaded CPU 1. Nov 7 21:34:13 serh kernel: [ 76.336022] Tty ptm0 FLUSHING CPU 0. Nov 7 21:34:13 serh kernel: [ 76.336024] Tty ptm0 FLUSHING multithreaded CPU 0. Nov 7 21:34:13 serh kernel: [ 76.336030] Tty ptm0 FLUSHING multithreaded CPU 1. Nov 7 21:34:13 serh kernel: [ 76.353134] Tty ptm0 FLUSHING CPU 0. Nov 7 21:34:13 serh kernel: [ 76.353136] Tty ptm0 FLUSHING multithreaded CPU 0. Nov 7 21:34:13 serh kernel: [ 76.353143] Tty ptm0 FLUSHING CPU 0. Nov 7 21:34:13 serh kernel: [ 76.353145] Tty ptm0 FLUSHING multithreaded CPU 0. Nov 7 21:34:13 serh kernel: [ 76.353148] Tty ptm0 FLUSHING multithreaded CPU 1.
......
......
Quoted 5 lines
Nested call
flush_to_ldisc()happened on different CPU only. It happens because one side pty call "schedule_work(&tty->buf.work)" on one CPU from "tty_flip_buffer_push()", the other side call "schedule_work(&tty->buf.work)" from n_tty's layer from "n_tty_set_room()" on different CPU. It also can happen in interrupt if interrupt handle on different CPU.Schedule_work()schedule work on CPU which was called(IMHO).
But the two sides are different ttys so that specific case shouldn't matter.
Something is clearly happening because you trigger the trace but I don't see why the paths you indicate cause it ? If you make mthread a cpu struct variable and add a WARN_ON(1) in the error case do you get any useful traces ?
Alan
Quoted 2 lines
see why the paths you indicate cause it ? If you make mthread a cpu struct variable and add a
WARN_ON(1)in the error case do you get any
I mean tty struct (ie tty->mthread) of course sorry