hello all
I am not subscribed to linux-kernel*; please CC any follow-ups to
[email protected] (I probably won't reply from 2000-12-22 to 2001-01-07)
I am running 2.4.0-test12-pre2
This snippet can prevent progress of any other processes that tries to do
a write to a pty:
============
#include <sys/fcntl.h>
#include <unistd.h>
int main()
{
int ptm;
ptm = open("/dev/ptmx", O_WRONLY);
while (1)
write(ptm, "hello, world!\n", 14);
}
============
With this running, and no process eating up the greetings, I can telnet to
my machine; the banner and login prompt appear. ps -alx at this point
reveals in.telnetd is in do_select(). As soon as I type even one char of
username, another ps reveals in.telnetd now stuck in __down_interruptible()
Lucky I usually leave a few logged in consoles lying around; xterm also
uses pty's!
Some observations:
2.2.12 behaves fine (telnet logins work fine)
2.2.12-2.4.0 diffs between tty_io.c show changes involving up/down/etc. in
do_tty_write().
Bernd Jendrissek
P.S. apologies to all for my void * arithmetic a few months ago; it was a
moment of eager-beaver weakness.
This 2.4 bug must be fixed.
[email protected] wrote:
>
> This snippet can prevent progress of any other processes that tries to do
> a write to a pty:
>
> ============
> #include <sys/fcntl.h>
> #include <unistd.h>
>
> int main()
> {
> int ptm;
> ptm = open("/dev/ptmx", O_WRONLY);
> while (1)
> write(ptm, "hello, world!\n", 14);
> }
> ============
>
Your program fills up the pty write buffer and then blocks.
This is fair enough. But it blocks while holding the /dev/ptmx
inode semaphore, so _all_ other users of /dev/ptmx are blocked.
Your xterms get stuck, telnet and ssh are dead. It's time to
find the car keys.
In 2.2, we have a per-tty semaphore. No problems. This
patch basically restores the 2.2 implementation and it
passes basic sanity testing.
But it's not my area. Is there really a _need_ to take the inode
semaphore while we're writing to the pty instance? I hope
not, because I took that out.
Perhaps it would be logical to clone the /dev/ptmx inode
in some manner when it's opened?
Anyone?
--- linux-2.4.0-test13-pre4/include/linux/tty.h Tue Dec 12 19:24:23 2000
+++ linux-akpm/include/linux/tty.h Fri Dec 22 21:44:47 2000
@@ -305,6 +305,7 @@
unsigned long canon_head;
unsigned int canon_column;
struct semaphore atomic_read;
+ struct semaphore atomic_write;
spinlock_t read_lock;
};
--- linux-2.4.0-test13-pre4/drivers/char/tty_io.c Tue Dec 12 19:24:17 2000
+++ linux-akpm/drivers/char/tty_io.c Fri Dec 22 21:48:22 2000
@@ -699,9 +699,8 @@
size_t count)
{
ssize_t ret = 0, written = 0;
- struct inode *inode = file->f_dentry->d_inode;
- if (down_interruptible(&inode->i_sem)) {
+ if (down_interruptible(&tty->atomic_write)) {
return -ERESTARTSYS;
}
if ( test_bit(TTY_NO_WRITE_SPLIT, &tty->flags) ) {
@@ -734,7 +733,7 @@
file->f_dentry->d_inode->i_mtime = CURRENT_TIME;
ret = written;
}
- up(&inode->i_sem);
+ up(&tty->atomic_write);
return ret;
}
@@ -1972,6 +1971,7 @@
tty->tq_hangup.routine = do_tty_hangup;
tty->tq_hangup.data = tty;
sema_init(&tty->atomic_read, 1);
+ sema_init(&tty->atomic_write, 1);
spin_lock_init(&tty->read_lock);
INIT_LIST_HEAD(&tty->tty_files);
}
-