Are calls to open()/close() serialized by tty layer?
Are calls to a tty driver's open()/close() methods serialized by the tty layer? If so, has this always been the case?
I'm working on converting an old tty driver over to using the "new" tty_port_* functions and the standard "struct tty_port" port data structure.
[I thought about converting it to a uart driver for serial_core, but AFAICT, that API doesn't provide any way to return errors for userspace open() and write() calls and would therefore break my user-land API.]
Quoted 1 line
From looking at example tty drivers in the 3.7.2 source tree, it
appears that they assume calls to open()/close() are serialized.
The old driver I'm working on had its own internal locking to protect race conditiions when incrementing and decrementing of a port's open count. That counter is being prelaced by tty->port->count, and in existing in-kernel tty drivers it doesn't look like there is any locking when that count is incremented during an open().
For example, from rocket.c:
891 static int rp_open(struct tty_struct *tty, struct file *filp)
892 {
[...]
924 tty->driver_data = info;
925 tty_port_tty_set(port, tty);
926
927 if (port->count++ == 0) {
928 atomic_inc(&rp_num_ports_open);
929
930 #ifdef ROCKET_DEBUG_OPEN
931 printk(KERN_INFO "rocket mod++ = %d...\n",
932 atomic_read(&rp_num_ports_open));
933 #endif
934 }
If calls to open()/close() aren't serialized, then line 927 would cause a race condition...
Signature · 3 lines
Grant Edwards grant.b.edwards Yow! Everybody gets free
at BORSCHT!
gmail.com Quoted 2 lines · On Wed, 16 Jan 2013 20:07:49 +0000 (UTC) Grant Edwards…
Are calls to a tty driver's open()/close() methods serialized by the tty layer? If so, has this always been the case?
They are not fully serialized - never have been. A lot of the drivers were always racy, some that have not been fully converted to the port helpers still are.
Quoted 8 lines
Quoted 1 line
From looking at example tty drivers in the 3.7.2 source tree, it
appears that they assume calls to open()/close() are serialized.
The old driver I'm working on had its own internal locking to protect race conditiions when incrementing and decrementing of a port's open count. That counter is being prelaced by
tty->port->count, and in existing in-kernel tty drivers it doesn't look like there is any locking when that count is incremented during anopen().
Use the tty_port helpers. They will also handle all the other detail you need. The helpers themselves provide serialized activate/shutdown callbacks on the port which are serialized against each other and hangups.
The port helpers will also ensure your refcounting is right, you don't have open/close/hangup races and that you get the required POSIX semantics.
The port helpers are designed so you can propogate your own error codes either by wrapping them or by returning your own error code from port->activate() which will be duly propogated to user space.
Alan
Quoted 8 lines · Alan Cox
Quoted 2 lines · Grant Edwards
Are calls to a tty driver's open()/close() methods serialized by the tty layer? If so, has this always been the case?
They are not fully serialized - never have been. A lot of the drivers were always racy, some that have not been fully converted to the port helpers still are.
That's what it looked like it to me, but I wasn't sure.
Quoted 8 lines
Quoted 6 lines
[...]
The old driver I'm working on had its own internal locking to protect race conditiions when incrementing and decrementing of a port's open count. That counter is being replaced bytty->port->count, and in existing in-kernel tty drivers it doesn't look like there is any locking when that count is incremented during anopen().Use the
tty_porthelpers.
That was the plan...
Quoted 3 lines
They will also handle all the other detail you need. The helpers themselves provide serialized activate/shutdown callbacks on the port which are serialized against each other and hangups.
Ah! None of the examples I looked at so far were using tty_port_open() and the port activate method, and I had somehow missed noticing their existence when I was looking through tty_port sources.
Quoted 7 lines
The port helpers will also ensure your refcounting is right, you don't have open/close/hangup races and that you get the required POSIX semantics.
The port helpers are designed so you can propogate your own error codes either by wrapping them or by returning your own error code from
port->activate()which will be duly propogated to user space.
Cool, thanks!
Which of the tty drivers is the best example to follow when it comes to use of the tty_port_ helpers?
It looks like the tty_port_ helpers only go back to 2.6.33, so supporting kernels older than that will require some sort of private, backported versions of the helpers. Maybe I can talk management and tech support into dropping support for kernels older than 2.6.33...
Signature · 3 lines
Grant Edwards grant.b.edwards Yow! I threw up on my
at window!
gmail.com