Messages in this thread | | | Date | Thu, 9 Dec 2021 22:18:36 +0900 | Subject | Re: [PATCH] tty: n_hdlc: make n_hdlc_tty_wakeup() asynchronous | From | Tetsuo Handa <> |
| |
On 2021/12/07 3:07, Linus Torvalds wrote: > On Mon, Dec 6, 2021 at 3:45 AM Tetsuo Handa > <penguin-kernel@i-love.sakura.ne.jp> wrote: >> >> Linus suspected that "struct tty_ldisc"->ops->write_wakeup() must not >> sleep, and Jiri confirmed it from include/linux/tty_ldisc.h. Thus, defer >> n_hdlc_send_frames() from n_hdlc_tty_wakeup() to a WQ context like >> net/nfc/nci/uart.c does. > > Thanks, this looks good to me. > > That said, I think there's pretty much the *exact* same pattern in > > drivers/net/caif/caif_serial.c: > write_wakeup() causes "handle_tx()", which then calls tty->ops->write(). > > drivers/net/hamradio/mkiss.c > mkiss_write_wakeup() -> tty->ops->write() > > drivers/tty/n_gsm.c: > gsmld_write_wakeup -> gsm_data_kick() -> gsmld_output -> > gsm->tty->ops->write() > > so this does seem to be a common bug pattern for code that has never > really seen a lot of testing.
Indeed.
> > The core tty stuff seems to get it right, but maybe I missed something > in my quick "grep and look for patterns".
handle_tx() in caif_serial.c has a line
/* skb_peek is safe because handle_tx is called after skb_queue_tail */
and I think that this comment is true only when handle_tx() is called from "struct net_device_ops"->ndo_start_xmit (== caif_xmit()). If handle_tx() is called from "struct tty_ldisc_ops"->write_wakeup (== ldisc_tx_wakeup()), handle_tx() might be called before skb_queue_tail() is called?
> > So I think this patch is good, but I do wonder if perhaps we should > move the "work_struct" into the tty layer itself, and do the whole > "schedule_work()" at that level.
I don't know about net_device_ops, but from synchronization point of view,
ser = tty->disc_data; BUG_ON(ser == NULL); WARN_ON(ser->tty != tty);
in ldisc_tx_wakeup() makes me feel uneasy, and I can't expect that ldisc_tx_wakeup() will do safe synchronization by moving the "work_struct" into the tty layer itself. That is, I think we somehow need to fix caif_serial.c after all.
> > Some code never wants it (most notably the regular n_tty one), but at > least n_tty doesn't really care, I suspect. n_tty is using > write_wakeup() literally just for fasync handling, so I suspect it's > not exactly going to be performance-critical. > > Of course, maybe the fix is to just fix caif_serial/mkiss and n_gsm. > Or mark them broken - does anybody use them?
I think that fixing individual driver sounds safer choice.
| |