Skip to contentSkip to conversation

HID: fix tty<->hid deadlock

PATCH In review Jiri Slaby 3 participants 4 messages Sep 4, 2008 – Oct 16, 2008 1 message not in the archive HID

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Jiri Slaby

author
#1 lore raw

PATCH 1/1 HID: fix tty<->hid deadlock

Quoted 51 lines · Jiri Slaby

[email protected] napsal(a):

Quoted 50 lines

(Adding Alan Cox to the cc: in case he can shed light on this one - it appears that HID_COMPAT only puts the bullet in the chamber, and doesn't actually cause the hang...)

Quoted 3 lines · Jiri Slaby
Quoted 2 lines · On 09/02/2008 06:11 PM, [email protected] wrote:

The following 3 lines don't go to console by default due to loglevel setting. So I'm not sure exactly where it hangs. But it's somewhere in here.

Right around here, we kick off a modprobe for 'hid_dummy'.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
[    1.959193] usbcore: registered new interface driver usbhid
[    1.973037] usbhid: v2.6:USB HID core driver

And since my initrd doesn't include any modules (since until now, I've had a kernel that can everything builtin so it can boot far enough to do the whole udev/modprobe off my root filesystem, and hid_dummy is a new one on me), this modprobe spits out a:

modprobe: FATAL: Could not open '/lib/modules/2.6.27-rc5-mmotm0829/modules.dep': No such file or directory

Well, yeah.. No modules on the initrd, so no modules.dep. But having spewed its error message, modprobe apparently decides to go off in a snit and hang. Eventually, the usermode_helper call does a wait() on the modprobe, and then *that* hangs because modprobe isn't returning. And eventually the whole level of initcalls comes to a screeching halt...

And here's the totally unexpected kernel traceback for the modprobe:

schedule_timeout+0x22/0xb4
? _raw_spin_lock+0xce/0x186
? _raw_spin_unlock+0xb7/0xe0
wait_for_common+-xb2/0xfb
? default_wake_function+0x0/0xf
wait_for_completion+0x18/0x1a
flush_cpu_workqueue+0x6b/0x77
? wq_barrior_func+0x0/0xf
flush_workqueue+0x4f/0x68
flush_scheduled_work+0x10/0x12
tty_ldisc_release+0x4a/0x21e
? _raw_pin_lock+0xce/0x186
? debug_mutex_unlock+0x127/0x14d
? mutex_unlock_slowpath+0x14a/0x15c
tty_release_dev+0x4da/0x508
? get_parent_ip+0x11/0x41
? get_parent_ip+0x11/0x41
tty_release+0x19/0x24
__fput+0xd9/0x198
fput+0x15/0x17
filp_close+0x67/0x72
sys_close+0xa9/0x104
system_call_fastpath+0x16/0x1b
Quoted 4 lines
Quoted 1 line

WTF? We hang trying to close a tty??!?

Hmm, *if* we stuck in request_module in hid, workqueue cannot be flushed and tty waits... Could you stick 2 printks into hid_compat_load if it finishes?

Ah, this is deadlock:
hid_compat_load() runs on the default workqueue, it request_module(), it execs modprobe, it exits, tty flushes default workqueue, it hangs, because we are still in it.

I haven't run into it since there always was at least one opener of the tty, I guess. Could you try the patch below?

--

---
 drivers/hid/hid-core.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)
drivers/hid/hid-core.c +10 -1
index 92c16e1..18d952a 100644
@@ -1663,6 +1663,7 @@ static void hid_compat_load(struct work_struct *ws)
 	request_module("hid-dummy");
 }
 static DECLARE_WORK(hid_compat_work, hid_compat_load);
+static struct workqueue_struct *hid_compat_wq;
 #endif
 
 static int __init hid_init(void)
@@ -1680,7 +1681,12 @@ static int __init hid_init(void)
 		goto err_bus;
 
 #ifdef CONFIG_HID_COMPAT
-	schedule_work(&hid_compat_work);
+	hid_compat_wq = create_workqueue("hid_compat");
+	if (!hid_compat_wq) {
+		hidraw_exit();
+		goto err;
+	}
+	queue_work(hid_compat_wq, &hid_compat_work);
 #endif
 
 	return 0;
@@ -1692,6 +1698,9 @@ err:
 
 static void __exit hid_exit(void)
 {
+#ifdef CONFIG_HID_COMPAT
+	destroy_workqueue(hid_compat_wq);
+#endif
 	hidraw_exit();
 	bus_unregister(&hid_bus_type);
 }
Signature · 1 line
1.6.0
#2 lore raw

drivers/hid/hid-core.c

Quoted 18 lines · Jiri Slaby

hid_compat_load() runs on the default workqueue, it request_module(), it execs modprobe, it exits, tty flushes default workqueue, it hangs, because we are still in it.

I haven't run into it since there always was at least one opener of the tty, I guess. Could you try the patch below?

--

---
 drivers/hid/hid-core.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)
drivers/hid/hid-core.c +0 -0
index 92c16e1..18d952a 100644

With this patch applied, the kernel behaves as expected for my config: it requests a 'modprobe hid_dummy', which prints a nasty message about being unable to find it on the initrd, and then we continue with the boot.

Thanks for the quick fix...

author
#3 lore raw

[email protected] napsal(a):

Quoted 16 lines · Jiri Slaby
Quoted 15 lines

hid_compat_load() runs on the default workqueue, it request_module(), it execs modprobe, it exits, tty flushes default workqueue, it hangs, because we are still in it.

I haven't run into it since there always was at least one opener of the tty, I guess. Could you try the patch below?

--

---
 drivers/hid/hid-core.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)
Quoted 8 lines

drivers/hid/hid-core.c

Quoted 4 lines
drivers/hid/hid-core.c +0 -0
index 92c16e1..18d952a 100644

With this patch applied, the kernel behaves as expected for my config: it requests a 'modprobe hid_dummy', which prints a nasty message about being unable to find it on the initrd, and then we continue with the boot.

Thanks for testing, I can't think of any method which would solve this problem. As this compat module is a temporary solution for those who have (very) old module init tools (not supporting aliases), I wouldn't do anything else until somebody complains.

#4 lore raw

hidraw_exit in discarded section (was: Re: [PATCH 1/1] HID: fix tty<->hid deadlock)

In reply to Jiri Slaby
Quoted 41 lines · Jiri Slaby

hid_compat_load() runs on the default workqueue, it request_module(), it execs modprobe, it exits, tty flushes default workqueue, it hangs, because we are still in it.

I haven't run into it since there always was at least one opener of the tty, I guess. Could you try the patch below?

--

---
 drivers/hid/hid-core.c |   11 ++++++++++-
 1 files changed, 10 insertions(+), 1 deletions(-)
drivers/hid/hid-core.c +7 -1
index 92c16e1..18d952a 100644
@@ -1663,6 +1663,7 @@ static void hid_compat_load(struct work_struct *ws)
 	request_module("hid-dummy");
 }
 static DECLARE_WORK(hid_compat_work, hid_compat_load);
+static struct workqueue_struct *hid_compat_wq;
 #endif
 
 static int __init hid_init(void)
@@ -1680,7 +1681,12 @@ static int __init hid_init(void)
 		goto err_bus;
 
 #ifdef CONFIG_HID_COMPAT
-	schedule_work(&hid_compat_work);
+	hid_compat_wq = create_workqueue("hid_compat");
+	if (!hid_compat_wq) {
+		hidraw_exit();
+		goto err;
+	}
+	queue_work(hid_compat_wq, &hid_compat_work);
 #endif
 
 	return 0;

drivers/hid/hid-core.c err: · line 1706

Quoted 9 lines
drivers/hid/hid-core.c +3 -0
@@ -1692,6 +1698,9 @@ err:
 
 static void __exit hid_exit(void)
 {
+#ifdef CONFIG_HID_COMPAT
+	destroy_workqueue(hid_compat_wq);
+#endif
 	hidraw_exit();
 	bus_unregister(&hid_bus_type);
 }

hid_init() is marked __init.
hidraw_exit() is marked __exit.

Hence I get:

hidraw_exit' referenced in section .init.text' of drivers/built-in.o: defined in discarded section `.exit.text' of drivers/built-in.o

and my `enable everything' m68k kernel fails to link.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds