Skip to contentSkip to conversation

usb: gadget: u_serial: Fix and cleanup

PATCH In review Ladislav Michl 2 participants 10 messages Jul 13, 2019 – Jul 15, 2019 USB

lore
Share & tools
lore
#0 lore raw

v2 cover usb: gadget: u_serial: Fix and cleanup

Following patchset makes console work (patch 1 and 4) for an AT91SAM9G20 board connected to xhci_hcd and does some cleanup.
Tested with "console=ttyS0,115200n8 console=ttyGS0" on kernel command line and following inittab:
console::respawn:/sbin/getty -L 115200 ttyS0 vt100 console::respawn:/sbin/getty -L 115200 ttyGS0 vt100

However there is an issue remaining:
Disconnect triggers WARN_ON in gs_close:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 501 at drivers/usb/gadget/function/u_serial.c:706 gs_close+0x3c/0x1e4
Modules linked in:
CPU: 0 PID: 501 Comm: getty Not tainted 5.2.0-rc7 #44
Hardware name: Atmel AT91SAM9
[<c0107514>] (unwind_backtrace) from [<c01051c0>] (show_stack+0x10/0x18)
[<c01051c0>] (show_stack) from [<c05465a8>] (dump_stack+0x18/0x24)
[<c05465a8>] (dump_stack) from [<c010fa80>] (__warn+0xcc/0xe4)
[<c010fa80>] (__warn) from [<c010fad0>] (warn_slowpath_null+0x38/0x48)
[<c010fad0>] (warn_slowpath_null) from [<c03b6648>] (gs_close+0x3c/0x1e4)
[<c03b6648>] (gs_close) from [<c03036b0>] (tty_release+0x1d4/0x460)
[<c03036b0>] (tty_release) from [<c01ce464>] (__fput+0xe4/0x1b0)
[<c01ce464>] (__fput) from [<c0124e1c>] (task_work_run+0x8c/0xa8)
[<c0124e1c>] (task_work_run) from [<c011155c>] (do_exit+0x354/0x814)
[<c011155c>] (do_exit) from [<c0111a9c>] (do_group_exit+0x54/0xb8)
[<c0111a9c>] (do_group_exit) from [<c011a190>] (get_signal+0x18c/0x658)
[<c011a190>] (get_signal) from [<c0104bbc>] (do_work_pending+0xe0/0x44c)
[<c0104bbc>] (do_work_pending) from [<c0101068>] (slow_work_pending+0xc/0x20)
Exception stack(0xc3797fb0 to 0xc3797ff8)
7fa0:                                     00000000 beb87d0c 00000001 00000000
7fc0: 0009a150 00000000 00099c04 00000003 0009a198 0007e049 00099bd4 0009a1e4
7fe0: b6e3f000 beb87cd8 00018210 b6dbcc40 60000010 00000000
---[ end trace 70af570fde0de49b ]---

Explanation lives in drivers/tty/serial/ip22zilog.c at line 751 (see also __tty_hangup), but naive approach adding

	if (test_bit(TTY_HUPPING, &tty->flags))
		return 0;

at the beginning of gs_close does not work as this way port->port.count never drops back to zero. Ideas welcome.

Btw, it would seem nicer to have single output kfifo for both kernel console and tty, but that involves greater changes and I'd like to have driver working in stable as well.

Changes:
- v2: Fixed locking, patches splitted, starving write fixed.

Ladislav Michl (5):
usb: gadget: u_serial: Fix console_req complete event race
usb: gadget: u_serial: Remove console specific alloc/free req
functions
usb: gadget: u_serial: Fix console_req access race
usb: gadget: u_serial: Fix starving write
usb: gadget: u_serial: Use bool for req_busy

 drivers/usb/gadget/function/u_serial.c | 120 +++++++++++--------------
 1 file changed, 51 insertions(+), 69 deletions(-)
Signature · 1 line
2.22.0
#1 lore raw

PATCH v2 1/5 usb: gadget: u_serial: Fix console_req complete event race

gs_complete_out might be called before con_lock following usb_ep_queue is locked, which prevents any future output on the console. Fix that by resetting req_busy only if usb_ep_queue fails. While there also put variable access we are racing with connection/disconnection events under con_lock as well.

Fixes a5beaaf39455 ("usb: gadget: Add the console support for usb-to-serial port")

---
Changes:
- v2: None

 drivers/usb/gadget/function/u_serial.c | 41 +++++++++++---------------
 1 file changed, 17 insertions(+), 24 deletions(-)
drivers/usb/gadget/function/u_serial.c +17 -24
index 65f634ec7fc2..876af4085cea 100644
@@ -984,47 +984,40 @@ static int gs_console_thread(void *data)
 	struct gs_port *port;
 	struct usb_request *req;
 	struct usb_ep *ep;
-	int xfer, ret, count, size;
+	int len, size, status;
 
+	spin_lock_irq(&info->con_lock);
 	do {
 		port = info->port;
-		set_current_state(TASK_INTERRUPTIBLE);
-		if (!port || !port->port_usb
-		    || !port->port_usb->in || !info->console_req)
+		if (!port || !port->port_usb || !info->console_req)
 			goto sched;
 
 		req = info->console_req;
 		ep = port->port_usb->in;
+		len = kfifo_len(&info->con_buf);
+		if (len > 0 && !info->req_busy) {
+			size = ep->maxpacket;
+			if (len < size)
+				size = len;
 
-		spin_lock_irq(&info->con_lock);
-		count = kfifo_len(&info->con_buf);
-		size = ep->maxpacket;
-
-		if (count > 0 && !info->req_busy) {
-			set_current_state(TASK_RUNNING);
-			if (count < size)
-				size = count;
-
-			xfer = kfifo_out(&info->con_buf, req->buf, size);
-			req->length = xfer;
-
-			spin_unlock(&info->con_lock);
-			ret = usb_ep_queue(ep, req, GFP_ATOMIC);
-			spin_lock(&info->con_lock);
-			if (ret < 0)
-				info->req_busy = 0;
-			else
-				info->req_busy = 1;
+			req->length = kfifo_out(&info->con_buf, req->buf, size);
+			info->req_busy = 1;
 
 			spin_unlock_irq(&info->con_lock);
+			status = usb_ep_queue(ep, req, GFP_ATOMIC);
+			spin_lock_irq(&info->con_lock);
+			if (status < 0)
+				info->req_busy = 0;
 		} else {
-			spin_unlock_irq(&info->con_lock);
 sched:
+			spin_unlock_irq(&info->con_lock);
 			if (kthread_should_stop()) {
 				set_current_state(TASK_RUNNING);
 				break;
 			}
 			schedule();
+			set_current_state(TASK_INTERRUPTIBLE);
+			spin_lock_irq(&info->con_lock);
 		}
 	} while (1);
 
Signature · 1 line
2.22.0
#2 lore raw

PATCH v2 2/5 usb: gadget: u_serial: Remove console specific alloc/free req functions

In reply to Ladislav Michl

Driver already contains request allocation and deallocation functions, so use them also for console_req.

---
Changes:
- v2: Move locking changes into separate patch

 drivers/usb/gadget/function/u_serial.c | 28 ++------------------------
 1 file changed, 2 insertions(+), 26 deletions(-)
drivers/usb/gadget/function/u_serial.c +2 -26
index 876af4085cea..cb88a640179e 100644
@@ -892,30 +892,6 @@ static struct tty_driver *gs_tty_driver;
 static struct gscons_info gscons_info;
 static struct console gserial_cons;
 
-static struct usb_request *gs_request_new(struct usb_ep *ep)
-{
-	struct usb_request *req = usb_ep_alloc_request(ep, GFP_ATOMIC);
-	if (!req)
-		return NULL;
-
-	req->buf = kmalloc(ep->maxpacket, GFP_ATOMIC);
-	if (!req->buf) {
-		usb_ep_free_request(ep, req);
-		return NULL;
-	}
-
-	return req;
-}
-
-static void gs_request_free(struct usb_request *req, struct usb_ep *ep)
-{
-	if (!req)
-		return;
-
-	kfree(req->buf);
-	usb_ep_free_request(ep, req);
-}
-
 static void gs_complete_out(struct usb_ep *ep, struct usb_request *req)
 {
 	struct gscons_info *info = &gscons_info;
@@ -955,7 +931,7 @@ static int gs_console_connect(int port_num)
 	port = ports[port_num].port;
 	ep = port->port_usb->in;
 	if (!info->console_req) {
-		info->console_req = gs_request_new(ep);
+		info->console_req = gs_alloc_req(ep, ep->maxpacket, GFP_ATOMIC);
 		if (!info->console_req)
 			return -ENOMEM;
 		info->console_req->complete = gs_complete_out;
@@ -974,7 +950,7 @@ static void gs_console_disconnect(struct usb_ep *ep)
 	struct gscons_info *info = &gscons_info;
 	struct usb_request *req = info->console_req;
 
-	gs_request_free(req, ep);
+	gs_free_req(ep, req);
 	info->console_req = NULL;
 }
 
Signature · 1 line
2.22.0
#3 lore raw

PATCH v2 3/5 usb: gadget: u_serial: Fix console_req access race

In reply to Ladislav Michl

console_req is always null when calling gs_console_connect, so remove check and put access under con_lock as we are racing with gs_console_thread.

---
Changes:
- v2: New patch

 drivers/usb/gadget/function/u_serial.c | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
drivers/usb/gadget/function/u_serial.c +11 -11
index cb88a640179e..558a6929ce68 100644
@@ -930,15 +930,12 @@ static int gs_console_connect(int port_num)
 
 	port = ports[port_num].port;
 	ep = port->port_usb->in;
-	if (!info->console_req) {
-		info->console_req = gs_alloc_req(ep, ep->maxpacket, GFP_ATOMIC);
-		if (!info->console_req)
-			return -ENOMEM;
-		info->console_req->complete = gs_complete_out;
-	}
-
-	info->port = port;
 	spin_lock(&info->con_lock);
+	info->console_req = gs_alloc_req(ep, ep->maxpacket, GFP_ATOMIC);
+	if (!info->console_req)
+		return -ENOMEM;
+	info->console_req->complete = gs_complete_out;
+	info->port = port;
 	info->req_busy = 0;
 	spin_unlock(&info->con_lock);
 	pr_vdebug("port[%d] console connect!\n", port_num);
@@ -948,10 +945,11 @@ static int gs_console_connect(int port_num)
 static void gs_console_disconnect(struct usb_ep *ep)
 {
 	struct gscons_info *info = &gscons_info;
-	struct usb_request *req = info->console_req;
 
-	gs_free_req(ep, req);
+	spin_lock(&info->con_lock);
+	gs_free_req(ep, info->console_req);
 	info->console_req = NULL;
+	spin_unlock(&info->con_lock);
 }
 
 static int gs_console_thread(void *data)
@@ -1008,7 +1006,6 @@ static int gs_console_setup(struct console *co, char *options)
 	info->port = NULL;
 	info->console_req = NULL;
 	info->req_busy = 0;
-	spin_lock_init(&info->con_lock);
 
 	status = kfifo_alloc(&info->con_buf, GS_CONSOLE_BUF_SIZE, GFP_KERNEL);
 	if (status) {
@@ -1064,6 +1061,9 @@ static struct console gserial_cons = {
 
 static void gserial_console_init(void)
 {
+	struct gscons_info *info = &gscons_info;
+
+	spin_lock_init(&info->con_lock);
 	register_console(&gserial_cons);
 }
 
Signature · 1 line
2.22.0
#4 lore raw

PATCH v2 4/5 usb: gadget: u_serial: Fix starving write

In reply to Ladislav Michl

Writing to ttyGS unconnected to host currently fills port_write_buf which then causes gs_write_room to return 0 and boot hangs waiting for connection.
Fix that by dropping previsous data if free space in port_write_buf reaches treshold and host is unconnected.

---
Changes:
- v2: New patch

 drivers/usb/gadget/function/u_serial.c | 26 ++++++++++++++++++++------
 1 file changed, 20 insertions(+), 6 deletions(-)
drivers/usb/gadget/function/u_serial.c +20 -6
index 558a6929ce68..2dd6e1211d4a 100644
@@ -79,6 +79,7 @@
  */
 #define QUEUE_SIZE		16
 #define WRITE_BUF_SIZE		8192		/* TX only */
+#define WRITE_BUF_TRESHOLD	1024
 #define GS_CONSOLE_BUF_SIZE	8192
 
 /* console info */
@@ -562,7 +563,7 @@ static int gs_start_io(struct gs_port *port)
 
 	/* unblock any pending writes into our circular buffer */
 	if (started) {
-		tty_wakeup(port->port.tty);
+		gs_start_tx(port);
 	} else {
 		gs_free_requests(ep, head, &port->read_allocated);
 		gs_free_requests(port->port_usb->in, &port->write_pool,
@@ -758,6 +759,7 @@ static int gs_write(struct tty_struct *tty, const unsigned char *buf, int count)
 {
 	struct gs_port	*port = tty->driver_data;
 	unsigned long	flags;
+	int		avail, cnt;
 
 	pr_vdebug("gs_write: ttyGS%d (%p) writing %d bytes\n",
 			port->port_num, tty, count);
@@ -766,8 +768,16 @@ static int gs_write(struct tty_struct *tty, const unsigned char *buf, int count)
 	if (count)
 		count = kfifo_in(&port->port_write_buf, buf, count);
 	/* treat count == 0 as flush_chars() */
-	if (port->port_usb)
+	if (port->port_usb) {
 		gs_start_tx(port);
+	} else {
+		avail = kfifo_avail(&port->port_write_buf);
+		if (avail < WRITE_BUF_TRESHOLD) {
+			cnt = WRITE_BUF_TRESHOLD - avail;
+			while (cnt--)
+				kfifo_skip(&port->port_write_buf);
+		}
+	}
 	spin_unlock_irqrestore(&port->port_lock, flags);
 
 	return count;
@@ -784,6 +794,9 @@ static int gs_put_char(struct tty_struct *tty, unsigned char ch)
 
 	spin_lock_irqsave(&port->port_lock, flags);
 	status = kfifo_put(&port->port_write_buf, ch);
+	if (!port->port_usb &&
+	     kfifo_avail(&port->port_write_buf) < WRITE_BUF_TRESHOLD)
+		kfifo_skip(&port->port_write_buf);
 	spin_unlock_irqrestore(&port->port_lock, flags);
 
 	return status;
@@ -799,6 +812,8 @@ static void gs_flush_chars(struct tty_struct *tty)
 	spin_lock_irqsave(&port->port_lock, flags);
 	if (port->port_usb)
 		gs_start_tx(port);
+	else
+		kfifo_reset_out(&port->port_write_buf);
 	spin_unlock_irqrestore(&port->port_lock, flags);
 }
 
@@ -806,11 +821,10 @@ static int gs_write_room(struct tty_struct *tty)
 {
 	struct gs_port	*port = tty->driver_data;
 	unsigned long	flags;
-	int		room = 0;
+	int		room;
 
 	spin_lock_irqsave(&port->port_lock, flags);
-	if (port->port_usb)
-		room = kfifo_avail(&port->port_write_buf);
+	room = kfifo_avail(&port->port_write_buf);
 	spin_unlock_irqrestore(&port->port_lock, flags);
 
 	pr_vdebug("gs_write_room: (%d,%p) room=%d\n",
@@ -823,7 +837,7 @@ static int gs_chars_in_buffer(struct tty_struct *tty)
 {
 	struct gs_port	*port = tty->driver_data;
 	unsigned long	flags;
-	int		chars = 0;
+	int		chars;
 
 	spin_lock_irqsave(&port->port_lock, flags);
 	chars = kfifo_len(&port->port_write_buf);
Signature · 1 line
2.22.0
#5 lore raw

PATCH v2 5/5 usb: gadget: u_serial: Use bool for req_busy

In reply to Ladislav Michl

Let's make code more consistent by using bool for req_busy as it is done for similar variables in struct gs_port.

---
Changes:
- v2: None

 drivers/usb/gadget/function/u_serial.c | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)
drivers/usb/gadget/function/u_serial.c +5 -6
index 2dd6e1211d4a..1b1359a168e7 100644
@@ -89,7 +89,7 @@ struct gscons_info {
 	struct kfifo		con_buf;
 	/* protect the buf and busy flag */
 	spinlock_t		con_lock;
-	int			req_busy;
+	bool			req_busy;
 	struct usb_request	*console_req;
 };
 
@@ -918,7 +918,7 @@ static void gs_complete_out(struct usb_ep *ep, struct usb_request *req)
 	case 0:
 		/* normal completion */
 		spin_lock(&info->con_lock);
-		info->req_busy = 0;
+		info->req_busy = false;
 		spin_unlock(&info->con_lock);
 
 		wake_up_process(info->console_thread);
@@ -950,7 +950,6 @@ static int gs_console_connect(int port_num)
 		return -ENOMEM;
 	info->console_req->complete = gs_complete_out;
 	info->port = port;
-	info->req_busy = 0;
 	spin_unlock(&info->con_lock);
 	pr_vdebug("port[%d] console connect!\n", port_num);
 	return 0;
@@ -989,13 +988,13 @@ static int gs_console_thread(void *data)
 				size = len;
 
 			req->length = kfifo_out(&info->con_buf, req->buf, size);
-			info->req_busy = 1;
+			info->req_busy = true;
 
 			spin_unlock_irq(&info->con_lock);
 			status = usb_ep_queue(ep, req, GFP_ATOMIC);
 			spin_lock_irq(&info->con_lock);
 			if (status < 0)
-				info->req_busy = 0;
+				info->req_busy = false;
 		} else {
 sched:
 			spin_unlock_irq(&info->con_lock);
@@ -1019,7 +1018,7 @@ static int gs_console_setup(struct console *co, char *options)
 
 	info->port = NULL;
 	info->console_req = NULL;
-	info->req_busy = 0;
+	info->req_busy = false;
 
 	status = kfifo_alloc(&info->con_buf, GS_CONSOLE_BUF_SIZE, GFP_KERNEL);
 	if (status) {
Signature · 1 line
2.22.0
#6 lore raw
In reply to Ladislav Michl
Quoted 9 lines · Ladislav Michl

Following patchset makes console work (patch 1 and 4) for an AT91SAM9G20 board connected to xhci_hcd and does some cleanup.
Tested with "console=ttyS0,115200n8 console=ttyGS0" on kernel command line and following inittab:
console::respawn:/sbin/getty -L 115200 ttyS0 vt100 console::respawn:/sbin/getty -L 115200 ttyGS0 vt100

However there is an issue remaining:
Disconnect triggers WARN_ON in gs_close:

[...]

Hi,

Can you try a patchset I sent some time ago [1] if it fixes your case?

Best Regards,
Michał Mirosław

[1] https://www.spinics.net/lists/linux-usb/msg177537.html

#7 lore raw
Quoted 8 lines · Michał Mirosław
Quoted 7 lines · Ladislav Michl

Following patchset makes console work (patch 1 and 4) for an AT91SAM9G20 board connected to xhci_hcd and does some cleanup.
Tested with "console=ttyS0,115200n8 console=ttyGS0" on kernel command line and following inittab:
console::respawn:/sbin/getty -L 115200 ttyS0 vt100 console::respawn:/sbin/getty -L 115200 ttyGS0 vt100

Quoted 7 lines
Quoted 2 lines

However there is an issue remaining:
Disconnect triggers WARN_ON in gs_close:

[...]

Hi,

Can you try a patchset I sent some time ago [1] if it fixes your case?

You can pull it from branch usb-console on https://rere.qmqm.pl/git/linux
(rebased on current usb-next tree).

Best Regards,
Michał Mirosław

#8 lore raw
Quoted 11 lines · Michał Mirosław
Quoted 10 lines · Michał Mirosław
Quoted 9 lines · Ladislav Michl

Following patchset makes console work (patch 1 and 4) for an AT91SAM9G20 board connected to xhci_hcd and does some cleanup.
Tested with "console=ttyS0,115200n8 console=ttyGS0" on kernel command line and following inittab:
console::respawn:/sbin/getty -L 115200 ttyS0 vt100 console::respawn:/sbin/getty -L 115200 ttyGS0 vt100

However there is an issue remaining:
Disconnect triggers WARN_ON in gs_close:

Quoted 8 lines
Quoted 5 lines

[...]

Hi,

Can you try a patchset I sent some time ago [1] if it fixes your case?

You can pull it from branch usb-console on https://rere.qmqm.pl/git/linux
(rebased on current usb-next tree).

Thank you! I has just in the middle of pulling patches from mail archive.

Your first two patches are definitely worth applying and I'll add my Tested-by after more testing. 3/5 is unfortunately "I do not care" as not having use case, 4/5 is something I do not see use case either, but I'm fine with that, except for #ifdefs, but 5/5 should be solved the other way around - just wait until everything is written out, of course only if USB is connected. That's what serial consoles does and that's what I expect from kernel console.

Your patch set does not attempt to solve tty write to unconnected USB and WARN_ON on gs_close, so I'll wait for comments to those two issues and redo fix on top of "reimplement console support" as getting rid of the kthread is definitely superior to fixing it :)

Best regards,
ladis

author Tested-by
#9 lore raw
In reply to Michał Mirosław

Hi Michał,

Quoted 8 lines · Michał Mirosław
Quoted 7 lines · Ladislav Michl

Following patchset makes console work (patch 1 and 4) for an AT91SAM9G20 board connected to xhci_hcd and does some cleanup.
Tested with "console=ttyS0,115200n8 console=ttyGS0" on kernel command line and following inittab:
console::respawn:/sbin/getty -L 115200 ttyS0 vt100 console::respawn:/sbin/getty -L 115200 ttyGS0 vt100

Quoted 7 lines
Quoted 2 lines

However there is an issue remaining:
Disconnect triggers WARN_ON in gs_close:

[...]

Hi,

Can you try a patchset I sent some time ago [1] if it fixes your case?

your "usb: gadget: u_serial: add missing port entry locking" and "usb: gadget: u_serial: reimplement console support" fixes every single problem I had with the driver (including that WARN_ON on gs_close), except starving console write.
Fortunately my patch "usb: gadget: u_serial: Fix starving write" applies on top of your first two patches.

So here's my

and also one nit for your second patch (there are people running cocci scripts, so save them some time ;-)):

drivers/usb/gadget/function/u_serial.c +3 -2
@@ -931,8 +931,8 @@ static void gs_console_complete_out(struct usb_ep *ep, struct usb_request *req)
 static void __gs_console_push(struct gs_console *cons)
 {
 	struct usb_request *req = cons->req;
-	struct usb_ep *ep = cons->console.data;
-	size_t size = 0;
+	struct usb_ep *ep;
+	size_t size;
 
 	if (!req)
 		return;	/* disconnected */
@@ -940,6 +940,7 @@ static void __gs_console_push(struct gs_console *cons)
 	if (req->length)
 		return;	/* busy */
 
+	ep = cons->console.data;
 	size = kfifo_out(&cons->buf, req->buf, ep->maxpacket);
 	if (!size)
 		return;

I'm dropping all my patches except the fourth one in favour of your solution as it is clearly superior. Care to resend it possibly with above fix and said fourth patch as a single serie?

Thank you,
ladis