Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Anirudh Rayabharam <mail@anirudhrb.com>
To:     Felipe Balbi <balbi@kernel.org>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Alan Stern <stern@rowland.harvard.edu>,
        Colin Ian King <colin.king@canonical.com>,
        "Gustavo A. R. Silva" <gustavoars@kernel.org>,
        Jules Irenge <jbi.octave@gmail.com>,
        Bui Quang Minh <minhquangbui99@gmail.com>,
        Anirudh Rayabharam <mail@anirudhrb.com>,
        Andrey Konovalov <andreyknvl@gmail.com>
Cc:     linux-kernel-mentees@lists.linuxfoundation.org,
        syzbot+eb4674092e6cc8d9e0bd@syzkaller.appspotmail.com,
        linux-usb@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: [PATCH v2] usb: gadget: dummy_hcd: fix gpf in gadget_setup
Date:   Mon, 19 Apr 2021 09:07:08 +0530
Message-Id: <20210419033713.3021-1-mail@anirudhrb.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Fix a general protection fault reported by syzbot due to a race between
gadget_setup() and gadget_unbind() in raw_gadget.

The gadget core is supposed to guarantee that there won't be any more
callbacks to the gadget driver once the driver's unbind routine is
called. That guarantee is enforced in usb_gadget_remove_driver as
follows:

        usb_gadget_disconnect(udc->gadget);
        if (udc->gadget->irq)
                synchronize_irq(udc->gadget->irq);
        udc->driver->unbind(udc->gadget);
        usb_gadget_udc_stop(udc);

usb_gadget_disconnect turns off the pullup resistor, telling the host
that the gadget is no longer connected and preventing the transmission
of any more USB packets. Any packets that have already been received
are sure to processed by the UDC driver's interrupt handler by the time
synchronize_irq returns.

But this doesn't work with dummy_hcd, because dummy_hcd doesn't use
interrupts; it uses a timer instead. It does have code to emulate the
effect of synchronize_irq, but that code doesn't get invoked at the
right time -- it currently runs in usb_gadget_udc_stop, after the unbind
callback instead of before. Indeed, there's no way for
usb_gadget_remove_driver to invoke this code before the unbind callback.

To fix this, move the synchronize_irq() emulation code to dummy_pullup
so that it runs before unbind. Also, add a comment explaining why it is
necessary to have it there.

Suggested-by: Alan Stern <stern@rowland.harvard.edu>
Reported-by: syzbot+eb4674092e6cc8d9e0bd@syzkaller.appspotmail.com
Acked-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Anirudh Rayabharam <mail@anirudhrb.com>
---

Changes in v2:
Improvements in the comment as suggested by Alan Stern.	

v1: https://lore.kernel.org/lkml/20210417125212.6274-1-mail@anirudhrb.com/

---
 drivers/usb/gadget/udc/dummy_hcd.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)

diff --git a/drivers/usb/gadget/udc/dummy_hcd.c b/drivers/usb/gadget/udc/dummy_hcd.c
index ce24d4f28f2a..7db773c87379 100644
--- a/drivers/usb/gadget/udc/dummy_hcd.c
+++ b/drivers/usb/gadget/udc/dummy_hcd.c
@@ -903,6 +903,21 @@ static int dummy_pullup(struct usb_gadget *_gadget, int value)
 	spin_lock_irqsave(&dum->lock, flags);
 	dum->pullup = (value != 0);
 	set_link_state(dum_hcd);
+	if (value == 0) {
+		/*
+		 * Emulate synchronize_irq(): wait for callbacks to finish.
+		 * This seems to be the best place to emulate the call to
+		 * synchronize_irq() that's in usb_gadget_remove_driver().
+		 * Doing it in dummy_udc_stop() would be too late since it
+		 * is called after the unbind callback and unbind shouldn't
+		 * be invoked until all the other callbacks are finished.
+		 */
+		while (dum->callback_usage > 0) {
+			spin_unlock_irqrestore(&dum->lock, flags);
+			usleep_range(1000, 2000);
+			spin_lock_irqsave(&dum->lock, flags);
+		}
+	}
 	spin_unlock_irqrestore(&dum->lock, flags);
 
 	usb_hcd_poll_rh_status(dummy_hcd_to_hcd(dum_hcd));
@@ -1004,14 +1019,6 @@ static int dummy_udc_stop(struct usb_gadget *g)
 	spin_lock_irq(&dum->lock);
 	dum->ints_enabled = 0;
 	stop_activity(dum);
-
-	/* emulate synchronize_irq(): wait for callbacks to finish */
-	while (dum->callback_usage > 0) {
-		spin_unlock_irq(&dum->lock);
-		usleep_range(1000, 2000);
-		spin_lock_irq(&dum->lock);
-	}
-
 	dum->driver = NULL;
 	spin_unlock_irq(&dum->lock);
 
-- 
2.26.2