Re: [PATCH] dwc3: gadget: Introduce dwc3_endpoint_xfer_xxx() to check endpoint type

From: Felipe Balbi
Date: Thu May 26 2016 - 03:52:15 EST



Hi,

Baolin Wang <baolin.wang@xxxxxxxxxx> writes:
> Hi Felipe,
>
> On 26 May 2016 at 14:22, Felipe Balbi <balbi@xxxxxxxxxx> wrote:
>>
>> Hi,
>>
>> Baolin Wang <baolin.wang@xxxxxxxxxx> writes:
>>> When handling the endpoint interrupt handler, it maybe disable the endpoint
>>> from another core user to set the USB endpoint descriptor pointor to be NULL
>>> while issuing usb_gadget_giveback_request() function to release lock. So it
>>> will be one bug to check the endpoint type by usb_endpoint_xfer_xxx() APIs with
>>> one NULL USB endpoint descriptor.
>>
>> too complex, Baolin :-) Can you see if this helps:
>>
>> https://git.kernel.org/cgit/linux/kernel/git/balbi/usb.git/commit/?id=88bf752cfb55e57a78e27c931c9fef63240c739a
>>
>> The only situation when that can happen is while we drop our lock on
>> dwc3_gadget_giveback().
>
> OK, But line 1974 and line 2025 as below may be at risk? So I think
> can we have a common place to solve the problem in case
> usb_endpoint_xfer_xxx() APIs are issued at this risk? What do you
> think? Thanks.
>
> 1956 static int dwc3_cleanup_done_reqs(struct dwc3 *dwc, struct dwc3_ep *dep,
> 1957 const struct dwc3_event_depevt *event, int status)
> 1958 {
> 1959 struct dwc3_request *req;
> 1960 struct dwc3_trb *trb;
> 1961 unsigned int slot;
> 1962 unsigned int i;
> 1963 int ret;
> 1964
> 1965 do {
> 1966 req = next_request(&dep->req_queued);
> 1967 if (WARN_ON_ONCE(!req))
> 1968 return 1;
> 1969
> 1970 i = 0;
> 1971 do {
> 1972 slot = req->start_slot + i;
> 1973 if ((slot == DWC3_TRB_NUM - 1) &&
> 1974 usb_endpoint_xfer_isoc(dep->endpoint.desc))

this is executed still with locks held.

> 1975 slot++;
> 1976 slot %= DWC3_TRB_NUM;
> 1977 trb = &dep->trb_pool[slot];
> 1978
> 1979 ret = __dwc3_cleanup_done_trbs(dwc, dep, req, trb,
> 1980 event, status);
> 1981 if (ret)
> 1982 break;
> 1983 } while (++i < req->request.num_mapped_sgs);
> 1984
> 1985 dwc3_gadget_giveback(dep, req, status);

the problem can only show up after this call

> 1986
> 1987 if (ret)
> 1988 break;
> 1989 } while (1);
> .......
>
> 2011 static void dwc3_endpoint_transfer_complete(struct dwc3 *dwc,
> 2012 struct dwc3_ep *dep, const struct dwc3_event_depevt *event)
> 2013 {
> 2014 unsigned status = 0;
> 2015 int clean_busy;
> 2016 u32 is_xfer_complete;
> 2017
> 2018 is_xfer_complete = (event->endpoint_event ==
> DWC3_DEPEVT_XFERCOMPLETE);
> 2019
> 2020 if (event->status & DEPEVT_STATUS_BUSERR)
> 2021 status = -ECONNRESET;
> 2022
> 2023 clean_busy = dwc3_cleanup_done_reqs(dwc, dep, event, status);
> 2024 if (clean_busy && (is_xfer_complete ||
> 2025

note the patch I linked you. This is where I added a bail out if the
descriptor is NULL. Here's the patch for reference:

commit 4d100e870616ccd30cf29abf21d437ec746e1f54
Author: Felipe Balbi <felipe.balbi@xxxxxxxxxxxxxxx>
Date: Wed May 18 12:37:21 2016 +0300

usb: dwc3: gadget: fix for possible endpoint disable race

when we call dwc3_gadget_giveback(), we end up
releasing our controller's lock. Another thread
could get scheduled and disable the endpoint,
subsequently setting dep->endpoint.desc to NULL.

In that case, we would end up dereferencing a NULL
pointer which would result in a Kernel Oops. Let's
avoid the problem by simply returning early if we
have a NULL descriptor.

Signed-off-by: Felipe Balbi <felipe.balbi@xxxxxxxxxxxxxxx>

diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
index f31a59cd5162..3d0573c74b13 100644
--- a/drivers/usb/dwc3/gadget.c
+++ b/drivers/usb/dwc3/gadget.c
@@ -2019,6 +2019,14 @@ static int dwc3_cleanup_done_reqs(struct dwc3 *dwc, struct dwc3_ep *dep,
break;
} while (1);

+ /*
+ * Our endpoint might get disabled by another thread during
+ * dwc3_gadget_giveback(). If that happens, we're just gonna return 1
+ * early on so DWC3_EP_BUSY flag gets cleared
+ */
+ if (!dep->endpoint.desc)
+ return 1;
+
if (usb_endpoint_xfer_isoc(dep->endpoint.desc) &&
list_empty(&dep->started_list)) {
if (list_empty(&dep->pending_list)) {
@@ -2085,6 +2093,14 @@ static void dwc3_endpoint_transfer_complete(struct dwc3 *dwc,
dwc->u1u2 = 0;
}

+ /*
+ * Our endpoint might get disabled by another thread during
+ * dwc3_gadget_giveback(). If that happens, we're just gonna return 1
+ * early on so DWC3_EP_BUSY flag gets cleared
+ */
+ if (!dep->endpoint.desc)
+ return;
+
if (!usb_endpoint_xfer_isoc(dep->endpoint.desc)) {
int ret;

Also note that the usb_endpoint_xfer_isoc() call on line 2067 of
gadget.c (as in my testing/next from today) won't even get executed, so
we're safe there.

--
balbi

Attachment: signature.asc
Description: PGP signature