lkml.org 
[lkml]   [2021]   [Jul]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] usb: dwc3: gadget: Use list_replace_init() before traversing lists
On Thu, Jul 29, 2021 at 11:09:57AM +0300, Felipe Balbi wrote:
>
> Hi,
>
> Wesley Cheng <wcheng@codeaurora.org> writes:
>
> > The list_for_each_entry_safe() macro saves the current item (n) and
> > the item after (n+1), so that n can be safely removed without
> > corrupting the list. However, when traversing the list and removing
> > items using gadget giveback, the DWC3 lock is briefly released,
> > allowing other routines to execute. There is a situation where, while
> > items are being removed from the cancelled_list using
> > dwc3_gadget_ep_cleanup_cancelled_requests(), the pullup disable
> > routine is running in parallel (due to UDC unbind). As the cleanup
> > routine removes n, and the pullup disable removes n+1, once the
> > cleanup retakes the DWC3 lock, it references a request who was already
> > removed/handled. With list debug enabled, this leads to a panic.
> > Ensure all instances of the macro are replaced where gadget giveback
> > is used.
> >
> > Example call stack:
> >
> > Thread#1:
> > __dwc3_gadget_ep_set_halt() - CLEAR HALT
> > -> dwc3_gadget_ep_cleanup_cancelled_requests()
> > ->list_for_each_entry_safe()
> > ->dwc3_gadget_giveback(n)
> > ->dwc3_gadget_del_and_unmap_request()- n deleted[cancelled_list]
> > ->spin_unlock
> > ->Thread#2 executes
> > ...
> > ->dwc3_gadget_giveback(n+1)
> > ->Already removed!
> >
> > Thread#2:
> > dwc3_gadget_pullup()
> > ->waiting for dwc3 spin_lock
> > ...
> > ->Thread#1 released lock
> > ->dwc3_stop_active_transfers()
> > ->dwc3_remove_requests()
> > ->fetches n+1 item from cancelled_list (n removed by Thread#1)
> > ->dwc3_gadget_giveback()
> > ->dwc3_gadget_del_and_unmap_request()- n+1
> > deleted[cancelled_list]
> > ->spin_unlock
> >
> > Fix this condition by utilizing list_replace_init(), and traversing
> > through a local copy of the current elements in the endpoint lists.
> > This will also set the parent list as empty, so if another thread is
> > also looping through the list, it will be empty on the next iteration.
> >
> > Fixes: d4f1afe5e896 ("usb: dwc3: gadget: move requests to cancelled_list")
> > Signed-off-by: Wesley Cheng <wcheng@codeaurora.org>
> >
> > ---
> > Previous patchset:
> > https://lore.kernel.org/linux-usb/1620716636-12422-1-git-send-email-wcheng@codeaurora.org/
> > ---
> > drivers/usb/dwc3/gadget.c | 18 ++++++++++++++++--
> > 1 file changed, 16 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/usb/dwc3/gadget.c b/drivers/usb/dwc3/gadget.c
> > index a29a4ca..3ce6ed9 100644
> > --- a/drivers/usb/dwc3/gadget.c
> > +++ b/drivers/usb/dwc3/gadget.c
> > @@ -1926,9 +1926,13 @@ static void dwc3_gadget_ep_cleanup_cancelled_requests(struct dwc3_ep *dep)
> > {
> > struct dwc3_request *req;
> > struct dwc3_request *tmp;
> > + struct list_head local;
> > struct dwc3 *dwc = dep->dwc;
> >
> > - list_for_each_entry_safe(req, tmp, &dep->cancelled_list, list) {
> > +restart:
> > + list_replace_init(&dep->cancelled_list, &local);
>
> hmm, if the lock is held and IRQs disabled when this runs, then no other
> threads will be able to append requests to the list which makes the
> "restart" label unnecessary, no?

As Wesley pointed out, the lock can be released during giveback and
requests can be added to the cancelled_list at that time.

On the other hand, if that happens, do you need to process those
requests in this function call? Will another cleanup iteration take
care of them later? (I don't know the driver well enough to answer
this.) If it will, you may not need to restart anything.

> I wonder if we should release the lock and reenable interrupts after
> replacing the head. The problem is that
> dwc3_gadget_ep_cleanup_cancelled_requests() can run from the IRQ
> handler.
>
> Alan, could you provide your insight here? Do you think we should defer
> this to a low priority tasklet or something along those lines?

I don't see why anything like that would be necessary. Giving back
cancelled requests isn't important enough to warrant special treatment.

An alternative approach, used by some other drivers, is to stick with
list_for_each_entry_safe as in the existing code, but go back to the
restart label immediately each time the lock is released and reacquired.

Also, if this loop always removes the entry it is processing from the
list (I don't know whether it does this), you don't have to use
list_for_each_entry_safe. You can simply use list_first_entry.

Alan Stern

> > + list_for_each_entry_safe(req, tmp, &local, list) {
> > dwc3_gadget_ep_skip_trbs(dep, req);
> > switch (req->status) {
> > case DWC3_REQUEST_STATUS_DISCONNECTED:
>
>
> --
> balbi

\
 
 \ /
  Last update: 2021-07-29 16:20    [W:0.121 / U:1.384 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site