lkml.org 
[lkml]   [2012]   [Jun]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] DMA: PL330: Fix racy mutex unlock
On Wed 13 Jun 2012 16:13:05 BST, Jassi Brar wrote:
> On 13 June 2012 19:37, Javi Merino<javi.merino@arm.com> wrote:
>> pl330_update() stores a pointer to the thrd->req that finished, which
>> contains a pointer to the corresponding pl330_req. This is done with
>> the pl330_lock held. Then, it iterates through the req_done list,
>> calling the callback for each of the requests that are done. The
>> problem is that the driver releases the lock before calling the
>> callback for each of the callbacks. pl330_submit_req() running in
>> another processor can then acquire the lock and insert another request
>> in one of the thrd->req that hasn't been processed yet, replacing the
>> pointer to pl330_req there. When the callback returns in
>> pl330_update() and the next rqdone is popped from the list, it
>> dereferences the pl330_req pointer to the just scheduled pl330_req,
>> instead of the one that has finished, calling pl330 with the wrong r.
>>
>> This patch fixes this by storing the pointer to pl330_req directly in
>> the list.
>>
> .....
>> @@ -1683,7 +1683,7 @@ static void pl330_dotask(unsigned long data)
>> /* Returns 1 if state was updated, 0 otherwise */
>> static int pl330_update(const struct pl330_info *pi)
>> {
>> - struct _pl330_req *rqdone;
>> + struct pl330_req *rqdone, *tmp;
>> struct pl330_dmac *pl330;
>> unsigned long flags;
>> void __iomem *regs;
>> @@ -1750,7 +1750,10 @@ static int pl330_update(const struct pl330_info *pi)
>> if (active == -1) /* Aborted */
>> continue;
>>
>> - rqdone =&thrd->req[active];
>> + /* Detach the req */
>> + rqdone = thrd->req[active].r;
>> + thrd->req[active].r = NULL;
>> +
> Doesn't this movement of "Detach the req" chunk effectively remain the
> same? Since that was already protected by the same lock. I thought I
> deliberately took care of that already.

Yes, but you release the lock before calling the first callback, so the
subsequent dereferences of rqdone->r are not protected by the lock.

> Do you see some real problem fixed by this patch? Info about that
> could help me better understand if I missed something here.

Ok, the description sucks. Let me try to describe it with the scenario
that failed:

Core 0:
- Two DMA transactions finish, in channels 0 and 1.
- pl330_update() is called, the "Event-Interrupt Status Register" (ES)
is 0x3.
- In the "for (ev = 0;..." loop
+ two pointers are stored in pl330->req_done:
pl330->channels[0]->req[0] and pl330->channels[1]->req[0]
- In the "while (!list_empty.." loop,
+ r = pl330->channels[0]->req[0]->r
+ Release the pl330_lock and call _callback()

Core 1:
- pl330_submit_req() for channel 1
- Grab the pl330_lock
- Hook a request in pl330->channels[1]->req[0]->r
- Release the pl330_lock

Core 0:
- _callback() returns
- Acquire the pl330_lock again
- second iteration of "while (!list_empty.." loop,
+ r = pl330->channels[1]->req[0]->r , but you get the r that has just
been scheduled, not the one that finished.

Hope it's now clearer,
Javi



\
 
 \ /
  Last update: 2012-06-13 21:01    [W:0.064 / U:0.280 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site