On Wed, Dec 29, 2010 at 11:42 AM, Felipe Contreras <felipe.contreras@xxxxxxxxx> wrote: > On Tue, Dec 28, 2010 at 4:48 PM, Ohad Ben-Cohen <ohad@xxxxxxxxxx> wrote: >> On Tue, Dec 28, 2010 at 2:12 PM, Felipe Contreras >> <felipe.contreras@xxxxxxxxx> wrote: >>> On Tue, Dec 28, 2010 at 12:56 PM, Ohad Ben-Cohen <ohad@xxxxxxxxxx> wrote: >>>> I still don't know how exactly you triggered the bug: is gst-dsp >>>> multithreaded ? and one of its threads invoked proc_un_map() while >>>> another thread called proc_begin_dma() ? >>> >>> I haven't investigated why that happens >> >> Btw, I still think you should look into this. >> >> The kernel panic will be solved, but you may still have a race there >> that can lead to data corruption: if proc_un_map will be fast enough, >> it will acquire the proc_lock mutex before proc_begin_dma(), and then >> you will miss a cache operation. > > Aquiring the lock is the first thing done; if proc_un_map() aquires > the lock first, it's because it was run first Not true. Again, we have two threads: T1 - called proc_begin_dma() T2 - called proc_un_map() Let's say T1 called proc_begin_dma() first, but it still didn't acquire the lock. At this point the scheduler let T2 run. It calls proc_un_map() and acquire the lock, and free all map_obj's. Then the scheduler let T1 continue to execute, but it can't find any map_obj, and the required cache operation is not performed. Of course the kernel has nothing to do with this, and don't care really. That's why I said it's a user space race - you need to make sure T2 will not call proc_un_map() before T1 (or any other thread you have) completed executing all DMA operations (that you care about). >, and thus a problem for > user-space. If user-space wants the cache operation, it must run > proc_begin_dma() first, there's nothing kernel-space can do to fix > that. > > -- > Felipe Contreras > -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
