Hi Felipe, >-----Original Message----- >From: Felipe Contreras [mailto:felipe.contreras@xxxxxxxxx] >Sent: Wednesday, March 24, 2010 12:15 PM >To: Guzman Lugo, Fernando >Cc: Hebbar, Shivananda; linux-omap@xxxxxxxxxxxxxxx; Hiroshi DOYU; Ameya >Palande; felipe.contreras@xxxxxxxxx >Subject: Re: [PATCH 2/2] DSPBRIDGE: DSP recovery feature > >On Tue, Mar 23, 2010 at 9:29 PM, Guzman Lugo, Fernando <x0095840@xxxxxx> >wrote: >> I have tested just using an application (base on bridged code) which is >registered for fatal events and once it receives an event it prints the >received event: > >Ok, I ran your test and it works, however, I tried to modify it to fit >what I'm doing in gst-dsp and it turns out it's very easy for this >code to reboot the device. I'm attaching the test. > >Also, I found the exact reason why my code fails but yours not. I do >something like this: > >while (!done) { > dsp_wait_for_events(&index); > > if (index == 0) { > /* node message */ > while (true) { > if (!dsp_node_get_message(&msg)) > break; > handle_message(msg); > } > } >} > >So, before your patch, the get_message() failed, and the next >wait_for_events() succeeded and returned the MMU fault. Now, the >get_message() fails, and so does the wait_for_events(). The think is that the now after fatal error and recovery process start all bridge ioctl's will fail with -EIO (that was the requirement when I did the code) @@ -496,6 +553,12 @@ static long bridge_ioctl(struct file *filp, unsigned int code, union Trapped_Args buf_in; DBC_REQUIRE(filp != NULL); +#ifdef CONFIG_BRIDGE_RECOVERY + if (recover) { + status = -EIO; + goto err; + } +#endif So you if you call dsp_node_get_message and then a fatal error happens, dsp_node_get_message will failed and if you what to call dsp_wait_for_events will fail because we are in recovery process (before it was succeeded because wait_for_events ioctl can go through and the MMUFault event was already signaled). To sum up: With this patch in order to be notified of fatal errors you need to register for those and call dsp_wait_for_events before the fatal error occurs. Otherwise dsp_wait_for_events will fail. So I think you need a dedicate thread to receive fatal errors. Regards, Fernando. > >The only way to make that code work with your patch is to remove the >inner while, so wait_for_events() and get_message() are always run one >after the other. > >That is breaking old behavior and should be fixed, right? > >-- >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
