DM6446 Preview Engine hangs the whole device.

Could you provide some more details, about the hang. You mention that no repsonse on UART. Can you enable logs to see the last known state and also some clues as to which software module causes the hang.

Hi Varada,

Thanks for the response. I'm working on this case now.  The test utility for previewer is mainly just an endless loop to call ioctl PREV_PREVIEW function:

while(1) {
        /* Call PREV_PREVIEW ioctl to do previewing */
        ret = ioctl(prevfd, PREV_PREVIEW, &convert);
}

while the udp_send test utility is just a generic program to generate UDP traffic.  When it happens, we lost control to the unit.  Our initial through is there might be some resource conflict in the kernel,  as it happens only when both test utilities are running at high speed.

Please let us know if you need more information.

David

I put some printk into kernel davinci_previewer.c previewer_ioctl() function:

    case PREV_PREVIEW:
        /* call preview function to do preview */
        if (copy_from_user(&conv, (struct prev_convert *)arg,
                   sizeof(struct prev_convert)))
            return -EFAULT;

        printk("before down\n");
        down_interruptible(&(device->sem));
        printk("after down\n");
        ret = preview(device, &conv);
        printk("before up\n");
        up(&(device->sem));
        printk("after up\n");
        break;

When the problem happens, it randomly stops at any position before or after down_interruptible()/up() functions and any position inside the preview() function.  Do you have any chance to reproduce the problem?  Thanks.

David - Thanks for the info. 

I have looped in other experts at TI.  They will get back with suggestions.

Regards

Varada

Varada,

Could you please update us on where this is sitting in the TI Expert's queue.

thanks,

Doug

I will follow up and let you know in a couple of days. Thanks

Varada

Hello All,

 I am actively looking into this issue and will inform about the findings/fix shortly.

-Manju

Would you please try adding the following in your preview_example.c  code?

..........

        if (ret < 0) {
                printf("\n error in setting parameters");
                goto out;
        }

        int prev_expand = 0x2c0;  <--//or any other value ( from 0 to 0x3FF) for tuning the preview data transfer throttling
        ret = ioctl(prevfd,PREV_SET_EXP,&prev_expand);
        if (ret < 0) {
               printf("\n error in setting expand parameters");
               goto out;
        }

.......
        /* request one input buffer of size height*width */
        reqbufs.buf_type = PREV_BUF_IN;
        reqbufs.size = HEIGHT * WIDTH;

Background:
Reg VPSS_SDR_REQ_EXP

Table 203. SDRAM Non-Real-Time Read Request Expand Register (SDR_REQ_EXP)
Field Descriptions
Bit
Field
Value
Description
31-30 Reserved 0 29-20 PRV_EXP 0-3FFh Preview read request expand. Delay to allow between consecutive read requests from the preview
                                     module. Units are in VPSS clock cycles (153/198 MHz in Normal/Turbo modes). Since the VPSS
                                    DMA priority is typically set to the highest for real-time requirements, this is for spreading any
                                   non-real-time reads with respect to the other traffic in the system. This minimizes the potential of
locking out other requests for the duration of a frame being read from DDR/SDR.

Although I still see the crash, yesterday night it appeared to take a longer time. Please try on your end.

-Manju

Issue still under debug.

Will keep the forum posted.

-Manju

Thanks for the update.  I tested the patch with different values and unfortunately it could still be reproduced less than one minute.

I might have found the solution. Will update with the fix after due testing.

-Manju

Please try the changes in the attachment and let me know the results7356.preview_fix.txt.

The file is the diff of the default driver file deletions(marked <) and new changes. (>)

Mainly:

1. Remove:

 struct prev_params params;
 struct prev_convert conv;

from prev_ioctl and directly use the arg.

2. Do not do copy_from_user

Thanks and Regards,

-Manju

Thanks a lot for the patch.  The initial test looks good and we are integrating it with production code and will do more load tests.  By the way, there is a small change need to be done in the patch:

<
<     /* indicate the completion ofr frame processing */
<     if (prevdevice)
<         complete(&(prevdevice->wfc));
<                              ^^- should be &(prevdevice.wfc)