TDA4VM: Access to shared mem through /dev/mem seems slow

Jan Hieber

Part Number: TDA4VM

I have a program running on A72 under Linux which exchanges data with the C7 DSP.

For this purpose I mapped a part of shared memory via /dev/mem and mmap().

Now when I have code on A72 operating on this shared memory it is somehow much slower compared to using local data.

My question is: Is this something with caches? How can I work around this? Is /dev/mem bad? Do I need the ION mem allocator maybe?

Thanks!
Br Jan

over 5 years ago

0 Keerthy J over 5 years ago

TI__Guru**** 162670 points

Hi Jan,

Can you share the exact steps or code you added to reproduce this behavior and also the SDK that you are using?

- Keerthy

0 Jan Hieber over 5 years ago in reply to Keerthy J

Intellectual 420 points

Actually I don't use the SDK.

I use the TI Linux kernel (same as from yocto/SDK) and a custom Linux.

Accessing shared memory is like this:

m_fd = open("/dev/mem", O_RDWR | O_SYNC);
void* virtAddr = mmap(nullptr, size, (PROT_READ | PROT_WRITE), MAP_SHARED, m_fd, static_cast<off_t>(_addr));

I compiled the Kernel with CONFIG_STRICT_DEVMEM and CONFIG_IO_STRICT_DEVMEM turned off.

I saw that you use the ION allocator in the SDK and wanted to ask why. What exactly is the downside of using /dev/mem?

0 Keerthy J over 5 years ago in reply to Jan Hieber

TI__Guru**** 162670 points

Hi Jan,

The recommendation from TI is use /dev/mem only for debug & not enabled on production systems.

2020 kernels do not use ION as well and the suggestion is to move to using dma-buf heaps.

Hope this answers your question on the slowness of ./dev/mem.

Best Regards,
Keerthy

0 Keerthy J over 5 years ago in reply to Keerthy J

TI__Guru**** 162670 points

Hi Jan,

If you have no further questions could you please resolve this thread?

- Keerthy

Processors

Processors forum

TDA4VM: Access to shared mem through /dev/mem seems slow