Messages in this thread | | | Date | Wed, 14 Jan 2009 17:22:39 -0800 | From | "Vineet Gupta" <> | Subject | dma_sync_api usage in drivers/ieee1394/sbp2.c |
| |
Hi,
While implementing dma map/sync APIs for a non-coherent DMA arch I ran into a use-case in IEEE1394 driver.
1. In sbp2util_create_command_orb_pool( ), dma_map_single(TO_DEVICE) is called for cmd->command_orb. Since this is ownership handover to driver, dcache wback #1 done for arch such as ARM]
2. In sbp2_create_command_orb ( ), it's ownership is transferred back to CPU so that driver can mod it by calling dma_sync_single_for_cpu ( TO_DEVICE ), [ this causes dcache wback #2]
3. at the end of same function, when finally driver is done with creating the cmd, it calls dma_sync_single_for_device (TO_DEVICE) [ dcache wback #3]
So dcache lines were flushed for the buffer 3 times whereas it was really required only once (maybe it was ok to do it at the time of map , but certainly for every sbp2_create_command_orb( ) it can be done only once instead of twice).
This is happening because ownership handover is being explicitly specified using _cpu( ) and _device( ) which has the upside that it makes the code really readable but the downside being it is not really efficient
As an optimization can sync_for_cpu (TO_DEVICE) be implemented as nop because that would REALLY mean cpu wants to own it. Similarly corollary case, sync_for_device (FROM_DEVICE) can be made a nop.
Do other people think the same too or am I missing something really fundamental here...
Thanks, Vineet
| |