Hi Trilok: > > I was just going this dspbridge code, and I was surprised to > find layers for un-necessary things, where we could have > simply called the linux kernel provided function. Here goes the list: > > 1. There is un-necessary "linux" directory inside mpu_driver > source codes. e.g src/osal/linux etc. I hope we are writing > this code for Linux kernel only. > 2. src/osal/linux > - As the name suggest for directory it is OS Abstraction > layer for Linux. I don't see any need for this at all. In > Linux kernel, we follow approach of directly calling the > provided functions and not to further wrap it until necessary. > > src/osal/list.c > > - This whole file tries implement doubly linked list code, I > am sure you can get away this with list.h itself. I don't see > any need of further wrapping it. > > src/osal/kfile.c > > - Don't use KENREL_VERSION comparision check unless very much > necessary, always try to keep driver in sync in latest kernel > version instead. > - Again there are wrappers for file reading functions > provided by kernel. > - If this functions are only used for doff loader, then > moving it to outside kernel space, then we can get away with > this functions. > > src/osal/mem.c > > - I see here that you guys are maintaining the pool of memory > allocated with dma_alloc_coherent, you guys should have used > existing mempool apis instead, look at dspgw code for example. > > src/osal/isr.c > > - Not required, please call request_irq/free_irq directly > please. I also see lots of #ifdef LINUX and #ifndef LINUX, > does TI also target to build this code for WinCE and Symbian > like OS. If yes, I don't see how we are going to maintain > this along with community. > > src/osal/dpc.c - please call tasklet functions directly instead. > > src/pmgr/linux/common/dmm.c - One of the goals of bridge is easier portability to other RTOS. You are right about the additional layer to call into Linux OS. This was intended to port this to other OS'es easily. > > Please correct if my undestanding is in-correct for dmm.c - > Here you guys are trying to manage the virtual memory of > target (dsp/iva) as linked list of node, so that reserved > memory from ARM can be mapped to DSP and then can be used as > shared memory. If we are trying manage virtual memory in > Gigabytes , e.g 2GB for IVA1.0, then doubly linked list may > not be optimal for searching the free node, where we can map > this buffer. > > You guys should have used lib/bitmap.c functions for creating > bitmap, once the bitmap is created use > bitmap_find_free_region, which is very fast. For example of > this function grep in SH architecture code. > > Your understanding is correct. The initial implementation was done for 32MB virtual address. But the linked lists will have performance impact for larger buffer sizes. We are exploring options to change this to a lookup table implementation. We will also investigate the bitmap functions as suggested by you. Thanks Vijay. -- 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
