On 01/03/2013 01:50 PM, Terje Bergström wrote:
On 03.01.2013 05:31, Mark Zhang wrote:
Sorry I didn't get it. Yes, in current design, you can pin all mem handles in one time but I haven't found anything related with "locking only once per submit".
My idea is:
- remove "job->addr_phys"
- assign "job->reloc_addr_phys" & "job->gather_addr_phys" separately
- In "pin_job_mem", just call "host1x_memmgr_pin_array_ids" twice to
fill the "reloc_addr_phy" & "gather_addr_phys".
Anything I misunderstood?
The current design uses CMA, which makes pinning basically a no-op. When we have IOMMU support, pinning requires calling into IOMMU. Updating SMMU tables requires locking, and probably maintenance before SMMU code also requires its own locked data tables. For example, preventing duplicate pinning might require a global table of handles.
Putting all of the handles in one table allows doing duplicate detection across cmdbuf and reloc tables. This allows pinning each buffer exactly once, which reduces number of calls to IOMMU.
Thanks Terje. Yes, I understand IOMMU will bring in more operations. But I'm not convinced that separating 2 arrays have lots of overheads than putting them into one array.
But it doesn't matter, after the IOMMU support version released, I can read this part of codes again. Let's talk about this at that time.
"host1x_cma_pin_array_ids" doesn't return negative value right now, so maybe you need to take a look at it.
True, and also a consequence of using CMA: pinning can never fail. With IOMMU, pinning can fail.
Got it. Agree.
Terje