Am 17.07.2018 um 10:30 schrieb Michel Dänzer: > On 2018-07-17 10:19 AM, Christian König wrote: >> Am 17.07.2018 um 10:03 schrieb Michel Dänzer: >>> On 2018-07-17 09:59 AM, Christian König wrote: >>>> Am 17.07.2018 um 09:46 schrieb Michel Dänzer: >>>>> On 2018-07-17 09:33 AM, Christian König wrote: >>>>>> Am 17.07.2018 um 09:26 schrieb Michel Dänzer: >>>>>> [SNIP] >>>>> All that should be needed is one struct list_head per BO, 16 bytes on >>>>> 64-bit. >>>> +malloc overhead and that for *every* BO the application/driver >>>> allocated. >>> The struct list_head can be stored in struct amdgpu_bo, no additional >>> malloc necessary. >> Well that sounds we are not talking about the same code, do we? >> >> IIRC the hashtable implementation in libdrm is using an ever growing >> array for the BOs and *NOT* a linked list. > So let's use something more suitable, e.g.: > > An array of 2^n struct list_head in struct amdgpu_device for the hash > buckets. The BO's handle is hashed to the bucket number > > handle & (2^n - 1) > > and linked in there via struct list_head in struct amdgpu_bo. > amdgpu_bo_alloc and amdgpu_create_bo_from_user_mem add the handle at the > end of the list, amdgpu_bo_import adds it at or moves it to the beginning. Yeah, that would certainly reduce the problem quite a bit and would allow us to get rid of the util_hash* implementation which to me always seemed to be a bit overkill. I actually don't see a reason why amdgpu_create_bo_from_user_mem() should add the handle at all, those BOs are not exportable. Christian.
