On Mon, Dec 18, 2017 at 10:33:00AM +0800, Wei Wang wrote: > > My only qualm is that I've been considering optimising the memory > > consumption when an entire 1024-bit chunk is full; instead of keeping a > > pointer to a 128-byte entry full of ones, store a special value in the > > radix tree which means "every bit is set". > > > > The downside is that we then have to pass GFP flags to xbit_clear() and > > xbit_zero(), and they can fail. It's not clear to me whether that's a > > good tradeoff. > > Yes, this will sacrifice performance. In many usages, users may set bits one > by one, and each time when a bit is set, it needs to scan the whole > ida_bitmap to see if all other bits are set, if so, it can free the > ida_bitmap. I think this extra scanning of the ida_bitmap would add a lot > overhead. Not a huge amount of overhead. An ida_bitmap is only two cachelines, and the loop is simply 'check each word against ~0ul', so up to 16 load/test/loop instructions. Plus we have to do that anyway to maintain the free tag for IDAs. > > But I need to get the XArray (which replaces the radix tree) finished first. > > OK. It seems the new implementation wouldn't be done shortly. > Other parts of this patch series are close to the end of review, and we hope > to make some progress soon. Would it be acceptable that we continue with the > basic xb_ implementation (e.g. as xbitmap 1.0) for this patch series? and > xbit_ implementation can come as xbitmap 2.0 in the future? Yes, absolutely, I don't want to hold you up behind the XArray.
