On Thu, Feb 09, 2017 at 08:31:21AM +0900, Minchan Kim wrote: > > Today, I did zram-lz4 performance test with fio in current mmotm and > found it makes regression about 20%. >
This may or may not be the cause of the specific regression you're observing, but I just noticed that the proposed patch drops a lot of FORCEINLINE annotations from upstream LZ4. The FORCEINLINE's are there for a reason, especially for the main decompression and compression functions which are basically "templates" that take in different sets of constant parameters, and should be left in. We should #define FORCEINLINE to __always_inline somewhere, or just do a s/FORCEINLINE/__always_inline/g. Note that the upstream LZ4 code is very carefully optimized, so we should not, in general, be changing things like when functions are force-inlined, what the hash table size is, etc. [Also, for some reason linux-crypto is apparently still not receiving patch 1/5 in the series. It's missing from the linux-crypto archive at http://www.spinics.net/lists/linux-crypto/, so it's not just me.] Thanks! Eric
