I found that asm version of csum_partial_copy_from_user() introduced in e9e016815f264227b6260f77ca84f1c43cf8b9bd was less effective. For csum_partial_copy_from_user() case, "both_aligned" 8-word copy/sum loop block is skipped to handle LOAD failure properly. So we should iterate 4-word copy/sum block for that case, otherwize we will loop at ineffective "less_than_4units" block. Signed-off-by: Atsushi Nemoto <anemo@xxxxxxxxxxxxx> --- diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S index ec0744d..0d6e9ae 100644 --- a/arch/mips/lib/csum_partial.S +++ b/arch/mips/lib/csum_partial.S @@ -488,8 +488,11 @@ EXC( STORE t2, UNIT(2)(dst), s_exc) ADDC(sum, t2) EXC( STORE t3, UNIT(3)(dst), s_exc) ADDC(sum, t3) - beqz len, done + /* If we skipped both_aligned 8-word loop, iterate here */ + bnez AT, cleanup_both_aligned ADD dst, dst, 4*NBYTES + beqz len, done + nop less_than_4units: /* * rem = len % NBYTES
