Eric Blake wrote: > According to Jim Meyering on 10/10/2009 1:35 PM: >> Those look fine. Thanks! >> I compared the performance of your modified touch >> on a tmpfs file system on a system running rawhide, >> creating 100,000 empty files (names 1..100,000), and found >> no significant difference when running touch with no options. > > Try it again with 'touch -a' or 'touch -m'. The savings come when > changing only one of the two timestamps: previously, touch had to stat() > to learn the time to preserve, as well as gettime() to learn what time to > change; but the new code can rely on the kernel to do both of those actions.
Same result. There is no significant difference on GNU/Linux. I do confirm (via strace) that the fstat syscall is gone, but figure we've merely moved the burden into futimens.
