On Mon, Jan 25, 2016 at 9:45 PM, James Bottomley <james.bottom...@hansenpartnership.com> wrote: > On Mon, 2016-01-25 at 21:28 +0200, Andy Shevchenko wrote: >> On Mon, Jan 25, 2016 at 8:56 PM, James Bottomley >> <james.bottom...@hansenpartnership.com> wrote: >> > On Mon, 2016-01-25 at 18:02 +0000, Elliott, Robert (Persistent >> > Memory) >> > wrote: >> >> > Using ffs leads to precision runaway >> >> How exactly?! > > Off by one. A size of 0xffffffffffffffff prints 18446744073709551615 B > rather than 20 GiB.
Because it's not a 20 GiB. It's exactly 20 GiB - 1 B. AFAIU, the intention was to show _exact_ size. >> It's good you are doing this better, but I still vote for __ffs64(), >> since it would be faster on binary units. > > Is speed of a start of day print a particular concern? If it's cheap to do, why not to do? >> Also, in one version I tried to convert couple of other users which >> are using only KM (in general whatever range it would be) units. Any >> ideas how to modify to support them? > > You mean units in odd increments of 6 digits (so K, M, T ...)? no. The > logarithmic reduction is done to the base of the unit increment (1000 > or 1024) so it doesn't really fit this case and it would be hard to > adjust because we don't have enough precision in the remainder. > However, unless there's a huge need to keep it, I'd just fit to the > closest 3 digit increment and then everything would work. KM case: K) if 1 MiB > value >= 0 — prints in KiB M) if ∞ > value >= 1 MiB — prints in MiB. -- With Best Regards, Andy Shevchenko
