On Thu, Mar 6, 2014 at 10:36 AM, Ingo Schwarze <schwa...@usta.de> wrote: > back in November 2013, following an idea by guenther@, > i cooked up another optimization for seekdir(3), > then failed to send out the patch. So here it is. ... > With this patch, rewinddir(3) changes from always being the worst > case to often being the best case (now always the best case for > directories with less than about 500 entries, or for rewinding > before reading less than about 500 entries), no other case becomes > any worse, and some other, less important edge cases improve, too. > > What the patch does is: > * Let opendir(3) save the cookie of the first entry of the directory. > * Let seekdir(3) use that saved cookie to move the internal pointers > to that entry when asked to do so, for example on rewinddir(3). > * Let seekdir(3) update that saved cookie when the buffer is exhausted > and when consequently, an lseek(2) had to be done. > * Let seekdir(3) invalidate the buffer by setting the pointer after > its end, instead of to its beginning. > * Have readdir(3) refrain from invalidating the buffer > when asked to return the first element from it. > > This comes with only one price: Adding another off_t to the > opaque struct _dirdesc (= DIR), which already contains another > off_t, two long, two int, and two pointers. That's not much > growth, especially since no sane code will allocate large > numbers of DIRs. > > As explained last year, this patch reduced typical execution > times of rewinddir(3) on my notebook for about 100 microseconds > to about 0.05 microseconds. > > The /usr/src/regress/lib/libc/telldir/ test suite is still happy, > and so are some other, manual tests i have done.
Looks good to me, with some nice simplifications thrown in. > Am i right that, the type DIR being opaque, this patch doesn't > require any libc bump? Right. > Comments, OKs? ok guenther@
