Re: [PATCH 2/2] git reflog expire

[Linus Torvalds <torvalds@osdl.org>]

On Tue, 19 Dec 2006, Shawn Pearce wrote:
> 
> Smaller mmap units may help on some larger projects. For example
> 'git whatchanged' shouldn't need to setup page table entries for
> the entire pack file when only the front and middle are needed to
> obtain the data for the most recent 25 commits, before the user
> kills the pager.

At least on Linux, we do NOT set up page table entries unnecessarily. The 
cost of a mmap() is largely independent of the size of the mmap, and the 
costs will be on page fault time (which is obviously O(n) in number of 
pages you need).

The same is _largely_ true of munmap() too. There's a small cost to large 
mmaps that were only sparsely accessed being unmapped (we have to look at 
the page tables closely, even if they end up being empty), but quite 
frankly, it's not going to be all that noticeable.

So mmap() costs end up largely being one (fairly small) fixed cost, plus a 
(fairly small) cost for each page beign demand-mapped in. 

Of course, Linux is just _better_ than the competition, so things that 
don't help Linux might still help other systems. Linux does both page 
faults and the mmap system call well, other systems will generally be an 
order of magnitude worse in both. And as we've seen, it can be more.

(Things that are really fast on Linux: VM manipulation, and filename path 
lookup. Those are both ops that Linux _really_ shines on. It may not sound 
like much, but when you do millions of them, it's the difference between 
seconds and hours).