[PATCH 3 of 3] revlog: read index data using mmap

[Jun Wu]

Excerpts from Yuya Nishihara's message of 2016-10-05 19:50:33 +0900:
> On Tue, 4 Oct 2016 15:59:45 +0100, Jun Wu wrote:
> > # HG changeset patch
> > # User Jun Wu <quark at fb.com>
> > # Date 1475550701 -3600
> > #      Tue Oct 04 04:11:41 2016 +0100
> > # Node ID 99e7b0589b916b83d825882e907dd79082f4a2d7
> > # Parent  5ebdfd4ffbedbfe66b7a36cbd06b1e8e624ae7ad
> > # Available At https://bitbucket.org/quark-zju/hg-draft 
> > #              hg pull https://bitbucket.org/quark-zju/hg-draft  -r 99e7b0589b91
> > revlog: read index data using mmap
> 
> What happens if revlog is stripped? Using mmap would introduce more error
> modes we'll need to test.

If revlog is truncated, I think we will have trouble without mmap, too - the
pre-truncated index was loaded into memory while the revlog date can be
truncated. Reading data of a truncated revision would cause an error. But
the error is catchable.

Possible solutions I can think of are:

  - Change strip to copy the changelog index and rename it, instead of
    truncating the original file directly. Seems like a reasonable solution
    to me.
  - Turn off mmap by default but gate it with a config option.

> FWIW, I know Python crashes if mmapped file is truncated to be empty.
> 
> with open('foo', 'wb') as fp:
>     fp.write('a')
> fp = open('foo', 'rb')
> s = mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ)
> with open('foo', 'wb') as fp2:
>     pass  # truncate
> print(s[0])