Re: [PATCH 2/3] mmap(win32): avoid copy-on-write when it is unnecessary

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Apr 27, 2016, 06:43 UTC
Message-ID: <alpine.DEB.2.20.1604270834440.2896@virtualbox>
In-Reply-To: <571FB923.9040808@kdbg.org>

Hi Hannes,

On Tue, 26 Apr 2016, Johannes Sixt wrote:

Show 29 quoted lines

> Am 22.04.2016 um 16:31 schrieb Johannes Schindelin:
> > Often we are mmap()ing read-only. In those cases, it is wasteful to map in
> > copy-on-write mode. Even worse: it can cause errors where we run out of
> > space in the page file.
> >
> > So let's be extra careful to map files in read-only mode whenever
> > possible.
> >
> > Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
> > ---
> >   compat/win32mmap.c | 5 +++--
> >   1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/compat/win32mmap.c b/compat/win32mmap.c
> > index 3a39f0f..b836169 100644
> > --- a/compat/win32mmap.c
> > +++ b/compat/win32mmap.c
> > @@ -22,14 +22,15 @@ void *git_mmap(void *start, size_t length, int prot, int
> > flags, int fd, off_t of
> >     die("Invalid usage of mmap when built with USE_WIN32_MMAP");
> >
> >   	hmap = CreateFileMapping((HANDLE)_get_osfhandle(fd), NULL,
> > -		PAGE_WRITECOPY, 0, 0, NULL);
> > +		prot == PROT_READ ? PAGE_READONLY : PAGE_WRITECOPY, 0, 0,
> > NULL);
> 
> As long as we use this implementation with MAP_PRIVATE, PAGE_WRITECOPY
> is the right setting. Should we insert a check for MAP_PRIVATE to catch
> unexpected use-cases (think of the index-helper daemon effort)?

I agree, we should have such a check. The line above the `die("Invalid usage ...")` that you can read as first line in above-quoted hunk reads:

	if (!(flags & MAP_PRIVATE))

So I think we're fine :-)

And yes, I am worrying about the index-helper support, too: I need this myself, so I will have to make mmap() work for that use case, too. But that is a story for another day ;-)

Show 10 quoted lines

> >    if (!hmap) {
> >     errno = EINVAL;
> >     return MAP_FAILED;
> >    }
> >
> > -	temp = MapViewOfFileEx(hmap, FILE_MAP_COPY, h, l, length, start);
> > +	temp = MapViewOfFileEx(hmap, prot == PROT_READ ?
> > +			FILE_MAP_READ : FILE_MAP_COPY, h, l, length, start);
> 
> Same here: FILE_MAP_COPY is the right choice for MAP_SHARED mmaps.

I agree ;-)

Show 5 quoted lines

> >    if (!CloseHandle(hmap))
> >     warning("unable to close file mapping handle");
> >
> 
> Except for these mental notes, I've no comments on this series. Looks good.

Thanks for reviewing!

Do you think we should add a note to the commit message that we'll have to do something about this when MAP_PRIVATE is not the only way mmap() will be used?

I am torn: on the one hand, it is the appropriate thing to do, on the other hand, it is easy to forget such notes in commit messages. On the third hand, I hope to find time to work on the index-helper this week, meaning that I will still know about this when touching compat/win32mmap.c.

So maybe I can just leave things as are here, and focus on the index-helper?

Ciao, Dscho