On Wed, 21 Nov 2007, David Glasser wrote:
> On Nov 21, 2007 5:20 PM, Daniel Rall <dlr@collab.net> wrote:
> >
> > On Wed, 21 Nov 2007, Daniel L. Rall wrote:
> >
> > > On Tue, 20 Nov 2007, Daniel L. Rall wrote:
> > >
> > > > On Tue, 20 Nov 2007, David Glasser wrote:
> > > >
> > > > > On Nov 19, 2007 4:59 PM, Dan Christian <dchristian@google.com> wrote:
> > > > > > Argh! Not this again. openssl has similar problems (under serf).
> > > > > >
> > > > > > On Nov 19, 2007 1:46 PM, David Glasser <glasser@davidglasser.net> wrote:
> > > > > > > Apparently SQLite requires you to compile it with -DSQLITE_THREADSAFE
> > > > > > > to be at all threadsafe. (There's a way to check at runtime
> > > > > > > (sqlite3_threadsafe()) if this is so, but that's an experimental API
> > > > > > > that may vanish.)
> > > ...
> > > > > > > Should we add something to INSTALL telling people they must build
> > > > > > > against a threadsafe SQLite if they are going to be using a thready
> > > > > > > server?
> > > > > >
> > > > > > YES! And configure should verify this. Otherwise, it looks like
> > > > > > subversion is flaky.
> > > > >
> > > > > I don't think it can verify it unless we require 3.5.
> > > >
> > > > Yup. We could use sqlite3_threadsafe() if we can safely detect it runtime
> > > > for dynamic builds of Subversion (e.g. via dlopen()), or via configure checks
> > > > for static Subversion builds. While not ideal, this would allow us to detect
> > > > cases where newer versions of SQLite which do provide sqlite3_threadsafe()
> > > > were not compiled with -DSQLITE_THREADSAFE=1.
> > >
> > > Attached is a (mostly untested) sketch of what I'm suggesting.
> >
> > Here's an updated version of this patch which uses apr_dso.h.
>
> I'm not really qualified to review most of the DSO and autoconf
> details, but it sure would nice to only have to run these checks once
> per process. I guess you just use a static sqlite_initialized boolean
> for this; hopefully that's threadsafe :)
Definitely. (Hadn't really gotten the patch to that point of polish yet.)
Thanks for the feedback; that's exactly what I was hoping for.
> (Also, I'm not quite sure about how DSOs work, but might your patch
> load in a DSO when you're already linked against sqlite? Then you'd
> have two copies of sqlite linked in?)
At least on Linux and Mac OS X, dlopen() is only going to load one copy.
After that, dlopen() just increments the reference count for the library.
- application/pgp-signature attachment: stored
Received on Thu Nov 22 04:38:09 2007