This is the mail archive of the cygwin-patches mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

RE: [PATCH] Preserve order of dlopen'd modules in dll_list::topsort


Hi Corinna,

Corinna Vinschen wrote:
> Hi David,
> 
> On Feb 25 16:27, David Allsopp wrote:
> > This patch (below - I hope I have managed to format this email
> > correctly) alters the behaviour of dll_list::topsort to preserve the
> > order of dlopen'd units.
> >
> > The load order of unrelated DLLs is reversed every time fork is
> > called, since dll_list::topsort finds the tail of the list and then
> > unwinds to reinsert items. My change takes advantage of what should be
> > undefined behaviour in dll_list::populate_deps (ndeps non-zero and
> > ndeps and deps not
> > initialised) to allow the deps field to be initialised prior to the
> > call and appended to, rather than overwritten.
> >
> > All DLLs which have been dlopen'd have their deps list initialised
> > with the list of all previously dlopen'd units. These extra
> > dependencies mean that the unwind preserves the order of dlopen'd units.
> >
> > The motivation for this is the FlexDLL linker used in OCaml. The
> > FlexDLL linker allows a dlopen'd unit to refer to symbols in
> > previously dlopen'd units and it resolves these symbols in DllMain
> > before anything else has initialised (including the Cygwin DLL). This
> > means that dependencies may exist between dlopen'd units (which the
> > OCaml runtime system understands) but which Windows is unaware of.
> > During fork, the process-level table which FlexDLL uses to get the
> > symbol table of each DLL is copied over but because the load order of
> > dlopen'd DLLs is reversed, it is possible for FlexDLL to attempt to
> > access memory in the DLL before it has been loaded and hence it fails
> > with an access violation. Because the list is reversed on each call to
> > fork, it means that a subsequent call to fork puts the DLLs back into
> > the correct order, hence "even" invocations of fork work!
> >
> > An interesting side-effect is that this only occurs if the DLLs load
> > at their preferred base address - if they have to be rebased, then
> > FlexDLL works because at the time that the dependent unit is loaded
> > out of order, there is still in memory the "dummy"
> > DONT_RESOLVE_DLL_REFERENCES version of the dependency which, as it
> > happens, will contain the correct symbol table in the data section.
> > For my tests, this initially appeared to be an x86-only problem, but
> > that was only because the two DLLs on x64 should have been rebased.
> >
> > I'm very happy to include the complete detail for this and, for the
> > extremely keen, the relevant Git branch in OCaml which demonstrates
> > this problem. Given the way in which FlexDLL operates, I would contend
> > that this is a sensible change of behaviour for the Cygwin DLL, though
> > not a bug fix.
> > I'd be extremely happy to see this patch integrated, as the workaround
> > necessary in FlexDLL to support Cygwin's fork is horrible (and
> > non-transparent to the library user).
> >
> > This patch is licensed under 2-clause BSD as per winsup/CONTRIBUTORS,
> > Copyright (c) 2017, MetaStack Solutions Ltd.
> 
> First of all, I think this makes perfect sense.  I just have a few
> questions in terms of the patch itself.
> 
> - Your browser inserts undesired line breaks, so the patch is broken.
>   Can you please resend the `git format-patch' output as attachment?

Darn - sorry about that (it's the first time I'd tried to send a format-patch, rather than as a PR). Patch attached.

> - While you're at it, please reformat your patch so the line length
>   is not longer than 80 chars.

Done - sorry, I'd inferred a longer length from a few other longer lines!

> - Last but not least.  You add code to topsort so the loaded DLLs
>   are handled first.  The subsequent code is untouched.  However,
>   shouldn't the next loop then restrict calling populate_deps to the
>   linked DLLs only, at least for performance?

Oops :$ That's an artefact of the "story" of the patch's development. As it happens, the first dlopen'd DLL would have been initialised in the second loop, not the first, but the presence of two loops like that was indeed mostly inefficient. I've kept the original one as a "fast path" for the case of no dlopen'd DLLs, though I don't know if that's a worthwhile optimisation. 

All best,


David

Attachment: 0001-Preserve-order-of-dlopen-d-modules-in-dll_list-topso.patch
Description: Binary data


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]