On Monday 24 July 2006 18:49, Daniel Eischen wrote: > On Mon, 24 Jul 2006, Tijl Coosemans wrote: > > On Monday 24 July 2006 17:39, Daniel Eischen wrote: > >> On Mon, 24 Jul 2006, Tijl Coosemans wrote: > >>> I've attached two patches that accomplish this, but this seems to > >>> trigger other problems, so use at your own risk. If you want to > >>> try them, place them in the port's files/ directory and add a > >>> line containing "USE_AUTOTOOLS= autoconf:259" to the Makefile. > >>> This seems to break wine+libpthread, so I've also changed the > >>> port to use libthr instead. > >>> > >>> For the libpthread experts, I haven't investigated that much > >>> further yet, but libpthread seems to fail in create_stack() from > >>> _pthread_create() from _thr_start_sig_daemon(). > >> > >> See my response to this in a previous reply to this thread. > >> libthr and libpthread use LDT's for TLS. WINE is stomping on them > >> because it doesn't properly create LDTs. This is not a problem > >> with either of the thread libraries and this issue has been known > >> ever since we implemented TLS years ago. > > > > And as I stated later on in that thread, I don't see where > > libpthread and libthr still use LDT entries. As far as I understand > > the code, instead of using an LDT entry per thread (as it sure used > > to be), only one single GDT entry is used whose base address is > > updated during a context switch. Looking at the cvs history, it has > > been working like this since a couple commits of Peter Wemm about a > > year ago. > > > > And if nothing but Wine uses the LDT, Wine's static allocation of > > LDT entries can't be the problem. > > Look, we use %gs for TLS, period. Go see > libpthread/arch/i386/i386/pthread_md.c for how libpthread does it. > TLS would not work without setting aside a register for the threads > library (and rtld) to use.
Aaarrrrgghhh :) What you say is true of course, but %gs points to a GDT entry, not LDT. libpthread and libthr no longer use LDT entries... There would be a problem of course if Wine or Windows programs change %gs. Wine does seem to touch %gs but I've never actually seen it change. It's always 0x001B, which is the correct value (GUGS_SEL). However, Wine/Windows uses %fs for TLS and it appears that the FreeBSD kernel doesn't preserve it. It always ends up pointing to GUDATA_SEL.
pgpMV6kS6NpT4.pgp
Description: PGP signature
