This is the mail archive of the mailing list for the glibc 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]

Disable all lock elision unless --enable-lock-elision=yes


The rwlock elision support you added in 2014-06-13
appears to unconditionally enable elision for rwlock's
if the RTM-bit indicates TSX is present. That isn't what
we do for mutexes, and isn't very conservative in that

Did I read the code right?

If I did, could we please put rwlock TSX usage under
the control of the --enable-lock-elision flag? I need
to easily disable TSX for all of glibc, until we can
get early microcode updates worked out, and I had to
hack things up because --enable-lock-elision no longer
controls all of the elision uses in glibc.

For example the following WIP patch does what I'm thinking.


diff -urN glibc-2.20.mod/sysdeps/unix/sysv/linux/x86/elision-conf.c glibc-2.20/sysdeps/unix/sysv/linux/x86/elision-conf.c
--- glibc-2.20.mod/sysdeps/unix/sysv/linux/x86/elision-conf.c	2014-09-27 00:25:46.443462345 -0400
+++ glibc-2.20/sysdeps/unix/sysv/linux/x86/elision-conf.c	2014-09-27 00:29:53.586615813 -0400
@@ -62,12 +62,16 @@
 	      char **argv  __attribute__ ((unused)),
 	      char **environ)
-  __elision_available = HAS_RTM;
+  __elision_available = HAS_RTM;
   __pthread_force_elision = __libc_enable_secure ? 0 : __elision_available;
   if (!HAS_RTM)
     __elision_aconf.retry_try_xbegin = 0; /* Disable elision on rwlocks */
+  __elision_available = 0;
+  __pthread_force_elision = 0;
+  __elision_aconf.retry_try_xbegin = 0;
 #ifdef SHARED
I should probably patch to change the description
of the enable option to talk about all locks.


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