[RFC] tunables: Add elision tunable

Message ID ae68f386-3824-9e6b-c4b3-56e470ef1fbe@linux.vnet.ibm.com
State Superseded
Headers

Commit Message

rcardoso@linux.vnet.ibm.com Sept. 13, 2017, 7:28 p.m. UTC
  Hi Stefan,

Thanks for your review. I attached a v2 patch with you fix suggestions.

 >In addition to Paul's patches, here is a link to a newer patch from 
 >Carlos:
 >"RFC: Always-on elision with per-process opt-in using tunables."
 >https://www.sourceware.org/ml/libc-alpha/2017-05/msg00335.html
 >
 >He also removes the --enable-lock-elision configure flag.
 >Is this also required for your patch?

I'm aware of Carlos patch. In fact I talk to him before send this mail 
since we are doing similar things. And we agree that I will send this 
patch and he will word around with changes (like enable-lock-elision flag).

 >I've also used a small s390 specific test program to check wether a 
mutex was elided or not:

Pretty neat test. I see you already sent the test on this thread on 
another reply. I will try to port it to power and x86 here.

Regards,
Rogerio
  

Comments

Stefan Liebler Sept. 14, 2017, 6:31 a.m. UTC | #1
Hi Rogerio,

On 09/13/2017 09:28 PM, rcardoso wrote:
> Hi Stefan,
> 
> Thanks for your review. I attached a v2 patch with you fix suggestions.
> 
>  >In addition to Paul's patches, here is a link to a newer patch from 
>  >Carlos:
>  >"RFC: Always-on elision with per-process opt-in using tunables."
>  >https://www.sourceware.org/ml/libc-alpha/2017-05/msg00335.html
>  >
>  >He also removes the --enable-lock-elision configure flag.
>  >Is this also required for your patch?
> 
> I'm aware of Carlos patch. In fact I talk to him before send this mail 
> since we are doing similar things. And we agree that I will send this 
> patch and he will word around with changes (like enable-lock-elision flag). >
Okay.
Please also change the order of the lines (in all elision-conf.c files):
+# include <elf/dl-tunables.h>
+# define TUNABLE_NAMESPACE elision

If TUNABLE_NAMESPACE is not defined before dl-tunables.h is included,
there will be a build error as TUNABLE_GET will require 5 instead of 
three arguments.


>  >I've also used a small s390 specific test program to check wether a 
> mutex was elided or not:
> 
> Pretty neat test. I see you already sent the test on this thread on 
> another reply. I will try to port it to power and x86 here.
>Testing the other variables is not as easy as en-/disabling lock-elision.

> Regards,
> Rogerio

Bye
Stefan
  

Patch

From a3d8a372bba91de2dd72f998393b08bf5390b9cf Mon Sep 17 00:00:00 2001
From: Rogerio Alves <rcardoso@linux.vnet.ibm.com>
Date: Thu, 6 Jul 2017 13:21:08 -0500
Subject: [PATCH v2] Add elision tunables.

This patch adds several new tunables to control the behavior of
elision on supported platforms.  This also disables elision
by default on powerpc.

This patch was initially proposed by Paul Murphy[1] but was
"staled" because the framework have changed since the patch was
originally proposed.

[1] https://patchwork.sourceware.org/patch/10342/

2017-06-06  Rogerio A. Cardoso  <rcardoso@linux.vnet.ibm.com>,
	    Paul E. Murphy  <murphyp@linux.vnet.ibm.com>

	* elf/dl-tunables.list: Add elision parameters.
	* manual/tunables.texi: Add entries about elision tunable.
	* sysdeps/unix/sysv/linux/powerpc/elision-conf.c:
	Add callback functions to dynamically enable/disable elision.
	Add multiple callbacks functions to set elision parameters.
	* sysdeps/unix/sysv/linux/s390/elision-conf.c: Likewise.
	* sysdeps/unix/sysv/linux/x86/elision-conf.c: Likewise.
---
 Changes in v2: Per Stefan Liebler review. Add missing `end dftp` to
manual. Change #ifdef to #if HAVE_TUNABLE.

 elf/dl-tunables.list                           | 34 ++++++++++++
 manual/tunables.texi                           | 65 +++++++++++++++++++++++
 sysdeps/unix/sysv/linux/powerpc/elision-conf.c | 72 +++++++++++++++++++++++---
 sysdeps/unix/sysv/linux/s390/elision-conf.c    | 69 ++++++++++++++++++++++--
 sysdeps/unix/sysv/linux/x86/elision-conf.c     | 71 ++++++++++++++++++++++---
 5 files changed, 294 insertions(+), 17 deletions(-)

diff --git a/elf/dl-tunables.list b/elf/dl-tunables.list
index c188c6a..77dfcb4 100644
--- a/elf/dl-tunables.list
+++ b/elf/dl-tunables.list
@@ -96,4 +96,38 @@  glibc {
       default: HWCAP_IMPORTANT
       }
   }
+
+  elision {
+    enable {
+      type: INT_32
+      minval: 0
+      maxval: 1
+      security_level: SXID_IGNORE
+    }
+    skip_lock_busy {
+      type: INT_32
+      default: 3
+      security_level: SXID_IGNORE
+    }
+    skip_lock_internal_abort {
+      type: INT_32
+      default: 3
+      security_level: SXID_IGNORE
+    }
+    skip_lock_after_retries {
+      type: INT_32
+      default: 3
+      security_level: SXID_IGNORE
+    }
+    tries {
+      type: INT_32
+      default: 3
+      security_level: SXID_IGNORE
+    }
+    skip_trylock_internal_abort {
+      type: INT_32
+      default: 3
+      security_level: SXID_IGNORE
+    }
+  }
 }
diff --git a/manual/tunables.texi b/manual/tunables.texi
index 3c19567..5f660e0 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -31,6 +31,7 @@  their own namespace.
 @menu
 * Tunable names::  The structure of a tunable name
 * Memory Allocation Tunables::  Tunables in the memory allocation subsystem
+* Elision Tunables::  Tunables in elision subsystem
 * Hardware Capability Tunables::  Tunables that modify the hardware
 				  capabilities seen by @theglibc{}
 @end menu
@@ -226,6 +227,70 @@  pre-fill the per-thread cache with.  The default, or when set to zero,
 is no limit.
 @end deftp
 
+@node Elision Tunables
+@section Elision Tunables
+@cindex elision tunables
+@cindex tunables, elision
+
+@deftp {Tunable namespace} glibc.elision
+Elision behavior can be modified by setting any of the following tunables in
+the @code{elision} namespace:
+@end deftp
+
+@deftp Tunable glibc.elision.enable
+The @code{glibc.elision.enable} tunable enables lock elision if the feature is
+supported by the hardware.  If elision is not supported by the hardware this
+tunable has no effect.
+
+Elision tunables are supported for x86-64, PowerPC, and S390 architectures.
+@end deftp
+
+@deftp Tunable glibc.elision.skip_lock_busy
+The @code{glibc.elision.skip_lock_busy} tunable sets how many times to use a
+non-transactional lock after a transactional failure has occurred because the
+lock is already acquired.  Expressed in number of lock acquisition attempts.
+
+The default value of this tunable is @samp{3}.
+@end deftp
+
+@deftp Tunable glibc.elision.skip_lock_internal_abort
+The @code{glibc.elision.skip_lock_internal_abort} tunable sets how many times
+to not attempt to use elision if a transaction aborted due to reasons other
+than other threads' memory accesses.  Expressed in number of lock acquisition
+attempts.
+
+The default value of this tunable is @samp{3}.
+@end deftp
+
+@deftp Tunable glibc.elision.skip_lock_after_retries
+The @code{glibc.elision.skip_lock_after_retries} tunable sets how my times to
+try to elide a lock with transactions that only fail due other threads' memory
+accesses, before falling back to regular lock.
+Expressed in number of lock elision attempts.
+
+This tunable is not supported by x86.
+
+The default value of this tunable is @samp{3}.
+@end deftp
+
+@deftp Tunable glibc.elision.tries
+The @code{glibc.elision.tries} sets how many times we retry using elision if
+there is chance for the transaction to finish execution (e.g., it wasn't aborted
+due to the lock being already acquired).  This tunable is set to @samp{0} if
+elision is not supported by the hardware to avoid retries.
+
+The default value of this tunable is @samp{3}.
+@end deftp
+
+@deftp Tunable glibc.elision.skip_trylock_internal_abort
+The @code{glibc.elision.skip_trylock_internal_abort} tunable sets how many times
+to not attempt to use elision for trylocks if a transaction aborted due to
+reasons other than other threads' memory accesses.  Expressed in number of try
+lock attempts.
+
+The default value of this tunable is @samp{3}.
+@end deftp
+
 @node Hardware Capability Tunables
 @section Hardware Capability Tunables
 @cindex hardware capability tunables
diff --git a/sysdeps/unix/sysv/linux/powerpc/elision-conf.c b/sysdeps/unix/sysv/linux/powerpc/elision-conf.c
index f631f0a..9effe09 100644
--- a/sysdeps/unix/sysv/linux/powerpc/elision-conf.c
+++ b/sysdeps/unix/sysv/linux/powerpc/elision-conf.c
@@ -21,6 +21,10 @@ 
 #include <elision-conf.h>
 #include <unistd.h>
 #include <dl-procinfo.h>
+#if HAVE_TUNABLES
+# include <elf/dl-tunables.h>
+# define TUNABLE_NAMESPACE elision
+#endif
 
 /* Reasonable initial tuning values, may be revised in the future.
    This is a conservative initial value.  */
@@ -50,7 +54,50 @@  struct elision_config __elision_aconf =
    DEFAULT locks should be automatically use elision in pthread_mutex_lock().
    Disabled for suid programs.  Only used when elision is available.  */
 
-int __pthread_force_elision attribute_hidden;
+int __pthread_force_elision attribute_hidden = 0;
+
+#if HAVE_TUNABLES
+static inline void
+__always_inline
+do_set_elision_enable (int32_t elision_enable)
+{
+  /* Enable elision if it's avaliable in hardware.  */
+  if (elision_enable && !__libc_enable_secure)
+    __pthread_force_elision = (GLRO (dl_hwcap2)
+			       & PPC_FEATURE2_HAS_HTM) ? 1 : 0;
+}
+
+/* The pthread->elision_enable tunable is 0 or 1 indicating that elision
+   should be disabled or enabled respectively.  The feature will only be used
+   if it's supported by the hardware.  */
+
+void
+TUNABLE_CALLBACK (set_elision_enable) (tunable_val_t *valp)
+{
+  int32_t elision_enable = (int32_t) valp->numval;
+  do_set_elision_enable (elision_enable);
+}
+
+#define TUNABLE_CALLBACK_FNDECL(__name, __type)			\
+static inline void						\
+__always_inline							\
+do_set_elision_ ## __name (__type value)			\
+{								\
+  __elision_aconf.__name = value;				\
+}								\
+void								\
+TUNABLE_CALLBACK (set_elision_ ## __name) (tunable_val_t *valp) \
+{								\
+  __type value = (__type) (valp)->numval;			\
+  do_set_elision_ ## __name (value);				\
+}
+
+TUNABLE_CALLBACK_FNDECL (skip_lock_busy, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_lock_internal_abort, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_lock_out_of_tbegin_retries, int32_t);
+TUNABLE_CALLBACK_FNDECL (try_tbegin, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_trylock_internal_abort, int32_t);
+#endif
 
 /* Initialize elision.  */
 
@@ -59,13 +106,26 @@  elision_init (int argc __attribute__ ((unused)),
 	      char **argv  __attribute__ ((unused)),
 	      char **environ)
 {
-#ifdef ENABLE_LOCK_ELISION
-  int elision_available = (GLRO (dl_hwcap2) & PPC_FEATURE2_HAS_HTM) ? 1 : 0;
-  __pthread_force_elision = __libc_enable_secure ? 0 : elision_available;
+#if HAVE_TUNABLES
+  /* Elision depends on tunables and must be explicitly turned on by setting
+     the appropriate tunable on a supported platform.  */
+
+  TUNABLE_GET (enable, int32_t,
+	       TUNABLE_CALLBACK (set_elision_enable));
+  TUNABLE_GET (skip_lock_busy, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_busy));
+  TUNABLE_GET (skip_lock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_internal_abort));
+  TUNABLE_GET (skip_lock_after_retries, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_out_of_tbegin_retries));
+  TUNABLE_GET (tries, int32_t,
+	       TUNABLE_CALLBACK (set_elision_try_tbegin));
+  TUNABLE_GET (skip_trylock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_trylock_internal_abort));
 #endif
+
   if (!__pthread_force_elision)
-    /* Disable elision on rwlocks.  */
-    __elision_aconf.try_tbegin = 0;
+    __elision_aconf.try_tbegin = 0; /* Disable elision on rwlocks.  */
 }
 
 #ifdef SHARED
diff --git a/sysdeps/unix/sysv/linux/s390/elision-conf.c b/sysdeps/unix/sysv/linux/s390/elision-conf.c
index cc0fdef..c6da0b3 100644
--- a/sysdeps/unix/sysv/linux/s390/elision-conf.c
+++ b/sysdeps/unix/sysv/linux/s390/elision-conf.c
@@ -21,6 +21,10 @@ 
 #include <elision-conf.h>
 #include <unistd.h>
 #include <dl-procinfo.h>
+#if HAVE_TUNABLES
+# include <elf/dl-tunables.h>
+# define TUNABLE_NAMESPACE elision
+#endif
 
 /* Reasonable initial tuning values, may be revised in the future.
    This is a conservative initial value.  */
@@ -53,6 +57,48 @@  struct elision_config __elision_aconf =
 
 int __pthread_force_elision attribute_hidden = 0;
 
+#if HAVE_TUNABLES
+static inline void
+__always_inline
+do_set_elision_enable (int32_t elision_enable)
+{
+  /* Enable elision if it's avaliable in hardware.  */
+  if (elision_enable && !__libc_enable_secure)
+    __pthread_force_elision = (GLRO (dl_hwcap) & HWCAP_S390_TE) ? 1 : 0;
+}
+
+/* The pthread->elision_enable tunable is 0 or 1 indicating that elision
+   should be disabled or enabled respectively.  The feature will only be used
+   if it's supported by the hardware.  */
+
+void
+TUNABLE_CALLBACK (set_elision_enable) (tunable_val_t *valp)
+{
+  int32_t elision_enable = (int32_t) valp->numval;
+  do_set_elision_enable (elision_enable);
+}
+
+#define TUNABLE_CALLBACK_FNDECL(__name, __type)			\
+static inline void						\
+__always_inline							\
+do_set_elision_ ## __name (__type value)			\
+{								\
+  __elision_aconf.__name = value;				\
+}								\
+void								\
+TUNABLE_CALLBACK (set_elision_ ## __name) (tunable_val_t *valp) \
+{								\
+  __type value = (__type) (valp)->numval;			\
+  do_set_elision_ ## __name (value);				\
+}
+
+TUNABLE_CALLBACK_FNDECL (skip_lock_busy, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_lock_internal_abort, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_lock_out_of_tbegin_retries, int32_t);
+TUNABLE_CALLBACK_FNDECL (try_tbegin, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_trylock_internal_abort, int32_t);
+#endif
+
 /* Initialize elison.  */
 
 static void
@@ -60,11 +106,26 @@  elision_init (int argc __attribute__ ((unused)),
 	      char **argv  __attribute__ ((unused)),
 	      char **environ)
 {
-  /* Set when the CPU and the kernel supports transactional execution.
-     When false elision is never attempted.  */
-  int elision_available = (GLRO (dl_hwcap) & HWCAP_S390_TE) ? 1 : 0;
+#if HAVE_TUNABLES
+  /* Elision depends on tunables and must be explicitly turned on by setting
+     the appropriate tunable on a supported platform.  */
+
+  TUNABLE_GET (enable, int32_t,
+	       TUNABLE_CALLBACK (set_elision_enable));
+  TUNABLE_GET (skip_lock_busy, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_busy));
+  TUNABLE_GET (skip_lock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_internal_abort));
+  TUNABLE_GET (skip_lock_after_retries, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_out_of_tbegin_retries));
+  TUNABLE_GET (tries, int32_t,
+	       TUNABLE_CALLBACK (set_elision_try_tbegin));
+  TUNABLE_GET (skip_trylock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_trylock_internal_abort));
+#endif
 
-  __pthread_force_elision = __libc_enable_secure ? 0 : elision_available;
+  if (!__pthread_force_elision)
+    __elision_aconf.try_tbegin = 0; /* Disable elision on rwlocks.  */
 }
 
 #ifdef SHARED
diff --git a/sysdeps/unix/sysv/linux/x86/elision-conf.c b/sysdeps/unix/sysv/linux/x86/elision-conf.c
index 673b000..0c027cb 100644
--- a/sysdeps/unix/sysv/linux/x86/elision-conf.c
+++ b/sysdeps/unix/sysv/linux/x86/elision-conf.c
@@ -21,6 +21,10 @@ 
 #include <init-arch.h>
 #include <elision-conf.h>
 #include <unistd.h>
+#if HAVE_TUNABLES
+# include <elf/dl-tunables.h>
+# define TUNABLE_NAMESPACE elision
+#endif
 
 /* Reasonable initial tuning values, may be revised in the future.
    This is a conservative initial value.  */
@@ -48,21 +52,74 @@  struct elision_config __elision_aconf =
    pthread_mutex_lock().  Disabled for suid programs.  Only used when elision
    is available.  */
 
-int __pthread_force_elision attribute_hidden;
+int __pthread_force_elision attribute_hidden = 0;
+
+#if HAVE_TUNABLES
+static inline void
+__always_inline
+do_set_elision_enable (int32_t elision_enable)
+{
+  /* Enable elision if it's avaliable in hardware.  */
+  if (elision_enable && !__libc_enable_secure)
+    __pthread_force_elision = HAS_CPU_FEATURE (RTM) ? 1 : 0;
+}
+
+/* The pthread->elision_enable tunable is 0 or 1 indicating that elision
+   should be disabled or enabled respectively.  The feature will only be used
+   if it's supported by the hardware.  */
+
+void
+TUNABLE_CALLBACK (set_elision_enable) (tunable_val_t *valp)
+{
+  int32_t elision_enable = (int32_t) valp->numval;
+  do_set_elision_enable (elision_enable);
+}
+
+#define TUNABLE_CALLBACK_FNDECL(__name, __type)			\
+static inline void						\
+__always_inline							\
+do_set_elision_ ## __name (__type value)			\
+{								\
+  __elision_aconf.__name = value;				\
+}								\
+void								\
+TUNABLE_CALLBACK (set_elision_ ## __name) (tunable_val_t *valp) \
+{								\
+  __type value = (__type) (valp)->numval;			\
+  do_set_elision_ ## __name (value);				\
+}
 
-/* Initialize elison.  */
+TUNABLE_CALLBACK_FNDECL (skip_lock_busy, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_lock_internal_abort, int32_t);
+TUNABLE_CALLBACK_FNDECL (retry_try_xbegin, int32_t);
+TUNABLE_CALLBACK_FNDECL (skip_trylock_internal_abort, int32_t);
+#endif
+
+/* Initialize elision.  */
 
 static void
 elision_init (int argc __attribute__ ((unused)),
 	      char **argv  __attribute__ ((unused)),
 	      char **environ)
 {
-  int elision_available = HAS_CPU_FEATURE (RTM);
-#ifdef ENABLE_LOCK_ELISION
-  __pthread_force_elision = __libc_enable_secure ? 0 : elision_available;
+#if HAVE_TUNABLES
+  /* Elision depends on tunables and must be explicitly turned on by setting
+     the appropriate tunable on a supported platform.  */
+
+  TUNABLE_GET (enable, int32_t,
+	       TUNABLE_CALLBACK (set_elision_enable));
+  TUNABLE_GET (skip_lock_busy, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_busy));
+  TUNABLE_GET (skip_lock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_lock_internal_abort));
+  TUNABLE_GET (tries, int32_t,
+	       TUNABLE_CALLBACK (set_elision_retry_try_xbegin));
+  TUNABLE_GET (skip_trylock_internal_abort, int32_t,
+	       TUNABLE_CALLBACK (set_elision_skip_trylock_internal_abort));
 #endif
-  if (!elision_available)
-    __elision_aconf.retry_try_xbegin = 0; /* Disable elision on rwlocks */
+
+  if (!__pthread_force_elision)
+    __elision_aconf.retry_try_xbegin = 0; /* Disable elision on rwlocks.  */
 }
 
 #ifdef SHARED
-- 
2.7.4