Mercurial > hg > graal-jvmci-8

diff src/os/linux/vm/os_linux.cpp @ 12176:88c255656030

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
8016155: SIGBUS when running Kitchensink with ParallelScavenge and ParallelOld Summary: When using NUMA and large pages we need to ease the requirement on which node the memory should be allocated on. To avoid the SIGBUS we now use the memory policy MPOL_PREFERRED, which prefers a certain node, instead of MPOL_BIND, which requires a certain node. Reviewed-by: jmasa, pliden Contributed-by: stefan.johansson@oracle.com
author mgerdin
date Thu, 22 Aug 2013 10:50:41 +0200
parents 59b052799158
children 0d59407e7e09
line wrap: on
line diff
--- a/src/os/linux/vm/os_linux.cpp	Fri Aug 23 15:59:20 2013 -0700
+++ b/src/os/linux/vm/os_linux.cpp	Thu Aug 22 10:50:41 2013 +0200
@@ -2796,7 +2796,19 @@
   Linux::numa_interleave_memory(addr, bytes);
 }
 
+// Define for numa_set_bind_policy(int). Setting the argument to 0 will set the
+// bind policy to MPOL_PREFERRED for the current thread.
+#define USE_MPOL_PREFERRED 0
+
 void os::numa_make_local(char *addr, size_t bytes, int lgrp_hint) {
+  // To make NUMA and large pages more robust when both enabled, we need to ease
+  // the requirements on where the memory should be allocated. MPOL_BIND is the
+  // default policy and it will force memory to be allocated on the specified
+  // node. Changing this to MPOL_PREFERRED will prefer to allocate the memory on
+  // the specified node, but will not force it. Using this policy will prevent
+  // getting SIGBUS when trying to allocate large pages on NUMA nodes with no
+  // free large pages.
+  Linux::numa_set_bind_policy(USE_MPOL_PREFERRED);
   Linux::numa_tonode_memory(addr, bytes, lgrp_hint);
 }
 
@@ -2898,6 +2910,8 @@
                                             libnuma_dlsym(handle, "numa_tonode_memory")));
       set_numa_interleave_memory(CAST_TO_FN_PTR(numa_interleave_memory_func_t,
                                             libnuma_dlsym(handle, "numa_interleave_memory")));
+      set_numa_set_bind_policy(CAST_TO_FN_PTR(numa_set_bind_policy_func_t,
+                                            libnuma_dlsym(handle, "numa_set_bind_policy")));
 
 
       if (numa_available() != -1) {
@@ -2964,6 +2978,7 @@
 os::Linux::numa_available_func_t os::Linux::_numa_available;
 os::Linux::numa_tonode_memory_func_t os::Linux::_numa_tonode_memory;
 os::Linux::numa_interleave_memory_func_t os::Linux::_numa_interleave_memory;
+os::Linux::numa_set_bind_policy_func_t os::Linux::_numa_set_bind_policy;
 unsigned long* os::Linux::_numa_all_nodes;
 
 bool os::pd_uncommit_memory(char* addr, size_t size) {