From: Ondrej Zajicek <santiago@crfreenet.org>
Date: Tue, 17 Jan 2023 16:13:50 +0000 (+0100)
Subject: Alloc: Disable transparent huge pages
X-Git-Tag: v3.0-alpha1~55^2
X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=928a1cb034e6f9e8edcdd1dc07264cd703e00827;p=thirdparty%2Fbird.git

Alloc: Disable transparent huge pages

The usage pattern implemented in allocator seems to be incompatible with
transparent huge pages, as memory released using madvise(MADV_DONTNEED)
with regular page size and alignment does not seem to trigger demotion
of huge pages back to regular pages, even when significant number of
pages is released. Even if demotion is triggered when system memory
is low, it still breaks memory accounting.
---

diff --git a/sysdep/cf/README b/sysdep/cf/README
index 68078bbe7..af65aaec4 100644
--- a/sysdep/cf/README
+++ b/sysdep/cf/README
@@ -14,3 +14,6 @@ CONFIG_DONTROUTE_UNICAST Use MSG_DONTROUTE flag for unicast packets (def for Fre
 CONFIG_USE_HDRINCL	Use IP_HDRINCL instead of control messages for source address on raw IP sockets.
 
 CONFIG_RESTRICTED_PRIVILEGES	Implements restricted privileges using drop_uid()
+
+CONFIG_MADV_DONTNEED_TO_FREE	To free pages, use MADV_DONTNEED instead of MADV_FREE (linux)
+CONFIG_DISABLE_THP	Disable transparent huge pages (linux)
diff --git a/sysdep/cf/linux.h b/sysdep/cf/linux.h
index 9c37dd8a6..5edc49695 100644
--- a/sysdep/cf/linux.h
+++ b/sysdep/cf/linux.h
@@ -24,6 +24,7 @@
 #define CONFIG_INCLUDE_SYSPRIV_H "sysdep/linux/syspriv.h"
 
 #define CONFIG_MADV_DONTNEED_TO_FREE
+#define CONFIG_DISABLE_THP
 
 #ifndef AF_MPLS
 #define AF_MPLS 28
diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c
index e7c4e6b0e..5d9c6bcd6 100644
--- a/sysdep/unix/alloc.c
+++ b/sysdep/unix/alloc.c
@@ -19,6 +19,10 @@
 #include <sys/mman.h>
 #endif
 
+#ifdef CONFIG_DISABLE_THP
+#include <sys/prctl.h>
+#endif
+
 long page_size = 0;
 
 #ifdef HAVE_MMAP
@@ -218,6 +222,12 @@ global_free_pages_cleanup_event(void *data UNUSED)
 void
 resource_sys_init(void)
 {
+#ifdef CONFIG_DISABLE_THP
+  /* Disable transparent huge pages, they do not work properly with madvice(MADV_DONTNEED) */
+  if (prctl(PR_SET_THP_DISABLE,  (unsigned long) 1,  (unsigned long) 0,  (unsigned long) 0,  (unsigned long) 0) < 0)
+    die("prctl(PR_SET_THP_DISABLE) failed: %m");
+#endif
+
 #ifdef HAVE_MMAP
   ASSERT_DIE(global_free_pages.cnt == 0);