Disable Transparent Hugepage

bin/varnishd/mgt/mgt.h

@@ -165,6 +165,8 @@ extern const struct jail_tech jail_tech_unix;
 extern const struct jail_tech jail_tech_solaris;
 extern const struct jail_tech jail_tech_linux;
 
+#define MATCH_JAIL_ARG(arg, option) (!strncmp(arg, option, sizeof(option) - 1))
+
 /* mgt_main.c */
 extern struct vsb *vident;
 extern struct VSC_mgt	*VSC_C_mgt;

bin/varnishd/mgt/mgt_jail_linux.c

@@ -48,11 +48,61 @@
 #include "mgt/mgt.h"
 #include "common/heritage.h"
 
+static int
+vjl_set_thp(const char *arg)
+{
+	int val;
+
+	if (!strcmp(arg, "ignore"))
+		return (0);
+	if (!strcmp(arg, "enable"))
+		val = 0;
+	else if (!strcmp(arg, "disable"))
+		val = 1;
+	else {
+		ARGV_ERR(
+		    "linux jail: unknown value '%s' for argument transparent_hugepage.\n",
+		    arg);
+	}
+	if (prctl(PR_SET_THP_DISABLE, val, 0, 0, 0) != 0) {
+		MGT_Complain(C_ERR,
+		    "Could not %s Transparent Hugepage: %s (%d)",
+		    arg, VAS_errtxt(errno), errno);
+	}
+	return (0);
+}
+
 static int
 vjl_init(char **args)
 {
+	char **unix_args;
+	int ret = 0;
+	size_t i;
+
+	if (args == NULL) {
+		/* Autoconfig */
+		if (vjl_set_thp("disable") != 0)
+			return (1);
+		return (jail_tech_unix.init(NULL));
+	}
 
-	return jail_tech_unix.init(args);
+	for (i = 0; args[i] != NULL; i++);
+	unix_args = calloc(i + 1, sizeof *unix_args);
+	AN(unix_args);
+
+	for (i = 0; *args != NULL && ret == 0; args++) {
+		if (MATCH_JAIL_ARG(*args, "transparent_hugepage=")) {
+			ret = vjl_set_thp((*args) + (sizeof("transparent_hugepage=") - 1));
+		} else {
+			unix_args[i] = *args;
+			i++;
+		}
+	}
+
+	if (ret == 0)
+		ret = jail_tech_unix.init(unix_args);
+	free(unix_args);
+	return (ret);
 }
 
 static void
@@ -94,17 +144,24 @@ vjl_make_workdir(const char *dname, const char *what, struct vsb *vsb)
 
 	vjl_master(JAIL_MASTER_FILE);
 	if (statfs(dname, &info) != 0) {
-		if (vsb)
-			VSB_printf(vsb, "Could not stat working directory '%s': %s (%d)\n", dname, VAS_errtxt(errno), errno);
-		else
-			MGT_Complain(C_ERR, "Could not stat working directory '%s': %s (%d)", dname, VAS_errtxt(errno), errno);
+		if (vsb) {
+			VSB_printf(vsb,
+			    "Could not stat working directory '%s': %s (%d)\n",
+			    dname, VAS_errtxt(errno), errno);
+		} else {
+			MGT_Complain(C_ERR,
+			    "Could not stat working directory '%s': %s (%d)",
+			    dname, VAS_errtxt(errno), errno);
+		}
 		return (1);
 	}
 	if (info.f_type != TMPFS_MAGIC) {
 		if (vsb != NULL)
-			VSB_printf(vsb, "Working directory not mounted on tmpfs partition\n");
+			VSB_printf(vsb,
+			    "Working directory not mounted on tmpfs partition\n");
 		else
-			MGT_Complain(C_INFO, "Working directory not mounted on tmpfs partition");
+			MGT_Complain(C_INFO,
+			    "Working directory not mounted on tmpfs partition");
 	}
 	vjl_master(JAIL_MASTER_LOW);
 	return (0);

bin/varnishd/mgt/mgt_jail_solaris.c

@@ -328,7 +328,7 @@ vjs_init(char **args)
 
 	if (args != NULL && *args != NULL) {
 		for (;*args != NULL; args++) {
-			if (!strncmp(*args, "worker=", 7)) {
+			if (MATCH_JAIL_ARG(*args, "worker=")) {
 				user = priv_str_to_set((*args) + 7, ",", &e);
 				if (user == NULL)
 					ARGV_ERR(

bin/varnishd/mgt/mgt_jail_unix.c

@@ -132,21 +132,21 @@ vju_init(char **args)
 			ARGV_ERR("Unix Jail: Must be root.\n");
 
 		for (;*args != NULL; args++) {
-			if (!strncmp(*args, "user=", 5)) {
+			if (MATCH_JAIL_ARG(*args, "user=")) {
 				if (vju_getuid((*args) + 5))
 					ARGV_ERR(
 					    "Unix jail: %s user not found.\n",
 					    (*args) + 5);
 				continue;
 			}
-			if (!strncmp(*args, "workuser=", 9)) {
+			if (MATCH_JAIL_ARG(*args, "workuser=")) {
 				if (vju_getwrkuid((*args) + 9))
 					ARGV_ERR(
 					    "Unix jail: %s user not found.\n",
 					    (*args) + 9);
 				continue;
 			}
-			if (!strncmp(*args, "ccgroup=", 8)) {
+			if (MATCH_JAIL_ARG(*args, "ccgroup=")) {
 				if (vju_getccgid((*args) + 8))
 					ARGV_ERR(
 					    "Unix jail: %s group not found.\n",

doc/sphinx/installation/platformnotes.rst

@@ -11,21 +11,29 @@ On some platforms it is necessary to adjust the operating system before running
 Varnish on it. The systems and steps known to us are described in this section.
 
 
-Transparent hugepages on Redhat Enterprise Linux 6
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Transparent Hugepage on Linux
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-On RHEL6 Transparent Hugepage kernel support is enabled by default.
-This is known to cause sporadic crashes of Varnish.
+On certain Linux distributions Transparent Hugepage kernel support is enabled
+by default. This is known to cause instabilities of Varnish.
 
-It is recommended to disable transparent hugepages on affected
-systems. This can be done with
-``echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled``
-(runtime) or by adding "transparent_hugepage=never" to the kernel boot
-line in the "/etc/grub.conf" file (persistent).
+It is recommended to disable Transparent Hugepage on affected systems.
+If Varnish is the only significant service running on this system, this can be
+done during runtime with::
 
-On Debian/Ubuntu systems running 3.2 kernels the default value is "madvise" and
-does not need to be changed.
+  echo never > /sys/kernel/mm/transparent_hugepage/enabled
 
+Alternatively, this can be persisted in your bootloader configuration by adding
+``transparent_hugepage=never`` to the kernel command line.
+
+On other systems the default value is ``madvise`` and does not need to be
+changed. Either way, the Linux :ref:`ref-varnishd-opt_j` will try to disable
+Transparent Hugepage in the ``varnishd`` process.
+
+The general recommendation is to mount the working directory in a ``tmpfs``
+partition, usually ``/var/lib/varnish``. By default tmpfs should be mounted with
+Transparent Hugepage disabled. Consider mounting the working directory with the
+``huge=never`` mount option if that is not the case with your OS vendor.
 
 OpenVZ
 ~~~~~~

doc/sphinx/reference/varnishd.rst

@@ -454,13 +454,18 @@ specific options. Available jails are:
 
     -j solaris,worker=basic
 
--j <linux[,user=`user`][,ccgroup=`group`][,workuser=`user`]>
+-j <linux[,transparent_hugepage=`action`][,<unix jail option>...]>
 
   Default on Linux platforms, it extends the UNIX jail with
   Linux-specific mechanisms:
 
   - It warns when *workdir* is not on a ``tmpfs``.
   - It tries to keep the process dumpable after dropping privileges.
+  - It tries to disable Transparent Hugepage in ``varnishd`` by default.
+
+The optional transparent_hugepage argument takes one action among enable,
+disable (default) and ignore. Failure to configure Transparent Hugepage is
+not a fatal error.
 
 -j <unix[,user=`user`][,ccgroup=`group`][,workuser=`user`]>