Resolves: RHEL-45756 - Varnish consumes significantly more memory

under RHEL8/9
2025-06-11 20:35:46 +02:00 · 2025-06-11 20:35:46 +02:00 · 19ddc52756
commit 19ddc52756
parent ccfa6b1033
5 changed files with 294 additions and 5 deletions
--- a/.gitignore
+++ b/.gitignore
@ -65,3 +65,4 @@ varnish-2.1.3.tar.gz
 /varnish-7.6.0.tgz
 /pkg-varnish-cache-7d90347.tar.gz
 /varnish-7.6.1.tgz
+/jemalloc-5.3.0.tar.bz2
--- a/jemalloc-5.3.0-aarch64-ts-segfault.patch
+++ b/jemalloc-5.3.0-aarch64-ts-segfault.patch
@ -0,0 +1,140 @@
+diff --git a/test/unit/psset.c b/test/unit/psset.c
+index 6ff7201..58b4a88 100644
+--- a/test/unit/psset.c
+++ b/test/unit/psset.c
+@@ -124,7 +124,7 @@ TEST_BEGIN(test_fill) {
+ 	hpdata_t pageslab;
+ 	hpdata_init(&pageslab, PAGESLAB_ADDR, PAGESLAB_AGE);
+ 
+-	edata_t alloc[HUGEPAGE_PAGES];
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	psset_t psset;
+ 	psset_init(&psset);
+@@ -147,6 +147,8 @@ TEST_BEGIN(test_fill) {
+ 	edata_init_test(&extra_alloc);
+ 	err = test_psset_alloc_reuse(&psset, &extra_alloc, PAGE);
+ 	expect_true(err, "Alloc succeeded even though psset should be empty");
+	
+	free(alloc);
+ }
+ TEST_END
+ 
+@@ -157,7 +159,7 @@ TEST_BEGIN(test_reuse) {
+ 	hpdata_t pageslab;
+ 	hpdata_init(&pageslab, PAGESLAB_ADDR, PAGESLAB_AGE);
+ 
+-	edata_t alloc[HUGEPAGE_PAGES];
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	psset_t psset;
+ 	psset_init(&psset);
+@@ -239,6 +241,8 @@ TEST_BEGIN(test_reuse) {
+ 	err = test_psset_alloc_reuse(&psset, &alloc[index_of_4], 4 * PAGE);
+ 	expect_false(err, "Should have been able to find alloc.");
+ 	edata_expect(&alloc[index_of_4], index_of_4, 4);
+	
+	free(alloc);
+ }
+ TEST_END
+ 
+@@ -249,7 +253,7 @@ TEST_BEGIN(test_evict) {
+ 	hpdata_t pageslab;
+ 	hpdata_init(&pageslab, PAGESLAB_ADDR, PAGESLAB_AGE);
+ 
+-	edata_t alloc[HUGEPAGE_PAGES];
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	psset_t psset;
+ 	psset_init(&psset);
+@@ -273,6 +277,8 @@ TEST_BEGIN(test_evict) {
+ 
+ 	err = test_psset_alloc_reuse(&psset, &alloc[0], PAGE);
+ 	expect_true(err, "psset should be empty.");
+	
+	free(alloc);
+ }
+ TEST_END
+ 
+@@ -286,7 +292,9 @@ TEST_BEGIN(test_multi_pageslab) {
+ 	    (void *)((uintptr_t)PAGESLAB_ADDR + HUGEPAGE),
+ 	    PAGESLAB_AGE + 1);
+ 
+-	edata_t alloc[2][HUGEPAGE_PAGES];
+	edata_t* alloc[2];
+	alloc[0] = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+	alloc[1] = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	psset_t psset;
+ 	psset_init(&psset);
+@@ -336,6 +344,9 @@ TEST_BEGIN(test_multi_pageslab) {
+ 	 */
+ 	err = test_psset_alloc_reuse(&psset, &alloc[1][0], 2 * PAGE);
+ 	expect_false(err, "Allocation should have succeeded");
+	
+	free(alloc[0]);
+	free(alloc[1]);
+ }
+ TEST_END
+ 
+@@ -385,7 +396,7 @@ TEST_BEGIN(test_stats) {
+ 	hpdata_t pageslab;
+ 	hpdata_init(&pageslab, PAGESLAB_ADDR, PAGESLAB_AGE);
+ 
+-	edata_t alloc[HUGEPAGE_PAGES];
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	psset_t psset;
+ 	psset_init(&psset);
+@@ -415,6 +426,8 @@ TEST_BEGIN(test_stats) {
+ 	stats_expect(&psset, 0);
+ 	psset_update_end(&psset, &pageslab);
+ 	stats_expect(&psset, 1);
+	
+	free(alloc);
+ }
+ TEST_END
+ 
+@@ -475,8 +488,8 @@ init_test_pageslabs(psset_t *psset, hpdata_t *pageslab,
+ 
+ TEST_BEGIN(test_oldest_fit) {
+ 	bool err;
+-	edata_t alloc[HUGEPAGE_PAGES];
+-	edata_t worse_alloc[HUGEPAGE_PAGES];
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+	edata_t *worse_alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+ 
+ 	hpdata_t pageslab;
+ 	hpdata_t worse_pageslab;
+@@ -493,14 +506,19 @@ TEST_BEGIN(test_oldest_fit) {
+ 	expect_false(err, "Nonempty psset failed page allocation");
+ 	expect_ptr_eq(&pageslab, edata_ps_get(&test_edata),
+ 	    "Allocated from the wrong pageslab");
+
+	free(alloc);
+	free(worse_alloc);
+ }
+ TEST_END
+ 
+ TEST_BEGIN(test_insert_remove) {
+ 	bool err;
+ 	hpdata_t *ps;
+-	edata_t alloc[HUGEPAGE_PAGES];
+-	edata_t worse_alloc[HUGEPAGE_PAGES];
+
+	edata_t *alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+	edata_t *worse_alloc = (edata_t *)malloc(sizeof(edata_t) * HUGEPAGE_PAGES);
+
+ 
+ 	hpdata_t pageslab;
+ 	hpdata_t worse_pageslab;
+@@ -539,6 +557,9 @@ TEST_BEGIN(test_insert_remove) {
+ 	psset_update_begin(&psset, &worse_pageslab);
+ 	err = test_psset_alloc_reuse(&psset, &alloc[HUGEPAGE_PAGES - 1], PAGE);
+ 	expect_true(err, "psset should be empty, but an alloc succeeded");
+
+	free(alloc);
+	free(worse_alloc);
+ }
+ TEST_END
+ 
--- a/jemalloc-5.3.0_fno-builtin.patch
+++ b/jemalloc-5.3.0_fno-builtin.patch
@ -0,0 +1,29 @@
+commit 3de0c24859f4413bf03448249078169bb50bda0f
+Author: divanorama <divanorama@gmail.com>
+Date:   Thu Sep 29 23:35:59 2022 +0200
+
+    Disable builtin malloc in tests
+    
+    With `--with-jemalloc-prefix=` and without `-fno-builtin` or `-O1` both clang and gcc may optimize out `malloc` calls
+    whose result is unused. Comparing result to NULL also doesn't necessarily count as being used.
+    
+    This won't be a problem in most client programs as this only concerns really unused pointers, but in
+    tests it's important to actually execute allocations.
+    `-fno-builtin` should disable this optimization for both gcc and clang, and applying it only to tests code shouldn't hopefully be an issue.
+    Another alternative is to force "use" of result but that'd require more changes and may miss some other optimization-related issues.
+    
+    This should resolve https://github.com/jemalloc/jemalloc/issues/2091
+
+diff --git a/Makefile.in b/Makefile.in
+index 6809fb29..a964f07e 100644
+--- a/Makefile.in
+++ b/Makefile.in
+@@ -458,6 +458,8 @@ $(TESTS_OBJS): $(objroot)test/%.$(O): $(srcroot)test/%.c
+ $(TESTS_CPP_OBJS): $(objroot)test/%.$(O): $(srcroot)test/%.cpp
+ $(TESTS_OBJS): CPPFLAGS += -I$(srcroot)test/include -I$(objroot)test/include
+ $(TESTS_CPP_OBJS): CPPFLAGS += -I$(srcroot)test/include -I$(objroot)test/include
+$(TESTS_OBJS): CFLAGS += -fno-builtin
+$(TESTS_CPP_OBJS): CPPFLAGS += -fno-builtin
+ ifneq ($(IMPORTLIB),$(SO))
+ $(CPP_OBJS) $(C_SYM_OBJS) $(C_OBJS) $(C_JET_SYM_OBJS) $(C_JET_OBJS): CPPFLAGS += -DDLLEXPORT
+ endif
--- a/1
+++ b/1
@ -1,2 +1,3 @@
 SHA512 (varnish-7.6.1.tgz) = a43ecdcc5a113b947d56a7f28d756199c82e702a0e98bbad635a5df4739c50aaf778143dee3acf57d586569b780615ed73996df71488e4f776fb515f206b7fca
 SHA512 (pkg-varnish-cache-7d90347.tar.gz) = c5bf026bb50b416001d0e22e56c2774c143dab1f4658f03f1a4e6578369b71cfda5854b7d6b580c43c2ab8e68bfb9033b56734adfd29ac0fddc61fd6b1b4b0c0
+SHA512 (jemalloc-5.3.0.tar.bz2) = 22907bb052096e2caffb6e4e23548aecc5cc9283dce476896a2b1127eee64170e3562fa2e7db9571298814a7a2c7df6e8d1fbe152bd3f3b0c1abec22a2de34b1
--- a/varnish.spec
+++ b/varnish.spec
@ -23,11 +23,24 @@
 # Default: Use jemalloc, as adviced by upstream project
 # Change to 1 to use system allocator (ie. glibc)
 %if 0%{?rhel}
-%bcond_without system_allocator
+    %if 0%{?rhel} > 9
+        # for rhel >= 10, use bundled jemalloc
+        %bcond_with system_allocator
+        %bcond_without bundled_jemalloc
+    %else
+        # for rhel <= 9, use system allocator
+        %bcond_without system_allocator
+        %bcond_with bundled_jemalloc
+    %endif
 %else
-%bcond_with system_allocator
+    # use jemalloc from repo
+    %bcond_with system_allocator
+    %bcond_with bundled_jemalloc
 %endif

+%define jemalloc_version 5.3.0
+%define jemalloc_prefix varnish_
+
 %if %{with system_allocator}
 # use _lto_cflags if present
 %else
@ -37,13 +50,20 @@
 Summary: High-performance HTTP accelerator
 Name: varnish
 Version: 7.6.1
-Release: 3%{?dist}
+Release: 4%{?dist}
 License: BSD-2-Clause AND (BSD-2-Clause-FreeBSD AND BSD-3-Clause AND LicenseRef-Fedora-Public-Domain AND Zlib)
 URL: https://www.varnish-cache.org/
 Source0: http://varnish-cache.org/_downloads/%{name}-%{version}.tgz
 Source1: https://github.com/varnishcache/pkg-varnish-cache/archive/%{commit1}.tar.gz#/pkg-varnish-cache-%{shortcommit1}.tar.gz
+Source2: https://github.com/jemalloc/jemalloc/releases/download/%{jemalloc_version}/jemalloc-%{jemalloc_version}.tar.bz2

 # Patches:
+%if %{with bundled_jemalloc}
+# bundled jemalloc patch
+Patch1: jemalloc-5.3.0_fno-builtin.patch
+Patch2: jemalloc-5.3.0-aarch64-ts-segfault.patch
+%endif
+
 # https://bugzilla.redhat.com/show_bug.cgi?id=2364235
 Patch100: varnish-7.6.1-CVE-2025-47905.patch

@ -64,6 +84,10 @@ Provides: vmod(unix)%{_isa} = %{version}-%{release}
 Provides: vmod(vtc)%{_isa} = %{version}-%{release}
 %endif

+%if %{with bundled_jemalloc}
+Provides: bundled(jemalloc)
+%endif
+

 %if 0%{?rhel} == 7
 BuildRequires: python34 python34-sphinx python34-docutils
@ -74,10 +98,12 @@ BuildRequires: gcc
 %if %{with system_allocator}
 # use glibc
 %else
+%if %{without bundled_jemalloc}
 %ifnarch aarch64
 BuildRequires: jemalloc-devel
 %endif
 %endif
+%endif

 BuildRequires: libedit-devel
 BuildRequires: make
@ -86,6 +112,11 @@ BuildRequires: pcre2-devel
 BuildRequires: pkgconfig
 BuildRequires: systemd-units

+%if %{with bundled_jemalloc}
+BuildRequires:  /usr/bin/xsltproc
+BuildRequires:  perl-generators
+%endif
+
 # Extra requirements for the build suite
 #   needs haproxy2
 %if 0%{?fedora} > 30 || 0%{?rhel} > 8
@ -109,8 +140,10 @@ Requires(postun): systemd-units
 %if %{with system_allocator}
 # use glibc
 %else
+%if %{without bundled_jemalloc}
 Requires: jemalloc
 %endif
+%endif

 %description
 This is Varnish Cache, a high-performance HTTP accelerator.
@ -152,7 +185,55 @@ sed -i 's,rst2man-3.6,rst2man-3.4,g; s,rst2html-3.6,rst2html-3.4,g; s,phinx-buil

 %patch 100 -p1 -b .CVE-2022-45060

+# jemalloc
+%if %{with bundled_jemalloc}
+tar xjf %SOURCE2
+sed -i '/^LIBPREFIX/s/@libprefix@/@libprefix@%{jemalloc_prefix}/' jemalloc*/Makefile.in
+pushd jemalloc*
+%patch 1 -p1 -b .jemalloc
+%patch 2 -p1 -b .ts-segfault
+popd
+
+# Override PAGESIZE, bz #1545539
+%ifarch %ix86 %arm x86_64 s390x riscv64
+%define lg_page --with-lg-page=12
+%endif
+
+%ifarch ppc64 ppc64le aarch64
+%define lg_page --with-lg-page=16
+%endif
+
+# Disable thp on systems not supporting this for now
+%ifarch %ix86 %arm aarch64 s390x
+%define disable_thp --disable-thp
+%endif
+%endif
+
 %build
+
+%if %{with bundled_jemalloc}
+# build bundled jemalloc first
+pushd jemalloc*
+
+echo "For debugging package builders"
+echo "What is the pagesize?"
+getconf PAGESIZE
+
+echo "What mm features are available?"
+ls /sys/kernel/mm
+ls /sys/kernel/mm/transparent_hugepage || true
+cat /sys/kernel/mm/transparent_hugepage/enabled || true
+
+echo "What kernel version and config is this?"
+uname -a
+
+%configure %{?disable_thp} %{?lg_page} --enable-prof
+make %{?_smp_mflags}
+popd
+%endif
+
+
+# varnish
 %if %{with system_allocator}
 export CFLAGS="%{optflags}"
 %else
@ -186,6 +267,14 @@ export RST2MAN=/bin/true
 # Explicit python, please
 export PYTHON=%{__python}

+for f in configure configure.ac; do
+  sed -i 's|ljemalloc|l%{jemalloc_prefix}jemalloc|g' $f
+done
+
+%if %{with bundled_jemalloc}
+export LDFLAGS="$LDFLAGS -L%{_builddir}/%{name}-%{version}/jemalloc-%{jemalloc_version}/lib"
+%endif
+
 %configure LT_SYS_LIBRARY_PATH=%_libdir \
 --disable-static \
  --localstatedir=/var/lib  \
@ -196,10 +285,13 @@ export PYTHON=%{__python}
  --enable-pcre2-jit=no \
 %endif
 %endif
-%if %{with system_allocator}
+%if %{with system_allocator} || %{without bundled_jemalloc}
  --with-jemalloc=no \
 %endif

+%if %{with bundled_jemalloc}
+export LD_LIBRARY_PATH=%{_builddir}/%{name}-%{version}/jemalloc-%{jemalloc_version}/lib
+%endif
 %make_build

 # One varnish user is enough
@ -209,23 +301,45 @@ sed -i 's,User=varnishlog,User=varnish,g;' redhat/varnishncsa.service
 rm -rf doc/html/_sources

 %check
+# check jemalloc first
+%if %{with bundled_jemalloc}
+pushd jemalloc*
+make %{?_smp_mflags} check
+popd
+%endif

 # Up the stack size in tests, necessary on secondary arches
 sed -i 's/thread_pool_stack 80k/thread_pool_stack 128k/g;' bin/varnishtest/tests/*.vtc
 sed -i 's/file,2M/file,8M/' bin/varnishtest/tests/r04036.vtc

+%if %{with bundled_jemalloc}
+export LD_LIBRARY_PATH=%{_builddir}/%{name}-%{version}/jemalloc-%{jemalloc_version}/lib
+%endif
+
 # Just a hack to avoid too high load on secondary arch builders
 %ifarch s390x ppc64le
 # This works when ran alone, but not in the whole suite. Load and/or timing issues
 rm bin/varnishtest/tests/t02014.vtc
 make -j2 check
 %else
-#make_build check
+%make_build check
 %endif

 %install
 rm -rf %{buildroot}

+# jemalloc
+%if %{with bundled_jemalloc}
+pushd jemalloc*
+make DESTDIR=%{buildroot} install_lib %{?_smp_mflags}
+
+find %{buildroot}%{_libdir}/ -name '*.a' -exec rm -vf {} ';'
+
+# we don't need .pc file
+rm  %{buildroot}%{_libdir}/pkgconfig/jemalloc.pc
+popd
+%endif
+
 # mock el7 defaults to LANG=C, which makes python3 fail when parsing utf8 text
 %if 0%{?rhel} == 7
 export LANG=en_US.UTF-8
@ -319,6 +433,10 @@ test -f /etc/varnish/secret || (uuidgen > /etc/varnish/secret && chmod 0600 /etc


 %changelog
+* Wed Jun 11 2025 Luboš Uhliarik <luhliari@redhat.com> - 7.6.1-4
+- Resolves: RHEL-45756 - Varnish consumes significantly more memory
+  under RHEL8/9
+
 * Tue May 20 2025 Luboš Uhliarik <luhliari@redhat.com> - 7.6.1-3
 - Resolves: RHEL-89691 - varnish: request smuggling attacks (CVE-2025-47905)