[pve-devel] [PATCH-SERIES qemu, qemu-server, manager 0/1] add tcmalloc support

[pve-devel] [PATCH-SERIES qemu, qemu-server, manager 0/1] add tcmalloc support

[Alexandre Derumier]

This patches series add support for tcmalloc allocator with a new optionnal
tuning option in qemu-server.

tcmalloc improve librbd performance by 30% in latency and iops.

some performance bench results:

malloc: 60k iops 4k randread
tcmalloc : 90k iops 4k randread


a qemu patch is provided, to detect if a different memory allocator than
glibc malloc is loaded through LD_PRELOAD, to avoid the call of malloc_trim().
I don't have seen crash without this patch, but it's safer with it,
and as it's called 3-4x by second, it'll avoid some unnecessary calls.

qemu:

Alexandre Derumier (1):
  add patch: add memory allocator detection and use malloc_trim for
    glibc only

 ...-and-call-malloc_trim-only-for-glibc.patch | 163 ++++++++++++++++++
 debian/patches/series                         |   1 +
 2 files changed, 164 insertions(+)
 create mode 100644 debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch

qemu-server:

Alexandre Derumier (1):
  qemu options: add tuning allocator

 PVE/QemuServer.pm | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

pve-manager:

Alexandre Derumier (1):
  ui: qemu : add tuning option

-- 
2.30.2

[pve-devel] [PATCH-SERIES qemu, qemu-server, manager 1/1] add patch: add memory allocator detection and use malloc_trim for glibc only

[Alexandre Derumier]

Detect if a different allocator than glibc malloc is linked with LD_PRELOAD,

and call malloc_trim() only for glibc malloc

This patch is mostly copy/paste from haproxy
https://github.com/haproxy/haproxy/blob/master/src/pool.c

Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
---
 ...-and-call-malloc_trim-only-for-glibc.patch | 163 ++++++++++++++++++
 debian/patches/series                         |   1 +
 2 files changed, 164 insertions(+)
 create mode 100644 debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch

diff --git a/debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch b/debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch
new file mode 100644
index 0000000..aacb6f4
--- /dev/null
+++ b/debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch
@@ -0,0 +1,163 @@
+From 9bd1fdf160b0c5801c7f09fc02d15024cd838b39 Mon Sep 17 00:00:00 2001
+From: Alexandre Derumier <aderumier@odiso.com>
+Date: Wed, 22 Mar 2023 17:51:35 +0100
+Subject: [PATCH] detect allocator and call malloc_trim only for glibc
+
+Allow to detect if a different allocator is dynamicaly loaded
+with LD_PRELOAD , and don't call malloc_trim in this case.
+
+mostly copy/paste from haproxy code
+
+https://github.com/haproxy/haproxy/blob/master/src/pool.c
+Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
+---
+ include/qemu/rcu.h | 19 ++++++++++++
+ util/rcu.c         | 76 +++++++++++++++++++++++++++++++++++++++++++++-
+ 2 files changed, 94 insertions(+), 1 deletion(-)
+
+diff --git a/include/qemu/rcu.h b/include/qemu/rcu.h
+index b063c6fde8..7ab8b899f6 100644
+--- a/include/qemu/rcu.h
++++ b/include/qemu/rcu.h
+@@ -196,6 +196,25 @@ G_DEFINE_AUTOPTR_CLEANUP_FUNC(RCUReadAuto, rcu_read_auto_unlock)
+ void rcu_add_force_rcu_notifier(Notifier *n);
+ void rcu_remove_force_rcu_notifier(Notifier *n);
+ 
++/* This macro may be used to block constant propagation that lets the compiler
++ * detect a possible NULL dereference on a variable resulting from an explicit
++ * assignment in an impossible check. Sometimes a function is called which does
++ * safety checks and returns NULL if safe conditions are not met. The place
++ * where it's called cannot hit this condition and dereferencing the pointer
++ * without first checking it will make the compiler emit a warning about a
++ * "potential null pointer dereference" which is hard to work around. This
++ * macro "washes" the pointer and prevents the compiler from emitting tests
++ * branching to undefined instructions. It may only be used when the developer
++ * is absolutely certain that the conditions are guaranteed and that the
++ * pointer passed in argument cannot be NULL by design.
++ */
++#define ALREADY_CHECKED(p) do { asm("" : "=rm"(p) : "0"(p)); } while (0)
++
++/* same as above but to be used to pass the input value to the output but
++ * without letting the compiler know about its initial properties.
++ */
++#define DISGUISE(v) ({ typeof(v) __v = (v); ALREADY_CHECKED(__v); __v; })
++
+ #ifdef __cplusplus
+ }
+ #endif
+diff --git a/util/rcu.c b/util/rcu.c
+index b6d6c71cff..5dc0fa5f58 100644
+--- a/util/rcu.c
++++ b/util/rcu.c
+@@ -35,6 +35,13 @@
+ #if defined(CONFIG_MALLOC_TRIM)
+ #include <malloc.h>
+ #endif
++#if (defined(__GNU_LIBRARY__) && (__GLIBC__ > 2 || __GLIBC__ == 2 && __GLIBC_MINOR__ >= 8))
++#define GLIBC_HAVE_MALLOC_TRIM
++#endif
++/* glibc 2.33 provides mallinfo2() that overcomes mallinfo()'s type limitations */
++#if (defined(__GNU_LIBRARY__) && (__GLIBC__ > 2 || __GLIBC__ == 2 && __GLIBC_MINOR__ >= 33))
++#define GLIBC_HAVE_MALLINFO2
++#endif
+ 
+ /*
+  * Global grace period counter.  Bit 0 is always one in rcu_gp_ctr.
+@@ -50,6 +57,9 @@ static int in_drain_call_rcu;
+ static QemuMutex rcu_registry_lock;
+ static QemuMutex rcu_sync_lock;
+ 
++static int using_default_allocator = 1;
++static int(*my_mallctl)(const char *, void *, size_t *, void *, size_t) = NULL;
++
+ /*
+  * Check whether a quiescent state was crossed between the beginning of
+  * update_counter_and_wait and now.
+@@ -256,7 +266,8 @@ static void *call_rcu_thread(void *opaque)
+                 n = qatomic_read(&rcu_call_count);
+                 if (n == 0) {
+ #if defined(CONFIG_MALLOC_TRIM)
+-                    malloc_trim(4 * 1024 * 1024);
++		    if (!using_default_allocator)
++			malloc_trim(4 * 1024 * 1024);
+ #endif
+                     qemu_event_wait(&rcu_call_ready_event);
+                 }
+@@ -445,11 +456,74 @@ static void rcu_init_child(void)
+ }
+ #endif
+ 
++/* Tries to retrieve the address of the first occurrence symbol <name>.
++ * Note that NULL in return is not always an error as a symbol may have that
++ * address in special situations.
++ */
++static void *get_sym_curr_addr(const char *name)
++{
++	void *ptr = NULL;
++
++#ifdef RTLD_DEFAULT
++	if (!build_is_static)
++		ptr = dlsym(RTLD_DEFAULT, name);
++#endif
++	return ptr;
++}
++
++/* check if we're using the same allocator as the one that provides
++ * malloc_trim() and mallinfo(). The principle is that on glibc, both
++ * malloc_trim() and mallinfo() are provided, and using mallinfo() we
++ * can check if malloc() is performed through glibc or any other one
++ * the executable was linked against (e.g. jemalloc). Prior to this we
++ * have to check whether we're running on jemalloc by verifying if the
++ * mallctl() function is provided. Its pointer will be used later.
++ */
++static void detect_allocator(void)
++{
++#if defined(__ELF__)
++	extern int mallctl(const char *, void *, size_t *, void *, size_t) __attribute__((weak));
++
++	my_mallctl = mallctl;
++#endif
++	if (!my_mallctl)
++		my_mallctl = get_sym_curr_addr("mallctl");
++
++	using_default_allocator = (my_mallctl == NULL);
++
++	if (!my_mallctl) {
++#if defined(GLIBC_HAVE_MALLOC_TRIM)
++
++#ifdef GLIBC_HAVE_MALLINFO2
++		struct mallinfo2 mi1, mi2;
++		void *ptr;
++		mi1 = mallinfo2();
++		ptr = DISGUISE(malloc(1));
++		mi2 = mallinfo2();
++
++#else
++		struct mallinfo mi1, mi2;
++		void *ptr;
++		mi1 = mallinfo();
++		ptr = DISGUISE(malloc(1));
++		mi2 = mallinfo();
++
++#endif
++
++		free(DISGUISE(ptr));
++		using_default_allocator = !!memcmp(&mi1, &mi2, sizeof(mi1));
++#endif
++	}
++}
++
+ static void __attribute__((__constructor__)) rcu_init(void)
+ {
+     smp_mb_global_init();
+ #ifdef CONFIG_POSIX
+     pthread_atfork(rcu_init_lock, rcu_init_unlock, rcu_init_child);
++#endif
++#if defined(CONFIG_MALLOC_TRIM)
++    detect_allocator();
+ #endif
+     rcu_init_complete();
+ }
+-- 
+2.30.2
+
diff --git a/debian/patches/series b/debian/patches/series
index 4e8ddd6..7715e13 100644
--- a/debian/patches/series
+++ b/debian/patches/series
@@ -94,3 +94,4 @@ pve/0059-vma-create-support-64KiB-unaligned-input-images.patch
 pve/0060-vma-create-avoid-triggering-assertion-in-error-case.patch
 pve/0061-block-alloc-track-avoid-premature-break.patch
 pve/0062-PVE-Backup-allow-passing-max-workers-performance-set.patch
+pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch
\ No newline at end of file
-- 
2.30.2

[pve-devel] [PATCH qemu-server 1/1] qemu options: add tuning allocator

[Alexandre Derumier]

Add a new tuning option with allocator property.

Available values:
- Default is 'system', aka glibc malloc
- tcmalloc (improve performance ceph librbd)

Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
---
 PVE/QemuServer.pm | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/PVE/QemuServer.pm b/PVE/QemuServer.pm
index 40be44d..761ab48 100644
--- a/PVE/QemuServer.pm
+++ b/PVE/QemuServer.pm
@@ -289,6 +289,19 @@ my $meta_info_fmt = {
     },
 };
 
+my $tuning_fmt = {
+    allocator => {
+	type => 'string',
+	enum => [ qw(system tcmalloc) ],
+	default => 'system',
+	description => "Override the memory allocator used in QEMU via LD_PRELOAD.",
+	verbose_description => "Override the memory allocator used in QEMU via LD_PRELOAD."
+	    ." Using tcmalloc might improve performance if ceph librbd is used."
+	    ." NOTE: you must install the libtcmalloc-minimal4 package first!",
+	optional => 1,
+    },
+};
+
 my $confdesc = {
     onboot => {
 	optional => 1,
@@ -723,6 +736,12 @@ EODESCR
 	description => "List of host cores used to execute guest processes, for example: 0,5,8-11",
 	optional => 1,
     },
+    tuning=> {
+	type => 'string',
+	format => $tuning_fmt,
+	description => "Tune some special features.",
+	optional => 1,
+    },
 };
 
 my $cicustom_fmt = {
@@ -2192,6 +2211,16 @@ sub parse_rng {
     return $res;
 }
 
+sub parse_tuning {
+    my ($value) = @_;
+
+    return if !$value;
+
+    my $res = eval { parse_property_string($tuning_fmt, $value) };
+    warn $@ if $@;
+    return $res;
+}
+
 sub parse_meta_info {
     my ($value) = @_;
 
@@ -5909,6 +5938,12 @@ sub vm_start_nolock {
 
     my $run_qemu = sub {
 	PVE::Tools::run_fork sub {
+
+	    my $tuning = $conf->{tuning} ? parse_tuning($conf->{tuning}) : undef;
+
+	    $ENV{LD_PRELOAD} = "/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4"
+		if $tuning && $tuning->{allocator} eq 'tcmalloc';
+
 	    PVE::Systemd::enter_systemd_scope($vmid, "Proxmox VE VM $vmid", %systemd_properties);
 
 	    my $tpmpid;
-- 
2.30.2

[pve-devel] [PATCH pve-manager 1/1] ui: qemu : add tuning option

[Alexandre Derumier]

with memory allocator property

Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
---
 www/manager6/Makefile               |  1 +
 www/manager6/Utils.js               | 13 +++++++++
 www/manager6/form/TuningSelector.js | 41 +++++++++++++++++++++++++++++
 www/manager6/qemu/Options.js        | 14 ++++++++++
 4 files changed, 69 insertions(+)
 create mode 100644 www/manager6/form/TuningSelector.js

diff --git a/www/manager6/Makefile b/www/manager6/Makefile
index b73b729a..c29bc87a 100644
--- a/www/manager6/Makefile
+++ b/www/manager6/Makefile
@@ -67,6 +67,7 @@ JSSRC= 							\
 	form/StorageSelector.js				\
 	form/TFASelector.js				\
 	form/TokenSelector.js 				\
+	form/TuningSelector.js 				\
 	form/USBSelector.js				\
 	form/UserSelector.js				\
 	form/VLanField.js				\
diff --git a/www/manager6/Utils.js b/www/manager6/Utils.js
index 7bf3955a..3c395a5c 100644
--- a/www/manager6/Utils.js
+++ b/www/manager6/Utils.js
@@ -545,6 +545,19 @@ Ext.define('PVE.Utils', {
 	return output.join(', ');
     },
 
+    render_tuning: function(values) {
+	let props = PVE.Parser.parsePropertyString(values);
+	if (Ext.Object.isEmpty(props)) {
+	    return Proxmox.Utils.noneText;
+	}
+
+	let output = [];
+	if (props.allocator === 'tcmalloc') {
+	    output.push('Memory allocator: ' + props.allocator);
+	}
+	return output.join(', ');
+    },
+
     // fixme: auto-generate this
     // for now, please keep in sync with PVE::Tools::kvmkeymaps
     kvm_keymaps: {
diff --git a/www/manager6/form/TuningSelector.js b/www/manager6/form/TuningSelector.js
new file mode 100644
index 00000000..d967dc29
--- /dev/null
+++ b/www/manager6/form/TuningSelector.js
@@ -0,0 +1,41 @@
+Ext.define('PVE.form.TuningSelector', {
+    extend: 'Proxmox.panel.InputPanel',
+    alias: 'widget.pveTuningSelector',
+
+    viewModel: {},
+
+    items: [
+	{
+	    xtype: 'proxmoxKVComboBox',
+	    itemId: 'allocator',
+	    name: 'allocator',
+	    value: 'system',
+	    fieldLabel: 'Memory Allocator',
+	    comboItems: [
+		['system', 'system'],
+		['tcmalloc', 'tcmalloc'],
+	    ],
+	},
+    ],
+
+    onGetValues: function(values) {
+	var ret = {};
+
+	if (values.allocator !== "system") {
+	    ret.allocator = values.allocator;
+	}
+
+	if (Ext.Object.isEmpty(ret)) {
+	    return { 'delete': 'tuning' };
+	}
+	var tuning_props = PVE.Parser.printPropertyString(ret);
+	return { tuning: tuning_props };
+    },
+
+    setValues: function(values) {
+	if (values.tuning) {
+	    var tuning = PVE.Parser.parsePropertyString(values.tuning);
+	    this.callParent([tuning]);
+	}
+    },
+});
diff --git a/www/manager6/qemu/Options.js b/www/manager6/qemu/Options.js
index 7b112400..53972d18 100644
--- a/www/manager6/qemu/Options.js
+++ b/www/manager6/qemu/Options.js
@@ -338,6 +338,20 @@ Ext.define('PVE.qemu.Options', {
 		    },
 		} : undefined,
 	    },
+	    tuning: {
+		header: gettext('Tuning'),
+		defaultValue: false,
+		renderer: PVE.Utils.render_tuning,
+		editor: caps.vms['VM.Config.Options'] ? {
+		    xtype: 'proxmoxWindowEdit',
+		    subject: gettext('Tuning'),
+		    onlineHelp: 'chapter_virtual_machines', // FIXME: use 'qm_tuning' once available
+		    items: {
+			xtype: 'pveTuningSelector',
+			name: 'tuning',
+		    },
+		} : undefined,
+	    },
 	    hookscript: {
 		header: gettext('Hookscript'),
 	    },
-- 
2.30.2

[pve-devel] [PATCH-SERIES v2 qemu, qemu-server, manager 0/1] add tcmalloc support

[Alexandre Derumier]

This patches series add support for tcmalloc allocator with a new optionnal
tuning option in qemu-server.

tcmalloc improve librbd performance by 30% in latency and iops.

some performance bench results:

malloc: 60k iops 4k randread
tcmalloc : 90k iops 4k randread


a qemu patch is provided, to detect if a different memory allocator than
glibc malloc is loaded through LD_PRELOAD, to avoid the call of malloc_trim().
I don't have seen crash without this patch, but it's safer with it,
and as it's called 3-4x by second, it'll avoid some unnecessary calls.


changelog v2:

- update qemu patch for qemu 8.0

qemu:

Alexandre Derumier (1):
  add patch: add memory allocator detection and use malloc_trim for
    glibc only

 ...-and-call-malloc_trim-only-for-glibc.patch | 163 ++++++++++++++++++
 debian/patches/series                         |   1 +
 2 files changed, 164 insertions(+)
 create mode 100644 debian/patches/pve/0063-detect-allocator-and-call-malloc_trim-only-for-glibc.patch

qemu-server:

Alexandre Derumier (1):
  qemu options: add tuning allocator

 PVE/QemuServer.pm | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

pve-manager:

Alexandre Derumier (1):
  ui: qemu : add tuning option

-- 
2.30.2

[pve-devel] [PATCH pve-manager 1/1] ui: qemu : add tuning option

[Alexandre Derumier]

with memory allocator property

Signed-off-by: Alexandre Derumier <aderumier@odiso.com>
---
 www/manager6/Makefile               |  1 +
 www/manager6/Utils.js               | 13 +++++++++
 www/manager6/form/TuningSelector.js | 41 +++++++++++++++++++++++++++++
 www/manager6/qemu/Options.js        | 14 ++++++++++
 4 files changed, 69 insertions(+)
 create mode 100644 www/manager6/form/TuningSelector.js

diff --git a/www/manager6/Makefile b/www/manager6/Makefile
index b73b729a..c29bc87a 100644
--- a/www/manager6/Makefile
+++ b/www/manager6/Makefile
@@ -67,6 +67,7 @@ JSSRC= 							\
 	form/StorageSelector.js				\
 	form/TFASelector.js				\
 	form/TokenSelector.js 				\
+	form/TuningSelector.js 				\
 	form/USBSelector.js				\
 	form/UserSelector.js				\
 	form/VLanField.js				\
diff --git a/www/manager6/Utils.js b/www/manager6/Utils.js
index 7bf3955a..3c395a5c 100644
--- a/www/manager6/Utils.js
+++ b/www/manager6/Utils.js
@@ -545,6 +545,19 @@ Ext.define('PVE.Utils', {
 	return output.join(', ');
     },
 
+    render_tuning: function(values) {
+	let props = PVE.Parser.parsePropertyString(values);
+	if (Ext.Object.isEmpty(props)) {
+	    return Proxmox.Utils.noneText;
+	}
+
+	let output = [];
+	if (props.allocator === 'tcmalloc') {
+	    output.push('Memory allocator: ' + props.allocator);
+	}
+	return output.join(', ');
+    },
+
     // fixme: auto-generate this
     // for now, please keep in sync with PVE::Tools::kvmkeymaps
     kvm_keymaps: {
diff --git a/www/manager6/form/TuningSelector.js b/www/manager6/form/TuningSelector.js
new file mode 100644
index 00000000..d967dc29
--- /dev/null
+++ b/www/manager6/form/TuningSelector.js
@@ -0,0 +1,41 @@
+Ext.define('PVE.form.TuningSelector', {
+    extend: 'Proxmox.panel.InputPanel',
+    alias: 'widget.pveTuningSelector',
+
+    viewModel: {},
+
+    items: [
+	{
+	    xtype: 'proxmoxKVComboBox',
+	    itemId: 'allocator',
+	    name: 'allocator',
+	    value: 'system',
+	    fieldLabel: 'Memory Allocator',
+	    comboItems: [
+		['system', 'system'],
+		['tcmalloc', 'tcmalloc'],
+	    ],
+	},
+    ],
+
+    onGetValues: function(values) {
+	var ret = {};
+
+	if (values.allocator !== "system") {
+	    ret.allocator = values.allocator;
+	}
+
+	if (Ext.Object.isEmpty(ret)) {
+	    return { 'delete': 'tuning' };
+	}
+	var tuning_props = PVE.Parser.printPropertyString(ret);
+	return { tuning: tuning_props };
+    },
+
+    setValues: function(values) {
+	if (values.tuning) {
+	    var tuning = PVE.Parser.parsePropertyString(values.tuning);
+	    this.callParent([tuning]);
+	}
+    },
+});
diff --git a/www/manager6/qemu/Options.js b/www/manager6/qemu/Options.js
index 7b112400..53972d18 100644
--- a/www/manager6/qemu/Options.js
+++ b/www/manager6/qemu/Options.js
@@ -338,6 +338,20 @@ Ext.define('PVE.qemu.Options', {
 		    },
 		} : undefined,
 	    },
+	    tuning: {
+		header: gettext('Tuning'),
+		defaultValue: false,
+		renderer: PVE.Utils.render_tuning,
+		editor: caps.vms['VM.Config.Options'] ? {
+		    xtype: 'proxmoxWindowEdit',
+		    subject: gettext('Tuning'),
+		    onlineHelp: 'chapter_virtual_machines', // FIXME: use 'qm_tuning' once available
+		    items: {
+			xtype: 'pveTuningSelector',
+			name: 'tuning',
+		    },
+		} : undefined,
+	    },
 	    hookscript: {
 		header: gettext('Hookscript'),
 	    },
-- 
2.30.2