* [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986
@ 2026-11-17 17:46 Denis V. Lunev
2026-11-17 17:46 ` [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986 Denis V. Lunev
` (2 more replies)
0 siblings, 3 replies; 6+ messages in thread
From: Denis V. Lunev @ 2026-11-17 17:46 UTC (permalink / raw)
To: svt-core; +Cc: andrey.drobyshev, den
HCI 8.1 guests run fork/exec heavy work about 10% slower than on 8.0.
perf-atomic showed it in all five 8.0 vs 8.1 comparisons on four HDD
nodes: unixbench execl -10.4% to -14.7%, syscall/sysexec (which are
really fork+exec) -8.6% to -12.1%, cpu-ins -5%. The host is not the
cause. 8.1 nova creates VMs with pc-q35-vz10.1.0 where 8.0 used
pc-q35-vz10.0.0, and the new machine type changes two things the guest
sees: CPUID leaf 2 reports 0xff, so glibc's cache queries at process
start cost twice the CPUID exits, and la57 is no longer hidden, so
Linux guests boot with 5-level paging.
The series turns both off in pc_vz_compat, which every Virtuozzo
machine type applies, so machine types added later keep them off as
well. The only type that changes now is pc-q35-vz10.1.0, which gets the
same CPU as pc-q35-vz10.0.0. All other VZ types, pc-i440fx-vz10.1.0
included, already had both off.
The trade-off: no Virtuozzo guest gets la57 on a fresh start, "-cpu
host" included, unless it asks for it with la57=on. Our guests do not
need a 57-bit address space.
This is for hci-8.1 only. hci-dev is identical to hci-8.1 at the moment,
so the patches apply there as is. The 8.0 QEMU defines pc-q35-vz10.1.0
too and keeps it unchanged, but nova on 8.0 pins pc-q35-vz10.0.0.
Tested with the qemu-kvm-10.1.0-20.vz10.23 SRPM rebuilt with v1 on
hw-hdd01 (HCI 8.1.0-227, Xeon Silver 4314). A minimal guest with the
perf guest kernel (6.12.0-211.74.2.el10_2), Icelake-Server, pinned to
one vCPU:
execl/s fork+exec/s insb/s CPUID exits per run
vz10.1.0, stock 8.1 3400 2675 217K 11.85M
vz10.1.0, patched 3890 2985 228K 6.83M
vz10.0.0, patched 3890 2990 229K 6.84M
The perf VM itself, hard rebooted through nova on the patched QEMU,
starts with la57 disabled in its live libvirt definition, boots with
48-bit virtual addresses and runs execl at ~3995/s against 3312/s
before (3667/s on the 8.0 node).
v2 sets exactly what v1 did: x-consistent-cache and x-no-la57-on-start,
read back over QMP for every pc-q35-vz* and pc-i440fx-vz* type, are the
same with v1 and v2.
RCA and all comparisons are in VSTOR-154986:
https://virtuozzo.atlassian.net/browse/VSTOR-154986
Changes since v1:
- move both properties into pc_vz_compat instead of a new
pc_vz_10_1_compat, so later machine types inherit them (Andrey)
- drop the version from the x-no-la57-on-start comment in cpu.h (Andrey)
- subjects now say "all Virtuozzo machine types"
Denis V. Lunev (2):
i386: turn off x-consistent-cache on all Virtuozzo machine types
#VSTOR-154986
i386: hide la57 at fresh start on all Virtuozzo machine types
#VSTOR-154986
hw/i386/pc.c | 7 ++-----
hw/i386/pc_piix.c | 2 +-
hw/i386/pc_q35.c | 2 +-
include/hw/i386/pc.h | 3 ---
target/i386/cpu.h | 4 ++--
5 files changed, 6 insertions(+), 12 deletions(-)
base-commit: ac2ba825418b2726a7a4f367f8ef6d69d8563e47
--
2.55.0
^ permalink raw reply [flat|nested] 6+ messages in thread
* [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986
2026-11-17 17:46 [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Denis V. Lunev
@ 2026-11-17 17:46 ` Denis V. Lunev
2026-11-17 17:48 ` no-reply
2026-11-17 17:46 ` [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start " Denis V. Lunev
2026-11-17 18:04 ` [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Andrey Drobyshev
2 siblings, 1 reply; 6+ messages in thread
From: Denis V. Lunev @ 2026-11-17 17:46 UTC (permalink / raw)
To: svt-core; +Cc: andrey.drobyshev, den
From: Denis V. Lunev <den@openvz.org>
pc-q35-vz10.1.0 is based on rhel10.2.0 and so takes the upstream default
of x-consistent-cache from 6699e5dc864 ("i386/cpu: Present same cache
model in CPUID 0x2 & 0x4"). CPUID leaf 2 then reports 0xff, "use leaf
4", instead of legacy descriptors. glibc reads leaf 2 for each of its
cache queries at every process start, and with 0xff each read walks the
leaf 4 subleaves. Under KVM every CPUID is a VM exit, so a vz10.1.0
guest takes many more CPUID exits per exec than a vz10.0.0 one, and
fork/exec heavy workloads (unixbench execl, spawn and shell) regress.
Neither cache model describes the host, both are synthetic, so the new
one buys the guest nothing. Turn it off in pc_vz_compat, which every
Virtuozzo machine type applies, so the types added later keep it off
too. Only pc-q35-vz10.1.0 changes, the older types and
pc-i440fx-vz10.1.0 already get it off from their RHEL base.
The guest reads leaf 2 at boot and glibc at process start, so a running
vz10.1.0 guest migrated to a node with this change keeps working, new
processes just see the old descriptors.
Signed-off-by: Denis V. Lunev <den@openvz.org>
---
hw/i386/pc.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/hw/i386/pc.c b/hw/i386/pc.c
index 2834982dab4..40127f70269 100644
--- a/hw/i386/pc.c
+++ b/hw/i386/pc.c
@@ -545,6 +545,7 @@ GlobalProperty pc_vz_compat[] = {
{ TYPE_X86_CPU, "kvm-pv-tlb-flush", "on" },
{ TYPE_X86_CPU, "kvm-pv-sched-yield", "on" },
{ TYPE_X86_CPU, "kvm-asyncpf-int", "on" },
+ { TYPE_X86_CPU, "x-consistent-cache", "false" },
};
const size_t pc_vz_compat_len = G_N_ELEMENTS(pc_vz_compat);
--
2.55.0
^ permalink raw reply [flat|nested] 6+ messages in thread
* [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start on all Virtuozzo machine types #VSTOR-154986
2026-11-17 17:46 [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Denis V. Lunev
2026-11-17 17:46 ` [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986 Denis V. Lunev
@ 2026-11-17 17:46 ` Denis V. Lunev
2026-11-17 17:49 ` no-reply
2026-11-17 18:04 ` [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Andrey Drobyshev
2 siblings, 1 reply; 6+ messages in thread
From: Denis V. Lunev @ 2026-11-17 17:46 UTC (permalink / raw)
To: svt-core; +Cc: andrey.drobyshev, den
From: Denis V. Lunev <den@openvz.org>
x-no-la57-on-start covers pc-q35-vz10.0.0 and older and every
pc-i440fx type, but not pc-q35-vz10.1.0. A Linux guest built with
CONFIG_X86_5LEVEL switches to 5-level paging whenever the CPU reports
la57, whatever its RAM size, so on hosts with la57 (Ice Lake SP and
newer, EPYC 9004/9005) every vz10.1.0 guest gets the extra page table
level. That lengthens the KVM software walk of guest page tables used
by emulated string I/O, and adds to page table setup and teardown in
every exec.
Our guests do not need a 57-bit address space. Move the property into
pc_vz_compat, which every Virtuozzo machine type applies, so the types
added later hide la57 too and pc_vz_10_0_compat goes away. Only
pc-q35-vz10.1.0 changes. A guest that needs la57 gets it only with
la57=on, "-cpu host" included. A running guest keeps la57 on incoming
migration.
Signed-off-by: Denis V. Lunev <den@openvz.org>
---
hw/i386/pc.c | 6 +-----
hw/i386/pc_piix.c | 2 +-
hw/i386/pc_q35.c | 2 +-
include/hw/i386/pc.h | 3 ---
target/i386/cpu.h | 4 ++--
5 files changed, 5 insertions(+), 12 deletions(-)
diff --git a/hw/i386/pc.c b/hw/i386/pc.c
index 40127f70269..5e194f43d40 100644
--- a/hw/i386/pc.c
+++ b/hw/i386/pc.c
@@ -546,13 +546,9 @@ GlobalProperty pc_vz_compat[] = {
{ TYPE_X86_CPU, "kvm-pv-sched-yield", "on" },
{ TYPE_X86_CPU, "kvm-asyncpf-int", "on" },
{ TYPE_X86_CPU, "x-consistent-cache", "false" },
-};
-const size_t pc_vz_compat_len = G_N_ELEMENTS(pc_vz_compat);
-
-GlobalProperty pc_vz_10_0_compat[] = {
{ TYPE_X86_CPU, "x-no-la57-on-start", "on" },
};
-const size_t pc_vz_10_0_compat_len = G_N_ELEMENTS(pc_vz_10_0_compat);
+const size_t pc_vz_compat_len = G_N_ELEMENTS(pc_vz_compat);
GlobalProperty pc_vz_9_2_compat[] = {
{ TYPE_X86_CPU, "x-tsx-compat", "on" },
diff --git a/hw/i386/pc_piix.c b/hw/i386/pc_piix.c
index 768c6e47f53..b1b595078eb 100644
--- a/hw/i386/pc_piix.c
+++ b/hw/i386/pc_piix.c
@@ -1056,7 +1056,7 @@ DEFINE_I440FX_RHEL_MACHINE(7, 2, 0);
/* maj min mic basemt isdefault */
#define VZ_PC_MT(X) \
- X(10, 1, 0, MT_VER_SYM(10, 0, 0), true, pc_vz_10_0_compat) \
+ X(10, 1, 0, MT_VER_SYM(10, 0, 0), true) \
X(10, 0, 0, MT_VER_SYM(10, 0, 0), false) \
X(7, 16, 0, MT_VER_SYM(7, 6, 0), false, pc_vz_9_2_compat, \
hw_compat_vz_9_1, \
diff --git a/hw/i386/pc_q35.c b/hw/i386/pc_q35.c
index 51539f22399..019239bec6a 100644
--- a/hw/i386/pc_q35.c
+++ b/hw/i386/pc_q35.c
@@ -973,7 +973,7 @@ DEFINE_Q35_RHEL_MACHINE(7, 3, 0);
/* maj min mic basemt isdefault */
#define VZ_Q35_MT(X) \
X(10, 1, 0, MT_VER_SYM(10, 2, 0), true) \
- X(10, 0, 0, MT_VER_SYM(10, 0, 0), false, pc_vz_10_0_compat) \
+ X(10, 0, 0, MT_VER_SYM(10, 0, 0), false) \
X(9, 2, 0, MT_VER_SYM(9, 6, 0), false, pc_vz_9_2_compat) \
X(9, 1, 0, MT_VER_SYM(9, 4, 0), false, hw_compat_vz_9_1) \
X(9, 0, 0, MT_VER_SYM(9, 2, 0), false) \
diff --git a/include/hw/i386/pc.h b/include/hw/i386/pc.h
index 021b139b872..bd73426644e 100644
--- a/include/hw/i386/pc.h
+++ b/include/hw/i386/pc.h
@@ -364,9 +364,6 @@ extern const size_t pc_rhel_7_2_compat_len;
extern GlobalProperty pc_vz_compat[];
extern const size_t pc_vz_compat_len;
-extern GlobalProperty pc_vz_10_0_compat[];
-extern const size_t pc_vz_10_0_compat_len;
-
extern GlobalProperty pc_vz_9_2_compat[];
extern const size_t pc_vz_9_2_compat_len;
diff --git a/target/i386/cpu.h b/target/i386/cpu.h
index cbea9cf63d9..bf19556aa38 100644
--- a/target/i386/cpu.h
+++ b/target/i386/cpu.h
@@ -2343,8 +2343,8 @@ struct ArchCPU {
bool svm_features_compat;
/*
- * Virtuozzo machine types up to vz10.0.0: hide la57 unless requested
- * explicitly when the VM starts fresh, keep it on incoming migration.
+ * Virtuozzo machine types: hide la57 unless requested explicitly when
+ * the VM starts fresh, keep it on incoming migration.
*/
bool no_la57_on_start;
--
2.55.0
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986
2026-11-17 17:46 ` [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986 Denis V. Lunev
@ 2026-11-17 17:48 ` no-reply
0 siblings, 0 replies; 6+ messages in thread
From: no-reply @ 2026-11-17 17:48 UTC (permalink / raw)
To: Denis V. Lunev, svt-core; +Cc: andrey.drobyshev
> From: Denis V. Lunev <den@openvz.org>
>
> pc-q35-vz10.1.0 is based on rhel10.2.0 and so takes the upstream default
> of x-consistent-cache from 6699e5dc864 ("i386/cpu: Present same cache
> model in CPUID 0x2 & 0x4"). CPUID leaf 2 then reports 0xff, "use leaf
> 4", instead of legacy descriptors. glibc reads leaf 2 for each of its
> cache queries at every process start, and with 0xff each read walks the
> leaf 4 subleaves. Under KVM every CPUID is a VM exit, so a vz10.1.0
> guest takes many more CPUID exits per exec than a vz10.0.0 one, and
> fork/exec heavy workloads (unixbench execl, spawn and shell) regress.
>
> Neither cache model describes the host, both are synthetic, so the new
> one buys the guest nothing. Turn it off in pc_vz_compat, which every
> Virtuozzo machine type applies, so the types added later keep it off
> too. Only pc-q35-vz10.1.0 changes, the older types and
> pc-i440fx-vz10.1.0 already get it off from their RHEL base.
>
> The guest reads leaf 2 at boot and glibc at process start, so a running
> vz10.1.0 guest migrated to a node with this change keeps working, new
> processes just see the old descriptors.
>
> Signed-off-by: Denis V. Lunev <den@openvz.org>
Sashiko has reviewed this patch and found no issues.
--
Sashiko AI review · https://sashiko.vzint.dev/#/patchset/20261117174610.1689714-1-den@openvz.org?part=1
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start on all Virtuozzo machine types #VSTOR-154986
2026-11-17 17:46 ` [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start " Denis V. Lunev
@ 2026-11-17 17:49 ` no-reply
0 siblings, 0 replies; 6+ messages in thread
From: no-reply @ 2026-11-17 17:49 UTC (permalink / raw)
To: Denis V. Lunev, svt-core; +Cc: andrey.drobyshev
> From: Denis V. Lunev <den@openvz.org>
>
> x-no-la57-on-start covers pc-q35-vz10.0.0 and older and every
> pc-i440fx type, but not pc-q35-vz10.1.0. A Linux guest built with
> CONFIG_X86_5LEVEL switches to 5-level paging whenever the CPU reports
> la57, whatever its RAM size, so on hosts with la57 (Ice Lake SP and
> newer, EPYC 9004/9005) every vz10.1.0 guest gets the extra page table
> level. That lengthens the KVM software walk of guest page tables used
> by emulated string I/O, and adds to page table setup and teardown in
> every exec.
>
> Our guests do not need a 57-bit address space. Move the property into
> pc_vz_compat, which every Virtuozzo machine type applies, so the types
> added later hide la57 too and pc_vz_10_0_compat goes away. Only
> pc-q35-vz10.1.0 changes. A guest that needs la57 gets it only with
> la57=on, "-cpu host" included. A running guest keeps la57 on incoming
> migration.
>
> Signed-off-by: Denis V. Lunev <den@openvz.org>
Sashiko has reviewed this patch and found no issues.
--
Sashiko AI review · https://sashiko.vzint.dev/#/patchset/20261117174610.1689714-1-den@openvz.org?part=2
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986
2026-11-17 17:46 [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Denis V. Lunev
2026-11-17 17:46 ` [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986 Denis V. Lunev
2026-11-17 17:46 ` [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start " Denis V. Lunev
@ 2026-11-17 18:04 ` Andrey Drobyshev
2 siblings, 0 replies; 6+ messages in thread
From: Andrey Drobyshev @ 2026-11-17 18:04 UTC (permalink / raw)
To: Denis V. Lunev, svt-core
Ack
Applied to hci-8.1
On 11/17/26 7:46 PM, Denis V. Lunev wrote:
> HCI 8.1 guests run fork/exec heavy work about 10% slower than on 8.0.
> perf-atomic showed it in all five 8.0 vs 8.1 comparisons on four HDD
> nodes: unixbench execl -10.4% to -14.7%, syscall/sysexec (which are
> really fork+exec) -8.6% to -12.1%, cpu-ins -5%. The host is not the
> cause. 8.1 nova creates VMs with pc-q35-vz10.1.0 where 8.0 used
> pc-q35-vz10.0.0, and the new machine type changes two things the guest
> sees: CPUID leaf 2 reports 0xff, so glibc's cache queries at process
> start cost twice the CPUID exits, and la57 is no longer hidden, so
> Linux guests boot with 5-level paging.
>
> The series turns both off in pc_vz_compat, which every Virtuozzo
> machine type applies, so machine types added later keep them off as
> well. The only type that changes now is pc-q35-vz10.1.0, which gets the
> same CPU as pc-q35-vz10.0.0. All other VZ types, pc-i440fx-vz10.1.0
> included, already had both off.
>
> The trade-off: no Virtuozzo guest gets la57 on a fresh start, "-cpu
> host" included, unless it asks for it with la57=on. Our guests do not
> need a 57-bit address space.
>
> This is for hci-8.1 only. hci-dev is identical to hci-8.1 at the moment,
> so the patches apply there as is. The 8.0 QEMU defines pc-q35-vz10.1.0
> too and keeps it unchanged, but nova on 8.0 pins pc-q35-vz10.0.0.
>
> Tested with the qemu-kvm-10.1.0-20.vz10.23 SRPM rebuilt with v1 on
> hw-hdd01 (HCI 8.1.0-227, Xeon Silver 4314). A minimal guest with the
> perf guest kernel (6.12.0-211.74.2.el10_2), Icelake-Server, pinned to
> one vCPU:
>
> execl/s fork+exec/s insb/s CPUID exits per run
> vz10.1.0, stock 8.1 3400 2675 217K 11.85M
> vz10.1.0, patched 3890 2985 228K 6.83M
> vz10.0.0, patched 3890 2990 229K 6.84M
>
> The perf VM itself, hard rebooted through nova on the patched QEMU,
> starts with la57 disabled in its live libvirt definition, boots with
> 48-bit virtual addresses and runs execl at ~3995/s against 3312/s
> before (3667/s on the 8.0 node).
>
> v2 sets exactly what v1 did: x-consistent-cache and x-no-la57-on-start,
> read back over QMP for every pc-q35-vz* and pc-i440fx-vz* type, are the
> same with v1 and v2.
>
> RCA and all comparisons are in VSTOR-154986:
>
> https://virtuozzo.atlassian.net/browse/VSTOR-154986
>
> Changes since v1:
> - move both properties into pc_vz_compat instead of a new
> pc_vz_10_1_compat, so later machine types inherit them (Andrey)
> - drop the version from the x-no-la57-on-start comment in cpu.h (Andrey)
> - subjects now say "all Virtuozzo machine types"
>
> Denis V. Lunev (2):
> i386: turn off x-consistent-cache on all Virtuozzo machine types
> #VSTOR-154986
> i386: hide la57 at fresh start on all Virtuozzo machine types
> #VSTOR-154986
>
> hw/i386/pc.c | 7 ++-----
> hw/i386/pc_piix.c | 2 +-
> hw/i386/pc_q35.c | 2 +-
> include/hw/i386/pc.h | 3 ---
> target/i386/cpu.h | 4 ++--
> 5 files changed, 6 insertions(+), 12 deletions(-)
>
>
> base-commit: ac2ba825418b2726a7a4f367f8ef6d69d8563e47
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2026-11-17 18:04 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-17 17:46 [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Denis V. Lunev
2026-11-17 17:46 ` [PATCH hci-8.1 v2 1/2] i386: turn off x-consistent-cache on all Virtuozzo machine types #VSTOR-154986 Denis V. Lunev
2026-11-17 17:48 ` no-reply
2026-11-17 17:46 ` [PATCH hci-8.1 v2 2/2] i386: hide la57 at fresh start " Denis V. Lunev
2026-11-17 17:49 ` no-reply
2026-11-17 18:04 ` [PATCH hci-8.1 v2 0/2] i386: fix the exec slowdown of pc-q35-vz10.1.0 guests #VSTOR-154986 Andrey Drobyshev
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.