Hi,
I'd like some input on some rather aggressive optimisations I'm after.
This is a WIP series (so, don't pay attention to bisectability just yet) to
perform aggressive DCE based on vendor checks. I've added some bloat-o-meter
outputs for the everything-enabled vs an amd-only builds for comparison. The
results are very promising.
Please focus on patch 3, as that's the key to the optimisations.
What I'm ultimately after is what I've come to call the "single-vendor
optimisation". In doing this it's fairly easy to optimise other cases too, such
as removing unused vendors and any branches they use exclusively. Please, tell
me what you think of the approach taken by the series.
Single-vendor optimisation
==========================
- If we compile Xen for a single-vendor AND
Xen validates it always boots on such vendor AND
We remove the default CPU path AND
Xen sanitises guest CPU policies so they always match the boot vendor, THEN
It's fair game to fold all vendor checks onto true/false depending on
whether you're testing against the single vendor you compiled-in or not.
Effectively ignoring the runtime component, and allowing DCE to perform
eliminations left, right and center. The results are fairly dramatic in
places like init_speculation_mitigations(). See bloat-o-meter below.
Compiled-out-vendor elimination
===============================
- If we have all possible x86 vendors described in Kconfig AND
We compile out a vendor from Xen THEN
It should be possible to remove such vendor from every vendor check in the
system. Either we didn't boot it or we're using the default CPU ops anyway.
This causes &-checks to turn into ==- checks (more size-efficient) for
multi-vendor checks turning into single-vendor checks, or even straight-up
"false" when all vendors in the check are compiled out.
This is key to enable transparent DCE-removals in multi-vendor settings.
This allows very transparent removal of all Hygon, Centaur and Shanghai code
in the system when it's not required, as it's most typically found in branches
around common x86 code gated by a runtime vendor check.
The secret sauce is a static always_inline function that operates on the
assumption that the second argument is always a constant and the first always
a variable (as it invairably is the case with these checks). It's written
carefully to fold everything into a constant for the appropriate cases. There's
a few more optimisations in the code, like a no-vendor optimisation where the
checks always return false when comparing against a non-unknown vendor when
no explicit vendor is compiled in.
Patch 1 adds the missing vendors and the default path to Kconfig, as they are
currently absent.
Patch 2 ensures consistency between host and guest CPU policies wrt CPU vendor
Patch 3 introduces the x86_vendor_is() function. This is the key.
Patch 4 migrates the early_cpu_init() vendor switch. It's a bit tricky because
it must be done compatibly with the single-vendor optimisation, but there's
nothing complex about it.
Patch 5 is simply code removal at the Makefile level for free thanks to DCE
Patches 6 through 11 is replacements of regular checks with x86_vendor_is(),
with 6 being the one with the most dramatic effect in the diffstat.
===============================================================================
Bloat-o-meter
===============================================================================
all-vendors+default-path
========================
add/remove: 0/1 grow/shrink: 12/10 up/down: 175/-266 (-91)
Function old new delta
start_vmx 1507 1582 +75
x86_cpu_policies_are_compatible 157 194 +37
xen_config_data 1479 1506 +27
early_cpu_init 948 958 +10
setup_apic_nmi_watchdog 977 986 +9
init_speculation_mitigations 9836 9841 +5
intel_mcheck_init 2398 2401 +3
set_cx_pminfo 1691 1693 +2
init_bsp_APIC 193 195 +2
guest_common_max_feature_adjustments 110 112 +2
disable_lapic_nmi_watchdog 119 121 +2
do_get_hw_residencies 1289 1290 +1
mce_firstbank 37 36 -1
mcheck_init 1227 1225 -2
hvm_vcpu_virtual_to_linear 631 628 -3
init_nonfatal_mce_checker 160 156 -4
domain_cpu_policy_changed 677 672 -5
recalculate_misc 898 890 -8
traps_init 543 527 -16
default_cpu 16 - -16
cpufreq_driver_init 468 441 -27
vmce_wrmsr 993 909 -84
vmce_rdmsr 1134 1034 -100
Total: Before=3726243, After=3726152, chg -0.00%
amd-only+no-default-path
========================
add/remove: 0/14 grow/shrink: 4/58 up/down: 93/-10948 (-10855)
Function old new delta
x86_cpu_policies_are_compatible 157 194 +37
amd_check_entrysign 807 829 +22
init_guest_cpu_policies 1364 1382 +18
xen_config_data 1471 1487 +16
opt_gds_mit 1 - -1
nmi_p6_event_width 4 - -4
nmi_p4_cccr_val 4 - -4
init_e820 1037 1033 -4
pci_cfg_ok 307 301 -6
get_hw_residencies 213 205 -8
recalculate_cpuid_policy 909 900 -9
init_amd 2508 2499 -9
dom0_setup_permissions 3809 3800 -9
arch_ioreq_server_get_type_addr 250 241 -9
cpu_has_amd_erratum 230 219 -11
parse_spec_ctrl 2321 2307 -14
amd_nonfatal_mcheck_init 192 177 -15
shanghai_cpu_dev 16 - -16
hygon_cpu_dev 16 - -16
default_init 16 - -16
default_cpu 16 - -16
centaur_cpu_dev 16 - -16
x86emul_0fae 2758 2741 -17
vmce_init_vcpu 153 136 -17
cpufreq_cpu_init 34 15 -19
nmi_watchdog_tick 534 514 -20
vmce_restore_vcpu 160 139 -21
init_nonfatal_mce_checker 142 120 -22
ucode_update_hcall_cont 888 865 -23
mce_firstbank 37 10 -27
init_shanghai 29 - -29
validate_gl4e 617 587 -30
l4e_propagate_from_guest 451 421 -30
guest_walk_tables_4_levels 3411 3381 -30
clear_msr_range 30 - -30
acpi_dead_idle 430 398 -32
print_mtrr_state 719 684 -35
amd_mcheck_init 451 416 -35
hvm_vcpu_virtual_to_linear 631 595 -36
do_IRQ 1783 1747 -36
init_bsp_APIC 193 149 -44
cpu_callback 4650 4600 -50
mc_memerr_dhandler 903 851 -52
mcheck_init 1187 1122 -65
microcode_nmi_callback 205 139 -66
disable_lapic_nmi_watchdog 119 49 -70
__start_xen 9448 9378 -70
alternative_instructions 154 82 -72
traps_init 543 468 -75
protmode_load_seg 1904 1829 -75
set_cx_pminfo 1691 1614 -77
init_intel_cacheinfo 1191 1111 -80
is_cpu_primary 93 - -93
do_mca 3181 3085 -96
guest_cpuid 2395 2292 -103
guest_common_max_feature_adjustments 110 - -110
read_msr 1431 1319 -112
x86emul_decode 12729 12597 -132
guest_common_default_feature_adjustments 232 62 -170
do_microcode_update 787 602 -185
cpufreq_driver_init 453 263 -190
vmce_wrmsr 967 768 -199
recalculate_misc 898 689 -209
vmce_rdmsr 1083 872 -211
early_cpu_init 948 732 -216
guest_wrmsr 2853 2622 -231
init_centaur 238 - -238
domain_cpu_policy_changed 677 408 -269
write_msr 1749 1465 -284
x86_emulate 222198 221891 -307
init_hygon 389 - -389
start_vmx 1507 1105 -402
guest_rdmsr 2308 1881 -427
setup_apic_nmi_watchdog 977 276 -701
do_get_hw_residencies 1289 9 -1280
init_speculation_mitigations 9714 6788 -2926
Total: Before=3679044, After=3668189, chg -0.30%
There's a few more patches needed to add conditional inclusion of amd.c and
intel.c at the Makefile level, but that can be done just as well. It just adds
5 patches worth of noise I don't want to discuss atm.
Just knowing x86_vendor_is() is "good to have" is good enough as it enables our
downstream to customise it with whatever optimisations we need.
I also suspect other areas of the hypervisor could benefit from this meld of
runtime+compiletime sort of checking, allowing transparent code removal.
I'm thinking DOM0LESS_BOOT vs DOM0_BOOT vs PVSHIM_BOOT, or AMD_SVM vs INTEL_VMX
in HVM-only builds, or family checks to have (i.e) a explicit "older-than-zen"
Kconfig option with a similar approach on a family range check.
This is maybe one of several such uses.
So... thoughts? I'm definitely fond of the single-vendor bloat-o-meter output.
Cheers,
Alejandro
Alejandro Vallejo (11):
x86: Add more granularity to the vendors in Kconfig
x86: Reject CPU policies with vendors other than the host's
x86: Add x86_vendor_is() by itself before using it
x86: Refactor early vendor lookup code to use x86_vendor_is()
x86: Conditionalise the inclusion of Hygon/Centaur/Shanghai cpu/ files
x86: Migrate switch-based vendor checks to x86_vendor_is()
x86: Migrate MSR handler vendor checks to x86_vendor_is()
x86: Migrate insn emulator to use x86_vendor_is()
x86: Migrate spec_ctrl vendor checks to x86_vendor_is()
x86: Migrate everything under cpu/ to use x86_vendor_is()
x86: Migrate every remaining raw vendor check to x86_vendor_is()
xen/arch/x86/Kconfig.cpu | 45 +++++++++++++++++++++
xen/arch/x86/acpi/cpu_idle.c | 19 ++++-----
xen/arch/x86/acpi/cpufreq/acpi.c | 2 +-
xen/arch/x86/acpi/cpufreq/cpufreq.c | 32 +++++----------
xen/arch/x86/alternative.c | 30 ++++++--------
xen/arch/x86/apic.c | 2 +-
xen/arch/x86/cpu-policy.c | 41 +++++++++----------
xen/arch/x86/cpu/Makefile | 6 +--
xen/arch/x86/cpu/amd.c | 6 +--
xen/arch/x86/cpu/common.c | 50 +++++++++++++++--------
xen/arch/x86/cpu/intel_cacheinfo.c | 5 +--
xen/arch/x86/cpu/mcheck/amd_nonfatal.c | 2 +-
xen/arch/x86/cpu/mcheck/mcaction.c | 3 +-
xen/arch/x86/cpu/mcheck/mce.c | 41 +++++--------------
xen/arch/x86/cpu/mcheck/mce.h | 20 +++++----
xen/arch/x86/cpu/mcheck/mce_amd.c | 6 +--
xen/arch/x86/cpu/mcheck/mce_intel.c | 6 +--
xen/arch/x86/cpu/mcheck/non-fatal.c | 20 +++------
xen/arch/x86/cpu/mcheck/vmce.c | 50 ++++++-----------------
xen/arch/x86/cpu/microcode/amd.c | 2 +-
xen/arch/x86/cpu/microcode/core.c | 2 +-
xen/arch/x86/cpu/mtrr/generic.c | 4 +-
xen/arch/x86/cpu/mwait-idle.c | 4 +-
xen/arch/x86/cpuid.c | 4 +-
xen/arch/x86/dom0_build.c | 3 +-
xen/arch/x86/domain.c | 37 ++++++++---------
xen/arch/x86/e820.c | 3 +-
xen/arch/x86/guest/xen/xen.c | 19 ++++-----
xen/arch/x86/hvm/hvm.c | 5 ++-
xen/arch/x86/hvm/ioreq.c | 2 +-
xen/arch/x86/hvm/vmx/vmx.c | 6 +--
xen/arch/x86/include/asm/cpuid.h | 56 +++++++++++++++++++++++++-
xen/arch/x86/include/asm/guest_pt.h | 4 +-
xen/arch/x86/irq.c | 4 +-
xen/arch/x86/msr.c | 41 ++++++++++---------
xen/arch/x86/nmi.c | 18 +++------
xen/arch/x86/pv/emul-priv-op.c | 24 +++++------
xen/arch/x86/setup.c | 2 +-
xen/arch/x86/spec_ctrl.c | 34 ++++++++--------
xen/arch/x86/traps-setup.c | 18 ++++-----
xen/arch/x86/x86_emulate/private.h | 4 +-
xen/arch/x86/x86_emulate/x86_emulate.c | 2 +-
xen/lib/x86/policy.c | 3 +-
43 files changed, 360 insertions(+), 327 deletions(-)
base-commit: fb0e37df71a31318c61e0715ffed3e149ca8a4aa
--
2.43.0
