| Message ID | 1522431741-4678-1-git-send-email-mst@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Friday, March 30, 2018 7:43:58 PM CEST Michael S. Tsirkin wrote: > As documented by commit b488f021 "ACPI: restore comment justifying > 'extra' P_LVLx access", Linux does an extra IO read after entering idle > because on (some) chipsets STPCLK# doesn't get asserted in time > to prevent further instruction processing. > > This can never be the case on KVM, and a timer read causes an expensive > VM exit in turn causing useless load on host system. Detect KVM and skip > the read. TODO: whitelist more hypervisors? > > Note: very lightly tested. Pls don't apply this yet, I am working on a > _CST implementation for KVM and will repost this without the RFC tag > when it's been tested properly. > > Posting now for early flames/feedback. > > Cc: Paolo Bonzini <pbonzini@redhat.com> > Cc: Radim Krčmář <rkrcmar@redhat.com> > Cc: kvm@vger.kernel.org > Signed-off-by: Michael S. Tsirkin <mst@redhat.com> > --- > drivers/acpi/processor_idle.c | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c > index abb559c..8ae28dc 100644 > --- a/drivers/acpi/processor_idle.c > +++ b/drivers/acpi/processor_idle.c > @@ -33,6 +33,7 @@ > #include <linux/cpuidle.h> > #include <linux/cpu.h> > #include <acpi/processor.h> > +#include <linux/kvm_para.h> > > /* > * Include the apic definitions for x86 to have the APIC timer related defines > @@ -665,7 +666,8 @@ static void __cpuidle acpi_idle_do_entry(struct acpi_processor_cx *cx) > /* Dummy wait op - must do something useless after P_LVL2 read > because chipsets cannot guarantee that STPCLK# signal > gets asserted in time to freeze execution properly. */ > - inl(acpi_gbl_FADT.xpm_timer_block.address); > + if (!kvm_para_available()) > + inl(acpi_gbl_FADT.xpm_timer_block.address); > } > } > > @@ -687,7 +689,8 @@ static int acpi_idle_play_dead(struct cpuidle_device *dev, int index) > else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) { > inb(cx->address); > /* See comment in acpi_idle_do_entry() */ > - inl(acpi_gbl_FADT.xpm_timer_block.address); > + if (!kvm_para_available()) > + inl(acpi_gbl_FADT.xpm_timer_block.address); > } else > return -ENODEV; > } > While I have no objections to this change from the ACPI side, I'd like someone from the KVM land to comment on this. Thanks, Rafael
On 23/04/2018 10:07, Rafael J. Wysocki wrote: > On Friday, March 30, 2018 7:43:58 PM CEST Michael S. Tsirkin wrote: >> As documented by commit b488f021 "ACPI: restore comment justifying >> 'extra' P_LVLx access", Linux does an extra IO read after entering idle >> because on (some) chipsets STPCLK# doesn't get asserted in time >> to prevent further instruction processing. >> >> This can never be the case on KVM, and a timer read causes an expensive >> VM exit in turn causing useless load on host system. Detect KVM and skip >> the read. TODO: whitelist more hypervisors? >> >> Note: very lightly tested. Pls don't apply this yet, I am working on a >> _CST implementation for KVM and will repost this without the RFC tag >> when it's been tested properly. >> >> Posting now for early flames/feedback. >> >> Cc: Paolo Bonzini <pbonzini@redhat.com> >> Cc: Radim Krčmář <rkrcmar@redhat.com> >> Cc: kvm@vger.kernel.org >> Signed-off-by: Michael S. Tsirkin <mst@redhat.com> >> --- >> drivers/acpi/processor_idle.c | 7 +++++-- >> 1 file changed, 5 insertions(+), 2 deletions(-) >> >> diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c >> index abb559c..8ae28dc 100644 >> --- a/drivers/acpi/processor_idle.c >> +++ b/drivers/acpi/processor_idle.c >> @@ -33,6 +33,7 @@ >> #include <linux/cpuidle.h> >> #include <linux/cpu.h> >> #include <acpi/processor.h> >> +#include <linux/kvm_para.h> >> >> /* >> * Include the apic definitions for x86 to have the APIC timer related defines >> @@ -665,7 +666,8 @@ static void __cpuidle acpi_idle_do_entry(struct acpi_processor_cx *cx) >> /* Dummy wait op - must do something useless after P_LVL2 read >> because chipsets cannot guarantee that STPCLK# signal >> gets asserted in time to freeze execution properly. */ >> - inl(acpi_gbl_FADT.xpm_timer_block.address); >> + if (!kvm_para_available()) >> + inl(acpi_gbl_FADT.xpm_timer_block.address); >> } >> } >> >> @@ -687,7 +689,8 @@ static int acpi_idle_play_dead(struct cpuidle_device *dev, int index) >> else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) { >> inb(cx->address); >> /* See comment in acpi_idle_do_entry() */ >> - inl(acpi_gbl_FADT.xpm_timer_block.address); >> + if (!kvm_para_available()) >> + inl(acpi_gbl_FADT.xpm_timer_block.address); >> } else >> return -ENODEV; >> } >> > > While I have no objections to this change from the ACPI side, I'd like someone > from the KVM land to comment on this. It would be nicer to blacklist the problematic chipsets, or as Michael said have a way to whitelist more hypervisors, but this is certainly okay from KVM land too. Thanks, paolo
diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c index abb559c..8ae28dc 100644 --- a/drivers/acpi/processor_idle.c +++ b/drivers/acpi/processor_idle.c @@ -33,6 +33,7 @@ #include <linux/cpuidle.h> #include <linux/cpu.h> #include <acpi/processor.h> +#include <linux/kvm_para.h> /* * Include the apic definitions for x86 to have the APIC timer related defines @@ -665,7 +666,8 @@ static void __cpuidle acpi_idle_do_entry(struct acpi_processor_cx *cx) /* Dummy wait op - must do something useless after P_LVL2 read because chipsets cannot guarantee that STPCLK# signal gets asserted in time to freeze execution properly. */ - inl(acpi_gbl_FADT.xpm_timer_block.address); + if (!kvm_para_available()) + inl(acpi_gbl_FADT.xpm_timer_block.address); } } @@ -687,7 +689,8 @@ static int acpi_idle_play_dead(struct cpuidle_device *dev, int index) else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) { inb(cx->address); /* See comment in acpi_idle_do_entry() */ - inl(acpi_gbl_FADT.xpm_timer_block.address); + if (!kvm_para_available()) + inl(acpi_gbl_FADT.xpm_timer_block.address); } else return -ENODEV; }
As documented by commit b488f021 "ACPI: restore comment justifying 'extra' P_LVLx access", Linux does an extra IO read after entering idle because on (some) chipsets STPCLK# doesn't get asserted in time to prevent further instruction processing. This can never be the case on KVM, and a timer read causes an expensive VM exit in turn causing useless load on host system. Detect KVM and skip the read. TODO: whitelist more hypervisors? Note: very lightly tested. Pls don't apply this yet, I am working on a _CST implementation for KVM and will repost this without the RFC tag when it's been tested properly. Posting now for early flames/feedback. Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Radim Krčmář <rkrcmar@redhat.com> Cc: kvm@vger.kernel.org Signed-off-by: Michael S. Tsirkin <mst@redhat.com> --- drivers/acpi/processor_idle.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)