This patchset enables the cpuidle-haltpoll driver and its namesake
governor on arm64. This is specifically interesting for KVM guests by
reducing the IPC latencies.
Comparing idle switching latencies on an arm64 KVM guest with
perf bench sched pipe:
usecs/op %stdev
no haltpoll (baseline) 13.48 +- 5.19%
with haltpoll 6.84 +- 22.07%
No change in performance for a similar test on x86:
usecs/op %stdev
haltpoll w/ cpu_relax() (baseline) 4.75 +- 1.76%
haltpoll w/ smp_cond_load_relaxed() 4.78 +- 2.31%
Both sets of tests were on otherwise idle systems with guest VCPUs
pinned to specific PCPUs. One reason for the higher stdev on arm64
is that trapping of the WFE instruction by the host KVM is contingent
on the number of tasks on the runqueue.
The patch series is organized in four parts:
- patches 1, 2 mangle the config option ARCH_HAS_CPU_RELAX, renaming
and moving it from x86 to common architectural code.
- next, patches 3-5, reorganize the haltpoll selection and init logic
to allow architecture code to select it.
- patch 6, reorganizes the poll_idle() loop, switching from using
cpu_relax() directly to smp_cond_load_relaxed().
- and finally, patches 7-9, add the bits for arm64 support.
What is still missing: this series largely completes the haltpoll side
of functionality for arm64. There are, however, a few related areas
that still need to be threshed out:
- WFET support: WFE on arm64 does not guarantee that poll_idle()
would terminate in halt_poll_ns. Using WFET would address this.
- KVM_NO_POLL support on arm64
- KVM TWED support on arm64: allow the host to limit time spent in
WFE.
Changelog:
v5:
- rework the poll_idle() loop around smp_cond_load_relaxed() (review
comment from Tomohiro Misono.)
- also rework selection of cpuidle-haltpoll. Now selected based
on the architectural selection of ARCH_CPUIDLE_HALTPOLL.
- arch_haltpoll_supported() (renamed from arch_haltpoll_want()) on
arm64 now depends on the event-stream being enabled.
- limit POLL_IDLE_RELAX_COUNT on arm64 (review comment from Haris Okanovic)
- ARCH_HAS_CPU_RELAX is now renamed to ARCH_HAS_OPTIMIZED_POLL.
v4 changes from v3:
- change 7/8 per Rafael input: drop the parens and use ret for the final check
- add 8/8 which renames the guard for building poll_state
v3 changes from v2:
- fix 1/7 per Petr Mladek - remove ARCH_HAS_CPU_RELAX from arch/x86/Kconfig
- add Ack-by from Rafael Wysocki on 2/7
v2 changes from v1:
- added patch 7 where we change cpu_relax with smp_cond_load_relaxed per PeterZ
(this improves by 50% at least the CPU cycles consumed in the tests above:
10,716,881,137 now vs 14,503,014,257 before)
- removed the ifdef from patch 1 per RafaelW
Ankur Arora (4):
cpuidle: rename ARCH_HAS_CPU_RELAX to ARCH_HAS_OPTIMIZED_POLL
cpuidle-haltpoll: condition on ARCH_CPUIDLE_HALTPOLL
arm64: support cpuidle-haltpoll
cpuidle/poll_state: limit POLL_IDLE_RELAX_COUNT on arm64
Joao Martins (4):
Kconfig: move ARCH_HAS_OPTIMIZED_POLL to arch/Kconfig
cpuidle-haltpoll: define arch_haltpoll_supported()
governors/haltpoll: drop kvm_para_available() check
arm64: define TIF_POLLING_NRFLAG
Mihai Carabas (1):
cpuidle/poll_state: poll via smp_cond_load_relaxed()
arch/Kconfig | 3 +++
arch/arm64/Kconfig | 10 ++++++++++
arch/arm64/include/asm/cpuidle_haltpoll.h | 21 +++++++++++++++++++++
arch/arm64/include/asm/thread_info.h | 2 ++
arch/x86/Kconfig | 4 +---
arch/x86/include/asm/cpuidle_haltpoll.h | 1 +
arch/x86/kernel/kvm.c | 10 ++++++++++
drivers/acpi/processor_idle.c | 4 ++--
drivers/cpuidle/Kconfig | 5 ++---
drivers/cpuidle/Makefile | 2 +-
drivers/cpuidle/cpuidle-haltpoll.c | 9 ++-------
drivers/cpuidle/governors/haltpoll.c | 6 +-----
drivers/cpuidle/poll_state.c | 21 ++++++++++++++++-----
drivers/idle/Kconfig | 1 +
include/linux/cpuidle.h | 2 +-
include/linux/cpuidle_haltpoll.h | 5 +++++
16 files changed, 79 insertions(+), 27 deletions(-)
create mode 100644 arch/arm64/include/asm/cpuidle_haltpoll.h
--
2.39.3
smp_cond_load_relaxed(), in its generic polling variant polls on the
loop condition, waiting for it to change, eventually exiting the loop
if the time limit has been exceeded.
To limit the frequency of the time check it is done only once every
POLL_IDLE_RELAX_COUNT iterations.
arm64, however uses an event based mechanism, where instead of polling,
we wait for store to a region.
Limit the POLL_IDLE_RELAX_COUNT to 1 for that case.
Suggested-by: Haris Okanovic <[email protected]>
Signed-off-by: Ankur Arora <[email protected]>
---
drivers/cpuidle/poll_state.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
index 532e4ed19e0f..b69fe7b67cb4 100644
--- a/drivers/cpuidle/poll_state.c
+++ b/drivers/cpuidle/poll_state.c
@@ -8,7 +8,18 @@
#include <linux/sched/clock.h>
#include <linux/sched/idle.h>
+#ifdef CONFIG_ARM64
+/*
+ * POLL_IDLE_RELAX_COUNT determines how often we check for timeout
+ * while polling for TIF_NEED_RESCHED in thread_info->flags.
+ *
+ * Set this to a low value since arm64, instead of polling, uses a
+ * event based mechanism.
+ */
+#define POLL_IDLE_RELAX_COUNT 1
+#else
#define POLL_IDLE_RELAX_COUNT 200
+#endif
static int __cpuidle poll_idle(struct cpuidle_device *dev,
struct cpuidle_driver *drv, int index)
--
2.39.3
The cpuidle-haltpoll driver and its namesake governor are selected
under KVM_GUEST on X86. In addition, KVM_GUEST also selects
ARCH_CPUIDLE_HALTPOLL and defines the requisite
arch_haltpoll_{enable,disable}() functions.
So remove the explicit dependence on KVM_GUEST, and instead use
ARCH_CPUIDLE_HALTPOLL as proxy for architectural support for
haltpoll.
While at it, change "halt poll" to "haltpoll" in one of the summary
clauses, since the second form is used everywhere else.
Signed-off-by: Ankur Arora <[email protected]>
---
drivers/cpuidle/Kconfig | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
diff --git a/drivers/cpuidle/Kconfig b/drivers/cpuidle/Kconfig
index 75f6e176bbc8..c1bebadf22bc 100644
--- a/drivers/cpuidle/Kconfig
+++ b/drivers/cpuidle/Kconfig
@@ -35,7 +35,6 @@ config CPU_IDLE_GOV_TEO
config CPU_IDLE_GOV_HALTPOLL
bool "Haltpoll governor (for virtualized systems)"
- depends on KVM_GUEST
help
This governor implements haltpoll idle state selection, to be
used in conjunction with the haltpoll cpuidle driver, allowing
@@ -72,8 +71,8 @@ source "drivers/cpuidle/Kconfig.riscv"
endmenu
config HALTPOLL_CPUIDLE
- tristate "Halt poll cpuidle driver"
- depends on X86 && KVM_GUEST && ARCH_HAS_OPTIMIZED_POLL
+ tristate "Haltpoll cpuidle driver"
+ depends on ARCH_CPUIDLE_HALTPOLL && ARCH_HAS_OPTIMIZED_POLL
select CPU_IDLE_GOV_HALTPOLL
default y
help
--
2.39.3
From: Joao Martins <[email protected]>
Right now kvm_para_has_hint(KVM_HINTS_REALTIME) is x86 only. In
pursuit of making cpuidle-haltpoll architecture independent, define
arch_haltpoll_supported() which handles the architectural check for
enabling haltpoll.
Move the (kvm_para_available() && kvm_para_has_hint(KVM_HINTS_REALTIME))
check to the x86 specific arch_haltpoll_supported().
Signed-off-by: Joao Martins <[email protected]>
Signed-off-by: Mihai Carabas <[email protected]>
Signed-off-by: Ankur Arora <[email protected]>
---
Changelog:
- s/arch_haltpoll_want/arch_haltpoll_supported/
- change the check in haltpoll_want() from:
(kvm_para_available() && arch_haltpoll_want()) || force;
to
arch_haltpoll_supported() || force;
Dropped Rafael's acked-by due to these changes.
---
arch/x86/include/asm/cpuidle_haltpoll.h | 1 +
arch/x86/kernel/kvm.c | 10 ++++++++++
drivers/cpuidle/cpuidle-haltpoll.c | 9 ++-------
include/linux/cpuidle_haltpoll.h | 5 +++++
4 files changed, 18 insertions(+), 7 deletions(-)
diff --git a/arch/x86/include/asm/cpuidle_haltpoll.h b/arch/x86/include/asm/cpuidle_haltpoll.h
index c8b39c6716ff..43ce79b88662 100644
--- a/arch/x86/include/asm/cpuidle_haltpoll.h
+++ b/arch/x86/include/asm/cpuidle_haltpoll.h
@@ -4,5 +4,6 @@
void arch_haltpoll_enable(unsigned int cpu);
void arch_haltpoll_disable(unsigned int cpu);
+bool arch_haltpoll_supported(void);
#endif
diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 7f0732bc0ccd..e4dcbe9acc07 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -1151,4 +1151,14 @@ void arch_haltpoll_disable(unsigned int cpu)
smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
}
EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
+
+bool arch_haltpoll_supported(void)
+{
+ /* Do not load haltpoll if idle= is passed */
+ if (boot_option_idle_override != IDLE_NO_OVERRIDE)
+ return false;
+
+ return kvm_para_available() && kvm_para_has_hint(KVM_HINTS_REALTIME);
+}
+EXPORT_SYMBOL_GPL(arch_haltpoll_supported);
#endif
diff --git a/drivers/cpuidle/cpuidle-haltpoll.c b/drivers/cpuidle/cpuidle-haltpoll.c
index d8515d5c0853..70f585383171 100644
--- a/drivers/cpuidle/cpuidle-haltpoll.c
+++ b/drivers/cpuidle/cpuidle-haltpoll.c
@@ -15,7 +15,6 @@
#include <linux/cpuidle.h>
#include <linux/module.h>
#include <linux/sched/idle.h>
-#include <linux/kvm_para.h>
#include <linux/cpuidle_haltpoll.h>
static bool force __read_mostly;
@@ -95,7 +94,7 @@ static void haltpoll_uninit(void)
static bool haltpoll_want(void)
{
- return kvm_para_has_hint(KVM_HINTS_REALTIME) || force;
+ return arch_haltpoll_supported() || force;
}
static int __init haltpoll_init(void)
@@ -103,11 +102,7 @@ static int __init haltpoll_init(void)
int ret;
struct cpuidle_driver *drv = &haltpoll_driver;
- /* Do not load haltpoll if idle= is passed */
- if (boot_option_idle_override != IDLE_NO_OVERRIDE)
- return -ENODEV;
-
- if (!kvm_para_available() || !haltpoll_want())
+ if (!haltpoll_want())
return -ENODEV;
cpuidle_poll_state_init(drv);
diff --git a/include/linux/cpuidle_haltpoll.h b/include/linux/cpuidle_haltpoll.h
index d50c1e0411a2..a3caf01d3f0e 100644
--- a/include/linux/cpuidle_haltpoll.h
+++ b/include/linux/cpuidle_haltpoll.h
@@ -12,5 +12,10 @@ static inline void arch_haltpoll_enable(unsigned int cpu)
static inline void arch_haltpoll_disable(unsigned int cpu)
{
}
+
+static inline bool arch_haltpoll_supported(void)
+{
+ return false;
+}
#endif
#endif
--
2.39.3
ARCH_HAS_CPU_RELAX is a bit of a misnomer since all architectures
define cpu_relax(). Not all, however, have a performant version, with
some only implementing it as a compiler barrier.
In contexts that this config option is used, it is expected to provide
an architectural primitive that can be used as part of a polling
mechanism -- one that would be cheaper than spinning in a tight loop.
Advertise the availability of such a primitive by renaming to
ARCH_HAS_OPTIMIZED_POLL. And, while at it, explicitly condition
cpuidle-haltpoll and intel-idle, both of which depend on a polling
state, on it.
Suggested-by: Will Deacon <[email protected]>
Signed-off-by: Ankur Arora <[email protected]>
---
arch/x86/Kconfig | 2 +-
drivers/acpi/processor_idle.c | 4 ++--
drivers/cpuidle/Kconfig | 2 +-
drivers/cpuidle/Makefile | 2 +-
drivers/idle/Kconfig | 1 +
include/linux/cpuidle.h | 2 +-
6 files changed, 7 insertions(+), 6 deletions(-)
diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 4474bf32d0a4..b238c874875a 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -368,7 +368,7 @@ config ARCH_MAY_HAVE_PC_FDC
config GENERIC_CALIBRATE_DELAY
def_bool y
-config ARCH_HAS_CPU_RELAX
+config ARCH_HAS_OPTIMIZED_POLL
def_bool y
config ARCH_HIBERNATION_POSSIBLE
diff --git a/drivers/acpi/processor_idle.c b/drivers/acpi/processor_idle.c
index bd6a7857ce05..ccef38410950 100644
--- a/drivers/acpi/processor_idle.c
+++ b/drivers/acpi/processor_idle.c
@@ -36,7 +36,7 @@
#include <asm/cpu.h>
#endif
-#define ACPI_IDLE_STATE_START (IS_ENABLED(CONFIG_ARCH_HAS_CPU_RELAX) ? 1 : 0)
+#define ACPI_IDLE_STATE_START (IS_ENABLED(CONFIG_ARCH_HAS_OPTIMIZED_POLL) ? 1 : 0)
static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
module_param(max_cstate, uint, 0400);
@@ -787,7 +787,7 @@ static int acpi_processor_setup_cstates(struct acpi_processor *pr)
if (max_cstate == 0)
max_cstate = 1;
- if (IS_ENABLED(CONFIG_ARCH_HAS_CPU_RELAX)) {
+ if (IS_ENABLED(CONFIG_ARCH_HAS_OPTIMIZED_POLL)) {
cpuidle_poll_state_init(drv);
count = 1;
} else {
diff --git a/drivers/cpuidle/Kconfig b/drivers/cpuidle/Kconfig
index cac5997dca50..75f6e176bbc8 100644
--- a/drivers/cpuidle/Kconfig
+++ b/drivers/cpuidle/Kconfig
@@ -73,7 +73,7 @@ endmenu
config HALTPOLL_CPUIDLE
tristate "Halt poll cpuidle driver"
- depends on X86 && KVM_GUEST
+ depends on X86 && KVM_GUEST && ARCH_HAS_OPTIMIZED_POLL
select CPU_IDLE_GOV_HALTPOLL
default y
help
diff --git a/drivers/cpuidle/Makefile b/drivers/cpuidle/Makefile
index d103342b7cfc..f29dfd1525b0 100644
--- a/drivers/cpuidle/Makefile
+++ b/drivers/cpuidle/Makefile
@@ -7,7 +7,7 @@ obj-y += cpuidle.o driver.o governor.o sysfs.o governors/
obj-$(CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED) += coupled.o
obj-$(CONFIG_DT_IDLE_STATES) += dt_idle_states.o
obj-$(CONFIG_DT_IDLE_GENPD) += dt_idle_genpd.o
-obj-$(CONFIG_ARCH_HAS_CPU_RELAX) += poll_state.o
+obj-$(CONFIG_ARCH_HAS_OPTIMIZED_POLL) += poll_state.o
obj-$(CONFIG_HALTPOLL_CPUIDLE) += cpuidle-haltpoll.o
##################################################################################
diff --git a/drivers/idle/Kconfig b/drivers/idle/Kconfig
index 6707d2539fc4..6f9b1d48fede 100644
--- a/drivers/idle/Kconfig
+++ b/drivers/idle/Kconfig
@@ -4,6 +4,7 @@ config INTEL_IDLE
depends on CPU_IDLE
depends on X86
depends on CPU_SUP_INTEL
+ depends on ARCH_HAS_OPTIMIZED_POLL
help
Enable intel_idle, a cpuidle driver that includes knowledge of
native Intel hardware idle features. The acpi_idle driver
diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
index 3183aeb7f5b4..7e7e58a17b07 100644
--- a/include/linux/cpuidle.h
+++ b/include/linux/cpuidle.h
@@ -275,7 +275,7 @@ static inline void cpuidle_coupled_parallel_barrier(struct cpuidle_device *dev,
}
#endif
-#if defined(CONFIG_CPU_IDLE) && defined(CONFIG_ARCH_HAS_CPU_RELAX)
+#if defined(CONFIG_CPU_IDLE) && defined(CONFIG_ARCH_HAS_OPTIMIZED_POLL)
void cpuidle_poll_state_init(struct cpuidle_driver *drv);
#else
static inline void cpuidle_poll_state_init(struct cpuidle_driver *drv) {}
--
2.39.3
From: Mihai Carabas <[email protected]>
The inner loop in poll_idle() polls up to POLL_IDLE_RELAX_COUNT times,
checking to see if the thread has the TIF_NEED_RESCHED bit set. The
loop exits once the condition is met, or if the poll time limit has
been exceeded.
The time check is done only infrequently (once in POLL_IDLE_RELAX_COUNT
iterations) so as to minimize the number of instructions executed in
each iteration. In addition, each loop iteration executes cpu_relax()
which on certain platforms provides a hint to the pipeline that the
loop is busy-waiting, thus allowing the processor to reduce power
consumption.
However, cpu_relax() is not defined optimally everywhere. In particular,
on arm64, it is implemented as a YIELD which merely serves as a hint to
prefer a different hardware thread if one is available.
arm64 exposes a better mechanism via smp_cond_load_relaxed() which uses
LDXR, WFE where the LDXR loads a memory region in exclusive state and
the WFE waits for any stores to the region.
So restructure the loop and fold both checks in smp_cond_load_relaxed().
Also, move the time check to the head of the loop so, once
TIF_NEED_RESCHED is set, we exit straight-away without doing an
unnecessary time check.
Suggested-by: Peter Zijlstra <[email protected]>
Signed-off-by: Mihai Carabas <[email protected]>
Signed-off-by: Ankur Arora <[email protected]>
---
Changelog:
- reorganized the loop to keep the original poll_idle() structure.
---
drivers/cpuidle/poll_state.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
index 9b6d90a72601..532e4ed19e0f 100644
--- a/drivers/cpuidle/poll_state.c
+++ b/drivers/cpuidle/poll_state.c
@@ -21,21 +21,21 @@ static int __cpuidle poll_idle(struct cpuidle_device *dev,
raw_local_irq_enable();
if (!current_set_polling_and_test()) {
- unsigned int loop_count = 0;
+ unsigned int loop_count;
u64 limit;
limit = cpuidle_poll_time(drv, dev);
while (!need_resched()) {
- cpu_relax();
- if (loop_count++ < POLL_IDLE_RELAX_COUNT)
- continue;
-
loop_count = 0;
if (local_clock_noinstr() - time_start > limit) {
dev->poll_time_limit = true;
break;
}
+
+ smp_cond_load_relaxed(¤t_thread_info()->flags,
+ VAL & _TIF_NEED_RESCHED ||
+ loop_count++ >= POLL_IDLE_RELAX_COUNT);
}
}
raw_local_irq_disable();
--
2.39.3
From: Joao Martins <[email protected]>
Commit 842514849a61 ("arm64: Remove TIF_POLLING_NRFLAG") had removed
TIF_POLLING_NRFLAG because arm64 only supported non-polled idling via
cpu_do_idle().
To add support for polling via cpuidle-haltpoll, we want to use the
standard poll_idle() interface, which sets TIF_POLLING_NRFLAG while
polling.
Reuse the same bit to define TIF_POLLING_NRFLAG.
Signed-off-by: Joao Martins <[email protected]>
Signed-off-by: Mihai Carabas <[email protected]>
Signed-off-by: Ankur Arora <[email protected]>
---
arch/arm64/include/asm/thread_info.h | 2 ++
1 file changed, 2 insertions(+)
diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h
index e72a3bf9e563..23ff72168e48 100644
--- a/arch/arm64/include/asm/thread_info.h
+++ b/arch/arm64/include/asm/thread_info.h
@@ -69,6 +69,7 @@ void arch_setup_new_exec(void);
#define TIF_SYSCALL_TRACEPOINT 10 /* syscall tracepoint for ftrace */
#define TIF_SECCOMP 11 /* syscall secure computing */
#define TIF_SYSCALL_EMU 12 /* syscall emulation active */
+#define TIF_POLLING_NRFLAG 16 /* set while polling in poll_idle() */
#define TIF_MEMDIE 18 /* is terminating due to OOM killer */
#define TIF_FREEZE 19
#define TIF_RESTORE_SIGMASK 20
@@ -91,6 +92,7 @@ void arch_setup_new_exec(void);
#define _TIF_SYSCALL_TRACEPOINT (1 << TIF_SYSCALL_TRACEPOINT)
#define _TIF_SECCOMP (1 << TIF_SECCOMP)
#define _TIF_SYSCALL_EMU (1 << TIF_SYSCALL_EMU)
+#define _TIF_POLLING_NRFLAG (1 << TIF_POLLING_NRFLAG)
#define _TIF_UPROBE (1 << TIF_UPROBE)
#define _TIF_SINGLESTEP (1 << TIF_SINGLESTEP)
#define _TIF_32BIT (1 << TIF_32BIT)
--
2.39.3
> Subject: Re: [PATCH 0/9] Enable haltpoll for arm64
A correction: please read the subject for the series as [PATCH v5] ...
Missed the version number it while sending out.
Thanks
Ankur
Ankur Arora <[email protected]> writes:
> This patchset enables the cpuidle-haltpoll driver and its namesake
> governor on arm64. This is specifically interesting for KVM guests by
> reducing the IPC latencies.
>
> Comparing idle switching latencies on an arm64 KVM guest with
> perf bench sched pipe:
>
> usecs/op %stdev
>
> no haltpoll (baseline) 13.48 +- 5.19%
> with haltpoll 6.84 +- 22.07%
>
>
> No change in performance for a similar test on x86:
>
> usecs/op %stdev
>
> haltpoll w/ cpu_relax() (baseline) 4.75 +- 1.76%
> haltpoll w/ smp_cond_load_relaxed() 4.78 +- 2.31%
>
> Both sets of tests were on otherwise idle systems with guest VCPUs
> pinned to specific PCPUs. One reason for the higher stdev on arm64
> is that trapping of the WFE instruction by the host KVM is contingent
> on the number of tasks on the runqueue.
>
>
> The patch series is organized in four parts:
> - patches 1, 2 mangle the config option ARCH_HAS_CPU_RELAX, renaming
> and moving it from x86 to common architectural code.
> - next, patches 3-5, reorganize the haltpoll selection and init logic
> to allow architecture code to select it.
> - patch 6, reorganizes the poll_idle() loop, switching from using
> cpu_relax() directly to smp_cond_load_relaxed().
> - and finally, patches 7-9, add the bits for arm64 support.
>
> What is still missing: this series largely completes the haltpoll side
> of functionality for arm64. There are, however, a few related areas
> that still need to be threshed out:
>
> - WFET support: WFE on arm64 does not guarantee that poll_idle()
> would terminate in halt_poll_ns. Using WFET would address this.
> - KVM_NO_POLL support on arm64
> - KVM TWED support on arm64: allow the host to limit time spent in
> WFE.
>
>
> Changelog:
>
> v5:
> - rework the poll_idle() loop around smp_cond_load_relaxed() (review
> comment from Tomohiro Misono.)
> - also rework selection of cpuidle-haltpoll. Now selected based
> on the architectural selection of ARCH_CPUIDLE_HALTPOLL.
> - arch_haltpoll_supported() (renamed from arch_haltpoll_want()) on
> arm64 now depends on the event-stream being enabled.
> - limit POLL_IDLE_RELAX_COUNT on arm64 (review comment from Haris Okanovic)
> - ARCH_HAS_CPU_RELAX is now renamed to ARCH_HAS_OPTIMIZED_POLL.
>
> v4 changes from v3:
> - change 7/8 per Rafael input: drop the parens and use ret for the final check
> - add 8/8 which renames the guard for building poll_state
>
> v3 changes from v2:
> - fix 1/7 per Petr Mladek - remove ARCH_HAS_CPU_RELAX from arch/x86/Kconfig
> - add Ack-by from Rafael Wysocki on 2/7
>
> v2 changes from v1:
> - added patch 7 where we change cpu_relax with smp_cond_load_relaxed per PeterZ
> (this improves by 50% at least the CPU cycles consumed in the tests above:
> 10,716,881,137 now vs 14,503,014,257 before)
> - removed the ifdef from patch 1 per RafaelW
>
> Ankur Arora (4):
> cpuidle: rename ARCH_HAS_CPU_RELAX to ARCH_HAS_OPTIMIZED_POLL
> cpuidle-haltpoll: condition on ARCH_CPUIDLE_HALTPOLL
> arm64: support cpuidle-haltpoll
> cpuidle/poll_state: limit POLL_IDLE_RELAX_COUNT on arm64
>
> Joao Martins (4):
> Kconfig: move ARCH_HAS_OPTIMIZED_POLL to arch/Kconfig
> cpuidle-haltpoll: define arch_haltpoll_supported()
> governors/haltpoll: drop kvm_para_available() check
> arm64: define TIF_POLLING_NRFLAG
>
> Mihai Carabas (1):
> cpuidle/poll_state: poll via smp_cond_load_relaxed()
>
> arch/Kconfig | 3 +++
> arch/arm64/Kconfig | 10 ++++++++++
> arch/arm64/include/asm/cpuidle_haltpoll.h | 21 +++++++++++++++++++++
> arch/arm64/include/asm/thread_info.h | 2 ++
> arch/x86/Kconfig | 4 +---
> arch/x86/include/asm/cpuidle_haltpoll.h | 1 +
> arch/x86/kernel/kvm.c | 10 ++++++++++
> drivers/acpi/processor_idle.c | 4 ++--
> drivers/cpuidle/Kconfig | 5 ++---
> drivers/cpuidle/Makefile | 2 +-
> drivers/cpuidle/cpuidle-haltpoll.c | 9 ++-------
> drivers/cpuidle/governors/haltpoll.c | 6 +-----
> drivers/cpuidle/poll_state.c | 21 ++++++++++++++++-----
> drivers/idle/Kconfig | 1 +
> include/linux/cpuidle.h | 2 +-
> include/linux/cpuidle_haltpoll.h | 5 +++++
> 16 files changed, 79 insertions(+), 27 deletions(-)
> create mode 100644 arch/arm64/include/asm/cpuidle_haltpoll.h
--
ankur
Hi Ankur,
kernel test robot noticed the following build errors:
[auto build test ERROR on rafael-pm/linux-next]
[also build test ERROR on rafael-pm/bleeding-edge tip/x86/core arm64/for-next/core linus/master v6.9-rc6 next-20240430]
[If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]
url: https://github.com/intel-lab-lkp/linux/commits/Ankur-Arora/cpuidle-rename-ARCH_HAS_CPU_RELAX-to-ARCH_HAS_OPTIMIZED_POLL/20240501-024252
base: https://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git linux-next
patch link: https://lore.kernel.org/r/20240430183730.561960-5-ankur.a.arora%40oracle.com
patch subject: [PATCH 4/9] cpuidle-haltpoll: define arch_haltpoll_supported()
config: i386-buildonly-randconfig-001-20240501 (https://download.01.org/0day-ci/archive/20240501/[email protected]/config)
compiler: clang version 18.1.4 (https://github.com/llvm/llvm-project e6c3289804a67ea0bb6a86fadbe454dd93b8d855)
reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20240501/[email protected]/reproduce)
If you fix the issue in a separate patch/commit (i.e. not just a new version of
the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/
All errors (new ones prefixed by >>):
>> ld.lld: error: undefined symbol: arch_haltpoll_enable
>>> referenced by cpuidle-haltpoll.c
>>> drivers/cpuidle/cpuidle-haltpoll.o:(haltpoll_cpu_online) in archive vmlinux.a
--
>> ld.lld: error: undefined symbol: arch_haltpoll_disable
>>> referenced by cpuidle-haltpoll.c
>>> drivers/cpuidle/cpuidle-haltpoll.o:(haltpoll_cpu_offline) in archive vmlinux.a
--
>> ld.lld: error: undefined symbol: arch_haltpoll_supported
>>> referenced by cpuidle-haltpoll.c
>>> drivers/cpuidle/cpuidle-haltpoll.o:(haltpoll_init) in archive vmlinux.a
--
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
On Tue, 30 Apr 2024, Ankur Arora wrote:
> ARCH_HAS_CPU_RELAX is a bit of a misnomer since all architectures
> define cpu_relax(). Not all, however, have a performant version, with
> some only implementing it as a compiler barrier.
>
> In contexts that this config option is used, it is expected to provide
> an architectural primitive that can be used as part of a polling
> mechanism -- one that would be cheaper than spinning in a tight loop.
The intend of cpu_relax() is not a polling mechanism. Initial AFAICT it
was introduced on x86 as the REP NOP instruction. Aka as PAUSE. And it was
part of a spin loop. So there was no connection to polling anything.
The intend was to make the processor aware that we are in a spin loop.
Various processors have different actions that they take upon encountering
such a cpu relax operation.
The polling (WFE/WFI) available on ARM (and potentially other platforms)
is a different mechanism that is actually intended to reduce the power
requirement of the processor until a certain condition is met and that
check is done in hardware.
These are not the same and I think we need both config options.
The issues that you have with WFET later in the patchset arise from not
making this distinction.
The polling (waiting for an event) could be implemented for a
processor not supporting that in hardware by using a loop that
checks for the condition and then does a cpu_relax().
With that you could f.e. support the existing cpu_relax() and also have
some form of cpu_poll() interface.
Christoph Lameter (Ampere) <[email protected]> writes:
> On Tue, 30 Apr 2024, Ankur Arora wrote:
>
>> ARCH_HAS_CPU_RELAX is a bit of a misnomer since all architectures
>> define cpu_relax(). Not all, however, have a performant version, with
>> some only implementing it as a compiler barrier.
>>
>> In contexts that this config option is used, it is expected to provide
>> an architectural primitive that can be used as part of a polling
>> mechanism -- one that would be cheaper than spinning in a tight loop.
>
> The intend of cpu_relax() is not a polling mechanism. Initial AFAICT it was
> introduced on x86 as the REP NOP instruction. Aka as PAUSE. And it was part of a
> spin loop. So there was no connection to polling anything.
Agreed, cpu_relax() is just a mechanism to tell the pipeline that
we are in a spin-loop.
> The intend was to make the processor aware that we are in a spin loop. Various
> processors have different actions that they take upon encountering such a cpu
> relax operation.
Sure, though most processors don't have a nice mechanism to do that.
x86 clearly has the REP; NOP thing. arm64 only has a YIELD which from my
measurements is basically a NOP when executed on a system without
hardware threads.
And that's why only x86 defines ARCH_HAS_CPU_RELAX.
> The polling (WFE/WFI) available on ARM (and potentially other platforms) is a
> different mechanism that is actually intended to reduce the power requirement of
> the processor until a certain condition is met and that check is done in
> hardware.
Sure. Which almost exactly fits the bill for the poll-idle loop -- except for the
timeout part.
> These are not the same and I think we need both config options.
My main concern is that poll_idle() conflates polling in idle with
ARCH_HAS_CPU_RELAX, when they aren't really related.
So, poll_idle(), and its users should depend on ARCH_HAS_OPTIMIZED_POLL
which, if defined by some architecture, means that poll_idle() would
be better than a spin-wait loop.
Beyond that I'm okay to keep ARCH_HAS_CPU_RELAX around.
That said, do you see a use for ARCH_HAS_CPU_RELAX? The only current
user is the poll-idle path.
> The issues that you have with WFET later in the patchset arise from not making
> this distinction.
Did you mean the issue with WFE? I'm not using WFET in this patchset at all.
With WFE, sure there's a problem in that you depend on an interrupt or
the event-stream to get out of the wait. And, so sometimes you would
overshoot the target poll timeout.
> The polling (waiting for an event) could be implemented for a processor not
> supporting that in hardware by using a loop that checks for the condition and
> then does a cpu_relax().
Yeah. That's exactly what patch-6 does. smp_cond_load_relaxed() uses
cpu_relax() internally in its spin-loop variant (non arm64).
On arm64, this would use LDXR; WFE. Or are you suggesting implementing
the arm64 loop via cpu_relax() (and thus YIELD?)
Ankur
> With that you could f.e. support the existing cpu_relax() and also have some
> form of cpu_poll() interface.
On Thu, 2 May 2024, Ankur Arora wrote:
>> The intend was to make the processor aware that we are in a spin loop. Various
>> processors have different actions that they take upon encountering such a cpu
>> relax operation.
>
> Sure, though most processors don't have a nice mechanism to do that.
> x86 clearly has the REP; NOP thing. arm64 only has a YIELD which from my
> measurements is basically a NOP when executed on a system without
> hardware threads.
>
> And that's why only x86 defines ARCH_HAS_CPU_RELAX.
My impression is that the use of arm YIELD has led cpu architects to
implement similar mechanisms to x86s PAUSE, This is not part of the spec
but it has been there for a long time. So I would rather leave it as is.
>> These are not the same and I think we need both config options.
>
> My main concern is that poll_idle() conflates polling in idle with
> ARCH_HAS_CPU_RELAX, when they aren't really related.
>
> So, poll_idle(), and its users should depend on ARCH_HAS_OPTIMIZED_POLL
> which, if defined by some architecture, means that poll_idle() would
> be better than a spin-wait loop.
>
> Beyond that I'm okay to keep ARCH_HAS_CPU_RELAX around.
>
> That said, do you see a use for ARCH_HAS_CPU_RELAX? The only current
> user is the poll-idle path.
I would think that we need a generic cpu_poll() mechanism that can fall
back to cpu_relax() on processors that do not offer such thing (x86?) and
if not even that is there fall back.
We already have something like that in the smp_cond_acquire mechanism (a
bit weird to put that in the barrier.h>).
So what if we had
void cpu_wait(unsigned flags, unsigned long timeout, void *cacheline);
With
#define CPU_POLL_INTERRUPT (1 << 0)
#define CPU_POLL_EVENT (1 << 1)
#define CPU_POLL_CACHELINE (1 << 2)
#define CPU_POLL_TIMEOUT (1 << 3)
#define CPU_POLL_BROADCAST_EVENT (1 << 4)
#define CPU_POLL_LOCAL_EVENT (1 << 5)
The cpu_poll() function coud be generically defined in asm-generic and
then arches could provide their own implementation optimizing the hardware
polling mechanisms.
Any number of flags could be specified simultaneously. On ARM this would
map then to SEVL SEV and WFI/WFE WFIT/WFET
So f.e.
cpu_wait(CPU_POLL_INTERUPT|CPU_POLL_EVENT|CPU_POLL_TIMEOUT|CPU_POLL_CACHELINE,
timeout, &mylock);
to wait on a change in a cacheline with a timeout.
In additional we could then think about making effective use of the
signaling mechanism provided by SEV in core logic of the kernel. Maybe
that is more effective then waiting for a cacheline in some situations.
> With WFE, sure there's a problem in that you depend on an interrupt or
> the event-stream to get out of the wait. And, so sometimes you would
> overshoot the target poll timeout.
Right. The dependence on the event stream makes this approach a bit
strange. Having some sort of generic cpu_wait() feature with timeout spec
could avoid that.
Christoph Lameter (Ampere) <[email protected]> writes:
> On Thu, 2 May 2024, Ankur Arora wrote:
>
>>> The intend was to make the processor aware that we are in a spin loop. Various
>>> processors have different actions that they take upon encountering such a cpu
>>> relax operation.
>>
>> Sure, though most processors don't have a nice mechanism to do that.
>> x86 clearly has the REP; NOP thing. arm64 only has a YIELD which from my
>> measurements is basically a NOP when executed on a system without
>> hardware threads.
>>
>> And that's why only x86 defines ARCH_HAS_CPU_RELAX.
>
> My impression is that the use of arm YIELD has led cpu architects to implement
> similar mechanisms to x86s PAUSE, This is not part of the spec but it has been
> there for a long time. So I would rather leave it as is.
>
>
>>> These are not the same and I think we need both config options.
>>
>> My main concern is that poll_idle() conflates polling in idle with
>> ARCH_HAS_CPU_RELAX, when they aren't really related.
>>
>> So, poll_idle(), and its users should depend on ARCH_HAS_OPTIMIZED_POLL
>> which, if defined by some architecture, means that poll_idle() would
>> be better than a spin-wait loop.
>>
>> Beyond that I'm okay to keep ARCH_HAS_CPU_RELAX around.
>>
>> That said, do you see a use for ARCH_HAS_CPU_RELAX? The only current
>> user is the poll-idle path.
>
> I would think that we need a generic cpu_poll() mechanism that can fall back to
> cpu_relax() on processors that do not offer such thing (x86?) and if not even
> that is there fall back.
>
> We already have something like that in the smp_cond_acquire mechanism (a bit
> weird to put that in the barrier.h>).
>
> So what if we had
>
> void cpu_wait(unsigned flags, unsigned long timeout, void *cacheline);
>
> With
>
> #define CPU_POLL_INTERRUPT (1 << 0)
> #define CPU_POLL_EVENT (1 << 1)
> #define CPU_POLL_CACHELINE (1 << 2)
> #define CPU_POLL_TIMEOUT (1 << 3)
> #define CPU_POLL_BROADCAST_EVENT (1 << 4)
> #define CPU_POLL_LOCAL_EVENT (1 << 5)
>
>
> The cpu_poll() function coud be generically defined in asm-generic and then
> arches could provide their own implementation optimizing the hardware polling
> mechanisms.
>
> Any number of flags could be specified simultaneously. On ARM this would map
> then to SEVL SEV and WFI/WFE WFIT/WFET
>
> So f.e.
>
> cpu_wait(CPU_POLL_INTERUPT|CPU_POLL_EVENT|CPU_POLL_TIMEOUT|CPU_POLL_CACHELINE,
> timeout, &mylock);
>
> to wait on a change in a cacheline with a timeout.
>
> In additional we could then think about making effective use of the signaling
> mechanism provided by SEV in core logic of the kernel. Maybe that is more
> effective then waiting for a cacheline in some situations.
>
>
>> With WFE, sure there's a problem in that you depend on an interrupt or
>> the event-stream to get out of the wait. And, so sometimes you would
>> overshoot the target poll timeout.
>
> Right. The dependence on the event stream makes this approach a bit strange.
> Having some sort of generic cpu_wait() feature with timeout spec could avoid
> that.
Thanks for the detailed comments. Helped me think through some of the details.
So, there are three issues that you bring up. Let me address each in turn.
1) A generic cpu_poll() mechanism that can fall back to cpu_relax().
> I would think that we need a generic cpu_poll() mechanism that can fall back to
> cpu_relax() on processors that do not offer such thing (x86?) and if not even
> that is there fall back.
>
> We already have something like that in the smp_cond_acquire mechanism (a bit
> weird to put that in the barrier.h>).
Isn't that exactly what this series does?
If you see patch-6, that gets rid of direct use of cpu_relax(), instead
using smp_cond_load_relaxed().
And smp_cond_load_relaxed(), in its generic variant (used everywhere but
arm64) uses cpu_relax() implicitly. Any architecture that override
this -- as arm64 does -- get their own optimizations.
(Maybe this patch would be clearer if it was sequenced after patch-6?)
2) That brings me back to your second point, about having a different
interface which allows for different optimizations.
> void cpu_wait(unsigned flags, unsigned long timeout, void *cacheline);
>
> With
>
> #define CPU_POLL_INTERRUPT (1 << 0)
> #define CPU_POLL_EVENT (1 << 1)
> #define CPU_POLL_CACHELINE (1 << 2)
> #define CPU_POLL_TIMEOUT (1 << 3)
> #define CPU_POLL_BROADCAST_EVENT (1 << 4)
> #define CPU_POLL_LOCAL_EVENT (1 << 5)
I agree with you that the polling logic does need to handle timeouts
but I don't think that we need a special interface.
Given that we are only concerned about poll_idle() here and that needs
to work with the scheduler's set_nr_if_polling() machinery to elide IPIs,
the polling on a cacheline is needed anyway. That also means we poll_idle()
doesn't need to handle interrupts.
For the rest, the architecture could internally choose whichever
variation they perform best at -- so long as they can either spin-wait
or have some kind of event driven mechanism (WFE/WFET, MONITOR/MWAIT[X])
The timeout is something that I plan to address separately. I think we can
straight-forwardly extend to smp_cond_load_timeout() to do that at least
for WFET supporting platforms where others depend on the event-stream.
#define smp_cond_load_relaxed_timeout(ptr, cond_expr, time_check_expr, \
time_limit, timeout) ({ \
typeof(ptr) __PTR = (ptr); \
__unqual_scalar_typeof(*ptr) VAL; \
unsigned int __count = 0; \
for (;;) { \
VAL = READ_ONCE(*__PTR); \
if (cond_expr) \
break; \
cpu_relax(); \
if (__count++ < smp_cond_time_check_count) \
continue; \
\
if ((time_check_expr) > time_limit) \
goto timeout; \
\
__count = 0; \
} \
(typeof(*ptr))VAL; \
})
arm64, for instance, can use alternatives to implement whichever variant
the processor supports (untested, also needs massaging).
static inline void __cmpwait_case_##sz(volatile void *ptr, \
unsigned long val, \
unsigned long etime) \
\
unsigned long tmp; \
\
const unsigned long ecycles = xloops_to_cycles(nsecs_to_xloops(etime)); \
asm volatile( \
" sevl\ n" \
" wfe\ n" \
" ldxr" #sfx "\ t%" #w "[tmp], %[v]\n" \
" eor %" #w "[tmp], %" #w "[tmp], %" #w "[val]\ n" \
" cbnz %" #w "[tmp], 1f\ n" \
ALTERNATIVE("wfe\ n", \
"msr s0_3_c1_c0_0, %[ecycles]\ n", \
ARM64_HAS_WFXT) \
"1:" \
: [tmp] "=&r" (tmp), [v] "+Q" (*(u##sz *)ptr) \
: [val] "r" (val), [ecycles] "r" (ecycles)); \
}
And, then poll_idle() only need be:
static int __cpuidle poll_idle(struct cpuidle_device *dev,
struct cpuidle_driver *drv, int index)
{
u64 time_start;
time_start = local_clock_noinstr();
dev->poll_time_limit = false;
raw_local_irq_enable();
if (!current_set_polling_and_test()) {
u64 time_limit;
time_limit = cpuidle_poll_time(drv, dev);
smp_cond_load_relaxed_timeout(¤t_thread_info()->flags,
VAL & _TIF_NEED_RESCHED,
local_clock_noinstr(),
time_start + time_limit,
timed_out);
}
...
On x86, this generates code pretty similar to the current version and
on arm64 similar to the WFE version.
3) Dependence on the event-stream for the WFE variants
> Right. The dependence on the event stream makes this approach a bit strange.
> Having some sort of generic cpu_wait() feature with timeout spec could avoid
> that.
Not sure I agree with that. Seems to me, the event-stream is present for
exactly that -- so we don't wait in a WFE or WFI forever. The spec
say (section D12.2.3)
"An implementation that includes the Generic Timer can use the system
counter to generate one or more event streams, to generate periodic
wakeup events as part of the mechanism described in Wait for Event.
An event stream might be used:
- To impose a time-out on a Wait For Event polling loop."
The overshoot is a problem, but I don't think it is a huge one. Most
of the time that haltpoll is in effect, it should wake up with work
to do in the guest_halt_poll_ns duration. The times, it doesn't it
would exit poll_idle() and exit to the hypervisor.
So, this is only an issue in the last iteration.
And, I'm not sure how a generic cpu_wait() would work? Either the
generic cpu_wait() spins in cpu_relax()/YIELD which is suboptimal
on arm64 or it uses WFE but sets a timer for 50us or whatever. Seems
like unnecessary overhead when the overshoot is relatively uncommon.
Or is there another mechanism you are thinking of for enforcing a
timeout?
Thanks
--
ankur