It seems gcc can automatically inline apbt_writel() for us, but apbt_real isn't inlined. This patch makes them inline to get a trivial performance improvement: 4096 rounds of __apbt_read_clocksource() call spend time on Marvell BG4CT platform: before the patch 1275240ns on average after the patch 1263240ns on average so we get 1% performance improvement. Signed-off-by: Jisheng Zhang <jszhang@marvell.com> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
It's safe to use the relaxed version. From another side, the relaxed io accessor macros are available on all architectures now, so we can use the relaxed versions to get a trivial system performance improvement, we measured time the following functions spent on Marvell BG4CT: 4096 rounds of __apbt_read_clocksource() call: before the patch: 1263240ns on average after the patch: 1250080ns on average improved by 1% 4096 rounds of apbt_eoi() call: before the patch: 1290960ns on average after the patch: 1248240ns on average 4096 rounds of apbt_next_event() call: before the patch: 3333660ns on average after the patch: 1322040ns on average improved by 60%! Signed-off-by: Jisheng Zhang <jszhang@marvell.com> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
On Marvell BG4CT platform, we observed the __apbt_read_clocksource() return wrong value: Let's assume the APBTMR_N_CURRENT_VALUE value is 0xf0000000, we got 0xffffffff0fffffff, but it should be 0xfffffff. This issue should be common on all 64bit platforms. We fix the issue by letting aptb_readl() return u32. apbt_writel() is also updated to write u32 val rather than unsigned long. Signed-off-by: Jisheng Zhang <jszhang@marvell.com> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Commit d2348fb6fdc6 ("tick: Dynamically set broadcast irq affinity")
adds one excellent feature CLOCK_EVT_FEAT_DYNIRQ to let the core set the
interrupt affinity of the broadcast interrupt to the cpu which has the
earliest expiry time. This patch adds CLOCK_EVT_FEAT_DYNIRQ flag to
avoid unnecessary wakeups and IPIs when the dw_apb_timer is used as
broadcast timer.
A simple test:
~ # rm /tmp/test.sh
~ # cat > /tmp/test.sh
cat /proc/interrupts
for i in `seq 10` ; do sleep $i; done
cat /proc/interrupts
~ # chmod +x /tmp/test.sh
~ # taskset 0x2 /tmp/test.sh
without the patch:
CPU0 CPU1
27: 115 36 GIC 27 arch_timer
45: 62 0 GIC 45 mmc0
160: 88 0 interrupt-controller 8 timer
227: 0 0 interrupt-controller 4 f7e81400.i2c
228: 0 0 interrupt-controller 5 f7e81800.i2c
229: 0 0 interrupt-controller 7 dw_spi65535
230: 0 0 interrupt-controller 21 f7e84000.i2c
231: 0 0 interrupt-controller 20 f7e84800.i2c
265: 445 0 interrupt-controller 8 serial
IPI0: 0 0 CPU wakeup interrupts
IPI1: 0 11 Timer broadcast interrupts
IPI2: 56 104 Rescheduling interrupts
IPI3: 0 0 Function call interrupts
IPI4: 0 4 Single function call interrupts
IPI5: 0 0 CPU stop interrupts
IPI6: 25 27 IRQ work interrupts
IPI7: 0 0 completion interrupts
IPI8: 0 0 CPU backtrace
Err: 0
CPU0 CPU1
27: 115 38 GIC 27 arch_timer
45: 62 0 GIC 45 mmc0
160: 160 0 interrupt-controller 8 timer
227: 0 0 interrupt-controller 4 f7e81400.i2c
228: 0 0 interrupt-controller 5 f7e81800.i2c
229: 0 0 interrupt-controller 7 dw_spi65535
230: 0 0 interrupt-controller 21 f7e84000.i2c
231: 0 0 interrupt-controller 20 f7e84800.i2c
265: 514 0 interrupt-controller 8 serial
IPI0: 0 0 CPU wakeup interrupts
IPI1: 0 83 Timer broadcast interrupts
IPI2: 56 104 Rescheduling interrupts
IPI3: 0 0 Function call interrupts
IPI4: 0 4 Single function call interrupts
IPI5: 0 0 CPU stop interrupts
IPI6: 25 46 IRQ work interrupts
IPI7: 0 0 completion interrupts
IPI8: 0 0 CPU backtrace
Err: 0
cpu0 get 160-88=72 timer interrupts, CPU1 got 83-11=72 broadcast timer
IPIs
So, overall system got 72+72=144 wake ups and 72 broadcast timer IPIs
With the patch:
CPU0 CPU1
27: 107 37 GIC 27 arch_timer
45: 62 0 GIC 45 mmc0
160: 66 7 interrupt-controller 8 timer
227: 0 0 interrupt-controller 4 f7e81400.i2c
228: 0 0 interrupt-controller 5 f7e81800.i2c
229: 0 0 interrupt-controller 7 dw_spi65535
230: 0 0 interrupt-controller 21 f7e84000.i2c
231: 0 0 interrupt-controller 20 f7e84800.i2c
265: 311 0 interrupt-controller 8 serial
IPI0: 0 0 CPU wakeup interrupts
IPI1: 2 4 Timer broadcast interrupts
IPI2: 58 100 Rescheduling interrupts
IPI3: 0 0 Function call interrupts
IPI4: 0 4 Single function call interrupts
IPI5: 0 0 CPU stop interrupts
IPI6: 21 24 IRQ work interrupts
IPI7: 0 0 completion interrupts
IPI8: 0 0 CPU backtrace
Err: 0
CPU0 CPU1
27: 107 39 GIC 27 arch_timer
45: 62 0 GIC 45 mmc0
160: 69 75 interrupt-controller 8 timer
227: 0 0 interrupt-controller 4 f7e81400.i2c
228: 0 0 interrupt-controller 5 f7e81800.i2c
229: 0 0 interrupt-controller 7 dw_spi65535
230: 0 0 interrupt-controller 21 f7e84000.i2c
231: 0 0 interrupt-controller 20 f7e84800.i2c
265: 380 0 interrupt-controller 8 serial
IPI0: 0 0 CPU wakeup interrupts
IPI1: 3 6 Timer broadcast interrupts
IPI2: 60 100 Rescheduling interrupts
IPI3: 0 0 Function call interrupts
IPI4: 0 4 Single function call interrupts
IPI5: 0 0 CPU stop interrupts
IPI6: 21 45 IRQ work interrupts
IPI7: 0 0 completion interrupts
IPI8: 0 0 CPU backtrace
Err: 0
cpu0 got 69-66=3, cpu1 got 75-7=68 timer interrupts. cpu0 got 3-2=1
broadcast timer IPIs, cpu1 got 6-4=2 broadcast timer IPIs.
So, overall system got 3+68+1+2=74 wakeups and 1+2=3 broadcast timer
IPIs
This patch removes 50% wakeups and almost 100% broadcast timer IPIs!
Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Migrate dw_apb driver to the new 'set-state' interface provided by clockevents core, the earlier 'set-mode' interface is marked obsolete now. This also enables us to implement callbacks for new states of clockevent devices, for example: ONESHOT_STOPPED. Cc: Jacob Pan <jacob.jun.pan@linux.intel.com> Cc: Jamie Iles <jamie@jamieiles.com> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Thanks to spatch, plus manual removal of "&*". Then a sweep for for_each_cpu_mask => for_each_cpu. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Jason Cooper <jason@lakedaemon.net> Cc: Chris Metcalf <cmetcalf@ezchip.com> Cc: netdev@vger.kernel.org
This patch removes the use of the IRQF_DISABLED flag It's a NOOP since 2.6.35 and it will be removed one day. [dlezcano] : slightly changed the changelog Signed-off-by: Michael Opdenacker <michael.opdenacker@free-electrons.com> Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
This reverts commit 55a68c23e0a675b2b8ac2656fd6edbf98b78e4c6. In order to avoid a collision with dw_apb_timer changes in the arm-soc tree, revert this change. I'm leaving it to the arm-soc folks to sort out if they want to keep the other side of the collision or if they're just going to back it all out and try again during the next release cycle. Reported-by: Dinh Nguyen <dinguyen@altera.com> Signed-off-by: John Stultz <john.stultz@linaro.org>
It seems we made a mistake when creating dw_apb_timer_of.c:
picoxcell sched_clock had parts that were not related to
dw_apb_timer, yet we moved them to dw_apb_timer_of, and tried to
use them on socfpga.
This results in system where user/system time is not measured
properly, as demonstrated by
time dd if=/dev/urandom of=/dev/zero bs=100000 count=100
So this patch switches sched_clock to hardware that exists on both
platforms, and adds missing of_node_put() in dw_apb_timer_init().
Signed-off-by: Pavel Machek <pavel@denx.de>
Acked-by: Jamie Iles <jamie@jamieiles.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: John Stultz <john.stultz@linaro.org> Cc: Magnus Damm <magnus.damm@gmail.com> Acked-by: Jamie Iles <jamie@jamieiles.com> Link: http://lkml.kernel.org/r/20130425143435.558006195@linutronix.de Signed-off-by: Thomas Gleixner <tglx@linutronix.de>