[PATCH v2] powerpc/perf: Use cpumask_last() to determine the designated cpu for nest/core units.

Nest and core imc(In-memory Collection counters) assigns a particular
cpu as the designated target for counter data collection.
During system boot, the first online cpu in a chip gets assigned as
the designated cpu for that chip(for nest-imc) and the first online cpu
in a core gets assigned as the designated cpu for that core(for core-imc).

If the designated cpu goes offline, the next online cpu from the same
chip(for nest-imc)/core(for core-imc) is assigned as the next target,
and the event context is migrated to the target cpu.
Currently, cpumask_any_but() function is used to find the target cpu.
Though this function is expected to return a `random` cpu, this always
returns the next online cpu.

If all cpus in a chip/core is offlined in a sequential manner, starting
from the first cpu, the event migration has to happen for all the cpus
which goes offline. Since the migration process involves a grace period,
the total time taken to offline all the cpus will be significantly high.

Example:
In a system which has 2 sockets, with
NUMA node0 CPU(s):     0-87
NUMA node8 CPU(s):     88-175

Time taken to offline cpu 88-175:
real    2m56.099s
user    0m0.191s
sys     0m0.000s

Use cpumask_last() to choose the target cpu, when the designated cpu
goes online, so the migration will happen only when the last_cpu in the
mask goes offline. This way the time taken to offline all cpus in a
chip/core can be reduced.

With the patch, 

Time taken  to offline cpu 88-175:
real    0m12.207s
user    0m0.171s
sys     0m0.000s


Offlining all cpus in reverse order is also taken care because,
cpumask_any_but() is used to find the designated cpu if the last cpu in
the mask goes offline. Since cpumask_any_but() always return the first
cpu in the mask, that becomes the designated cpu and migration will happen
only when the first_cpu in the mask goes offline.

Example:
With the patch,

Time taken to offline cpu from 175-88:
real    0m9.330s
user    0m0.110s
sys     0m0.000s

Signed-off-by: Anju T Sudhakar <redacted>
Reviewed-by: Madhavan Srinivasan <redacted>
---

Changes from v1:
	Modified the commit log with more info.
---

 arch/powerpc/perf/imc-pmu.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c
index 31fa753..fbfd6e7 100644
--- a/arch/powerpc/perf/imc-pmu.c
+++ b/arch/powerpc/perf/imc-pmu.c

@@ -366,7 +366,14 @@ static int ppc_nest_imc_cpu_offline(unsigned int cpu)
 	 */
 	nid = cpu_to_node(cpu);
 	l_cpumask = cpumask_of_node(nid);
-	target = cpumask_any_but(l_cpumask, cpu);
+	target = cpumask_last(l_cpumask);
+
+	/*
+	 * If this(target) is the last cpu in the cpumask for this chip,
+	 * check for any possible online cpu in the chip.
+	 */
+	if (unlikely(target == cpu))
+		target = cpumask_any_but(l_cpumask, cpu);
 
 	/*
 	 * Update the cpumask with the target cpu and

@@ -671,7 +678,10 @@ static int ppc_core_imc_cpu_offline(unsigned int cpu)
 		return 0;
 
 	/* Find any online cpu in that core except the current "cpu" */
-	ncpu = cpumask_any_but(cpu_sibling_mask(cpu), cpu);
+	ncpu = cpumask_last(cpu_sibling_mask(cpu));
+
+	if (unlikely(ncpu == cpu))
+		ncpu = cpumask_any_but(cpu_sibling_mask(cpu), cpu);
 
 	if (ncpu >= 0 && ncpu < nr_cpu_ids) {
 		cpumask_set_cpu(ncpu, &core_imc_cpumask);

-- 
1.8.3.1

On Mon, 2019-06-10 at 12:02 +0530, Anju T Sudhakar wrote:

Nest and core imc(In-memory Collection counters) assigns a particular
cpu as the designated target for counter data collection.
During system boot, the first online cpu in a chip gets assigned as
the designated cpu for that chip(for nest-imc) and the first online cpu
in a core gets assigned as the designated cpu for that core(for core-imc).

If the designated cpu goes offline, the next online cpu from the same
chip(for nest-imc)/core(for core-imc) is assigned as the next target,
and the event context is migrated to the target cpu.
Currently, cpumask_any_but() function is used to find the target cpu.
Though this function is expected to return a `random` cpu, this always
returns the next online cpu.

If all cpus in a chip/core is offlined in a sequential manner, starting
from the first cpu, the event migration has to happen for all the cpus
which goes offline. Since the migration process involves a grace period,
the total time taken to offline all the cpus will be significantly high.

Example:
In a system which has 2 sockets, with
NUMA node0 CPU(s):     0-87
NUMA node8 CPU(s):     88-175

Time taken to offline cpu 88-175:
real    2m56.099s
user    0m0.191s
sys     0m0.000s

Use cpumask_last() to choose the target cpu, when the designated cpu
goes online, so the migration will happen only when the last_cpu in the
mask goes offline. This way the time taken to offline all cpus in a
chip/core can be reduced.

With the patch, 

Time taken  to offline cpu 88-175:
real    0m12.207s
user    0m0.171s
sys     0m0.000s


Offlining all cpus in reverse order is also taken care because,
cpumask_any_but() is used to find the designated cpu if the last cpu in
the mask goes offline. Since cpumask_any_but() always return the first
cpu in the mask, that becomes the designated cpu and migration will happen
only when the first_cpu in the mask goes offline.

Example:
With the patch,

Time taken to offline cpu from 175-88:
real    0m9.330s
user    0m0.110s
sys     0m0.000s

Seems like a very interesting work.
Out of curiosity, have you used 'chcpu -d' to create your benchmark?

Hi Leonardo,

On 6/11/19 12:17 AM, Leonardo Bras wrote:

On Mon, 2019-06-10 at 12:02 +0530, Anju T Sudhakar wrote:

quoted

Nest and core imc(In-memory Collection counters) assigns a particular
cpu as the designated target for counter data collection.
During system boot, the first online cpu in a chip gets assigned as
the designated cpu for that chip(for nest-imc) and the first online cpu
in a core gets assigned as the designated cpu for that core(for core-imc).

If the designated cpu goes offline, the next online cpu from the same
chip(for nest-imc)/core(for core-imc) is assigned as the next target,
and the event context is migrated to the target cpu.
Currently, cpumask_any_but() function is used to find the target cpu.
Though this function is expected to return a `random` cpu, this always
returns the next online cpu.

If all cpus in a chip/core is offlined in a sequential manner, starting
from the first cpu, the event migration has to happen for all the cpus
which goes offline. Since the migration process involves a grace period,
the total time taken to offline all the cpus will be significantly high.

Seems like a very interesting work.
Out of curiosity, have you used 'chcpu -d' to create your benchmark?

Here I did not use chcpu to disable the cpu.

I used a script which will offline cpus 88-175 by echoing  `0` to

/sys/devices/system/cpu/cpu*/online.


Regards,

Anju

On Mon, 2019-06-10 at 06:32:29 UTC, Anju T Sudhakar wrote:

Nest and core imc(In-memory Collection counters) assigns a particular
cpu as the designated target for counter data collection.
During system boot, the first online cpu in a chip gets assigned as
the designated cpu for that chip(for nest-imc) and the first online cpu
in a core gets assigned as the designated cpu for that core(for core-imc).

If the designated cpu goes offline, the next online cpu from the same
chip(for nest-imc)/core(for core-imc) is assigned as the next target,
and the event context is migrated to the target cpu.
Currently, cpumask_any_but() function is used to find the target cpu.
Though this function is expected to return a `random` cpu, this always
returns the next online cpu.

If all cpus in a chip/core is offlined in a sequential manner, starting
from the first cpu, the event migration has to happen for all the cpus
which goes offline. Since the migration process involves a grace period,
the total time taken to offline all the cpus will be significantly high.

Example:
In a system which has 2 sockets, with
NUMA node0 CPU(s):     0-87
NUMA node8 CPU(s):     88-175

Time taken to offline cpu 88-175:
real    2m56.099s
user    0m0.191s
sys     0m0.000s

Use cpumask_last() to choose the target cpu, when the designated cpu
goes online, so the migration will happen only when the last_cpu in the
mask goes offline. This way the time taken to offline all cpus in a
chip/core can be reduced.

With the patch, 

Time taken  to offline cpu 88-175:
real    0m12.207s
user    0m0.171s
sys     0m0.000s


Offlining all cpus in reverse order is also taken care because,
cpumask_any_but() is used to find the designated cpu if the last cpu in
the mask goes offline. Since cpumask_any_but() always return the first
cpu in the mask, that becomes the designated cpu and migration will happen
only when the first_cpu in the mask goes offline.

Example:
With the patch,

Time taken to offline cpu from 175-88:
real    0m9.330s
user    0m0.110s
sys     0m0.000s

Signed-off-by: Anju T Sudhakar <redacted>
Reviewed-by: Madhavan Srinivasan <redacted>

Applied to powerpc next, thanks.

https://git.kernel.org/powerpc/c/9c9f8fb71feed144973a70455e0a4ee3da57ed2a

cheers