Skip to contentSkip to conversation

Optimization to improve CPU online/offline on Powerpc

PATCH New Srikar Dronamraju 1 participant 12 messages Oct 7, 2020 PowerPC

lore
Share & tools
lore
#0 lore raw

v3 cover Optimization to improve CPU online/offline on Powerpc

Changelog v2->v3:
v1 link: https://lore.kernel.org/linuxppc-dev/[email protected]/t/#u
Use GFP_ATOMIC instead of GFP_KERNEL since allocations need to
atomic at the time of CPU HotPlug.
Reported by Qian Cai <[email protected]>
Only changes in Patch 09 and Patch 11.

Changelog v1->v2:
v1 link: https://lore.kernel.org/linuxppc-dev/[email protected]/t/#u
Added five more patches on top of Seven.
Rebased to 19th Sept 2020 powerpc/next (based on v5.9-rc2)

Here are some optimizations and fixes to make CPU online/offline faster and hence result in faster bootup.

Its based on top of my v5 coregroup support patchset.
https://lore.kernel.org/linuxppc-dev/[email protected]/t/#u

Anton reported that his 4096 cpu (1024 cores in a socket) was taking too long to boot. He also analyzed that most of the time was being spent on updating cpu_core_mask.

The first two patches should solve Anton's immediate problem. On the unofficial patches, Anton reported that the boot time came from 30 mins to 6 seconds. (Basically a high core count in a single socket configuration). Satheesh also reported similar numbers.

The rest are cleanups/optimizations.

Since cpu_core_mask is an exported symbol for a long duration, lets retain as a snapshot of cpumask_of_node.

$ lscpu
Architecture:        ppc64le
Byte Order:          Little Endian
CPU(s):              1024
On-line CPU(s) list: 0-1023
Thread(s) per core:  8
Core(s) per socket:  8
Socket(s):           16
NUMA node(s):        16
Model:               2.0 (pvr 004d 0200)
Model name:          POWER8 (architected), altivec supported
Hypervisor vendor:   pHyp
Virtualization type: para
L1d cache:           64K
L1i cache:           32K
L2 cache:            512K
L3 cache:            8192K
NUMA node0 CPU(s):   0-63
NUMA node1 CPU(s):   64-127
NUMA node2 CPU(s):   128-191
NUMA node3 CPU(s):   192-255
NUMA node4 CPU(s):   256-319
NUMA node5 CPU(s):   320-383
NUMA node6 CPU(s):   384-447
NUMA node7 CPU(s):   448-511
NUMA node8 CPU(s):   512-575
NUMA node9 CPU(s):   576-639
NUMA node10 CPU(s):  640-703
NUMA node11 CPU(s):  704-767
NUMA node12 CPU(s):  768-831
NUMA node13 CPU(s):  832-895
NUMA node14 CPU(s):  896-959
NUMA node15 CPU(s):  960-1023
$ dmesg -k | grep -i -e Bringing -e Brought -e sysrq -e bug
With powerp/next
[    0.000000] printk: debug: ignoring loglevel setting.
[    0.354971] smp: Bringing up secondary CPUs ...
[  233.354676] smp: Brought up 16 nodes, 1024 CPUs
[  330.023073] sysrq: Changing Loglevel
[  330.023101] sysrq: Loglevel set to 9

With +patchset

[    0.000000] printk: debug: ignoring loglevel setting.
[    0.351703] smp: Bringing up secondary CPUs ...
[    4.059859] smp: Brought up 16 nodes, 1024 CPUs
[   98.309015] sysrq: Changing Loglevel
[   98.309044] sysrq: Loglevel set to 9

Observations:
CPU bringup time reduced to 4 seconds from 233 seconds on this 1024 CPU system. This resulted in System boot up time reducing to 98 seconds from 330 seconds. The actual improvement would depend on your system topology.

Topology verification post patchset on a 2 node Power9 PowerVM LPAR

powerpc/next +patchset
------------ ---------

$ lscpu
Architecture:        ppc64le                                        Architecture:        ppc64le
Byte Order:          Little Endian                                  Byte Order:          Little Endian
CPU(s):              128                                            CPU(s):              128
On-line CPU(s) list: 0-127                                          On-line CPU(s) list: 0-127
Thread(s) per core:  8                                              Thread(s) per core:  8
Core(s) per socket:  8                                              Core(s) per socket:  8
Socket(s):           2                                              Socket(s):           2
NUMA node(s):        2                                              NUMA node(s):        2
Model:               2.2 (pvr 004e 0202)                            Model:               2.2 (pvr 004e 0202)
Model name:          POWER9 (architected), altivec supported        Model name:          POWER9 (architected), altivec supported
Hypervisor vendor:   pHyp                                           Hypervisor vendor:   pHyp
Virtualization type: para                                           Virtualization type: para
L1d cache:           32K                                            L1d cache:           32K
L1i cache:           32K                                            L1i cache:           32K
L2 cache:            512K                                           L2 cache:            512K
L3 cache:            10240K                                         L3 cache:            10240K
NUMA node0 CPU(s):   0-63                                           NUMA node0 CPU(s):   0-63
NUMA node1 CPU(s):   64-127                                         NUMA node1 CPU(s):   64-127
$ tail -f /proc/cpuinfo
processor	: 127                                               processor	: 127
cpu		: POWER9 (architected), altivec supported           cpu		: POWER9 (architected), altivec supported
clock		: 3000.000000MHz                                    clock		: 3000.000000MHz
revision	: 2.2 (pvr 004e 0202)                               revision	: 2.2 (pvr 004e 0202)

timebase : 512000000 timebase : 512000000
platform : pSeries platform : pSeries
model : IBM,9008-22L model : IBM,9008-22L
machine : CHRP IBM,9008-22L machine : CHRP IBM,9008-22L
MMU : Radix MMU : Radix

$ grep . /proc/sys/kernel/sched_domain/cpu0/domain*/name
--------------------------------------------------------
/proc/sys/kernel/sched_domain/cpu0/domain0/name:SMT                 /proc/sys/kernel/sched_domain/cpu0/domain0/name:SMT
/proc/sys/kernel/sched_domain/cpu0/domain1/name:CACHE               /proc/sys/kernel/sched_domain/cpu0/domain1/name:CACHE
/proc/sys/kernel/sched_domain/cpu0/domain2/name:DIE                 /proc/sys/kernel/sched_domain/cpu0/domain2/name:DIE
/proc/sys/kernel/sched_domain/cpu0/domain3/name:NUMA                /proc/sys/kernel/sched_domain/cpu0/domain3/name:NUMA
$ grep . /proc/sys/kernel/sched_domain/cpu0/domain*/flags
---------------------------------------------------------
/proc/sys/kernel/sched_domain/cpu0/domain0/flags:2391               /proc/sys/kernel/sched_domain/cpu0/domain0/flags:2391
/proc/sys/kernel/sched_domain/cpu0/domain1/flags:2327               /proc/sys/kernel/sched_domain/cpu0/domain1/flags:2327
/proc/sys/kernel/sched_domain/cpu0/domain2/flags:2071               /proc/sys/kernel/sched_domain/cpu0/domain2/flags:2071
/proc/sys/kernel/sched_domain/cpu0/domain3/flags:12801              /proc/sys/kernel/sched_domain/cpu0/domain3/flags:12801

Post ppc64_cpu --smt=1

$ tail -f /proc/cpuinfo
processor	: 120                                               processor	: 120
cpu		: POWER9 (architected), altivec supported           cpu		: POWER9 (architected), altivec supported
clock		: 3000.000000MHz                                    clock		: 3000.000000MHz
revision	: 2.2 (pvr 004e 0202)                               revision	: 2.2 (pvr 004e 0202)

timebase : 512000000 timebase : 512000000
platform : pSeries platform : pSeries
model : IBM,9008-22L model : IBM,9008-22L
machine : CHRP IBM,9008-22L machine : CHRP IBM,9008-22L
MMU : Radix MMU : Radix

$ grep . /proc/sys/kernel/sched_domain/cpu0/domain*/name
--------------------------------------------------------
/proc/sys/kernel/sched_domain/cpu0/domain0/name:DIE                 /proc/sys/kernel/sched_domain/cpu0/domain0/name:DIE
/proc/sys/kernel/sched_domain/cpu0/domain1/name:NUMA                /proc/sys/kernel/sched_domain/cpu0/domain1/name:NUMA
$ grep . /proc/sys/kernel/sched_domain/cpu0/domain*/flags
---------------------------------------------------------
/proc/sys/kernel/sched_domain/cpu0/domain0/flags:2071               /proc/sys/kernel/sched_domain/cpu0/domain0/flags:2071
/proc/sys/kernel/sched_domain/cpu0/domain1/flags:12801              /proc/sys/kernel/sched_domain/cpu0/domain1/flags:12801

Performance impact post +patchset
---------------------------------
100 iterations of ebizzy
Units: Records/second : higher is better
-----------------------------------------

kernel        N    Min     Max     Median  Avg        Stddev
powerpc/next  100  753917  870520  819054  817636.56  22649.7
+patchset     100  746258  874984  816681  813876.74  26424.351

100 iterations of perf bench sched pipe -l 10000000 (aka Hackbench)
units: usec/ops: lesser is better
--------------------------------

kernel        N    Min        Max        Median     Avg        Stddev
powerpc/next  100  13.845834  14.569539  14.06263   14.086167  0.17512607
+patchset     100  13.637611  18.097744  13.862656  13.9257    0.43872453

schbench Latency percentiles (usec)
units: usec : lesser is better
-----------------------------------

powerpc/next      	+patchset
50.0000th: 48     	50.0000th: 49
75.0000th: 65     	75.0000th: 66
90.0000th: 77     	90.0000th: 79
95.0000th: 84     	95.0000th: 85
*99.0000th: 101   	*99.0000th: 99
99.5000th: 113    	99.5000th: 104
99.9000th: 159    	99.9000th: 129
min=0, max=15221  	min=0, max=7666

100 interations of ppc64_cpu --smt=1 / ppc64_cpu --smt=8
Units: seconds : lesser is better
---------------------------------
ppc64_cpu --smt=1

kernel        N    Min    Max    Median  Avg      Stddev
powerpc/next  100  13.39  17.55  14.71   14.7658  0.69184745
+patchset     100  13.3   16.27  14.33   14.4179  0.5427433

ppc64_cpu --smt=8

kernel        N    Min    Max    Median  Avg      Stddev
powerpc/next  100  21.65  26.17  23.71   23.7111  0.8589786
+patchset     100  21.88  25.79  23.16   23.2945  0.86394839

Observations:
Performance of ebizzy/ perf_sched_bench / schbench remain the same with and without the patchset.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>

Srikar Dronamraju (11):
powerpc/topology: Update topology_core_cpumask powerpc/smp: Stop updating cpu_core_mask powerpc/smp: Remove get_physical_package_id powerpc/smp: Optimize remove_cpu_from_masks powerpc/smp: Limit CPUs traversed to within a node. powerpc/smp: Stop passing mask to update_mask_by_l2 powerpc/smp: Depend on cpu_l1_cache_map when adding CPUs powerpc/smp: Check for duplicate topologies and consolidate powerpc/smp: Optimize update_mask_by_l2 powerpc/smp: Move coregroup mask updation to a new function powerpc/smp: Optimize update_coregroup_mask

 arch/powerpc/include/asm/smp.h      |   5 -
 arch/powerpc/include/asm/topology.h |   7 +-
 arch/powerpc/kernel/smp.c           | 188 +++++++++++++++++++++++-------------
 3 files changed, 122 insertions(+), 78 deletions(-)
Signature · 1 line
2.17.1
#1 lore raw

PATCH v3 1/11 powerpc/topology: Update topology_core_cpumask

On Power, cpu_core_mask and cpu_cpu_mask refer to the same set of CPUs. cpu_cpu_mask is needed by scheduler, hence look at deprecating cpu_core_mask. Before deleting the cpu_core_mask, ensure its only user is moved to cpu_cpu_mask.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/include/asm/topology.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/include/asm/topology.h +1 -1
index 6609174918ab..e0f232533c9d 100644
@@ -122,7 +122,7 @@ int get_physical_package_id(int cpu);
 #endif
 
 #define topology_sibling_cpumask(cpu)	(per_cpu(cpu_sibling_map, cpu))
-#define topology_core_cpumask(cpu)	(per_cpu(cpu_core_map, cpu))
+#define topology_core_cpumask(cpu)	(cpu_cpu_mask(cpu))
 #define topology_core_id(cpu)		(cpu_to_core_id(cpu))
 
 #endif
Signature · 1 line
2.17.1
#2 lore raw

PATCH v3 2/11 powerpc/smp: Stop updating cpu_core_mask

In reply to Srikar Dronamraju

Anton Blanchard reported that his 4096 vcpu KVM guest took around 30 minutes to boot. He also analyzed it to the time taken to iterate while setting the cpu_core_mask.

Further analysis shows that cpu_core_mask and cpu_cpu_mask for any CPU would be equal on Power. However updating cpu_core_mask took forever to update as its a per cpu cpumask variable. Instead cpu_cpu_mask was a per NODE /per DIE cpumask that was shared by all the respective CPUs.

Also cpu_cpu_mask is needed from a scheduler perspective. However cpu_core_map is an exported symbol. Hence stop updating cpu_core_map and make it point to cpu_cpu_mask.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/include/asm/smp.h |  5 -----
 arch/powerpc/kernel/smp.c      | 33 +++++++--------------------------
 2 files changed, 7 insertions(+), 31 deletions(-)
arch/powerpc/include/asm/smp.h +0 -5
index 041f0b97c45b..40e121dd16af 100644
@@ -119,11 +119,6 @@ static inline struct cpumask *cpu_sibling_mask(int cpu)
 	return per_cpu(cpu_sibling_map, cpu);
 }
 
-static inline struct cpumask *cpu_core_mask(int cpu)
-{
-	return per_cpu(cpu_core_map, cpu);
-}
-
 static inline struct cpumask *cpu_l2_cache_mask(int cpu)
 {
 	return per_cpu(cpu_l2_cache_map, cpu);
arch/powerpc/kernel/smp.c +7 -26
index 3d96752d6570..ec41491beca4 100644
@@ -953,12 +953,17 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 				local_memory_node(numa_cpu_lookup_table[cpu]));
 		}
 #endif
+		/*
+		 * cpu_core_map is now more updated and exists only since
+		 * its been exported for long. It only will have a snapshot
+		 * of cpu_cpu_mask.
+		 */
+		cpumask_copy(per_cpu(cpu_core_map, cpu), cpu_cpu_mask(cpu));
 	}
 
 	/* Init the cpumasks so the boot CPU is related to itself */
 	cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
 	cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
-	cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
 
 	if (has_coregroup_support())
 		cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
@@ -1260,9 +1265,7 @@ static void remove_cpu_from_masks(int cpu)
 {
 	int i;
 
-	/* NB: cpu_core_mask is a superset of the others */
-	for_each_cpu(i, cpu_core_mask(cpu)) {
-		set_cpus_unrelated(cpu, i, cpu_core_mask);
+	for_each_cpu(i, cpu_cpu_mask(cpu)) {
 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
 		if (has_big_cores)
@@ -1312,7 +1315,6 @@ EXPORT_SYMBOL_GPL(get_physical_package_id);
 static void add_cpu_to_masks(int cpu)
 {
 	int first_thread = cpu_first_thread_sibling(cpu);
-	int pkg_id = get_physical_package_id(cpu);
 	int i;
 
 	/*
@@ -1320,7 +1322,6 @@ static void add_cpu_to_masks(int cpu)
 	 * add it to it's own thread sibling mask.
 	 */
 	cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
-	cpumask_set_cpu(cpu, cpu_core_mask(cpu));
 
 	for (i = first_thread; i < first_thread + threads_per_core; i++)
 		if (cpu_online(i))
@@ -1342,26 +1343,6 @@ static void add_cpu_to_masks(int cpu)
 				set_cpus_related(cpu, i, cpu_coregroup_mask);
 		}
 	}
-
-	if (pkg_id == -1) {
-		struct cpumask *(*mask)(int) = cpu_sibling_mask;
-
-		/*
-		 * Copy the sibling mask into core sibling mask and
-		 * mark any CPUs on the same chip as this CPU.
-		 */
-		if (shared_caches)
-			mask = cpu_l2_cache_mask;
-
-		for_each_cpu(i, mask(cpu))
-			set_cpus_related(cpu, i, cpu_core_mask);
-
-		return;
-	}
-
-	for_each_cpu(i, cpu_online_mask)
-		if (get_physical_package_id(i) == pkg_id)
-			set_cpus_related(cpu, i, cpu_core_mask);
 }
 
 /* Activate a secondary processor. */
Signature · 1 line
2.17.1
#3 lore raw

PATCH v3 3/11 powerpc/smp: Remove get_physical_package_id

In reply to Srikar Dronamraju

Now that cpu_core_mask has been removed and topology_core_cpumask has been updated to use cpu_cpu_mask, we no more need get_physical_package_id.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/include/asm/topology.h |  5 -----
 arch/powerpc/kernel/smp.c           | 20 --------------------
 2 files changed, 25 deletions(-)
arch/powerpc/include/asm/topology.h +0 -5
index e0f232533c9d..e45219f74be0 100644
@@ -114,12 +114,7 @@ static inline int cpu_to_coregroup_id(int cpu)
 #ifdef CONFIG_PPC64
 #include <asm/smp.h>
 
-#ifdef CONFIG_PPC_SPLPAR
-int get_physical_package_id(int cpu);
-#define topology_physical_package_id(cpu)	(get_physical_package_id(cpu))
-#else
 #define topology_physical_package_id(cpu)	(cpu_to_chip_id(cpu))
-#endif
 
 #define topology_sibling_cpumask(cpu)	(per_cpu(cpu_sibling_map, cpu))
 #define topology_core_cpumask(cpu)	(cpu_cpu_mask(cpu))
arch/powerpc/kernel/smp.c +0 -20
index ec41491beca4..8c095fe237b2 100644
@@ -1292,26 +1292,6 @@ static inline void add_cpu_to_smallcore_masks(int cpu)
 	}
 }
 
-int get_physical_package_id(int cpu)
-{
-	int pkg_id = cpu_to_chip_id(cpu);
-
-	/*
-	 * If the platform is PowerNV or Guest on KVM, ibm,chip-id is
-	 * defined. Hence we would return the chip-id as the result of
-	 * get_physical_package_id.
-	 */
-	if (pkg_id == -1 && firmware_has_feature(FW_FEATURE_LPAR) &&
-	    IS_ENABLED(CONFIG_PPC_SPLPAR)) {
-		struct device_node *np = of_get_cpu_node(cpu, NULL);
-		pkg_id = of_node_to_nid(np);
-		of_node_put(np);
-	}
-
-	return pkg_id;
-}
-EXPORT_SYMBOL_GPL(get_physical_package_id);
-
 static void add_cpu_to_masks(int cpu)
 {
 	int first_thread = cpu_first_thread_sibling(cpu);
Signature · 1 line
2.17.1
#4 lore raw

PATCH v3 4/11 powerpc/smp: Optimize remove_cpu_from_masks

In reply to Srikar Dronamraju

While offlining a CPU, system currently iterate through all the CPUs in the DIE to clear sibling, l2_cache and smallcore maps. However if there are more cores in a DIE, system can end up spending more time iterating through CPUs which are completely unrelated.

Optimize this by only iterating through smaller but relevant cpumap. If shared_cache is set, cpu_l2_cache_map should be relevant else cpu_sibling_map would be relevant.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
arch/powerpc/kernel/smp.c +9 -2
index 8c095fe237b2..2e61a81aad88 100644
@@ -1263,14 +1263,21 @@ static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
 #ifdef CONFIG_HOTPLUG_CPU
 static void remove_cpu_from_masks(int cpu)
 {
+	struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
 	int i;
 
-	for_each_cpu(i, cpu_cpu_mask(cpu)) {
+	if (shared_caches)
+		mask_fn = cpu_l2_cache_mask;
+
+	for_each_cpu(i, mask_fn(cpu)) {
 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
 		if (has_big_cores)
 			set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
-		if (has_coregroup_support())
+	}
+
+	if (has_coregroup_support()) {
+		for_each_cpu(i, cpu_coregroup_mask(cpu))
 			set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
 	}
 }
Signature · 1 line
2.17.1
#5 lore raw

PATCH v3 5/11 powerpc/smp: Limit CPUs traversed to within a node.

In reply to Srikar Dronamraju

All the arch specific topology cpumasks are within a node/DIE. However when setting these per CPU cpumasks, system traverses through all the online CPUs. This is redundant.

Reduce the traversal to only CPUs that are online in the node to which the CPU belongs to.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/kernel/smp.c +1 -1
index 2e61a81aad88..c860c4950c9f 100644
@@ -1241,7 +1241,7 @@ static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
 	}
 
 	cpumask_set_cpu(cpu, mask_fn(cpu));
-	for_each_cpu(i, cpu_online_mask) {
+	for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
 		/*
 		 * when updating the marks the current CPU has not been marked
 		 * online, but we need to update the cache masks
Signature · 1 line
2.17.1
#6 lore raw

PATCH v3 6/11 powerpc/smp: Stop passing mask to update_mask_by_l2

In reply to Srikar Dronamraju

update_mask_by_l2 is called only once. But it passes cpu_l2_cache_mask as parameter. Instead of passing cpu_l2_cache_mask, use it directly in update_mask_by_l2.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
arch/powerpc/kernel/smp.c +4 -4
index c860c4950c9f..441c9c64b1e3 100644
@@ -1218,7 +1218,7 @@ static struct device_node *cpu_to_l2cache(int cpu)
 	return cache;
 }
 
-static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
+static bool update_mask_by_l2(int cpu)
 {
 	struct device_node *l2_cache, *np;
 	int i;
@@ -1240,7 +1240,7 @@ static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
 		return false;
 	}
 
-	cpumask_set_cpu(cpu, mask_fn(cpu));
+	cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
 	for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
 		/*
 		 * when updating the marks the current CPU has not been marked
@@ -1251,7 +1251,7 @@ static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
 			continue;
 
 		if (np == l2_cache)
-			set_cpus_related(cpu, i, mask_fn);
+			set_cpus_related(cpu, i, cpu_l2_cache_mask);
 
 		of_node_put(np);
 	}
@@ -1315,7 +1315,7 @@ static void add_cpu_to_masks(int cpu)
 			set_cpus_related(i, cpu, cpu_sibling_mask);
 
 	add_cpu_to_smallcore_masks(cpu);
-	update_mask_by_l2(cpu, cpu_l2_cache_mask);
+	update_mask_by_l2(cpu);
 
 	if (has_coregroup_support()) {
 		int coregroup_id = cpu_to_coregroup_id(cpu);
Signature · 1 line
2.17.1
#7 lore raw

PATCH v3 7/11 powerpc/smp: Depend on cpu_l1_cache_map when adding CPUs

In reply to Srikar Dronamraju

Currently on hotplug/hotunplug, CPU iterates through all the CPUs in its core to find threads in its thread group. However this info is already captured in cpu_l1_cache_map. Hence reduce iterations and cleanup add_cpu_to_smallcore_masks function.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
arch/powerpc/kernel/smp.c +3 -4
index 441c9c64b1e3..aeb219a4bf7a 100644
@@ -1285,16 +1285,15 @@ static void remove_cpu_from_masks(int cpu)
 
 static inline void add_cpu_to_smallcore_masks(int cpu)
 {
-	struct cpumask *this_l1_cache_map = per_cpu(cpu_l1_cache_map, cpu);
-	int i, first_thread = cpu_first_thread_sibling(cpu);
+	int i;
 
 	if (!has_big_cores)
 		return;
 
 	cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
 
-	for (i = first_thread; i < first_thread + threads_per_core; i++) {
-		if (cpu_online(i) && cpumask_test_cpu(i, this_l1_cache_map))
+	for_each_cpu(i, per_cpu(cpu_l1_cache_map, cpu)) {
+		if (cpu_online(i))
 			set_cpus_related(i, cpu, cpu_smallcore_mask);
 	}
 }
Signature · 1 line
2.17.1
#8 lore raw

PATCH v3 8/11 powerpc/smp: Check for duplicate topologies and consolidate

In reply to Srikar Dronamraju

CACHE and COREGROUP domains are now part of default topology. However on systems that don't support CACHE or COREGROUP, these domains will eventually be degenerated. The degeneration happens per CPU. Do note the current fixup_topology() logic ensures that mask of a domain that is not supported on the current platform is set to the previous domain.

Instead of waiting for the scheduler to degenerated try to consolidate based on their masks and sd_flags. This is done just before setting the scheduler topology.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)
arch/powerpc/kernel/smp.c +26 -0
index aeb219a4bf7a..6f866e6b12f8 100644
@@ -1401,6 +1401,8 @@ int setup_profiling_timer(unsigned int multiplier)
 
 static void fixup_topology(void)
 {
+	int i;
+
 #ifdef CONFIG_SCHED_SMT
 	if (has_big_cores) {
 		pr_info("Big cores detected but using small core scheduling\n");
@@ -1410,6 +1412,30 @@ static void fixup_topology(void)
 
 	if (!has_coregroup_support())
 		powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
+
+	/*
+	 * Try to consolidate topology levels here instead of
+	 * allowing scheduler to degenerate.
+	 * - Dont consolidate if masks are different.
+	 * - Dont consolidate if sd_flags exists and are different.
+	 */
+	for (i = 1; i <= die_idx; i++) {
+		if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
+			continue;
+
+		if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
+				powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
+			continue;
+
+		if (!powerpc_topology[i - 1].sd_flags)
+			powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
+
+		powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
+		powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
+#ifdef CONFIG_SCHED_DEBUG
+		powerpc_topology[i].name = powerpc_topology[i + 1].name;
+#endif
+	}
 }
 
 void __init smp_cpus_done(unsigned int max_cpus)
Signature · 1 line
2.17.1
#9 lore raw

PATCH v3 9/11 powerpc/smp: Optimize update_mask_by_l2

In reply to Srikar Dronamraju

All threads of a SMT4 core can either be part of this CPU's l2-cache mask or not related to this CPU l2-cache mask. Use this relation to reduce the number of iterations needed to find all the CPUs that share the same l2-cache.

Use a temporary mask to iterate through the CPUs that may share l2_cache mask. Also instead of setting one CPU at a time into cpu_l2_cache_mask, copy the SMT4/sub mask at one shot.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>

---
Changelog v2->v3:
Use GFP_ATOMIC instead of GFP_KERNEL since allocations need to atomic at the time of CPU HotPlug
Reported by Qian Cai <[email protected]>

 arch/powerpc/kernel/smp.c | 52 +++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 46 insertions(+), 6 deletions(-)
arch/powerpc/kernel/smp.c +46 -6
index 6f866e6b12f8..17e90c2414af 100644
@@ -671,6 +671,28 @@ static void set_cpus_unrelated(int i, int j,
 #endif
 
 /*
+ * Extends set_cpus_related. Instead of setting one CPU at a time in
+ * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
+ */
+static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
+				struct cpumask *(*dstmask)(int))
+{
+	struct cpumask *mask;
+	int k;
+
+	mask = srcmask(j);
+	for_each_cpu(k, srcmask(i))
+		cpumask_or(dstmask(k), dstmask(k), mask);
+
+	if (i == j)
+		return;
+
+	mask = srcmask(i);
+	for_each_cpu(k, srcmask(j))
+		cpumask_or(dstmask(k), dstmask(k), mask);
+}
+
+/*
  * parse_thread_groups: Parses the "ibm,thread-groups" device tree
  *                      property for the CPU device node @dn and stores
  *                      the parsed output in the thread_groups
@@ -1220,7 +1242,9 @@ static struct device_node *cpu_to_l2cache(int cpu)
 
 static bool update_mask_by_l2(int cpu)
 {
+	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
 	struct device_node *l2_cache, *np;
+	cpumask_var_t mask;
 	int i;
 
 	l2_cache = cpu_to_l2cache(cpu);
@@ -1240,22 +1264,38 @@ static bool update_mask_by_l2(int cpu)
 		return false;
 	}
 
-	cpumask_set_cpu(cpu, cpu_l2_cache_mask(cpu));
-	for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
+	/* In CPU-hotplug path, hence use GFP_ATOMIC */
+	alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
+	cpumask_and(mask, cpu_online_mask, cpu_cpu_mask(cpu));
+
+	if (has_big_cores)
+		submask_fn = cpu_smallcore_mask;
+
+	/* Update l2-cache mask with all the CPUs that are part of submask */
+	or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
+
+	/* Skip all CPUs already part of current CPU l2-cache mask */
+	cpumask_andnot(mask, mask, cpu_l2_cache_mask(cpu));
+
+	for_each_cpu(i, mask) {
 		/*
 		 * when updating the marks the current CPU has not been marked
 		 * online, but we need to update the cache masks
 		 */
 		np = cpu_to_l2cache(i);
-		if (!np)
-			continue;
 
-		if (np == l2_cache)
-			set_cpus_related(cpu, i, cpu_l2_cache_mask);
+		/* Skip all CPUs already part of current CPU l2-cache */
+		if (np == l2_cache) {
+			or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
+			cpumask_andnot(mask, mask, submask_fn(i));
+		} else {
+			cpumask_andnot(mask, mask, cpu_l2_cache_mask(i));
+		}
 
 		of_node_put(np);
 	}
 	of_node_put(l2_cache);
+	free_cpumask_var(mask);
 
 	return true;
 }
Signature · 1 line
2.17.1
#10 lore raw

PATCH v3 10/11 powerpc/smp: Move coregroup mask updation to a new function

In reply to Srikar Dronamraju

Move the logic for updating the coregroup mask of a CPU to its own function. This will help in reworking the updation of coregroup mask in subsequent patch.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>
Cc Qian Cai <[email protected]>
---
 arch/powerpc/kernel/smp.c | 32 +++++++++++++++++++-------------
 1 file changed, 19 insertions(+), 13 deletions(-)
arch/powerpc/kernel/smp.c +19 -13
index 17e90c2414af..b48ae4e306d3 100644
@@ -1338,6 +1338,23 @@ static inline void add_cpu_to_smallcore_masks(int cpu)
 	}
 }
 
+static void update_coregroup_mask(int cpu)
+{
+	int first_thread = cpu_first_thread_sibling(cpu);
+	int coregroup_id = cpu_to_coregroup_id(cpu);
+	int i;
+
+	cpumask_set_cpu(cpu, cpu_coregroup_mask(cpu));
+	for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
+		int fcpu = cpu_first_thread_sibling(i);
+
+		if (fcpu == first_thread)
+			set_cpus_related(cpu, i, cpu_coregroup_mask);
+		else if (coregroup_id == cpu_to_coregroup_id(i))
+			set_cpus_related(cpu, i, cpu_coregroup_mask);
+	}
+}
+
 static void add_cpu_to_masks(int cpu)
 {
 	int first_thread = cpu_first_thread_sibling(cpu);
@@ -1356,19 +1373,8 @@ static void add_cpu_to_masks(int cpu)
 	add_cpu_to_smallcore_masks(cpu);
 	update_mask_by_l2(cpu);
 
-	if (has_coregroup_support()) {
-		int coregroup_id = cpu_to_coregroup_id(cpu);
-
-		cpumask_set_cpu(cpu, cpu_coregroup_mask(cpu));
-		for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
-			int fcpu = cpu_first_thread_sibling(i);
-
-			if (fcpu == first_thread)
-				set_cpus_related(cpu, i, cpu_coregroup_mask);
-			else if (coregroup_id == cpu_to_coregroup_id(i))
-				set_cpus_related(cpu, i, cpu_coregroup_mask);
-		}
-	}
+	if (has_coregroup_support())
+		update_coregroup_mask(cpu);
 }
 
 /* Activate a secondary processor. */
Signature · 1 line
2.17.1
#11 lore raw

PATCH v3 11/11 powerpc/smp: Optimize update_coregroup_mask

In reply to Srikar Dronamraju

All threads of a SMT4/SMT8 core can either be part of CPU's coregroup mask or outside the coregroup. Use this relation to reduce the number of iterations needed to find all the CPUs that share the same coregroup

Use a temporary mask to iterate through the CPUs that may share coregroup mask. Also instead of setting one CPU at a time into cpu_coregroup_mask, copy the SMT4/SMT8/submask at one shot.

Cc linuxppc-dev <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Nicholas Piggin <[email protected]>
Cc Anton Blanchard <[email protected]>
Cc Oliver O'Halloran <[email protected]>
Cc Nathan Lynch <[email protected]>
Cc Michael Neuling <[email protected]>
Cc Gautham R Shenoy <[email protected]>
Cc Satheesh Rajendran <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Valentin Schneider <[email protected]>

---
Changelog v2->v3:
Use GFP_ATOMIC instead of GFP_KERNEL since allocations need to atomic at the time of CPU HotPlug
Reported by Qian Cai <[email protected]>

 arch/powerpc/kernel/smp.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)
arch/powerpc/kernel/smp.c +23 -8
index b48ae4e306d3..bbaea93dc558 100644
@@ -1340,19 +1340,34 @@ static inline void add_cpu_to_smallcore_masks(int cpu)
 
 static void update_coregroup_mask(int cpu)
 {
-	int first_thread = cpu_first_thread_sibling(cpu);
+	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
+	cpumask_var_t mask;
 	int coregroup_id = cpu_to_coregroup_id(cpu);
 	int i;
 
-	cpumask_set_cpu(cpu, cpu_coregroup_mask(cpu));
-	for_each_cpu_and(i, cpu_online_mask, cpu_cpu_mask(cpu)) {
-		int fcpu = cpu_first_thread_sibling(i);
+	/* In CPU-hotplug path, hence use GFP_ATOMIC */
+	alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
+	cpumask_and(mask, cpu_online_mask, cpu_cpu_mask(cpu));
+
+	if (shared_caches)
+		submask_fn = cpu_l2_cache_mask;
 
-		if (fcpu == first_thread)
-			set_cpus_related(cpu, i, cpu_coregroup_mask);
-		else if (coregroup_id == cpu_to_coregroup_id(i))
-			set_cpus_related(cpu, i, cpu_coregroup_mask);
+	/* Update coregroup mask with all the CPUs that are part of submask */
+	or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
+
+	/* Skip all CPUs already part of coregroup mask */
+	cpumask_andnot(mask, mask, cpu_coregroup_mask(cpu));
+
+	for_each_cpu(i, mask) {
+		/* Skip all CPUs not part of this coregroup */
+		if (coregroup_id == cpu_to_coregroup_id(i)) {
+			or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
+			cpumask_andnot(mask, mask, submask_fn(i));
+		} else {
+			cpumask_andnot(mask, mask, cpu_coregroup_mask(i));
+		}
 	}
+	free_cpumask_var(mask);
 }
 
 static void add_cpu_to_masks(int cpu)
Signature · 1 line
2.17.1