| Message ID | 20200203235107.190609-1-gshan@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | arm64: Dereference CPU operations indirectly | expand |
On 2/4/20 10:51 AM, Gavin Shan wrote: > In current implementation, an array (@cpu_ops[NR_CPUS]) is maintained > to dereference the CPU operations. 2KB memory are consumed when NR_CPUS > is configured to 256. It's too much than what I expected. This series > reworks the implementation to dereference the CPU operations indirectly, > so that less memory (64 bytes) will be consumed for the same purpose. > The optimization bases on the assumption: these CPU operations aren't > dereferenced in hot path. > > PATCH[1/4] isn't too much relevant, to declare ACPI parking protocol only > when CONFIG_ARM64_ACPI_PARKING_PROTOCOL has been enabled. PATCH[2/4] > renames cpu_read_ops() to init_cpu_ops(), which is obviously more precise > because it's initializing the CPU operations. PATCH[3/4] introduces > get_cpu_ops(), preparing for dereferencing CPU operations indirectly. > PATCH[4/4] makes dereferencing CPU operations indirectly happen. > > Changelog > ========= > v2: > * Pack 4 CPUs' indexes into one byte. 64 bytes are consumed in order > to get the CPU operations (Robin Murphy) > * Use ARRAY_SIZE() to iterate @cpu_ops[] (Robin Murphy) > * Make index-0 valid (Robin Murphy) > Robin, Please let me know if you have more comments on this series, thanks! Thanks, Gavin > Gavin Shan (4): > arm64: Declare ACPI parking protocol CPU operation if needed > arm64: Rename cpu_read_ops() to init_cpu_ops() > arm64: Introduce get_cpu_ops() helper function > arm64: Dereference CPU operations indirectly > > arch/arm64/include/asm/cpu_ops.h | 8 ++--- > arch/arm64/kernel/cpu_ops.c | 56 ++++++++++++++++-------------- > arch/arm64/kernel/cpuidle.c | 9 ++--- > arch/arm64/kernel/setup.c | 8 +++-- > arch/arm64/kernel/smp.c | 59 ++++++++++++++++++++++---------- > 5 files changed, 85 insertions(+), 55 deletions(-) >
In current implementation, an array (@cpu_ops[NR_CPUS]) is maintained to dereference the CPU operations. 2KB memory are consumed when NR_CPUS is configured to 256. It's too much than what I expected. This series reworks the implementation to dereference the CPU operations indirectly, so that less memory (64 bytes) will be consumed for the same purpose. The optimization bases on the assumption: these CPU operations aren't dereferenced in hot path. PATCH[1/4] isn't too much relevant, to declare ACPI parking protocol only when CONFIG_ARM64_ACPI_PARKING_PROTOCOL has been enabled. PATCH[2/4] renames cpu_read_ops() to init_cpu_ops(), which is obviously more precise because it's initializing the CPU operations. PATCH[3/4] introduces get_cpu_ops(), preparing for dereferencing CPU operations indirectly. PATCH[4/4] makes dereferencing CPU operations indirectly happen. Changelog ========= v2: * Pack 4 CPUs' indexes into one byte. 64 bytes are consumed in order to get the CPU operations (Robin Murphy) * Use ARRAY_SIZE() to iterate @cpu_ops[] (Robin Murphy) * Make index-0 valid (Robin Murphy) Gavin Shan (4): arm64: Declare ACPI parking protocol CPU operation if needed arm64: Rename cpu_read_ops() to init_cpu_ops() arm64: Introduce get_cpu_ops() helper function arm64: Dereference CPU operations indirectly arch/arm64/include/asm/cpu_ops.h | 8 ++--- arch/arm64/kernel/cpu_ops.c | 56 ++++++++++++++++-------------- arch/arm64/kernel/cpuidle.c | 9 ++--- arch/arm64/kernel/setup.c | 8 +++-- arch/arm64/kernel/smp.c | 59 ++++++++++++++++++++++---------- 5 files changed, 85 insertions(+), 55 deletions(-)