Change Details

This change adds support for transparent superpages for PowerPC64 systems using Hashed Page Tables (HPT). All pmap operations are supported. The changes were inspired by RISC-V implementation of superpages, by @markj (r344106), but heavily adapted to fit PPC64 HPT architecture and existing MMU OEA64 code. While these changes are not better tested, superpages support is disabled by default. To enable it, use `vm.pmap.sp_enabled=1`. The focus of this initial implementation is on correctness, thus, performance is not very good (yet). Below are the buildworld times of a POWER8 machine with 32GB RAM, with CURRENT kernel (r362045) using GENERIC config: ``` * Without D25237: >>> World built in 6031 seconds, ncpu: 80, make -j32 * With D25237 and vm.pmap.sp_enabled=1: >>> World built in 6742 seconds, ncpu: 80, make -j32 Slowdown: ~11.7% Counters: vm.pmap.sp.promotions: 1800 vm.pmap.sp.p_failures: 648294440 vm.pmap.sp.mappings: 1800 vm.pmap.sp.demotions: 0 vm.pmap.sp_enabled: 1 * With D25237 and vm.pmap.sp_enabled=0: >>> World built in 6105 seconds, ncpu: 80, make -j32 Slowdown: ~1.2% ``` The high count of promotion failures seem to suggest that something could be done to reduce the number of promotion failures or to speed up the code that checks if promotion is possible.

This change adds support for transparent superpages for PowerPC64 systems using Hashed Page Tables (HPT). All pmap operations are supported. The changes were inspired by RISC-V implementation of superpages, by @markj (r344106), but heavily adapted to fit PPC64 HPT architecture and existing MMU OEA64 code. While these changes are not better tested, superpages support is disabled by default. To enable it, use `vm.pmap.pg_ps_enabled=1`. The focus of this initial implementation is on correctness, thus, performance is not very good (yet). Below are the buildworld times of a POWER8 machine with 32GB RAM, with CURRENT kernel (r362045) using GENERIC config: ``` * Without D25237: >>> World built in 6031 seconds, ncpu: 80, make -j32 * With D25237 and vm.pmap.pg_ps_enabled=1: >>> World built in 6742 seconds, ncpu: 80, make -j32 Slowdown: ~11.7% Counters: vm.pmap.sp.promotions: 1800 vm.pmap.sp.p_failures: 648294440 vm.pmap.sp.mappings: 1800 vm.pmap.sp.demotions: 0 vm.pmap.pg_ps_enabled: 1 * With D25237 and vm.pmap.pg_ps_enabled=0: >>> World built in 6105 seconds, ncpu: 80, make -j32 Slowdown: ~1.2% ``` The high count of promotion failures seem to suggest that something could be done to reduce the number of promotion failures or to speed up the code that checks if promotion is possible.

This change adds support for transparent superpages for PowerPC64 systems using Hashed Page Tables (HPT). All pmap operations are supported. The changes were inspired by RISC-V implementation of superpages, by @markj (r344106), but heavily adapted to fit PPC64 HPT architecture and existing MMU OEA64 code. While these changes are not better tested, superpages support is disabled by default. To enable it, use `vm.pmap.sppg_ps_enabled=1`. The focus of this initial implementation is on correctness, thus, performance is not very good (yet). Below are the buildworld times of a POWER8 machine with 32GB RAM, with CURRENT kernel (r362045) using GENERIC config: ``` * Without D25237: >>> World built in 6031 seconds, ncpu: 80, make -j32 * With D25237 and vm.pmap.sppg_ps_enabled=1: >>> World built in 6742 seconds, ncpu: 80, make -j32 Slowdown: ~11.7% Counters: vm.pmap.sp.promotions: 1800 vm.pmap.sp.p_failures: 648294440 vm.pmap.sp.mappings: 1800 vm.pmap.sp.demotions: 0 vm.pmap.sppg_ps_enabled: 1 * With D25237 and vm.pmap.sppg_ps_enabled=0: >>> World built in 6105 seconds, ncpu: 80, make -j32 Slowdown: ~1.2% ``` The high count of promotion failures seem to suggest that something could be done to reduce the number of promotion failures or to speed up the code that checks if promotion is possible.