Messages in this thread | | | Date | Fri, 01 Feb 2008 11:14:14 -0800 | From | travis@sgi ... | Subject | [PATCH 0/4] percpu: Optimize percpu accesses |
| |
This patchset provides the following:
* Generic: Percpu infrastructure to rebase the per cpu area to zero
This provides for the capability of accessing the percpu variables using a local register instead of having to go through a table on node 0 to find this cpu specific offsets. It also would allow atomic operations on percpu variables to reduce required locking.
* Init: Move setup of nr_cpu_ids to as early as possible for usage by early boot functions.
* x86_64: Fold pda into per cpu area
Declare the pda as a per cpu variable. This will move the pda area to an address accessible by the x86_64 per cpu macros. Subtraction of __per_cpu_start will make the offset based from the beginning of the per cpu area. Since %gs is pointing to the pda, it will then also point to the per cpu variables and can be accessed thusly:
%gs:[&per_cpu_xxxx - __per_cpu_start]
* x86_64: Rebase per cpu variables to zero
Take advantage of the zero-based per cpu area provided above. Then we can directly use the x86_32 percpu operations. x86_32 offsets %fs by __per_cpu_start. x86_64 has %gs pointing directly to the pda and the per cpu area thereby allowing access to the pda with the x86_64 pda operations and access to the per cpu variables using x86_32 percpu operations. After rebasing the access now becomes:
%gs:[&per_cpu_xxxx]
Introduces a new DEFINE_PER_CPU_FIRST to locate the percpu variable (pda in this case) at the beginning of the percpu .data section.
* x86_64: Cleanup non-smp usage of cpu maps
Cleanup references to the early cpu maps for the non-SMP configuration and remove some functions called for SMP configurations only.
Based on linux-2.6.git + x86.git
Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Mike Travis <travis@sgi.com> --- Notes:
(1 - had to disable CONFIG_SIS190 to build) (2 - no modules)
Configs built and booted:
x86_64-default x86_64-defconfig (2) x86_64-nonuma (2) x86_64-nosmp (2) x86_64-"Ingo Stress Test" (1,2)
Configs built with no errors:
arm-default i386-allyesconfig (1) i386-allmodconfig (1) i386-defconfig i386-nosmp ppc-pmac32 ppc-smp sparc64-default sparc64-smp x86_64-allmodconfig (1) x86_64-allyesconfig (1) x86_64-maxsmp (NR_CPUS=4k, MAXNODES=512)
Configs with errors prior to patch (preventing full build checkout):
ia64-sn2: undefined reference to `mem_map' (more) ia64-default: (same error) ia64-nosmp: `per_cpu__kstat' truncated in .bss (more) s390-default: implicit declaration of '__raw_spin_is_contended' sparc-default: include/asm/pgtable.h: syntax '___f___swp_entry'
Memory Effects (using x86_64-maxsmp config):
Note that 1/2MB has been moved from permanent data to the init data section, (which is removed after bootup), while the per cpu section is only increased by 128 bytes per cpu. Also text size is reduced increasing cache performance.
4k-cpus-before 4k-cpus-after 6588928 .data.cacheline_alig -524288 -7% 48072 .data.percpu +128 +0% 4804576 .data.read_mostly -32656 +0% 854048 .init.data +557056 +65% 160382 .init.text +62 +0% 1254214 .rodata +274 +0% 3915552 .text -1632 +0% 11040 __param -272 -2%
3915552 Text -1632 +0% 1085440 InitData +557056 +51% 11454056 OtherData -557056 -4% 48072 PerCpu +128 +0% 20459748 Total -1330 +0%
--
| |