Local atomic operations are fast and highly reentrant per CPU counters. Used for percpu variable updates. Local atomic operations only guarantee variable modification atomicity wrt the CPU which owns the data and these needs to be executed in a preemption safe way.
Here is the design of the patchset. Since local_* operations are only need to be atomic to interrupts (IIUC), we have two options. Either replay the "op" if interrupted or replay the interrupt after the "op". Initial patchset posted was based on implementing local_* operation based on CR5 which replay's the "op". Patchset had issues in case of rewinding the address pointor from an array. This make the slow patch really slow. Since CR5 based implementation proposed using __ex_table to find the rewind addressr, this rasied concerns about size of __ex_table and vmlinux. https://lists.ozlabs.org/pipermail/linuxppc-dev/2014-December/123115.html But this patchset uses Benjamin Herrenschmidt suggestion of using arch_local_irq_disable_var() to soft_disable interrupts (including PMIs). After finishing the "op", arch_local_irq_restore() called and correspondingly interrupts are replayed if any occured. patch re-write the current local_* functions to use arch_local_irq_disbale. Base flow for each function is { arch_local_irq_disable_var(2) load .. store arch_local_irq_restore() } Currently only asm/local.h has been rewrite, and also the entire change is tested only in PPC64 (pseries guest) Reason for the approach is that, currently l[w/d]arx/st[w/d]cx. instruction pair is used for local_* operations, which are heavy on cycle count and they dont support a local variant. So to see whether the new implementation helps, used a modified version of Rusty's benchmark code on local_t. https://lkml.org/lkml/2008/12/16/450 Modifications to Rusty's benchmark code: - Executed only local_t test Here are the values with the patch. Time in ns per iteration Local_t Without Patch With Patch _inc 28 8 _add 28 8 _read 3 3 _add_return 28 7 First four are the clean up patches which lays the foundation to make things easier. Fifth patch in the patchset reverse the current soft_enabled logic and commit message details the reason and need for this change. Sixth patch holds the changes needed for reversing logic. Rest of the patches are to add support for maskable PMI and implementation of local_t using arch_local_disable_*(). Since the patchset is experimental, changes made are focused on pseries and powernv platforms only. Would really like to know comments for this approach before extending to other powerpc platforms. Tested the patchset in a - pSeries LPAR (with perf record). - Ran kernbench with perf record for 24 hours. - More testing needed. Signed-off-by: Madhavan Srinivasan <ma...@linux.vnet.ibm.com> Madhavan Srinivasan (9): Add #defs for paca->soft_enabled flags Cleanup to use LAZY_INTERRUPT_* macros for paca->soft_enabled update powerpc: move set_soft_enabled() powerpc: Use set_soft_enabled api to update paca->soft_enabled powerpc: reverse the soft_enable logic powerpc: modify __SOFTEN_TEST to support tri-state soft_enabled flag powerpc: Add support to mask perf interrupts powerpc: Support to replay PMIs powerpc: rewrite local_t using soft_irq arch/powerpc/include/asm/exception-64s.h | 24 +++++++-- arch/powerpc/include/asm/hw_irq.h | 43 ++++++++++++--- arch/powerpc/include/asm/irqflags.h | 8 +-- arch/powerpc/include/asm/kvm_ppc.h | 2 +- arch/powerpc/include/asm/local.h | 91 ++++++++++++++++++++++---------- arch/powerpc/kernel/entry_64.S | 16 +++--- arch/powerpc/kernel/exceptions-64s.S | 27 ++++++++-- arch/powerpc/kernel/head_64.S | 3 +- arch/powerpc/kernel/idle_power4.S | 3 +- arch/powerpc/kernel/irq.c | 24 +++++---- arch/powerpc/kernel/process.c | 3 +- arch/powerpc/kernel/setup_64.c | 4 ++ arch/powerpc/kernel/time.c | 4 +- arch/powerpc/mm/hugetlbpage.c | 2 +- arch/powerpc/perf/core-book3s.c | 2 +- 15 files changed, 184 insertions(+), 72 deletions(-) -- 2.7.4 _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
