Skip to content

Clean up SMP IPI support

Russell King - ARM Linux linux-arm-kernel 1 version New

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest9 patches 0 repliesNew

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/9ARM: remove obsolete smp_cross_call_done()arch/arm/mach-tegra/include/mach/smp.h+0 -70 replies
2/9ARM: smp: pass an ipi number to smp_cross_call()arch/arm/include/asm/smp.h arch/arm/kernel/entry-armv.S arch/arm/kernel/smp.c +12 more+24 -301 reply
3/9ARM: smp: avoid using bitmasks and locks for IPIs, use hardware insteadarch/arm/kernel/smp.c+26 -611 reply
3/9ARM: smp: avoid using bitmasks and locks for IPIs, use hardware instead+0 -01 reply
4/9ARM: include local timer irq stats only when local timers configuredarch/arm/include/asm/hardirq.h arch/arm/kernel/irq.c arch/arm/kernel/smp.c+16 -120 replies
5/9ARM: smp: provide accessors for irq_stat dataarch/arm/include/asm/hardirq.h arch/arm/kernel/smp.c+5 -20 replies
6/9ARM: smp: move ipi_count into irq_stat structurearch/arm/include/asm/hardirq.h arch/arm/kernel/smp.c+5 -120 replies
7/9ARM: fix /proc/interrupts formattingarch/arm/include/asm/mach/irq.h arch/arm/include/asm/smp.h arch/arm/kernel/fiq.c +2 more+23 -180 replies
8/9ARM: smp: provide individual IPI interrupt statisticsarch/arm/include/asm/hardirq.h arch/arm/kernel/irq.c arch/arm/kernel/smp.c+27 -70 replies
9/9ARM: smp: collect IPI and local timer IRQs for /proc/statarch/arm/include/asm/hardirq.h arch/arm/kernel/smp.c+23 -00 replies

Reviews trailers per patch

None of the 10 patches has a Reviewed-by or Acked-by of their own; the series has one for all of them.

PatchCatalin Marinas
SeriesR-b
1/9
2/9
3/9
3/9
4/9
5/9
6/9
7/9
8/9
9/9

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes.

Read the v1 thread

Cover letter [PATCH 0/9] Clean up SMP IPI support

This series is aimed at cleaning up our SMP IPI support. As has been recently pointed out, we have a spinlock in percpu data, which is not very nice.

If we switch to using the GIC SGIs, one for each IPI function, then we don't need the mask of ipi bits indicating what work needs to be done. This allows us to get rid of the troublesome spinlock in send_ipi_message().

However, we end up with slightly more expense on the IPI receiver side, as they no longer have a word to read to process a set of pending IPIs. This isn't much of a down-side as you hardly end up with masses of pending IPIs. Modern Cortex A9 platforms usually only clock up the reschedule IPI.

This leaves us with just ipi_count in the ipi per-cpu data structure. Other platforms put this in the global irq_stat array, and break this down into individual IPIs. So, let's do the same.

One down-side of this patchset is that platforms which use SGI 1 to wake their secondary CPUs will receive an "Unknown IPI message" kernel critical message at boot time - as SGI 1 is no longer recognised as a Linux IPI interrupt. We _could_ make this 'IPI_BOOT', something which is only used for booting and therefore simply ignored.