Messages in this thread | | | From | Arnd Bergmann <> | Date | Wed, 21 Oct 2020 17:59:48 +0200 | Subject | Re: [PATCH 2/3] arm: introduce IRQ stacks |
| |
On Wed, Oct 21, 2020 at 2:57 PM Russell King - ARM Linux admin <linux@armlinux.org.uk> wrote: > On Wed, Oct 21, 2020 at 01:45:42PM +0100, Russell King - ARM Linux admin wrote: > > > > - define 'current' as 'this_cpu_read_stable(current_task);' > > > > - convert to CONFIG_THREAD_INFO_IN_TASK > > > > That means we need to also code that up in assembly - remember, we > > need to access thread_info from assembly code. > > Note also that there is a circular dependency involved. If you make > thread_info accessible via per-cpu, then: > > #ifndef __my_cpu_offset > #define __my_cpu_offset per_cpu_offset(raw_smp_processor_id()) > #endif > #ifdef CONFIG_DEBUG_PREEMPT > #define my_cpu_offset per_cpu_offset(smp_processor_id()) > #else > #define my_cpu_offset __my_cpu_offset > #endif
Right, I had missed the fallback path using asm-generic/percpu.h that is used with CONFIG_SMP && CONFIG_CPU_V6 Almost everything either uses fixed percpu data (on UP builds) or TPIDRPRW when building a v7-only or v6k/v7 kernel without v6 support.
> smp_processor_id() ultimately ends up as raw_smp_processor_id() which > is: > > #define raw_smp_processor_id() (current_thread_info()->cpu) > > and if current_thread_info() itself involves reading from per-cpu data, > we end up recursing... infinitely. > > This is why I said in the other thread: > > "We don't do it because we don't have a separate register to be able > to store the thread_info pointer, and copying that lump between the > SVC and IRQ stack will add massively to IRQ latency, especially for > older machines."
As discussed on IRC, I think it can still be done in one of these ways, though admittedly none of them are perfect:
a) add runtime patching for __my_cpu_offset() when CONFIG_SMP_ON_UP is set. This adds complexity but avoids the fallback for for SMP&&CPU_V6. It possibly also speeds up running on single-cpu systems if the TPIDRPRW access adds any measurable runtime overhead compared to patching it out.
b) If irq stacks are left as a compile-time option, that could be made conditional on "!(SMP&&CPU_V6)". Presumably very few people still run kernels built that way any more. The only supported platforms are i.MX3, OMAP2 and Realview-eb, all of which are fairly uncommon these days and would usually run v6-only non-SMP kernels.
c) If we decide that we no longer care about that configuration at all, we could decide to just make SMP depend on !CPU_V6, and possibly kill off the entire SMP_ON_UP patching logic. I suspect we still want to keep SMP_ON_UP for performance reasons, but I don't know how significant they are to start with.
Arnd
| |