Index | Thread | Search

From:
Greg Schaefer <gsgs7878@proton.me>
Subject:
Re: Mostly-Inverted CPU-Core ordering
To:
Stefan Fritsch <sf@sfritsch.de>
Cc:
Mark Kettenis <mark.kettenis@xs4all.nl>, tech@openbsd.org, visa@openbsd.org
Date:
Wed, 14 Jan 2026 16:09:54 +0000

Download raw body.

Thread


> > > visa@ Theo suggested to cc you because of mips64
> > > 
> > > On Tue, 13 Jan 2026, Stefan Fritsch wrote:
> > > 
> > > > On Mon, 12 Jan 2026, Greg Schaefer wrote:
> > > > 
> > > > > While debugging another issue, I noticed that amd64/cpu_attach (and others
> > > > > it appears) reorder the ACPI x2APIC CPU list. In a hybrid world, think that
> > > > > Intel recommends x2APIC ordering by "CPU power" (P-cores first, E-cores next,
> > > > > LP-cores last). cpu_info_list contains a static primary (boot-CPU) record
> > > > > and a linked-list to the remaining application processors. The boot core is
> > > > > copied to the static allocation, but the remaining cores are head-inserted
> > > > > (reversed) into the linked-list.
> > > > > 
> > > > > case CPU_ROLE_AP:
> > > > > #if defined(MULTIPROCESSOR)
> > > > > cpu_intr_init(ci);
> > > > > cpu_start_secondary(ci);
> > > > > clockqueue_init(&ci->ci_queue);
> > > > > sched_init_cpu(ci);
> > > > > ncpus++;
> > > > > if (ci->ci_flags & CPUF_PRESENT) {
> > > > > ** ci->ci_next = cpu_info_list->ci_next;
> > > > > ** cpu_info_list->ci_next = ci;
> > > > > 
> > > > > Using the example of a hybrid process with 2 P-cores, 2 E-cores and 2 LP-cores,
> > > > > the resulting ordering is P-Core 0, LP-Core 1, LP-Core 0, E-Core 1, E-core 0,
> > > > > P-core 0. When NICs setup their interrupt queues (often power-of-two), via
> > > > > intrmap_create(), they are striped across the cores in that mostly-reversed
> > > > > order.
> > > > > 
> > > > > Unclear what, if any, performance impact this has, but it can make debugging
> > > > > interesting. The simple fix is appending to cpu_info_list instead of reverse
> > > > > head-inserting. Please ignore if a known/non-issue.
> > > > 
> > > > I can confirm your analysis. I don't have access to a hybrid CPU system
> > > > at the moment, but I have Linux cpuinfo and acpi table infos from two
> > > > Dell laptops with Core i7-1270P and Core i7-1370P and they seem to have
> > > > the P-cores before the E-cores in the MADT. Also, if I enable INTRDEBUG
> > > > in amd64/intr.c, I see that CPU_INFO_FOREACH traverses the CPUs in the
> > > > order 0,5,4,3,2,1 on a 6 VCPU VM and intrmap_create() uses
> > > > CPU_INFO_FOREACH.
> > > > 
> > > > I also agree that one probably wants to have multi-queue NICs use the
> > > > CPUs with the highest performance and re-ordering the cpu_info_list
> > > > seems to have that effect at least on intel. The amd64 diff below puts
> > > > them in the correct order and seems to work. But the same code seems to
> > > > be in most (all?) other MP architectures.
> > > > 
> > > > I don't know what effect the different ordering could have on non-hybrid
> > > > systems. On multi-socket systems, the interrupt distribution over the
> > > > sockets may change, but I don't see a problem with that. The order in
> > > > which IPIs are send will also change, and the scheduler will likely also
> > > > do things slightly differently, but I have no idea if that would have a
> > > > measurable impact. Anything else?
> > > 
> > > Updated diff below that should fix all architectures. sparc64 already has
> > > the correct ordering. Boot tested on amd64, i386, arm64.
> > 
> > On arm64, the slow cores are typically (but not always) listed before
> > the fast cores. Don't think that is a good reason not to change the
> > order of the list to be more sensible.
> 
> 
> Does this mean that the boot cpu is then a slow core, too?
> 
> In the long term, one probably wants to have some field in cpu_info that
> denotes the core type and then make intrmap_create() use that info to
> prefer the fastest cores.

intrmap_cpus_get() builds the cpu-map that intrmap_create() uses. It already
has some notion of excluding cores (I filtered out LP-cores there). In a hybrid
world, 8 E-core queues might perform better than 4 P-core queues. Rather than
try to guess, seems like a configuration to either include or exclude by core
type would is cleaner (of course predicated on having a notion of "core type").

> As noted by tb@, the previous diff broke !MULTIPROCESSOR configurations.
> New diff attached.
> 
> diff --git a/sys/arch/alpha/alpha/cpu.c b/sys/arch/alpha/alpha/cpu.c
> index 23b25c94219..38425f1d905 100644
> --- a/sys/arch/alpha/alpha/cpu.c
> +++ b/sys/arch/alpha/alpha/cpu.c
> @@ -419,7 +419,7 @@ cpu_announce_extensions(struct cpu_info *ci)
> void
> cpu_boot_secondary_processors(void)
> {
> - struct cpu_info *ci;
> + struct cpu_info *ci, ci_last;
> u_long i;
> 
> for (i = 0; i < ALPHA_MAXPROCS; i++) {
> @@ -434,8 +434,10 @@ cpu_boot_secondary_processors(void)
> /
> * Link the processor into the list, and launch it.
> */
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> atomic_setbits_ulong(&ci->ci_flags, CPUF_RUNNING);
> 
> atomic_setbits_ulong(&cpus_running, (1UL << i));
> ncpus++;
> diff --git a/sys/arch/amd64/amd64/cpu.c b/sys/arch/amd64/amd64/cpu.c
> index 18cfb8d6377..4d56f320c33 100644
> --- a/sys/arch/amd64/amd64/cpu.c
> +++ b/sys/arch/amd64/amd64/cpu.c
> @@ -729,8 +729,11 @@ cpu_attach(struct device *parent, struct device *self, void *aux)
> sched_init_cpu(ci);
> ncpus++;
> if (ci->ci_flags & CPUF_PRESENT) {
> 
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + struct cpu_info *ci_last = cpu_info_list;
> +
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> }
> #else
> printf("%s: not started\n", sc->sc_dev.dv_xname);
> 
> diff --git a/sys/arch/arm/arm/cpu.c b/sys/arch/arm/arm/cpu.c
> index f51404f73be..005211ba35f 100644
> --- a/sys/arch/arm/arm/cpu.c
> +++ b/sys/arch/arm/arm/cpu.c
> @@ -354,10 +354,14 @@ cpu_attach(struct device *parent, struct device *dev, void *aux)
> ci = &cpu_info_primary;
> ci->ci_flags |= CPUF_RUNNING | CPUF_PRESENT | CPUF_PRIMARY;
> 
> } else {
> + struct cpu_info *ci_last;
> +
> ci = malloc(sizeof(*ci), M_DEVBUF, M_WAITOK | M_ZERO);
> cpu_info[dev->dv_unit] = ci;
> 
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> ci->ci_flags |= CPUF_AP;
> 
> ncpus++;
> }
> diff --git a/sys/arch/arm64/arm64/cpu.c b/sys/arch/arm64/arm64/cpu.c
> index f9784e432db..8f59fc03e0f 100644
> --- a/sys/arch/arm64/arm64/cpu.c
> +++ b/sys/arch/arm64/arm64/cpu.c
> @@ -1481,6 +1481,7 @@ cpu_attach(struct device *parent, struct device *dev, void *aux)
> struct cpu_info *ci;
> void *kstack;
> #ifdef MULTIPROCESSOR
> + struct cpu_info *ci_last;
> uint64_t mpidr = READ_SPECIALREG(mpidr_el1);
> #endif
> uint32_t opp;
> @@ -1494,8 +1495,10 @@ cpu_attach(struct device *parent, struct device *dev, void *aux)
> } else {
> ci = malloc(sizeof(*ci), M_DEVBUF, M_WAITOK | M_ZERO);
> cpu_info[dev->dv_unit] = ci;
> 
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> ci->ci_flags |= CPUF_AP;
> 
> ncpus++;
> }
> diff --git a/sys/arch/i386/i386/cpu.c b/sys/arch/i386/i386/cpu.c
> index c5f18dd4433..e3b4d7cc589 100644
> --- a/sys/arch/i386/i386/cpu.c
> +++ b/sys/arch/i386/i386/cpu.c
> @@ -234,6 +234,7 @@ cpu_attach(struct device *parent, struct device *self, void *aux)
> struct cpu_info *ci;
> 
> #ifdef MULTIPROCESSOR
> + struct cpu_info *ci_last;
> int cpunum = sc->sc_dev.dv_unit;
> 
> vaddr_t kstack;
> struct pcb *pcb;
> @@ -365,8 +366,10 @@ cpu_attach(struct device *parent, struct device *self, void *aux)
> identifycpu(ci);
> clockqueue_init(&ci->ci_queue);
> 
> sched_init_cpu(ci);
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> ncpus++;
> #endif
> break;
> diff --git a/sys/arch/mips64/mips64/cpu.c b/sys/arch/mips64/mips64/cpu.c
> index 040669d397b..fe45ce0970c 100644
> --- a/sys/arch/mips64/mips64/cpu.c
> +++ b/sys/arch/mips64/mips64/cpu.c
> @@ -95,9 +95,13 @@ cpuattach(struct device *parent, struct device *dev, void *aux)
> panic("unable to allocate cpu_info");
> }
> } else {
> + struct cpu_info *ci_last;
> +
> ci = &cpu_info_secondaries[cpuno - 1];
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> ci->ci_flags |= CPUF_PRESENT;
> 
> }
> #else
> diff --git a/sys/arch/riscv64/riscv64/cpu.c b/sys/arch/riscv64/riscv64/cpu.c
> index e8400d4eed3..1c2c2a1341c 100644
> --- a/sys/arch/riscv64/riscv64/cpu.c
> +++ b/sys/arch/riscv64/riscv64/cpu.c
> @@ -195,10 +195,14 @@ cpu_attach(struct device *parent, struct device *dev, void *aux)
> ci->ci_flags |= CPUF_RUNNING | CPUF_PRESENT | CPUF_PRIMARY;
> 
> csr_set(sie, SIE_SSIE);
> } else {
> + struct cpu_info *ci_last;
> +
> ci = malloc(sizeof(*ci), M_DEVBUF, M_WAITOK | M_ZERO);
> cpu_info[dev->dv_unit] = ci;
> 
> - ci->ci_next = cpu_info_list->ci_next;
> 
> - cpu_info_list->ci_next = ci;
> 
> + ci_last = cpu_info_list;
> + while (ci_last->ci_next != NULL)
> 
> + ci_last = ci_last->ci_next;
> 
> + ci_last->ci_next = ci;
> 
> ci->ci_flags |= CPUF_AP;
> 
> ncpus++;
> }