powerpc/64s: cpufeatures: add initial implementation for cpufeatures
Subsystems (3)
Nicholas Piggin
reviewerRFC PATCH 1/1 powerpc/64s: cpufeatures: add initial implementation for cpufeatures
The /cpus/features dt binding describes architected CPU features along with some compatibility, privilege, and enablement properties that allow flexibility with discovering and enabling capabilities.
Presence of this feature implies a base level of functionality, then additional feature nodes advertise the presence of new features.
Using this, we can do CPU setup without matching CPU tables on PVR.
A given feature and its setup procedure is defined once and used by all CPUs which are compatible by that feature. Features that follow a supported "prescription" can be enabled by a hypervisor or OS that does not understand them natively.
Not ready for merge yet, but this gets a POWER8 booting on mambo without using PVR/cputables, couple of small bits missing, but concept seems to work.
Comments?
Thanks,
Nick
--- .../devicetree/bindings/powerpc/cpufeatures.txt | 249 +++++++++++++ arch/powerpc/include/asm/cpu_has_feature.h | 4 +- arch/powerpc/include/asm/cpufeatures.h | 53 +++ arch/powerpc/include/asm/cputable.h | 1 + arch/powerpc/kernel/Makefile | 1 + arch/powerpc/kernel/cpufeatures.c | 404 +++++++++++++++++++++ arch/powerpc/kernel/cputable.c | 14 +- arch/powerpc/kernel/prom.c | 186 +++++++++- arch/powerpc/kernel/setup-common.c | 2 +- arch/powerpc/kernel/setup_64.c | 18 +- drivers/of/fdt.c | 31 ++ include/linux/of_fdt.h | 5 + 12 files changed, 948 insertions(+), 20 deletions(-) create mode 100644 Documentation/devicetree/bindings/powerpc/cpufeatures.txt create mode 100644 arch/powerpc/include/asm/cpufeatures.h create mode 100644 arch/powerpc/kernel/cpufeatures.c
Documentation/devicetree/bindings/powerpc/cpufeatures.txt new file +249 -0
@@ -0,0 +1,249 @@ +powerpc cpu features binding +============================ + +The device tree describes supported CPU features as nodes containing +compatibility and enablement information as properties. + +The binding specifies features common to all CPUs in the system. +Heterogeneous CPU features are not supported at present (such could be added +by providing nodes with additional features and linking those to particular +CPUs). + +This binding is intended to provide fine grained control of CPU features at +all levels of the stack (firmware, hypervisor, OS, userspace), with the +ability for new CPU features to be used by some components without all +components being upgraded (e.g., a new floating point instruction could be +used by userspace math library without upgrading kernel and hypervisor). + +The binding is passed to the hypervisor by firmware. The hypervisor must +remove any features that require hypervisor enablement but that it does not +enable. It must remove any features that depend on removed features. It may +pass remaining features usable to the OS and PR to guests, depending on +configuration policy (not specified here). + +The modified binding is passed to the guest by hypervisor, with HV bit +cleared from the usable-mask and the hv-support and hfscr-bit properties +removed. The guest must similarly rmeove features that require OS enablement +that it does not enable. The OS may pass PR usable features to userspace via +ELF AUX vectors AT_HWCAP, AT_HWCAP2, AT_HWCAP3, etc., or use some other +method (outside the scope of this specification). + +The binding will specify a "base" level of features that will be present +when the cpu features binding exists. Additional features will be explicitly +specified. + +/cpus/features node binding +--------------------------- + +Node: features + +Description: Container of CPU feature nodes. + +The node name must be "features" and it must be a child of the node "/cpus". + +The node is optional but should be provided by new firmware. + +Each child node of cpufeatures represents an architected CPU feature (e.g., +a new set of vector instructions) or an important CPU performance +characteristic (e.g., fast unaligned memory operations). The specification +of each feature (instructions, registers, exceptions, etc.) will be +documented with device tree bindings. + +As a rough guide, features should be based on functional groups of changes, +those that share common enablement requirements, that share particular +registers or functionality. For example, "POWER9" would be too general, and +a new feature for every instruction would be too specific. The "summary of +changes" preface in Power ISA specification is a good guideline for the +architected features. + +Features should be "positive" where possible. For example the presence of a +feature node should indicate the presence of a new CPU feature (e.g., a +new instruction or register). An errata workaround for example would then +remove that feature from the device tree to disable it. "Negative" features +may be unavoidable in some cases. + +Properties: + +- device_type + Usage: required + Value type: string + Definition: "cpu-features" + + +/cpus/features/feature node bindings +------------------------------------ + +Node: A string describing an architected CPU feature, e.g., "vsx". + +Description: An architected feature supported by the CPUs. + +The name of the node will follow a convention such that software will +match known features by a string comparison with the node name. Presence +of the node indicates the feature is available to use (XXX: could +advertise all supported by hardware, with disabled/enabled status +property). + +The name of the child node corresponds to the name of the feature. +Software will detect known features by string matching. + +Properties: + +- isa + Usage: required + Value type: <u32> + Definition: + + First level of the Power ISA that the feature appears in. + Software should filter out features when constraining the + environment to a particular ISA version. + + Value corresponds to the "Power ISA Version" multiplied by 1000. + For example, <3000> corresponds to Version 3.0, <2070> to Version 2.07. + The minor digit is available for revisions. + +- usable-mask + Usage: required + Value type: <u32> bit mask + Definition: + bit 0 - PR (problem state / user mode) + bit 1 - OS (privileged state) + bit 2 - HV (hypervisor state) + All other bits reserved and should be zero. + + This property describes the privilege levels and/or software components + that can use the feature. + + If bit 0 is set, then the hwcap-bit-nr property will exist. + + Reserved bits must be treated as the feature being unsupported. + +- hv-support + Usage: optional + Value type: <u32> + Definition: + 1 - Custom + 2 - HFSCR + Other values reserved. + + This property describes the HV privilege state support required to + enable the feature. If the property does not exist then no support is + required. + + If the value of this property is 1, then the hypervisor must have + explicit support for this feature. + + If the value of this property is 2, then the hfscr-bit-nr property + will exist. + + Reserved values must be treated as the feature being unsupported. + +- os-support + Usage: optional + Value type: <u32> + Definition: + 1 - Custom + 2 - FSCR + Other values reserved. + + This property describes the OS privilege state support required to + enable the feature. If the property does not exist then no support is + required. + + If the value of this property is 1, then the operating system must + have explicit support for this feature. + + If the value of this property is 2, then the fscr-bit-nr property will + exist. + + Reserved values must be treated as the feature being unsupported. + +- hfscr-bit-nr + Usage: optional + Value type: <u32> + Definition: + + This property exists when the hv-support property value is 2. This + property describes the bit number in the HFSCR register that the + hypervisor must set in order to enable this feature. + +- fscr-bit-nr + Usage: optional + Value type: <u32> + Definition: + + This property exists when the os-support property value is 2. This + property describes the bit number in the FSCR register that the + operating system must set in order to enable this feature. + +- hwcap-bit-nr + Usage: optional + Value type: <u32> + Definition: + + This property may exist when the usable-mask property value has bit 0 + (PR) set. This property describes the bit number that should be set in + the ELF AUX hardware capability vectors in order to advertise this + feature to userspace. Bits 0-31 correspond to bits 0-31 in AT_HWCAP + vector. Bits 32-63 correspond to 0-31 in AT_HWCAP2 vector, and so on. + Missing AT_HWCAPx vectors implies that the feature is not enabled or + can not be advertised. Operating systems may provide a number of + unassigned hardware capability bits to allow for new features to be + advertised. + + Some properties representing features created before this binding are + advertised to userspace without a one-to-one hwcap bit number may not + specify this bit. Operating system will handle those bits specifically. + All new features usable by userspace will have a hwcap-bit-nr property. + +- dependencies + Usage: optional + Value type: <prop-encoded-array> + Definition: + + If this property exists then it is a list of phandles to cpu feature + nodes that must be enabled for this feature to be enabled. + + +Example +------- + + /cpus/features { + + darn { + isa = <3000>; + usable-mask = <1 | 2 | 4>; + hwcap-bit-nr = <xx>; + }; + + scv { + isa = <3000>; + usable-mask = <1 | 2>; + os-support = <1>; + hwcap-bit-nr = <xx>; + }; + + stop { + isa = <3000>; + usable-mask = <2 | 4>; + hv-support = <1>; + os-support = <1>; + }; + + vsx2 (hypothetical) { + isa = <3010>; + usable-mask = <1 | 2 | 4>; + hv-support = <1>; + os-support = <1>; + hwcap-bit-nr = <xx>; + }; + + vsx2-newinsns { + isa = <3020>; + usable-mask = <1 | 2 | 4>; + os-support = <2>; + fscr-bit-nr = <xx>; + hwcap-bit-nr = <xx>; + dependencies = <&vsx2>; + }; + + };
arch/powerpc/include/asm/cpu_has_feature.h +2 -2
@@ -1,5 +1,5 @@ -#ifndef __ASM_POWERPC_CPUFEATURES_H -#define __ASM_POWERPC_CPUFEATURES_H +#ifndef __ASM_POWERPC_CPU_HAS_FEATURE_H +#define __ASM_POWERPC_CPU_HAS_FEATURE_H#ifndef __ASSEMBLY__
arch/powerpc/include/asm/cpufeatures.h new file +53 -0
53 lines not on this page · raw message
arch/powerpc/include/asm/cputable.h +1 -0
@@ -118,6 +118,7 @@ extern struct cpu_spec *cur_cpu_spec;extern unsigned int __start___ftr_fixup, __stop___ftr_fixup; +extern void set_cur_cpu_spec (struct cpu_spec *s);extern struct cpu_spec *identify_cpu (unsigned long offset,unsigned int pvr);extern void do_feature_fixups (unsigned long value,void *fixup_start,void *fixup_end);
arch/powerpc/kernel/Makefile +1 -0
@@ -44,6 +44,7 @@ obj-$(CONFIG_PPC64) += setup_64.o sys_ppc32.o \ obj-$(CONFIG_VDSO32) += vdso32/ obj-$(CONFIG_HAVE_HW_BREAKPOINT) += hw_breakpoint.o obj-$(CONFIG_PPC_BOOK3S_64) += cpu_setup_ppc970.o cpu_setup_pa6t.o +obj-$(CONFIG_PPC_BOOK3S_64) += cpufeatures.o obj-$(CONFIG_PPC_BOOK3S_64) += cpu_setup_power.o obj-$(CONFIG_PPC_BOOK3S_64) += mce.o mce_power.o obj-$(CONFIG_PPC_BOOK3E_64) += exceptions-64e.o idle_book3e.o
arch/powerpc/kernel/cpufeatures.c new file +404 -0
404 lines not on this page · raw message
arch/powerpc/kernel/cputable.c +12 -2
@@ -23,7 +23,9 @@#include <asm/mmu.h> #include <asm/setup.h> -struct cpu_spec * cur_cpu_spec =NULL ; +static struct cpu_spec the_cpu_spec__read_mostly ; + +struct cpu_spec * cur_cpu_spec__read_mostly =NULL ;EXPORT_SYMBOL (cur_cpu_spec);/* The platform string corresponding to the real PVR */
@@ -2159,7 +2161,15 @@ static struct cpu_spec __initdata cpu_specs[] = {#endif /* CONFIG_E500 */ }; -static struct cpu_spec the_cpu_spec; +void __init set_cur_cpu_spec (struct cpu_spec *s) +{ +struct cpu_spec *t = &the_cpu_spec; + + t =PTRRELOC (t); + *t = *s; + + *PTRRELOC (&cur_cpu_spec) = &the_cpu_spec; +}static struct cpu_spec *__init setup_cpu_spec (unsigned long offset,struct cpu_spec *s)
arch/powerpc/kernel/prom.c +173 -13
232 lines not on this page · raw message
arch/powerpc/kernel/setup-common.c +1 -1
@@ -247,7 +247,7 @@ static int show_cpuinfo(struct seq_file *m, void *v)seq_printf (m,"processor\t: %lu\n" , cpu_id);seq_printf (m,"cpu\t\t: " ); -if (cur_cpu_spec->pvr_mask) +if (cur_cpu_spec->pvr_mask && cur_cpu_spec->cpu_name)seq_printf (m,"%s" , cur_cpu_spec->cpu_name);else seq_printf (m,"unknown (%08x)" , pvr);
arch/powerpc/kernel/setup_64.c +16 -2
@@ -49,6 +49,7 @@#include <asm/paca.h> #include <asm/time.h> #include <asm/cputable.h> +#include <asm/cpufeatures.h> #include <asm/sections.h> #include <asm/btext.h> #include <asm/nvram.h>
@@ -271,11 +272,22 @@ static void cpu_ready_for_interrupts(void)void __init early_setup (unsigned long dt_ptr) {static __initdata struct paca_struct boot_paca; +int cpufeatures =0 ;/* -------- printk is _NOT_ safe to use here ! ------- */ -/* Identify CPU type */ -identify_cpu (0 ,mfspr (SPRN_PVR )); +#ifdef CONFIG_PPC_BOOK3S_64 +/* Setup flat device-tree pointer */ + initial_boot_params =__va (dt_ptr); +if (early_init_devtree_check_cpu_features_exists ()) { + cpufeatures =1 ; +cpufeatures_setup_cpu (); + } else +#endif + { +/* Legacy table-based approach when /cpus/features is missing */ +identify_cpu (0 ,mfspr (SPRN_PVR )); + }/* Assume we're on cpu 0 for now. Don't write to the paca yet! */ initialise_paca (&boot_paca,0 );
@@ -289,6 +301,8 @@ void __init early_setup(unsigned long dt_ptr)DBG (" -> early_setup(), dt_ptr: 0x%lx\n" , dt_ptr); +printk ("cpufeatures:%d\n" , cpufeatures); +/* * Do early initialization using the flattened device * tree, such as retrieving the physical memory map or
drivers/of/fdt.c +31 -0
@@ -751,6 +751,37 @@ int __init of_scan_flat_dt(int (*it)(unsigned long node, }/** +* of_scan_flat_dt_subnodes - scan sub-nodes of a node call callback on each. +* @it: callback function +* @data: context data pointer +* +* This function is used to scan sub-nodes of a node. +*/ +int __init of_scan_flat_dt_subnodes (unsigned long node, +int (*it)(unsigned long node, +const char *uname, +void *data), +void *data) +{ +const void *blob = initial_boot_params; +const char *pathp; +int offset, rc =0 ; + + offset = node; +for (offset =fdt_first_subnode (blob, offset); + offset >=0 && !rc; + offset =fdt_next_subnode (blob, offset)) { + + pathp =fdt_get_name (blob, offset,NULL ); +if (*pathp =='/' ) + pathp =kbasename (pathp); + rc =it (offset, pathp, data); + } +return rc; +} + + +/** * of_get_flat_dt_subnode_by_name - get the subnode by given name * * @node: the parent node
include/linux/of_fdt.h +5 -0
@@ -54,6 +54,11 @@ extern char __dtb_end[];extern int of_scan_flat_dt (int (*it)(unsigned long node,const char *uname,int depth,void *data),void *data); +extern int of_scan_flat_dt_subnodes (unsigned long node, +int (*it)(unsigned long node, +const char *uname, +void *data), +void *data);extern int of_get_flat_dt_subnode_by_name (unsigned long node,const char *uname); extern const void *of_get_flat_dt_prop(unsigned long node, const char *name,
Signature · 1 line
2.11.0
Nicholas Piggin
reviewerRFC PATCH 1/1 cpufeatures: add base and POWER8 dt
With this patch and the Linux one, I can boot (in mambo) a POWER8 without looking up any cpu tables. It's not quite complete, but getting close.
Main questions are what features do we add, and what "base" feature set do we start with. Where to discuss and document these features. OPAL firmware group may become the "owner" of specification, but we will have to involve userspace, the proprietary stack (if we can convince them), etc.
For the base feature set I've started with ISAv207 minus some things:
a few large chunks like TM and FP and MMU, things that have FSCR bits, CPU_FTR bits, things that have been missing from DD1 cores, etc.
Not that there are plans to drop any of these things that I know of, but the more such "optional" things we have, the easier it will be for different feature sets or quirks to be backward compatible -- whereas adding a new negative feature to switch off something is not backward compatible.
A few immediate pain points getting the Linux side going are:
- The PMU, that's for another day. We can cope with booting without a known PMU, or provide some subset functionality.
- TLB invalidate size, which we should be able to deal with using the tlb-congurence-classes property.
- The machine check handler. Moving most of that to OPAL will solve it. For POWER8 and possibly 9, processor specific MCE can be advertised for existing CPUs.
- Would be nice to avoid OPAL_REINIT_CPUS requiring the host to re-setup system registers (FSCR, LPCR, etc). If firmware could always save/restore, it would make life easier (for now I've ignored that on the Linux side). Ben said this only affects POWER8 DD1 and could be changed.
Thanks,
Nick
--- core/device.c | 7 ++ core/init.c | 3 + hdata/cpu-common.c | 305 +++++++++++++++++++++++++++++++++++++++++++++++++++++ include/device.h | 1 + 4 files changed, 316 insertions(+)
core/device.c +7 -0
@@ -548,6 +548,13 @@ u32 dt_property_get_cell(const struct dt_property *prop, u32 index)return fdt32_to_cpu (((const u32 *)prop->prop)[index]); } +void dt_property_set_cell (struct dt_property *prop,u32 index,u32 val) +{ +assert (prop->len >= (index+1 )*sizeof (u32)); +/* Always aligned, so this works. */ + ((u32 *)prop->prop)[index] =cpu_to_fdt32 (val); +} +/* First child of this node. */ struct dt_node *dt_first (const struct dt_node *root) {
core/init.c +3 -0
@@ -703,6 +703,8 @@ static void per_thread_sanity_checks(void)/* Called from head.S, thus no prototype. */ void main_cpu_entry (const void *fdt); +extern void mambo_add_cpu_features (struct dt_node *root); + void __noreturn __nomcount main_cpu_entry(const void *fdt) {/*
@@ -774,6 +776,7 @@ void __noreturn __nomcount main_cpu_entry(const void *fdt)abort (); }else {dt_expand (fdt); +mambo_add_cpu_features (dt_root); }/* Now that we have a full devicetree, verify that we aren't on fire. */
hdata/cpu-common.c +305 -0
338 lines not on this page · raw message
include/device.h +1 -0
@@ -125,6 +125,7 @@ void dt_check_del_prop(struct dt_node *node, const char *name);/* Warning: moves *prop! */ void dt_resize_property (struct dt_property **prop,size_t len); +void dt_property_set_cell (struct dt_property *prop,u32 index,u32 val);u32 dt_property_get_cell (const struct dt_property *prop,u32 index);/* First child of this node. */
Signature · 1 line
2.11.0