Now that all the Linux specifics have been factored out of the load programs that fp-stress uses and we have a simple VMM we can build versions for running with that VMM. Provide an asm-utils-kvm.S with startup code that's easier in the guest and implementations of the functions that the programs call, writing output to the fixed address expected by the VMM and using PSCI to shut down on error.
The programs are built into flat binaries to simplify the VMM, this is a bit more Makefile effort but it's all fairly standard for bare metal programming. We install single instruction BRK handlers for all exceptions with the expectation that the VMM will trap BRKs and log exception state, we do not expect to trigger any exceptions in normal operation. We log the current EL to help with diagnostics. Since we're constantly reading and writing the contents of the floating point registers to and from memory there is fairly heavy memory traffic. We do not currently turn the MMU on so on hardware the number of iterations a test does will be very much lower that that seen on the host, this can be addressed in future. Signed-off-by: Mark Brown <[email protected]> --- tools/testing/selftests/arm64/fp/.gitignore | 3 + tools/testing/selftests/arm64/fp/Makefile | 40 +++++++++- tools/testing/selftests/arm64/fp/asm-utils-kvm.S | 86 ++++++++++++++++++++++ .../testing/selftests/arm64/fp/fp-stress-kvm.lds.S | 58 +++++++++++++++ 4 files changed, 186 insertions(+), 1 deletion(-) diff --git a/tools/testing/selftests/arm64/fp/.gitignore b/tools/testing/selftests/arm64/fp/.gitignore index 229a5c65b65f..7a788b23ae76 100644 --- a/tools/testing/selftests/arm64/fp/.gitignore +++ b/tools/testing/selftests/arm64/fp/.gitignore @@ -1,6 +1,7 @@ fp-pidbench fp-ptrace fp-stress +fp-stress-kvm.lds fp-stress-vmm fpsimd-test kernel-test @@ -17,3 +18,5 @@ za-ptrace za-test zt-ptrace zt-test +*-kvm.bin +*-kvm.elf diff --git a/tools/testing/selftests/arm64/fp/Makefile b/tools/testing/selftests/arm64/fp/Makefile index 3c4e17a34776..5322cb66f0a9 100644 --- a/tools/testing/selftests/arm64/fp/Makefile +++ b/tools/testing/selftests/arm64/fp/Makefile @@ -21,17 +21,41 @@ TEST_GEN_PROGS_EXTENDED := fp-pidbench fp-stress-vmm fpsimd-test \ zt-test \ vlset TEST_PROGS_EXTENDED := fpsimd-stress sve-stress ssve-stress za-stress +TEST_GEN_FILES := \ + fpsimd-test-kvm.bin \ + sve-test-kvm.bin \ + ssve-test-kvm.bin \ + za-test-kvm.bin \ + zt-test-kvm.bin EXTRA_CLEAN += $(OUTPUT)/asm-utils.o $(OUTPUT)/asm-utils-linux.o \ - $(OUTPUT)/rdvl.o $(OUTPUT)/za-fork-asm.o + $(OUTPUT)/rdvl.o $(OUTPUT)/za-fork-asm.o \ + $(patsubst %.bin,%.elf,$(TEST_GEN_FILES)) \ + $(KVM_LDS) ASM_UTILS_LINUX := $(OUTPUT)/asm-utils.o $(OUTPUT)/asm-utils-linux.o +ASM_UTILS_KVM := asm-utils.S asm-utils-kvm.S + +OBJCOPY ?= $(CROSS_COMPILE)objcopy # For sysreg definitons ARCH_INCLUDES := $(KHDR_INCLUDES) -I$(top_srcdir)/tools/arch/arm64/include \ -I$(top_srcdir)/tools/include \ -I$(top_srcdir)/tools/include/generated +# The guests run with no OS under them at all, link them as flat binaries +# with the link script placing them where fp-stress-vmm puts their RAM. +KVM_LDS := $(OUTPUT)/fp-stress-kvm.lds +KVM_LDFLAGS := -nostdlib -static $(ARCH_INCLUDES) -D__ASSEMBLY__ \ + -Wl,-T,$(KVM_LDS) + +# Supported/required from GNU ld 2.39, not confirmed LLD but after 22.1.2 +COMMA=, +KVM_LDFLAGS += $(call try-run,$(LD) --warn-rwx-segments -v,-Wl$(COMMA)--warn-rwx-segments) + +$(KVM_LDS): fp-stress-kvm.lds.S fp-stress-vmm.h + $(CC) -E -P -x c $< -o $@ + $(OUTPUT)/fp-stress-vmm: CFLAGS += $(ARCH_INCLUDES) # Build with nolibc to avoid effects due to libc's clone() support @@ -61,4 +85,18 @@ $(OUTPUT)/zt-ptrace: zt-ptrace.c $(OUTPUT)/zt-test: zt-test.S $(ASM_UTILS_LINUX) $(CC) -nostdlib $^ -o $@ +$(OUTPUT)/fpsimd-test-kvm.elf: fpsimd-test.S $(ASM_UTILS_KVM) $(KVM_LDS) + $(CC) $(KVM_LDFLAGS) $(filter %.S,$^) -o $@ +$(OUTPUT)/sve-test-kvm.elf: sve-test.S $(ASM_UTILS_KVM) $(KVM_LDS) + $(CC) $(KVM_LDFLAGS) $(filter %.S,$^) -o $@ +$(OUTPUT)/ssve-test-kvm.elf: sve-test.S $(ASM_UTILS_KVM) $(KVM_LDS) + $(CC) -DSSVE $(KVM_LDFLAGS) $(filter %.S,$^) -o $@ +$(OUTPUT)/za-test-kvm.elf: za-test.S $(ASM_UTILS_KVM) $(KVM_LDS) + $(CC) $(KVM_LDFLAGS) $(filter %.S,$^) -o $@ +$(OUTPUT)/zt-test-kvm.elf: zt-test.S $(ASM_UTILS_KVM) $(KVM_LDS) + $(CC) $(KVM_LDFLAGS) $(filter %.S,$^) -o $@ + +$(OUTPUT)/%-kvm.bin: $(OUTPUT)/%-kvm.elf + $(OBJCOPY) -O binary $< $@ + include ../../lib.mk diff --git a/tools/testing/selftests/arm64/fp/asm-utils-kvm.S b/tools/testing/selftests/arm64/fp/asm-utils-kvm.S new file mode 100644 index 000000000000..422d6425ab56 --- /dev/null +++ b/tools/testing/selftests/arm64/fp/asm-utils-kvm.S @@ -0,0 +1,86 @@ +// SPDX-License-Identifier: GPL-2.0-only +// +// Utility functions for assembly code programs run under fp-stress-vmm. +// +// Copyright 2026, Arm Ltd + +#include <asm/sysreg.h> + +#include "assembler.h" +#include "asm-offsets.h" +#include "fp-stress-vmm.h" + +#define PSCI_0_2_FN_SYSTEM_OFF 0x84000008 + +// Print a single character x0 to the console +// Clobbers x1 +function putc + movz x1, #(CONSOLE_BASE >> 16), lsl #16 + strb w0, [x1] + ret +endfunction +.globl putc + +// Print a \0 terminated string starting at address x0 to the console. +// Doesn't use putc as a slight optimisation. +// Clobbers x0-x3 +function puts + mov x2, x0 + movz x3, #(CONSOLE_BASE >> 16), lsl #16 + +0: ldrb w0, [x2], #1 + cbz w0, 1f + strb w0, [x3] + b 0b + +1: ret +endfunction +.globl puts + +// Report errors to the host by shutting down, there is no other reason one +// of these test programs should exit. +function exit_error + movz x0, #(PSCI_0_2_FN_SYSTEM_OFF >> 16), lsl #16 + movk x0, #(PSCI_0_2_FN_SYSTEM_OFF & 0xffff) + hvc #0 + +0: b 0b // PSCI should not have returned +endfunction +.globl exit_error + +// Exception vectors. We have no handlers for anything so each entry +// just BRKs with its own number. + .balign 2048 +vectors: + .irp vector, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 + .balign 0x80 + brk #\vector + .endr + +// Main program entry point. +// +// The VMM starts us at the very start of the image so the link script +// places .text.boot first. + .section .text.boot, "ax" +.globl _start +function _start + // Vectors that report any exception we take to the VMM. + adr x0, vectors + msr VBAR_EL1, x0 + + // The VMM sizes our RAM as the image plus GUEST_EXTRA_RAM, put the + // stack at the top of that so it grows down away from the image. + ldr x0, =_end + ldr x1, =GUEST_EXTRA_RAM + add x0, x0, x1 + and x0, x0, #~15 + mov sp, x0 + + // Log the current EL for diagnostic purposes. + puts "EL:\t\tEL" + mrs x0, CurrentEL + ubfx x0, x0, #2, #2 + bl putdecn + + b main +endfunction diff --git a/tools/testing/selftests/arm64/fp/fp-stress-kvm.lds.S b/tools/testing/selftests/arm64/fp/fp-stress-kvm.lds.S new file mode 100644 index 000000000000..f1d44f0ae49c --- /dev/null +++ b/tools/testing/selftests/arm64/fp/fp-stress-kvm.lds.S @@ -0,0 +1,58 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * Link script for the flat binary builds of the fp-stress loads which run + * as guests under fp-stress-vmm. + * + * The VMM loads the image at the start of the guest RAM and starts the + * guest there so the startup code has to be first in the image, it is in + * .text.boot for that. + */ + +#include "fp-stress-vmm.h" + +OUTPUT_ARCH(aarch64) +ENTRY(_start) + +SECTIONS +{ + . = GUEST_RAM_BASE; + + .text : { + *(.text.boot) + *(.text) + *(.text.*) + } + + .rodata : { + *(.rodata) + *(.rodata.*) + } + + .data : { + *(.data) + *(.data.*) + } + + .bss : { + *(.bss) + *(.bss.*) + *(COMMON) + } + + . = ALIGN(16); + _end = .; + + /DISCARD/ : { + *(.note*) + *(.comment) + *(.eh_frame*) + } +} + +/* + * The VMM sizes the guest RAM from the size of the image file and the + * guest puts its stack immediately above _end, anything allocated but not + * present in the file would leave the two disagreeing about where the RAM + * ends. + */ +ASSERT(SIZEOF(.bss) == 0, "fp-stress guests must not have a .bss") -- 2.47.3
