[PATCH 3/3] riscv: Add ISA double trap extension testing

Clément Léger cleger at rivosinc.com
Tue Jun 3 02:39:37 PDT 2025



On 03/06/2025 11:09, Andrew Jones wrote:
> On Fri, May 23, 2025 at 09:53:10AM +0200, Clément Léger wrote:
>> This test allows to test the double trap implementation of hardware as
>> well as the SBI FWFT and SSE support for double trap. The tests will try
>> to trigger double trap using various sequences and will test to receive
>> the SSE double trap event if supported.
>>
>> It is provided as a separate test from the SBI one for two reasons:
>> - It isn't specifically testing SBI "per se".
>> - It ends up by trying to crash into in M-mode.
>>
>> Currently, the test uses a page fault to raise a trap programatically.
>> Some concern was raised by a github user on the original branch [1]
>> saying that the spec doesn't mandate any trap to be delegatable and that
>> we would need a way to detect which ones are delegatable. I think we can
>> safely assume that PAGE FAULT is delagatable and if a hardware that does
>> not have support comes up then it will probably be the vendor
>> responsibility to provide a way to do so.
>>
>> Link: https://github.com/clementleger/kvm-unit-tests/issues/1 [1]
>> Signed-off-by: Clément Léger <cleger at rivosinc.com>
>> ---
>>  riscv/Makefile      |   1 +
>>  lib/riscv/asm/csr.h |   1 +
>>  riscv/isa-dbltrp.c  | 189 ++++++++++++++++++++++++++++++++++++++++++++
>>  riscv/unittests.cfg |   5 ++
>>  4 files changed, 196 insertions(+)
>>  create mode 100644 riscv/isa-dbltrp.c
>>
>> diff --git a/riscv/Makefile b/riscv/Makefile
>> index 11e68eae..d71c9d2e 100644
>> --- a/riscv/Makefile
>> +++ b/riscv/Makefile
>> @@ -14,6 +14,7 @@ tests =
>>  tests += $(TEST_DIR)/sbi.$(exe)
>>  tests += $(TEST_DIR)/selftest.$(exe)
>>  tests += $(TEST_DIR)/sieve.$(exe)
>> +tests += $(TEST_DIR)/isa-dbltrp.$(exe)
>>  
>>  all: $(tests)
>>  
>> diff --git a/lib/riscv/asm/csr.h b/lib/riscv/asm/csr.h
>> index 3e4b5fca..6a8e0578 100644
>> --- a/lib/riscv/asm/csr.h
>> +++ b/lib/riscv/asm/csr.h
>> @@ -18,6 +18,7 @@
>>  
>>  #define SR_SIE			_AC(0x00000002, UL)
>>  #define SR_SPP			_AC(0x00000100, UL)
>> +#define SR_SDT			_AC(0x01000000, UL) /* Supervisor Double Trap */
>>  
>>  /* Exception cause high bit - is an interrupt if set */
>>  #define CAUSE_IRQ_FLAG		(_AC(1, UL) << (__riscv_xlen - 1))
>> diff --git a/riscv/isa-dbltrp.c b/riscv/isa-dbltrp.c
>> new file mode 100644
>> index 00000000..174aee2a
>> --- /dev/null
>> +++ b/riscv/isa-dbltrp.c
>> @@ -0,0 +1,189 @@
>> +// SPDX-License-Identifier: GPL-2.0-only
>> +/*
>> + * SBI verification
>> + *
>> + * Copyright (C) 2025, Rivos Inc., Clément Léger <cleger at rivosinc.com>
>> + */
>> +#include <alloc.h>
>> +#include <alloc_page.h>
>> +#include <libcflat.h>
>> +#include <stdlib.h>
>> +
>> +#include <asm/csr.h>
>> +#include <asm/page.h>
>> +#include <asm/processor.h>
>> +#include <asm/ptrace.h>
>> +#include <asm/sbi.h>
>> +
>> +#include <sbi-tests.h>
>> +
>> +static bool double_trap;
>> +static bool set_sdt = true;
>> +
>> +#define GEN_TRAP()								\
>> +do {										\
>> +	void *ptr = NULL;							\
>> +	unsigned long value = 0;						\
>> +	asm volatile(								\
>> +	"	.option push\n"							\
>> +	"	.option arch,-c\n"						\
>> +	"	sw %0, 0(%1)\n"							\
>> +	"	.option pop\n"							\
>> +	: : "r"(value), "r"(ptr) : "memory");					\
> 
> nit: need some spaces
> 
>  "r" (value), "r" (ptr)
> 
>> +} while (0)
>> +
>> +static void syscall_trap_handler(struct pt_regs *regs)
> 
> This is a page fault handler, not a syscall.
> >> +{
>> +	if (set_sdt)
>> +		csr_set(CSR_SSTATUS, SR_SDT);
>> +
>> +	if (double_trap) {
>> +		double_trap = false;
>> +		GEN_TRAP();
>> +	}
>> +
>> +	/* Skip trapping instruction */
>> +	regs->epc += 4;
>> +}
>> +
>> +static bool sse_dbltrp_called;
>> +
>> +static void sse_dbltrp_handler(void *data, struct pt_regs *regs, unsigned int hartid)
>> +{
>> +	struct sbiret ret;
>> +	unsigned long flags;
>> +	unsigned long expected_flags = SBI_SSE_ATTR_INTERRUPTED_FLAGS_SSTATUS_SPP |
>> +				       SBI_SSE_ATTR_INTERRUPTED_FLAGS_SSTATUS_SDT;
>> +
>> +	ret = sbi_sse_read_attrs(SBI_SSE_EVENT_LOCAL_DOUBLE_TRAP, SBI_SSE_ATTR_INTERRUPTED_FLAGS, 1,
>> +				 &flags);
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "Get double trap event flags");
>> +	report(flags == expected_flags, "SSE flags == 0x%lx", expected_flags);
>> +
>> +	sse_dbltrp_called = true;
>> +
>> +	/* Skip trapping instruction */
>> +	regs->epc += 4;
>> +}
>> +
>> +static void sse_double_trap(void)
>> +{
>> +	struct sbiret ret;
>> +
>> +	struct sbi_sse_handler_arg handler_arg = {
>> +		.handler = sse_dbltrp_handler,
>> +		.stack = alloc_page() + PAGE_SIZE,
>> +	};
>> +
>> +	report_prefix_push("sse");
>> +
>> +	ret = sbi_sse_hart_unmask();
>> +	if (!sbiret_report_error(&ret, SBI_SUCCESS, "SSE hart unmask ok"))
>> +		goto out;
> 
> The unmasking failed, but the out label takes us to a mask.

Hi Drew,

Yeah masking it even if unmask wasn't so important since it's the boot
state. I'll add a separate label though.

> 
>> +
>> +	ret = sbi_sse_register(SBI_SSE_EVENT_LOCAL_DOUBLE_TRAP, &handler_arg);
>> +	if (ret.error == SBI_ERR_NOT_SUPPORTED) {
>> +		report_skip("SSE double trap event is not supported");
>> +		goto out;
>> +	}
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "SSE double trap register");
>> +
>> +	ret = sbi_sse_enable(SBI_SSE_EVENT_LOCAL_DOUBLE_TRAP);
>> +	if (!sbiret_report_error(&ret, SBI_SUCCESS, "SSE double trap enable"))
>> +		goto out_unregister;
>> +
>> +	/*
>> +	 * Generate a double crash so that an SSE event should be generated. The SPEC (ISA nor SBI)
>> +	 * does not explicitly tell that if supported it should generate an SSE event but that's
>> +	 * a reasonable assumption to do so if both FWFT and SSE are supported.
> 
> This is something to raise in a tech-prs call, because it sounds like we
> need another paragraph for FWFT which states when DOUBLE_TRAP is enabled
> and SSE is supported that SSE local double trap events will be raised. Or,
> we need another FWFT feature that allows S-mode to request that behavior
> be enabled/disabled when SSE is supported (and M-mode can decide yes/no
> to that request).

That's a good point, I'll submit a patch to modify the SBI doc in that way.


> 
>> +	 */
>> +	set_sdt = true;
>> +	double_trap = true;
> 
> WRITE_ONCE(set_sdt, true);
> WRITE_ONCE(double_trap, true);
> 
>> +	GEN_TRAP();
>> +
>> +	report(sse_dbltrp_called, "SSE double trap event generated");
> 
> READ_ONCE(sse_dbltrp_called)
> 
>> +
>> +	ret = sbi_sse_disable(SBI_SSE_EVENT_LOCAL_DOUBLE_TRAP);
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "SSE double trap disable");
>> +out_unregister:
>> +	ret = sbi_sse_unregister(SBI_SSE_EVENT_LOCAL_DOUBLE_TRAP);
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "SSE double trap unregister");
>> +
>> +out:
>> +	sbi_sse_hart_mask();
>> +	free_page(handler_arg.stack - PAGE_SIZE);
>> +
>> +	report_prefix_pop();
>> +}
>> +
>> +static void check_double_trap(void)
>> +{
>> +	struct sbiret ret;
>> +
>> +	/* Disable double trap */
>> +	ret = sbi_fwft_set(SBI_FWFT_DOUBLE_TRAP, 0, 0);
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "Set double trap enable feature value == 0");
>> +	ret = sbi_fwft_get(SBI_FWFT_DOUBLE_TRAP);
>> +	sbiret_report(&ret, SBI_SUCCESS, 0, "Get double trap enable feature value == 0");
>> +
>> +	install_exception_handler(EXC_STORE_PAGE_FAULT, syscall_trap_handler);
>> +
>> +	double_trap = true;
> 
> WRITE_ONCE(double_trap, true);
> 
>> +	GEN_TRAP();
>> +	report_pass("Double trap disabled, trap first time ok");
>> +
>> +	/* Enable double trap */
>> +	ret = sbi_fwft_set(SBI_FWFT_DOUBLE_TRAP, 1, 1);
> 
> Why lock it?

That's a mistake.

> 
>> +	sbiret_report_error(&ret, SBI_SUCCESS, "Set double trap enable feature value == 1");
>> +	ret = sbi_fwft_get(SBI_FWFT_DOUBLE_TRAP);
>> +	if (!sbiret_report(&ret, SBI_SUCCESS, 1, "Get double trap enable feature value == 1"))
>> +		return;
>> +
>> +	/* First time, clear the double trap flag (SDT) so that it doesn't generate a double trap */
>> +	set_sdt = false;
>> +	double_trap = true;
> 
> WRITE_ONCE(set_sdt, true);
> WRITE_ONCE(double_trap, true);
> 
>> +	GEN_TRAP();
>> +	report_pass("Trapped twice allowed ok");
>> +
>> +	if (sbi_probe(SBI_EXT_SSE))
>> +		sse_double_trap();
>> +	else
>> +		report_skip("SSE double trap event will not be tested, extension is not available");
>> +
>> +	/*
>> +	 * Second time, keep the double trap flag (SDT) and generate another trap, this should
> 
> Third time if we did the SSE test.
> 
>> +	 * generate a double trap. Since there is no SSE handler registered, it should crash to
>> +	 * M-mode.
> 
> No SSE handler that we know of... sse_double_trap() should return
> some error if it fails to unregister and then we should skip this
> test in that case.

Indeed, good catch.

> 
>> +	 */
>> +	set_sdt = true;
>> +	double_trap = true;
> 
> WRITE_ONCE(set_sdt, true);
> WRITE_ONCE(double_trap, true);
> 
>> +	report_info("Should generate a double trap and crash !");
>> +	GEN_TRAP();
>> +	report_fail("Should have crashed !");
> 
> nit: Put the '!' next to the last word.
> 
> I think this M-mode crash test should be optional. We can make it optional
> on an environment variable since we already use environment variables for
> other optional tests. We even have env_or_skip() in riscv/sbi-tests.h for
> that purpose.

Acked, I'll use env_or_skip().

Thanks,

Clément

> 
>> +}
>> +
>> +int main(int argc, char **argv)
>> +{
>> +	struct sbiret ret;
>> +
>> +	report_prefix_push("dbltrp");
>> +
>> +	if (!sbi_probe(SBI_EXT_FWFT)) {
>> +		report_skip("FWFT extension is not available, can not enable double traps");
>> +		goto out;
>> +	}
>> +
>> +	ret = sbi_fwft_get(SBI_FWFT_DOUBLE_TRAP);
>> +	if (ret.value == SBI_ERR_NOT_SUPPORTED) {
>> +		report_skip("SBI_FWFT_DOUBLE_TRAP is not supported !");
> 
> nit: Put the '!' next to the last word.
> 
>> +		goto out;
>> +	}
>> +
>> +	if (sbiret_report_error(&ret, SBI_SUCCESS, "SBI_FWFT_DOUBLE_TRAP get value"))
>> +		check_double_trap();
>> +
>> +out:
>> +	report_prefix_pop();
>> +
>> +	return report_summary();
>> +}
>> diff --git a/riscv/unittests.cfg b/riscv/unittests.cfg
>> index 2eb760ec..757e6027 100644
>> --- a/riscv/unittests.cfg
>> +++ b/riscv/unittests.cfg
>> @@ -18,3 +18,8 @@ groups = selftest
>>  file = sbi.flat
>>  smp = $MAX_SMP
>>  groups = sbi
>> +
>> +[dbltrp]
>> +file = isa-dbltrp.flat
>> +smp = $MAX_SMP
> 
> The test doesn't appear to require multiple harts.
> 
>> +groups = isa
> 
> groups = isa sbi
> 
>> -- 
>> 2.49.0
>>
> 
> Thanks,
> drew




More information about the kvm-riscv mailing list