When we reconfigure the SVE vector length we discard the backing storage for the SVE vectors and then reallocate on next SVE use, leaving the SME specific state alone. This means that we do not enable SME traps if they were already disabled. That means that userspace code can enter streaming mode without trapping, putting the task in a state where if we try to save the state of the task we will fault.

Since the ABI does not specify that changing the SVE vector length disturbs SME state, and since SVE code may not be aware of SME code in the process, we shouldn't simply discard any ZA state. Instead immediately reallocate the storage for SVE if SME is active, and disable SME if we change the SVE vector length while there is no SME state active.

Disabling SME traps on SVE vector length changes would make the overall code more complex since we would have a state where we have valid SME state stored but might get a SME trap.

Fixes: 9e4ab6c89109 ("arm64/sme: Implement vector length configuration prctl()s") Reported-by: David Spickett David.Spickett@arm.com Signed-off-by: Mark Brown broonie@kernel.org Cc: stable@vger.kernel.org --- arch/arm64/kernel/fpsimd.c | 32 +++++++++++++++++++++++++------- 1 file changed, 25 insertions(+), 7 deletions(-)

diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c index 7a1aeb95d7c3..a527b95c06e7 100644 --- a/arch/arm64/kernel/fpsimd.c +++ b/arch/arm64/kernel/fpsimd.c @@ -847,6 +847,9 @@ void sve_sync_from_fpsimd_zeropad(struct task_struct *task) int vec_set_vector_length(struct task_struct *task, enum vec_type type, unsigned long vl, unsigned long flags) { + bool free_sme = false; + bool alloc_sve = true; + if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT | PR_SVE_SET_VL_ONEXEC)) return -EINVAL; @@ -897,22 +900,37 @@ int vec_set_vector_length(struct task_struct *task, enum vec_type type, task->thread.fp_type = FP_STATE_FPSIMD; }

- if (system_supports_sme() && type == ARM64_VEC_SME) { - task->thread.svcr &= ~(SVCR_SM_MASK | - SVCR_ZA_MASK); - clear_thread_flag(TIF_SME); + if (system_supports_sme()) { + if (type == ARM64_VEC_SME || + !(task->thread.svcr & (SVCR_SM_MASK | SVCR_ZA_MASK))) { + /* + * We are changing the SME VL or weren't using + * SME anyway, discard the state and force a + * reallocation. + */ + task->thread.svcr &= ~(SVCR_SM_MASK | + SVCR_ZA_MASK); + clear_thread_flag(TIF_SME); + free_sme = true; + } else { + alloc_sve = true; + } }

if (task == current) put_cpu_fpsimd_context();

/* - * Force reallocation of task SVE and SME state to the correct - * size on next use: + * Free the changed states if they are not in use, they will + * be reallocated to the correct size on next use. If we need + * SVE state due to having untouched SME state then reallocate + * it immediately. */ sve_free(task); - if (system_supports_sme() && type == ARM64_VEC_SME) + if (free_sme) sme_free(task); + if (alloc_sve) + sve_alloc(task, true);

task_set_vl(task, type, vl);

We just fixed an issue where changing the SVE VL while SME was active could result in us attempting to save the streaming mode SVE vectors without any backing storage. Add a test case which provokes that issue, ideally we should also verify that the contents of ZA are unaffected by any of what we did.

Note that since we need to keep streaming mode enabled we can't use any syscalls to trigger the issue, we have to sit in a loop in usersapce and hope to be preempted. The chosen numbers trigger with defconfig on all the virtual platforms for me, this won't be 100% on all systems but avoid an overcomplicated test implementation.

Signed-off-by: Mark Brown broonie@kernel.org --- tools/testing/selftests/arm64/fp/vec-syscfg.c | 105 +++++++++++++++++++++++++- 1 file changed, 102 insertions(+), 3 deletions(-)

diff --git a/tools/testing/selftests/arm64/fp/vec-syscfg.c b/tools/testing/selftests/arm64/fp/vec-syscfg.c index 9bcfcdc34ee9..58ea4bde5be7 100644 --- a/tools/testing/selftests/arm64/fp/vec-syscfg.c +++ b/tools/testing/selftests/arm64/fp/vec-syscfg.c @@ -6,6 +6,7 @@ #include <assert.h> #include <errno.h> #include <fcntl.h> +#include <stdbool.h> #include <stddef.h> #include <stdio.h> #include <stdlib.h> @@ -39,9 +40,11 @@ struct vec_data { int max_vl; };

+#define VEC_SVE 0 +#define VEC_SME 1

static struct vec_data vec_data[] = { - { + [VEC_SVE] = { .name = "SVE", .hwcap_type = AT_HWCAP, .hwcap = HWCAP_SVE, @@ -51,7 +54,7 @@ static struct vec_data vec_data[] = { .prctl_set = PR_SVE_SET_VL, .default_vl_file = "/proc/sys/abi/sve_default_vector_length", }, - { + [VEC_SME] = { .name = "SME", .hwcap_type = AT_HWCAP2, .hwcap = HWCAP2_SME, @@ -644,18 +647,107 @@ static const test_type tests[] = { prctl_set_all_vqs, };

+static inline void smstart(void) +{ + asm volatile("msr S0_3_C4_C7_3, xzr"); +} + +static inline void smstart_sm(void) +{ + asm volatile("msr S0_3_C4_C3_3, xzr"); +} + +static inline void smstop(void) +{ + asm volatile("msr S0_3_C4_C6_3, xzr"); +} + + +/* + * Verify we can change the SVE vector length while SME is active and + * continue to use SME afterwards. + */ +static void change_sve_with_za(void) +{ + struct vec_data *sve_data = &vec_data[VEC_SVE]; + bool pass = true; + int ret, i; + + if (sve_data->min_vl == sve_data->max_vl) { + ksft_print_msg("Only one SVE VL supported, can't change\n"); + ksft_test_result_skip("change_sve_while_sme\n"); + return; + } + + /* Ensure we will trigger a change when we set the maximum */ + ret = prctl(sve_data->prctl_set, sve_data->min_vl); + if (ret != sve_data->min_vl) { + ksft_print_msg("Failed to set SVE VL %d: %d\n", + sve_data->min_vl, ret); + pass = false; + } + + /* Enable SM and ZA */ + smstart(); + + /* Trigger another VL change */ + ret = prctl(sve_data->prctl_set, sve_data->max_vl); + if (ret != sve_data->max_vl) { + ksft_print_msg("Failed to set SVE VL %d: %d\n", + sve_data->max_vl, ret); + pass = false; + } + + /* + * Spin for a bit with SM enabled to try to trigger another + * save/restore. We can't use syscalls without exiting + * streaming mode. + */ + for (i = 0; i < 100000000; i++) + smstart_sm(); + + /* + * TODO: Verify that ZA was preserved over the VL change and + * spin. + */ + + /* Clean up after ourselves */ + smstop(); + ret = prctl(sve_data->prctl_set, sve_data->default_vl); + if (ret != sve_data->default_vl) { + ksft_print_msg("Failed to restore SVE VL %d: %d\n", + sve_data->default_vl, ret); + pass = false; + } + + ksft_test_result(pass, "change_sve_with_za\n"); +} + +typedef void (*test_all_type)(void); + +static const struct { + const char *name; + test_all_type test; +} all_types_tests[] = { + { "change_sve_with_za", change_sve_with_za }, +}; + int main(void) { + bool all_supported = true; int i, j;

ksft_print_header(); - ksft_set_plan(ARRAY_SIZE(tests) * ARRAY_SIZE(vec_data)); + ksft_set_plan(ARRAY_SIZE(tests) * ARRAY_SIZE(vec_data) + + ARRAY_SIZE(all_types_tests));

for (i = 0; i < ARRAY_SIZE(vec_data); i++) { struct vec_data *data = &vec_data[i]; unsigned long supported;

supported = getauxval(data->hwcap_type) & data->hwcap; + if (!supported) + all_supported = false;

for (j = 0; j < ARRAY_SIZE(tests); j++) { if (supported) @@ -666,5 +758,12 @@ int main(void) } }

+ for (i = 0; i < ARRAY_SIZE(all_types_tests); i++) { + if (all_supported) + all_types_tests[i].test(); + else + ksft_test_result_skip("%s\n", all_types_tests[i].name); + } + ksft_exit_pass(); }