kselftest/arm64: Exit streaming mode after collecting signal context

When we collect a signal context with one of the SME modes enabled we will
have enabled that mode behind the compiler and libc's back so they may
issue some instructions not valid in streaming mode, causing spurious
failures.

For the code prior to issuing the BRK to trigger signal handling we need to
stay in streaming mode if we were already there since that's a part of the
signal context the caller is trying to collect. Unfortunately this code
includes a memset() which is likely to be heavily optimised and is likely
to use FP instructions incompatible with streaming mode. We can avoid this
happening by open coding the memset(), inserting a volatile assembly
statement to avoid the compiler recognising what's being done and doing
something in optimisation. This code is not performance critical so the
inefficiency should not be an issue.

After collecting the context we can simply exit streaming mode, avoiding
these issues. Use a full SMSTOP for safety to prevent any issues appearing
with ZA.

Reported-by: Will Deacon <will@kernel.org>
Signed-off-by: Mark Brown <broonie@kernel.org>
Link: https://lore.kernel.org/r/20230728-arm64-signal-memcpy-fix-v4-1-0c1290db5d46@kernel.org
Signed-off-by: Will Deacon <will@kernel.org>

1 files changed, 24 insertions, 1 deletions

diff --git a/tools/testing/selftests/arm64/signal/test_signals_utils.h b/tools/testing/selftests/arm64/signal/test_signals_utils.h
index 222093f51b67..c7f5627171dd 100644
--- a/tools/testing/selftests/arm64/signal/test_signals_utils.h
+++ b/tools/testing/selftests/arm64/signal/test_signals_utils.h
@@ -60,13 +60,25 @@ static __always_inline bool get_current_context(struct tdescr *td,
 						size_t dest_sz)
 {
 	static volatile bool seen_already;
+	int i;
+	char *uc = (char *)dest_uc;
 
 	assert(td && dest_uc);
 	/* it's a genuine invocation..reinit */
 	seen_already = 0;
 	td->live_uc_valid = 0;
 	td->live_sz = dest_sz;
-	memset(dest_uc, 0x00, td->live_sz);
+
+	/*
+	 * This is a memset() but we don't want the compiler to
+	 * optimise it into either instructions or a library call
+	 * which might be incompatible with streaming mode.
+	 */
+	for (i = 0; i < td->live_sz; i++) {
+		uc[i] = 0;
+		__asm__ ("" : "=r" (uc[i]) : "0" (uc[i]));
+	}
+
 	td->live_uc = dest_uc;
 	/*
 	 * Grab ucontext_t triggering a SIGTRAP.
@@ -104,6 +116,17 @@ static __always_inline bool get_current_context(struct tdescr *td,
 		      : "memory");
 
 	/*
+	 * If we were grabbing a streaming mode context then we may
+	 * have entered streaming mode behind the system's back and
+	 * libc or compiler generated code might decide to do
+	 * something invalid in streaming mode, or potentially even
+	 * the state of ZA.  Issue a SMSTOP to exit both now we have
+	 * grabbed the state.
+	 */
+	if (td->feats_supported & FEAT_SME)
+		asm volatile("msr S0_3_C4_C6_3, xzr");
+
+	/*
 	 * If we get here with seen_already==1 it implies the td->live_uc
 	 * context has been used to get back here....this probably means
 	 * a test has failed to cause a SEGV...anyway live_uc does not