Skip to contentSkip to conversation

parisc: Fix double SIGFPE crash

PATCH In review [email protected] 2 participants 2 messages May 3, 2025 PA-RISC architecture

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 parisc: Fix double SIGFPE crash

From: Helge Deller <[email protected]>

Camm noticed that on parisc a SIGFPE exception may triggger a second SIGFPE in the glibc part of the signal handler which then finally leads to a crash of the program.
This can easily be reproduced with this test program which will abort in the second SIGFPE.

root@parisc:~# cat fpe.c
static void fpe_func(int sig, siginfo_t *i, void *v) {
        sigset_t set;
        sigemptyset(&set);
        sigaddset(&set, SIGFPE);
        sigprocmask(SIG_UNBLOCK, &set, NULL);
        printf("GOT signal %d with si_code %ld\n", sig, i->si_code);
}
int main() {
        struct sigaction action = {
                .sa_sigaction = fpe_func,
                .sa_flags = SA_RESTART|SA_SIGINFO };
        sigaction(SIGFPE, &action, 0);
        feenableexcept(FE_OVERFLOW);
        return printf("%lf\n",1.7976931348623158E308*1.7976931348623158E308);
}
root@parisc:~# gcc fpe.c -lm
root@parisc:~# ./a.out
Floating point exception
root@parisc:~# strace -f ./a.out
execve("./a.out", ["./a.out"], 0xf9ac7034 /* 20 vars */) = 0
getrlimit(RLIMIT_STACK, {rlim_cur=8192*1024, rlim_max=RLIM_INFINITY}) = 0
...
rt_sigaction(SIGFPE, {sa_handler=0x1110a, sa_mask=[], sa_flags=SA_RESTART|SA_SIGINFO}, NULL, 8) = 0
--- SIGFPE {si_signo=SIGFPE, si_code=FPE_FLTOVF, si_addr=0x1078f} ---
--- SIGFPE {si_signo=SIGFPE, si_code=FPE_FLTOVF, si_addr=0xf8f21237} ---
+++ killed by SIGFPE +++
Floating point exception

Fix this by clearing the Trap (T) bit in the FP status register before returning to the signal handler in userspace.

---
 arch/parisc/math-emu/driver.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
arch/parisc/math-emu/driver.c +13 -3
index 34495446e051..15f183cbe915 100644
@@ -97,9 +97,19 @@ handle_fpe(struct pt_regs *regs)
 
 	memcpy(regs->fr, frcopy, sizeof regs->fr);
 	if (signalcode != 0) {
-	    force_sig_fault(signalcode >> 24, signalcode & 0xffffff,
-			    (void __user *) regs->iaoq[0]);
-	    return -1;
+		int sig = signalcode >> 24;
+
+		if (sig == SIGFPE) {
+			/*
+			 * Clear floating point trap bit to avoid trapping
+			 * again on the first fstd/fld access in the userspace
+			 * signal handler.
+			 */
+			regs->fr[0] &= ~(1ULL << 38);
+		}
+		force_sig_fault(sig, signalcode & 0xffffff,
+				(void __user *) regs->iaoq[0]);
+		return -1;
 	}
 
 	return signalcode ? -1 : 0;
Signature · 1 line
2.47.0
#1 lore raw

Hi Helge,

Maybe the comment can be improved.

The signal handler crashes with a second SIGFPE because glibc uses a double-word floating-point store to atomically update function descriptors. As a result of lazy binding, we hit a floating-point store in fpe_func almost immediately.

When the T bit is set, an assist exception trap occurs when when the co-processor encounters *any* floating-point instruction except for a double store of register %fr0. The latter cancels all pending traps.

Dave

Quoted 78 lines · On 2025-05-03 12:38 p.m., [email protected] wrote:

From: Helge Deller <[email protected]>

Camm noticed that on parisc a SIGFPE exception may triggger a second SIGFPE in the glibc part of the signal handler which then finally leads to a crash of the program.
This can easily be reproduced with this test program which will abort in the second SIGFPE.

root@parisc:~# cat fpe.c
static void fpe_func(int sig, siginfo_t *i, void *v) {
        sigset_t set;
        sigemptyset(&set);
        sigaddset(&set, SIGFPE);
        sigprocmask(SIG_UNBLOCK, &set, NULL);
        printf("GOT signal %d with si_code %ld\n", sig, i->si_code);
}
int main() {
        struct sigaction action = {
                .sa_sigaction = fpe_func,
                .sa_flags = SA_RESTART|SA_SIGINFO };
        sigaction(SIGFPE, &action, 0);
        feenableexcept(FE_OVERFLOW);
        return printf("%lf\n",1.7976931348623158E308*1.7976931348623158E308);
}
root@parisc:~# gcc fpe.c -lm
root@parisc:~# ./a.out
Floating point exception
root@parisc:~# strace -f ./a.out
execve("./a.out", ["./a.out"], 0xf9ac7034 /* 20 vars */) = 0
getrlimit(RLIMIT_STACK, {rlim_cur=8192*1024, rlim_max=RLIM_INFINITY}) = 0
...
rt_sigaction(SIGFPE, {sa_handler=0x1110a, sa_mask=[], sa_flags=SA_RESTART|SA_SIGINFO}, NULL, 8) = 0
--- SIGFPE {si_signo=SIGFPE, si_code=FPE_FLTOVF, si_addr=0x1078f} ---
killed by SIGFPE +++ renamed from SIGFPE {si_signo=SIGFPE, si_code=FPE_FLTOVF, si_addr=0xf8f21237} --- +0 -0

Floating point exception

Fix this by clearing the Trap (T) bit in the FP status register before returning to the signal handler in userspace.

---
 arch/parisc/math-emu/driver.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
arch/parisc/math-emu/driver.c +13 -3
index 34495446e051..15f183cbe915 100644
@@ -97,9 +97,19 @@ handle_fpe(struct pt_regs *regs)
 
 	memcpy(regs->fr, frcopy, sizeof regs->fr);
 	if (signalcode != 0) {
-	    force_sig_fault(signalcode >> 24, signalcode & 0xffffff,
-			    (void __user *) regs->iaoq[0]);
-	    return -1;
+		int sig = signalcode >> 24;
+
+		if (sig == SIGFPE) {
+			/*
+			 * Clear floating point trap bit to avoid trapping
+			 * again on the first fstd/fld access in the userspace
+			 * signal handler.
+			 */
+			regs->fr[0] &= ~(1ULL << 38);
+		}
+		force_sig_fault(sig, signalcode & 0xffffff,
+				(void __user *) regs->iaoq[0]);
+		return -1;
 	}
 
 	return signalcode ? -1 : 0;
Signature · 1 line
John David Anglin  [email protected]