Skip to contentSkip to conversation

linux-user/i386: report the real trap number in sigcontext.trapno

PATCH In review Nat Brown 2 participants 2 messages Aug 24 – Sep 9

lore
Share & tools
lore

Nat Brown

author
#0 lore raw

PATCH 1/1 linux-user/i386: report the real trap number in sigcontext.trapno

setup_sigcontext() fills sigcontext.trapno from CPUState::exception_index, but by the time a signal frame is built that field has already been reset to -1:
cpu_exec() copies the exception number into its return value and clears exception_index before returning to cpu_loop(), which is where the signal is queued. Guests therefore see trapno == -1 for every cpu exception.

  #include <signal.h>
  #include <stdio.h>
  #include <ucontext.h>

  static void h(int sig, siginfo_t *si, void *uc_)
  {
      ucontext_t *uc = uc_;
      printf("sig=%d trapno=%d err=%#x\n", sig,
             (int)uc->uc_mcontext.gregs[REG_TRAPNO],
             (unsigned)uc->uc_mcontext.gregs[REG_ERR]);
      fflush(stdout);
      _exit(0);
  }

  int main(void)
  {
      struct sigaction sa = { .sa_sigaction = h, .sa_flags = SA_SIGINFO };
      sigaction(SIGSEGV, &sa, 0);
      *(volatile int *)0 = 1;
  }

Natively this prints trapno=14 err=0x6; under qemu it prints trapno=-1 err=0x6. A SIGILL likewise reports -1 instead of 6, and SIGFPE -1 instead of 0. Both qemu-i386 and qemu-x86_64 are affected, and both report the expected 14/6/0 with this patch applied. Only trapno is wrong; error_code lives in CPUX86State and survives.

x86_cpu_record_sigsegv() already notes the coupling, observing that we cannot let the caller clobber exception_index "short of inventing a new place to store the trapno". Invent it: record the exception in env->trap_nr, mirroring linux's thread.trap_nr, which is exactly what the kernel reports in sigcontext.trapno and which likewise persists beyond the exception that set it. Only hardware vectors are recorded; the EXCP_* values at 0x100 and above are emulation internals and never appear in a signal frame.

This matters to wine, whose segv_handler() dispatches on the trap number and cannot service a fault it sees as -1, so 32-bit module loading fails under qemu-i386. The equivalent 64-bit handler dispatches the same way, so the same failure is expected under qemu-x86_64. Discussed at
https://gitlab.winehq.org/wine/wine/-/merge_requests/11737 , where the suggestion was that qemu is the right place to fix this.

---
 linux-user/i386/cpu_loop.c         |  9 +++++++++
 linux-user/i386/signal.c           |  6 ++----
 target/i386/cpu.h                  | 10 ++++++++++
 target/i386/tcg/user/excp_helper.c |  8 ++++----
 4 files changed, 25 insertions(+), 8 deletions(-)
linux-user/i386/cpu_loop.c +9 -0
index fe922fceb5..24fdcc9b24 100644
@@ -218,6 +218,15 @@ void cpu_loop(CPUX86State *env)
         cpu_exec_end(cs);
         qemu_process_cpu_events(cs);
 
+        /*
+         * Remember the exception for sigcontext.trapno, as linux does in
+         * thread.trap_nr.  Only hardware exception vectors qualify; the
+         * EXCP_* values at 0x100 and above are emulation internals.
+         */
+        if (trapnr >= EXCP00_DIVZ && trapnr <= EXCP12_MCHK) {
+            env->trap_nr = trapnr;
+        }
+
         switch(trapnr) {
         case 0x80:
 #ifndef TARGET_X86_64
linux-user/i386/signal.c +2 -4
index b646fde431..9650fac940 100644
@@ -367,8 +367,6 @@ static void setup_sigcontext(CPUX86State *env,
                              abi_ptr fxstate_addr,
                              abi_ptr fpend_addr)
 {
-    CPUState *cs = env_cpu(env);
-
 #ifndef TARGET_X86_64
     uint16_t magic;
 
@@ -385,7 +383,7 @@ static void setup_sigcontext(CPUX86State *env,
     __put_user(env->regs[R_EDX], &sc->edx);
     __put_user(env->regs[R_ECX], &sc->ecx);
     __put_user(env->regs[R_EAX], &sc->eax);
-    __put_user(cs->exception_index, &sc->trapno);
+    __put_user(env->trap_nr, &sc->trapno);
     __put_user(env->error_code, &sc->err);
     __put_user(env->eip, &sc->eip);
     __put_user(env->segs[R_CS].selector, (uint32_t *)&sc->cs);
@@ -416,7 +414,7 @@ static void setup_sigcontext(CPUX86State *env,
     __put_user(env->regs[14], &sc->r14);
     __put_user(env->regs[15], &sc->r15);
 
-    __put_user(cs->exception_index, &sc->trapno);
+    __put_user(env->trap_nr, &sc->trapno);
     __put_user(env->error_code, &sc->err);
     __put_user(env->eip, &sc->rip);
 
target/i386/cpu.h +10 -0
index 641f3ee5c2..cf727b2a56 100644
@@ -2188,6 +2188,16 @@ typedef struct CPUArchState {
 
     /* exception/interrupt handling */
     int error_code;
+#ifdef CONFIG_USER_ONLY
+    /*
+     * The number of the last cpu exception taken by this thread, mirroring
+     * linux's thread.trap_nr, which is what the kernel reports in
+     * sigcontext.trapno.  CPUState::exception_index cannot be used for this:
+     * cpu_exec() resets it before returning, long before the signal frame is
+     * built during delivery.
+     */
+    int trap_nr;
+#endif
     int exception_is_int;
     target_ulong exception_next_eip;
     target_ulong dr[8]; /* debug registers; note dr4 and dr5 are unused */
target/i386/tcg/user/excp_helper.c +4 -4
index 0957ad2e9e..26a4c0ef92 100644
@@ -31,10 +31,10 @@ void x86_cpu_record_sigsegv(CPUState *cs, vaddr addr,
 
     /*
      * The error_code that hw reports as part of the exception frame
-     * is copied to linux sigcontext.err.  The exception_index is
-     * copied to linux sigcontext.trapno.  Short of inventing a new
-     * place to store the trapno, we cannot let our caller raise the
-     * signal and set exception_index to EXCP_INTERRUPT.
+     * is copied to linux sigcontext.err.  The trapno reported in
+     * linux sigcontext.trapno is recorded separately in env->trap_nr
+     * by cpu_loop(), since cpu_exec() clears exception_index before
+     * the signal frame is built.
      */
     env->cr[2] = addr;
     env->error_code = (maperr ? 0 : PG_ERROR_P_MASK)
Signature · 1 line
2.50.1 (Apple Git-155)
Reviewed-by on v1 1/1
#1 lore raw
Quoted 48 lines · Nat Brown

setup_sigcontext() fills sigcontext.trapno from CPUState::exception_index, but by the time a signal frame is built that field has already been reset to -1:
cpu_exec() copies the exception number into its return value and clears exception_index before returning to cpu_loop(), which is where the signal is queued. Guests therefore see trapno == -1 for every cpu exception.

   #include <signal.h>
   #include <stdio.h>
   #include <ucontext.h>

   static void h(int sig, siginfo_t *si, void *uc_)
   {
       ucontext_t *uc = uc_;
       printf("sig=%d trapno=%d err=%#x\n", sig,
              (int)uc->uc_mcontext.gregs[REG_TRAPNO],
              (unsigned)uc->uc_mcontext.gregs[REG_ERR]);
       fflush(stdout);
       _exit(0);
   }

   int main(void)
   {
       struct sigaction sa = { .sa_sigaction = h, .sa_flags = SA_SIGINFO };
       sigaction(SIGSEGV, &sa, 0);
       *(volatile int *)0 = 1;
   }

Natively this prints trapno=14 err=0x6; under qemu it prints trapno=-1 err=0x6. A SIGILL likewise reports -1 instead of 6, and SIGFPE -1 instead of 0. Both qemu-i386 and qemu-x86_64 are affected, and both report the expected 14/6/0 with this patch applied. Only trapno is wrong; error_code lives in CPUX86State and survives.

x86_cpu_record_sigsegv() already notes the coupling, observing that we cannot let the caller clobber exception_index "short of inventing a new place to store the trapno". Invent it: record the exception in env->trap_nr, mirroring linux's thread.trap_nr, which is exactly what the kernel reports in sigcontext.trapno and which likewise persists beyond the exception that set it. Only hardware vectors are recorded; the EXCP_* values at 0x100 and above are emulation internals and never appear in a signal frame.

This matters to wine, whose segv_handler() dispatches on the trap number and cannot service a fault it sees as -1, so 32-bit module loading fails under qemu-i386. The equivalent 64-bit handler dispatches the same way, so the same failure is expected under qemu-x86_64. Discussed at
https://gitlab.winehq.org/wine/wine/-/merge_requests/11737 , where the suggestion was that qemu is the right place to fix this.

Quoted 7 lines
---
  linux-user/i386/cpu_loop.c         |  9 +++++++++
  linux-user/i386/signal.c           |  6 ++----
  target/i386/cpu.h                  | 10 ++++++++++
  target/i386/tcg/user/excp_helper.c |  8 ++++----
  4 files changed, 25 insertions(+), 8 deletions(-)