target/i386/tcg: Fix masking of real-mode addresses with A20 bit

The correct A20 masking is done if paging is enabled (protected mode) but it
seems to have been forgotten in real mode. For example from the AMD64 APM Vol. 2
section 1.2.4:

> If the sum of the segment base and effective address carries over into bit 20,
> that bit can be optionally truncated to mimic the 20-bit address wrapping of the
> 8086 processor by using the A20M# input signal to mask the A20 address bit.

Most BIOSes will enable the A20 line on boot, but I found by disabling the A20 line
afterwards, the correct wrapping wasn't taking place.

`handle_mmu_fault' in target/i386/tcg/sysemu/excp_helper.c seems to be the culprit.
In real mode, it fills the TLB with the raw unmasked address. However, for the
protected mode, the `mmu_translate' function does the correct A20 masking.

The fix then should be to just apply the A20 mask in the first branch of the if
statement.

Signed-off-by: Stephen Michael Jothen <sjothen@gmail.com>
Message-Id: <Yo5MUMSz80jXtvt9@air-old.local>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/target/i386/tcg/sysemu/excp_helper.c b/target/i386/tcg/sysemu/excp_helper.c
index e1b6d88683385ff2ac55f822478409b4bfa6c7c9..48feba7e759d5e2f1d61d9cb9f8e5c85eeaacc8b 100644 (file)
--- a/target/i386/tcg/sysemu/excp_helper.c
+++ b/target/i386/tcg/sysemu/excp_helper.c
@@ -359,6 +359,7 @@ static int handle_mmu_fault(CPUState *cs, vaddr addr, int size,
     CPUX86State *env = &cpu->env;
     int error_code = PG_ERROR_OK;
     int pg_mode, prot, page_size;
+    int32_t a20_mask;
     hwaddr paddr;
     hwaddr vaddr;
 
@@ -368,7 +369,8 @@ static int handle_mmu_fault(CPUState *cs, vaddr addr, int size,
 #endif
 
     if (!(env->cr[0] & CR0_PG_MASK)) {
-        paddr = addr;
+        a20_mask = x86_get_a20_mask(env);
+        paddr = addr & a20_mask;
 #ifdef TARGET_X86_64
         if (!(env->hflags & HF_LMA_MASK)) {
             /* Without long mode we can only address 32bits in real mode */