Re: [Qemu-devel] [PATCH v3 1/2] scsi: lsi: exit infinite loop while executing script (CVE-2019-12068)

[Paolo Bonzini]

On 13/08/19 12:31, Philippe Mathieu-Daudé wrote:
>> |  
>> |      s->istat1 |= LSI_ISTAT1_SRUN;
>> |  again:
>> | -    insn_processed++;
>> | +    if (++insn_processed > LSI_MAX_INSN) {
>> | +        trace_lsi_execute_script_tc_illegal();
>> | +        lsi_script_dma_interrupt(s, LSI_DSTAT_IID);
>> | +        lsi_disconnect(s);
>> | +        trace_lsi_execute_script_stop();
>> | +        return;
> My understanding of Marcelo's explanation
> (https://lists.gnu.org/archive/html/qemu-devel/2019-08/msg01427.html) is
> we can kill insn_processed.
> 

All zeros is not an illegal instruction, it's just a block move with 
zero transfer count.  It's not clear to me from the spec that the 
behavior of QEMU, skipping the second word, is correct, but I do not 
really dare changing it.

After the first instruction is processed, "again" is only reached if 
s->waiting == LSI_NOWAIT.  Therefore, we could move the Windows hack to 
the beginning and remove the s->waiting condition.  The only change 
would be that it would also be triggered by all zero instructions, like this:

diff --git a/hw/scsi/lsi53c895a.c b/hw/scsi/lsi53c895a.c
index 10468c1..9d714af 100644
--- a/hw/scsi/lsi53c895a.c
+++ b/hw/scsi/lsi53c895a.c
@@ -185,6 +185,9 @@ static const char *names[] = {
 /* Flag set if this is a tagged command.  */
 #define LSI_TAG_VALID     (1 << 16)
 
+/* Maximum instructions to process. */
+#define LSI_MAX_INSN    10000
+
 typedef struct lsi_request {
     SCSIRequest *req;
     uint32_t tag;
@@ -1132,7 +1135,19 @@ static void lsi_execute_script(LSIState *s)
 
     s->istat1 |= LSI_ISTAT1_SRUN;
 again:
-    insn_processed++;
+    if (++insn_processed > LSI_MAX_INSN) {
+        /* Some windows drivers make the device spin waiting for a memory
+           location to change.  If we have been executed a lot of code then
+           assume this is the case and force an unexpected device disconnect.
+           This is apparently sufficient to beat the drivers into submission.
+         */
+        if (!(s->sien0 & LSI_SIST0_UDC)) {
+            qemu_log_mask(LOG_GUEST_ERROR,
+                          "lsi_scsi: inf. loop with UDC masked");
+        }
+        lsi_script_scsi_interrupt(s, LSI_SIST0_UDC, 0);
+        lsi_disconnect(s);
+    }
     insn = read_dword(s, s->dsp);
     if (!insn) {
         /* If we receive an empty opcode increment the DSP by 4 bytes
@@ -1569,19 +1584,7 @@ again:
             }
         }
     }
-    if (insn_processed > 10000 && s->waiting == LSI_NOWAIT) {
-        /* Some windows drivers make the device spin waiting for a memory
-           location to change.  If we have been executed a lot of code then
-           assume this is the case and force an unexpected device disconnect.
-           This is apparently sufficient to beat the drivers into submission.
-         */
-        if (!(s->sien0 & LSI_SIST0_UDC)) {
-            qemu_log_mask(LOG_GUEST_ERROR,
-                          "lsi_scsi: inf. loop with UDC masked");
-        }
-        lsi_script_scsi_interrupt(s, LSI_SIST0_UDC, 0);
-        lsi_disconnect(s);
-    } else if (s->istat1 & LSI_ISTAT1_SRUN && s->waiting == LSI_NOWAIT) {
+    if (s->istat1 & LSI_ISTAT1_SRUN && s->waiting == LSI_NOWAIT) {
         if (s->dcntl & LSI_DCNTL_SSM) {
             lsi_script_dma_interrupt(s, LSI_DSTAT_SSI);
         } else {
@@ -1969,6 +1972,10 @@ static void lsi_reg_writeb(LSIState *s, int offset, 
uint8_t val)
     case 0x2f: /* DSP[24:31] */
         s->dsp &= 0x00ffffff;
         s->dsp |= val << 24;
+        /*
+         * FIXME: if s->waiting != LSI_NOWAIT, this will only execute one
+         * instruction.  Is this correct?
+         */
         if ((s->dmode & LSI_DMODE_MAN) == 0
             && (s->istat1 & LSI_ISTAT1_SRUN) == 0)
             lsi_execute_script(s);
@@ -1987,6 +1994,10 @@ static void lsi_reg_writeb(LSIState *s, int offset, 
uint8_t val)
         break;
     case 0x3b: /* DCNTL */
         s->dcntl = val & ~(LSI_DCNTL_PFF | LSI_DCNTL_STD);
+        /*
+         * FIXME: if s->waiting != LSI_NOWAIT, this will only execute one
+         * instruction.  Is this correct?
+         */
         if ((val & LSI_DCNTL_STD) && (s->istat1 & LSI_ISTAT1_SRUN) == 0)
             lsi_execute_script(s);
         break;

Does it make sense?  Do you have a reproducer, and does the above
patch work?  Also, can the reproducer be modified into a qtest test
case?

Thanks,

Paolo