[PATCH v5 16/24] tests/avocado: reverse_debugging.py verify addresses between record and replay

[Nicholas Piggin]

gdb is only attached in the replay phase, and the recorded trace is
single-stepped to build up an address map that is used to then verify
reverse-step and reverse-continue.

This change attaches gdb in the record phase, and builds the address
map by single-stepping the live machine. That address map is then
verified by single-stepping the replay, and then used for the same
reverse debug verification.

This is a stronger verifcation between the record and the replay phases,
as well as providing some testing of gdb operation in record mode.

x86_64 has problems with taking the first step, it does not advance the
icount, so a workaround is added for it.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 tests/avocado/reverse_debugging.py | 68 +++++++++++++++++++++++++-----
 1 file changed, 58 insertions(+), 10 deletions(-)

diff --git a/tests/avocado/reverse_debugging.py 
b/tests/avocado/reverse_debugging.py
index e76cf54ad7..8a9578ad5a 100644
--- a/tests/avocado/reverse_debugging.py
+++ b/tests/avocado/reverse_debugging.py
@@ -35,6 +35,10 @@ class ReverseDebugging(LinuxKernelTest):
     STEPS = 10
     endian_is_le = True
 
+    # If first_step_workaround is true, check whether the first step moved
+    # icount, and if not then step again.
+    first_step_workaround = False
+
     def run_vm(self, record, shift, args, replay_path, image_path, port):
         logger = logging.getLogger('replay')
         vm = self.get_vm()
@@ -45,7 +49,7 @@ def run_vm(self, record, shift, args, replay_path, 
image_path, port):
         else:
             logger.info('replaying the execution...')
             mode = 'replay'
-            vm.add_args('-gdb', 'tcp::%d' % port, '-S')
+        vm.add_args('-gdb', 'tcp::%d' % port, '-S')
         vm.add_args('-icount', 'shift=%s,rr=%s,rrfile=%s,rrsnapshot=init' %
                     (shift, mode, replay_path),
                     '-net', 'none')
@@ -87,6 +91,17 @@ def check_pc(self, g, addr):
         if pc != addr:
             self.fail('Invalid PC (read %x instead of %x)' % (pc, addr))
 
+    @staticmethod
+    def gdb_cont(g):
+        g.cmd(b'c')
+
+    @staticmethod
+    def gdb_cont_nowait(g):
+        # The avocado GDBRemote does not have a good way to disconnect or
+        # continue without waiting for a reply, so open-code our own here.
+        data = g.encode(b'c')
+        g._socket.send(data)
+
     @staticmethod
     def gdb_step(g):
         g.cmd(b's', b'T05thread:01;')
@@ -117,11 +132,40 @@ def reverse_debugging(self, shift=7, args=None):
         replay_path = os.path.join(self.workdir, 'replay.bin')
         port = find_free_port()
 
+        steps = []
+
         # record the log
         vm = self.run_vm(True, shift, args, replay_path, image_path, port)
+        logger.info('connecting to gdbstub')
+        g = gdb.GDBRemote('127.0.0.1', port, False, False)
+        g.connect()
+        r = g.cmd(b'qSupported')
+        if b'qXfer:features:read+' in r:
+            g.cmd(b'qXfer:features:read:target.xml:0,ffb')
+
+        if self.vm_get_icount(vm) != 0:
+            self.fail('icount does not start at zero')
+
+        # save the addresses of the first STEPS instructions executed
+        logger.info('stepping forward')
+        for i in range(self.STEPS):
+            pc = self.get_pc(g)
+            logger.info('saving position %x' % pc)
+            steps.append(pc)
+            self.gdb_step(g)
+            if self.first_step_workaround and i == 0 and 
self.vm_get_icount(vm) == 0:
+                logger.warn('failed to take first step, stepping again')
+                self.gdb_step(g)
+        if self.vm_get_icount(vm) != self.STEPS:
+            self.fail('icount (%d) does not match number of instructions 
stepped' % self.vm_get_icount(vm))
+
+        logger.info('continue running')
+        self.gdb_cont_nowait(g)
+
         while self.vm_get_icount(vm) <= self.STEPS:
             pass
         last_icount = self.vm_get_icount(vm)
+        logger.info('shutdown...')
         vm.shutdown()
 
         logger.info("recorded log with %s+ steps" % last_icount)
@@ -139,23 +183,23 @@ def reverse_debugging(self, shift=7, args=None):
         if b'ReverseContinue+' not in r:
             self.fail('Reverse continue is not supported by QEMU')
 
+        # Try single stepping
         logger.info('stepping forward')
-        steps = []
-        # record first instruction addresses
-        for _ in range(self.STEPS):
-            pc = self.get_pc(g)
-            logger.info('saving position %x' % pc)
-            steps.append(pc)
+        for addr in steps:
+            # verify addresses match what initial execution saw
+            self.check_pc(g, addr)
             self.gdb_step(g)
+            logger.info('found position %x' % addr)
 
-        # visit the recorded instruction in reverse order
+        # Try reverse stepping
         logger.info('stepping backward')
         for addr in steps[::-1]:
             self.gdb_bstep(g)
+            # verify addresses match what initial execution saw
             self.check_pc(g, addr)
             logger.info('found position %x' % addr)
 
-        # visit the recorded instruction in forward order
+        # Step forward again
         logger.info('stepping forward')
         for addr in steps:
             self.check_pc(g, addr)
@@ -175,7 +219,7 @@ def reverse_debugging(self, shift=7, args=None):
         # continue - will return after pausing
         # This could stop at the end and get a T02 return, or by
         # re-executing one of the breakpoints and get a T05 return.
-        g.cmd(b'c')
+        self.gdb_cont(g)
         if self.vm_get_icount(vm) == last_icount - 1:
             logger.info('reached the end (icount %s)' % (last_icount - 1))
         else:
@@ -201,6 +245,10 @@ class ReverseDebugging_X86_64(ReverseDebugging):
 
     REG_PC = 0x10
     REG_CS = 0x12
+
+    # The initial step does not change pc on x86 for some reason.
+    first_step_workaround = True
+
     def get_pc(self, g):
         return self.get_reg_le(g, self.REG_PC) \
             + self.get_reg_le(g, self.REG_CS) * 0x10
-- 
2.42.0