This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PREC/RFA] Add not_replay to make precord support release memory better


Hui Zhu wrote:
On Sun, Jul 26, 2009 at 04:41, Michael Snyder<msnyder@vmware.com> wrote:
Hui Zhu wrote:
On Fri, Jul 24, 2009 at 10:10, Michael Snyder<msnyder@vmware.com> wrote:
1) During the recording "pass", there's no change.
2) During the reverse-replay pass, if the memory is
not readable or writable, we will set this flag to TRUE.
3) Finally, during the forward-replay pass, if the flag
has previously been set to TRUE, we will skip this entry
(and set the flag to FALSE.)

So my question is -- is there any reason to set it to FALSE here?
Is there any way that the memory can ever become readable again?

Seems to me, once it is set TRUE, we might as well just leave it TRUE.
Am I right?
I thought about it too.  I think if we don't need this entry.  We can
delete it directly.
But there is a special status, after release memory, inferior alloc
some memory and its address is same with the memory that released
memory.  Then the memory entry will can be use in this status.  User
can get the right value of memory before this entry.  So I think maybe
we can keep it.

What do you think about it?
Let's say a program does something like this:

buf = mmap (...);
munmap (...);
buf = mmap (...);
munmap (...);
buf = mmap (...);

and so on.  And suppose that, for whatever reason, mmap always
returns the same address.

Then it seems to me (and please correct me if I am wrong), that
it all depends on where we stop the recording.

If we stop the recording after mmap, but before munmap, then
the memory will be readable throughout the ENTIRE recording.

But if we stop the recording after munmap, but before mmap, then
the memory will NOT be readable (again for the entire recording).

So as you replay backward and forward through the recording, the
readability state of the memory location will never change -- it
will remain either readable, or unreadable, depending only on
the mapped-state when the recording ended.

The only way for it to change, I think, would be if we could
resume the process and add some more execution to the end of
the existing recording.


Agree with you. We can do more thing around release memory.


But this patch make prec work OK even if inferior release some memory.
 Of course, the released memory we still can not access.  But other
memory is OK.

This is a low cost idea for release memory.  And we still can add
other thing about  release memory.

Yes, I like the patch, and I want to see it go in, but you haven't really answered my question:

Once we have set this flag to TRUE in a recording entry,
why would we ever need to set it to FALSE?

The patch is simpler and easier to understand if we simply
leave the flag TRUE.  It worries me to see it toggling back
and forth between TRUE and FALSE every time we play through
the recording, if there is no benefit to doing that.  Plus
it means that we keep trying to read the target memory even
though we know it will fail.

I'm attaching a diff to show what I mean, in case it isn't clear.
This diff gives me the same behavior with your munmap test case
as your diff does.



Index: record.c
===================================================================
RCS file: /cvs/src/src/gdb/record.c,v
retrieving revision 1.9
diff -u -p -r1.9 record.c
--- record.c	22 Jul 2009 05:31:26 -0000	1.9
+++ record.c	1 Aug 2009 22:59:55 -0000
@@ -51,6 +51,7 @@ struct record_mem_entry
 {
   CORE_ADDR addr;
   int len;
+  int mem_entry_not_accessible;
   gdb_byte *val;
 };
 
@@ -275,6 +276,7 @@ record_arch_list_add_mem (CORE_ADDR addr
   rec->type = record_mem;
   rec->u.mem.addr = addr;
   rec->u.mem.len = len;
+  rec->u.mem.mem_entry_not_accessible = 0;
 
   if (target_read_memory (addr, rec->u.mem.val, len))
     {
@@ -727,32 +729,52 @@ record_wait (struct target_ops *ops,
 	  else if (record_list->type == record_mem)
 	    {
 	      /* mem */
-	      gdb_byte *mem = alloca (record_list->u.mem.len);
-	      if (record_debug > 1)
-		fprintf_unfiltered (gdb_stdlog,
-				    "Process record: record_mem %s to "
-				    "inferior addr = %s len = %d.\n",
-				    host_address_to_string (record_list),
-				    paddress (gdbarch, record_list->u.mem.addr),
-				    record_list->u.mem.len);
-
-	      if (target_read_memory
-		  (record_list->u.mem.addr, mem, record_list->u.mem.len))
-		error (_("Process record: error reading memory at "
-			 "addr = %s len = %d."),
-		       paddress (gdbarch, record_list->u.mem.addr),
-		       record_list->u.mem.len);
-
-	      if (target_write_memory
-		  (record_list->u.mem.addr, record_list->u.mem.val,
-		   record_list->u.mem.len))
-		error (_
-		       ("Process record: error writing memory at "
-			"addr = %s len = %d."),
-		       paddress (gdbarch, record_list->u.mem.addr),
-		       record_list->u.mem.len);
-
-	      memcpy (record_list->u.mem.val, mem, record_list->u.mem.len);
+	      if (record_list->u.mem.mem_entry_not_accessible == 0)
+		{
+		  gdb_byte *mem = alloca (record_list->u.mem.len);
+		  if (record_debug > 1)
+		    fprintf_unfiltered (gdb_stdlog,
+				        "Process record: record_mem %s to "
+				        "inferior addr = %s len = %d.\n",
+				        host_address_to_string (record_list),
+				        paddress (gdbarch,
+					          record_list->u.mem.addr),
+				        record_list->u.mem.len);
+
+		  if (target_read_memory (record_list->u.mem.addr, mem,
+		                          record_list->u.mem.len))
+	            {
+		      if (execution_direction != EXEC_REVERSE)
+		        error (_("Process record: error reading memory at "
+			         "addr = %s len = %d."),
+		               paddress (gdbarch, record_list->u.mem.addr),
+		               record_list->u.mem.len);
+		      else
+		        record_list->u.mem.mem_entry_not_accessible = 1;
+		    }
+		  else
+		    {
+		      if (target_write_memory (record_list->u.mem.addr,
+			                       record_list->u.mem.val,
+		                               record_list->u.mem.len))
+	                {
+			  if (execution_direction != EXEC_REVERSE)
+			    error (_("Process record: error writing memory at "
+			             "addr = %s len = %d."),
+		                   paddress (gdbarch, record_list->u.mem.addr),
+		                   record_list->u.mem.len);
+			  else
+			    {
+			      record_list->u.mem.mem_entry_not_accessible = 1;
+			    }
+			}
+		      else
+		        {
+			  memcpy (record_list->u.mem.val, mem,
+				  record_list->u.mem.len);
+			}
+		    }
+		}
 	    }
 	  else
 	    {

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]