Re: [PATCH] Delegate to target_ops->beneath to read cache lines

[Pedro Alves <palves at redhat dot com>]

On 11/27/2013 12:20 PM, Yao Qi wrote:
> GDB on x86_64-linux is unable to disassemble on core-file target.
> 
> $ ./gdb ./testsuite/gdb.base/corefile
> (gdb) core-file ./testsuite/gdb.base/corefile.core
> (gdb) disassemble main
> Dump of assembler code for function main:
>    0x0000000000400976 <+0>:	Cannot access memory at address 0x400976
> 
> However, it works if we turn code-cache off.
> 
> (gdb) set code-cache off
> (gdb) disassemble main,+4
> Dump of assembler code from 0x400976 to 0x40097a:
>    0x0000000000400976 <main+0>:	push   %rbp
>    0x0000000000400977 <main+1>:	mov    %rsp,%rbp
> End of assembler dump.
> 
> When code-cache is off, GDB will iterate target_ops from top and call
> to_xfer_partial.  When current_target is "core", it will call
> to_xfer_partial of target "exec", which reads the contents for
> disassemble.  However, dcache doesn't have such mechanism, and that is
> the cause for the error.

This points out that we end up caching core and exec code, which
isn't really necessary.  We could limit it to has_all_memory targets,
I think.  Not sure if that'd complicate things.

> This patch adds something similar in dcache_read_line to go through
> target_ops from top to bottom, and call to_xfer_partial.
> The original code uses TARGET_OBJECT_RAW_MEMORY, which is replaced
> by TARGET_OBJECT_MEMORY in target_xfer_partial,
> 
>       enum target_object raw_object = object;
> 
>       /* If this is a raw memory transfer, request the normal
> 	 memory object from other layers.  */
>       if (raw_object == TARGET_OBJECT_RAW_MEMORY)
> 	raw_object = TARGET_OBJECT_MEMORY;
> 
> so we can use TARGET_OBJECT_MEMORY here.  Regression tested on
> x86_64-linux.

The dcache is the sole user of TARGET_OBJECT_RAW_MEMORY.  This shows
that if we want to code from lower targets too, then this sole user also
wants to do the top to bottom delegation that memory_xfer_partial_1 does.

So this can all be done within target.c.  Factor out the
memory_xfer_partial_1 top to bottom memory read code to a separate
raw_memory_xfer_partial function (despite the name, it'd request
TARGET_OBJECT_MEMORY from the targets), and make target_xfer_partial
call that for TARGET_OBJECT_RAW_MEMORY:

LONGEST
target_xfer_partial (struct target_ops *ops,
		     enum target_object object, const char *annex,
		     void *readbuf, const void *writebuf,
		     ULONGEST offset, LONGEST len)
{
  LONGEST retval;

  /* If this is a memory transfer, let the memory-specific code
     have a look at it instead.  Memory transfers are more
     complicated.  */
  if (object == TARGET_OBJECT_MEMORY || object == TARGET_OBJECT_STACK_MEMORY
      || object == TARGET_OBJECT_CODE_MEMORY)
    retval = memory_xfer_partial (ops, object, readbuf,
				  writebuf, offset, len);
  else if (object == TARGET_OBJECT_RAW_MEMORY)
    {
      /* Request the normal memory object from other layers.  */
      retval = raw_memory_xfer_partial (ops, readbuf, writebuf, offset, len);
    }
  else
    {
      retval = ops->to_xfer_partial (ops, raw_object, annex, readbuf,
				     writebuf, offset, len);
    }

Then dcache.c doesn't need to change, and doesn't need to inline that
top to bottom dance.

-- 
Pedro Alves