Re: [PATCH 2/2] Documentation and testcase

[Pedro Alves <palves@redhat.com>]

On 03/19/2015 11:22 PM, Sergio Durigan Junior wrote:

> ---
>  gdb/doc/gdb.texinfo                        |  33 ++++++++
>  gdb/testsuite/gdb.base/coredump-filter.c   |  61 ++++++++++++++
>  gdb/testsuite/gdb.base/coredump-filter.exp | 128 +++++++++++++++++++++++++++++
>  3 files changed, 222 insertions(+)
>  create mode 100644 gdb/testsuite/gdb.base/coredump-filter.c
>  create mode 100644 gdb/testsuite/gdb.base/coredump-filter.exp
> 
> diff --git a/gdb/doc/gdb.texinfo b/gdb/doc/gdb.texinfo
> index 552da31..5382e91 100644
> --- a/gdb/doc/gdb.texinfo
> +++ b/gdb/doc/gdb.texinfo
> @@ -10952,6 +10952,39 @@ specified, the file name defaults to @file{core.@var{pid}}, where
>  
>  Note that this command is implemented only for some systems (as of
>  this writing, @sc{gnu}/Linux, FreeBSD, Solaris, and S390).
> +
> +On @sc{gnu}/Linux, this command can take into account the value of the
> +file @file{/proc/@var{pid}/coredump_filter} when generating the core
> +dump (@pxref{set use-coredump-filter}).
> +
> +@kindex set use-coredump-filter
> +@anchor{set use-coredump-filter}
> +@item set use-coredump-filter on
> +@itemx set use-coredump-filter off
> +Enable or disable the use of the file
> +@file{/proc/@var{pid}/coredump_filter} when generating core dump
> +files.  This file is used by the Linux kernel to decide what types of
> +memory mappings will be dumped or ignored when generating a core dump
> +file.  @var{pid} is the process ID of a currently running process.
> +
> +
> +To make use of this feature, you have to write in the
> +@file{/proc/@var{pid}/coredump_filter} file a value, in hexadecimal,
> +which is a bit mask representing the memory mapping types.  If a bit
> +is set in the bit mask, then the memory mappings of the corresponding
> +types will be dumped; otherwise, they will be ignored.  For more
> +information about the bits that can be set in the
> +@file{/proc/@var{pid}/coredump_filter} file, please refer to the
> +manpage of @code{core(5)}.

Might be good to mention that the settings are inherited by child
processes.  Reading this, I thought "wow, do I really need to
set every time I'm debugging a new pid/process?"

> +    # The variables are 'char', and using it here would be OK because
> +    # GDB actually reads the contents of the memory (i.e., it
> +    # dereferences the pointer).  However, to make it clear that we
> +    # are interested not in the pointer itself, but in the memory it
> +    # points to, we are using '*(unsigned int *)'.
> +    gdb_test "print *(unsigned int *) $addr($var)" "\(\\\$$decimal = <error: \)?Cannot access memory at address $hex\(>\)?" \
> +	"printing $var when core is loaded (should not work)"
> +    gdb_test "print/x *(unsigned int *) $addr($working_var)" " = $working_value.*" \
> +	"print/x *$working_var ( = $working_value)"

This comment still gave me pause.  The variables are
'char *' not 'char':

  char *private_anon, *shared_anon;
  char *dont_dump;

so I guess you're referring to the issue that plain "print" would
assume they are strings and thus deference the pointer, right?

I honestly think that all that just distracts from what
we're doing.  Why not just:

   # Access the memory the addresses point to.
   gdb_test "print *(char *) $addr($var)" "\(\\\$$decimal = <error: \)?Cannot access memory at address $hex\(>\)?" \

I would never ever think to do:

   gdb_test "print (char *) $addr($var)"

to test the contents of what addr points to.  IOW, reading

   # Access the memory the addresses point to.
   gdb_test "print *(char *) $addr($var)" ...

I'd never really wonder why the leftmost '*' is in there.  It's super
obvious.

Maybe even throw in an /x for super clarity:

   gdb_test "print /x *(char *) $addr($var)" ...


> +set all_corefiles { { "non-Private-Anonymous" "0x7e" \
> +			  $non_private_anon_core \
> +			  "private_anon" \
> +			  "shared_anon" "0x22" }
> +    { "non-Shared-Anonymous" "0x7d" \
> +	  $non_shared_anon_core "shared_anon" \
> +	  "private_anon" "0x11" }
> +    { "DoNotDump" "0x33" \
> +	  $dont_dump_core "dont_dump" \
> +	  "shared_anon" "0x22" } }

Does this cover the case of making sure we don't dump file-based
regions?  That's important.

If not (I assume not), we could test that by loading the core
into gdb, but _not_ the program, and then disassembling a function's
address.  It should fail.  Then load the program and disassemble
again.  It should work now.  Or something along those lines.

Thanks,
Pedro Alves