Re: [PATCH v2] Search for DWZ files in debug-file-directories as well

[Sergio Durigan Junior via Gdb-patches <gdb-patches@sourceware.org>]

On Thursday, November 26 2020, Simon Marchi wrote:

> On 2020-11-18 9:27 p.m., Sergio Durigan Junior via Gdb-patches wrote:
>> Changes from v1:
>>
>> - Addressed Simon's comments (new comment explaining how we try to
>>   match for the current debug-file-directory; properly use
>>   .erase/.insert methods -- keeping in mind that they modify the
>>   string in-place).
>>
>>
>> When Debian (and Ubuntu) builds its binaries, it (still) doesn't use
>> dwz's "--relative" option.  This causes their debuginfo files to
>> carry a .gnu_debugaltlink section containing a full pathname to the
>> DWZ alt debug file, like this:
>>
>>   $ readelf -wk /usr/bin/cat
>>   Contents of the .gnu_debugaltlink section:
>>
>>     Separate debug info file: /usr/lib/debug/.dwz/x86_64-linux-gnu/coreutils.debug
>>     Build-ID (0x14 bytes):
>>    ee 76 5d 71 97 37 ce 46 99 44 32 bb e8 a9 1a ef 99 96 88 db
>>
>>   Contents of the .gnu_debuglink section:
>>
>>     Separate debug info file: 06d3bee37b8c7e67b31cb2689cb351102ae73b.debug
>>     CRC value: 0x53267655
>>
>> This usually works OK, because most of the debuginfo files installed
>> via apt will be present in /usr/lib/debug anyway.  However, imagine
>> the following scenario:
>>
>> - You are using /usr/bin/cat, it crashes on you and generates a
>>   corefile.
>>
>> - You don't want/need to "apt install" the debuginfo file for
>>   coreutils from the repositories.  Instead, you already have the
>>   debuginfo files in a separate directory (e.g., $HOME/dbgsym).
>>
>> - You start GDB and "set debug-file-directory $HOME/dbgsym".
>
> Should the above read $HOME/dbgsym/usr/lib/debug to be consistent with
> the example below?

Sure.

>> diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
>> index 3c59826291..c462b9bb2c 100644
>> --- a/gdb/dwarf2/read.c
>> +++ b/gdb/dwarf2/read.c
>> @@ -2190,7 +2190,7 @@ locate_dwz_sections (bfd *abfd, asection *sectp, dwz_file *dwz_file)
>>  struct dwz_file *
>>  dwarf2_get_dwz_file (dwarf2_per_bfd *per_bfd)
>>  {
>> -  const char *filename;
>> +  std::string filename;
>>    bfd_size_type buildid_len_arg;
>>    size_t buildid_len;
>>    bfd_byte *buildid;
>> @@ -2216,19 +2216,17 @@ dwarf2_get_dwz_file (dwarf2_per_bfd *per_bfd)
>>
>>    filename = data.get ();
>
> Declare the filename variable here.

Done.

>>
>> -  std::string abs_storage;
>> -  if (!IS_ABSOLUTE_PATH (filename))
>> +  if (!IS_ABSOLUTE_PATH (filename.c_str ()))
>>      {
>>        gdb::unique_xmalloc_ptr<char> abs
>>  	= gdb_realpath (bfd_get_filename (per_bfd->obfd));
>>
>> -      abs_storage = ldirname (abs.get ()) + SLASH_STRING + filename;
>> -      filename = abs_storage.c_str ();
>> +      filename = ldirname (abs.get ()) + SLASH_STRING + filename;
>>      }
>>
>>    /* First try the file name given in the section.  If that doesn't
>>       work, try to use the build-id instead.  */
>> -  gdb_bfd_ref_ptr dwz_bfd (gdb_bfd_open (filename, gnutarget));
>> +  gdb_bfd_ref_ptr dwz_bfd (gdb_bfd_open (filename.c_str (), gnutarget));
>>    if (dwz_bfd != NULL)
>>      {
>>        if (!build_id_verify (dwz_bfd.get (), buildid_len, buildid))
>> @@ -2238,6 +2236,69 @@ dwarf2_get_dwz_file (dwarf2_per_bfd *per_bfd)
>>    if (dwz_bfd == NULL)
>>      dwz_bfd = build_id_to_debug_bfd (buildid_len, buildid);
>>
>> +  if (dwz_bfd == nullptr)
>> +    {
>> +      /* If the user has provided us with different
>> +	 debug-file-directories, we can try them in order.  */
>> +      size_t dwz_pos = filename.find ("/.dwz/");
>> +
>> +      if (dwz_pos != std::string::npos)
>> +	{
>> +	  std::vector<gdb::unique_xmalloc_ptr<char>> debugdir_vec
>> +	    = dirnames_to_char_ptr_vec (debug_file_directory);
>> +
>> +	  for (const gdb::unique_xmalloc_ptr<char> &debugdir : debugdir_vec)
>> +	    {
>> +	      /* The idea is to iterate over the
>> +		 debug-file-directories provided by the user and
>> +		 replace the hard-coded path in the "filename" by each
>> +		 debug-file-directory.
>> +
>> +		 For example, suppose that filename is:
>> +
>> +		   /usr/lib/debug/.dwz/foo.dwz
>> +
>> +		 And suppose that we have "$HOME/bar" as the
>> +		 debug-file-directory.  We would then adjust filename
>> +		 to look like:
>> +
>> +		   $HOME/bar/.dwz/foo.dwz
>> +
>> +		 which would hopefully allow us to find the alt debug
>> +		 file.  */
>> +	      std::string ddir = debugdir.get ();
>> +
>> +	      /* Check whether the beginning of FILENAME is DDIR.  If
>> +		 it is, then we are dealing with a file which we
>> +		 already attempted to open before, so we just skip it
>> +		 and continue processing the reamining
>> +		 debug-file-directories.  */
>> +	      if (filename.size () > ddir.size ()
>> +		  && filename.compare (0, ddir.size (), ddir) == 0)
>> +		continue;
>
> Corner case, but what if file name is /usr/lib/abcde/.dwz/foo.dwz and
> ddir is /usr/lib/abc?  We will wrongfully skip that debug dir I guess?
>
> Filesystem paths are hard :(.  Is there some "is parent of" function we
> can use?  Worst case, we skip that check and do an unnecessary attempt.

So, this case is only problematic if the debug-file-directory in
question doesn't end with DIR_SEPARATOR.

What I can do is check whether ddir ends with DIR_SEPARATOR, and if not,
I can add one.  This would make sure that we always check for the full
directory name, instead of the incomplete path.

>> +
>> +	      /* Replace FILENAME's default debug-file-directory with
>> +		 DDIR.  */
>> +	      std::string new_filename = filename;
>> +	      new_filename.erase (0, dwz_pos);
>> +	      new_filename.insert (0, ddir);
>
> I think it would be more readable and efficient (less bytes copied
> around) to do just build the new_filename string with something like:
>
>   std::string new_filename = debugdir + &filename[dwz_pos];
>
> This is untested, but I think you'll get the point.  &filename[dwz_pos]
> could also be put into a variable with a meaningful name, for clarity
> (though I can't think of a good name right now).

It is more efficient, but using .erase + .insert can be more readable.
But OK, I don't have a strong opinion here.

> And we probably don't need the allocated `ddir` std::string, you
> construct it but never really use it (you could just refer to debugdir).

I'll need it for the new version of the code, which will implement the
idea I gave above (always guaranteeing that ddir ends with
DIR_SEPARATOR).

>> +
>> +	      dwz_bfd = gdb_bfd_open (new_filename.c_str (), gnutarget);
>> +
>> +	      if (dwz_bfd != nullptr)
>> +		{
>> +		  if (!build_id_verify (dwz_bfd.get (), buildid_len, buildid))
>> +		    {
>> +		      dwz_bfd.reset (nullptr);
>
> Just ".reset ()".

OK.

>> +		      continue;
>> +		    }
>> +		  /* Found it.  */
>> +		  break;
>> +		}
>
> Try to minimize the indentation levels where possible, by using early
> returns or continues.  For example, here:
>
>   if (dwz_bfd == nullptr)
>     continue;
>
>   if (!build_id_verify (...))
>     {
>       dwz_bfd.reset ();
>       continue;
>     }
>
>   /* Found it.  */
>   break;

Done.

Thanks,

-- 
Sergio
GPG key ID: 237A 54B1 0287 28BF 00EF  31F4 D0EB 7628 65FC 5E36
Please send encrypted e-mail if possible
https://sergiodj.net/