[PATCH 22/28] elf: Add extension mechanism to ld.so.cache

[Adhemerval Zanella]

On 30/10/2020 09:22, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>>> +/* Print the extension information at the cache at start address
>>> +   FILE_BASE, of ltength FILE_SIZE bytes.  The new-format cache header
>>
>> s/ltength/length
> 
> Fixed.
> 
>>> +   is at CACHE, and the file name for diagnostics is CACHE_NAME.  */
>>> +static void
>>> +print_extensions (struct cache_extension_all_loaded *ext)
>>> +{
>>> +  if (ext->sections[cache_extension_tag_generator].base != NULL)
>>> +    {
>>> +      fputs (_("Cache generated by: "), stdout);
>>> +      fwrite (ext->sections[cache_extension_tag_generator].base, 1,
>>> +	      ext->sections[cache_extension_tag_generator].size, stdout);
>>> +      putchar ('\n');
>>> +    }
>>> +}
>>> +
>>
>> Ok.  Will be the extension tag data always comprised of ascii printable
>> characters?
> 
> For the generator extension: Yes.  But not for other extensions.
> This code only deals with the generator extension.
> 
>>> +/* Size of the cache extension directory.  All tags are assumed to be
>>> +   present.  */
>>> +enum
>>> +  {
>>> +   cache_extension_size = (offsetof (struct cache_extension, sections)
>>> +			   + (cache_extension_count
>>> +			      * sizeof (struct cache_extension_section)))
>>> +  };
>>> +
>>> +/* Write the cache extensions to FD.  The extension directory is
>>> +   assumed to be located at CACHE_EXTENSION_OFFSET.  */
>>> +static void
>>> +write_extensions (int fd, uint32_t cache_extension_offset)
>>> +{
>>> +  assert ((cache_extension_offset % 4) == 0);
>>
>> Maybe a proper error msg instead of an assert here?
> 
> This is in the cache generator, so it's a code bug if the assert fires,
> not corrupted input.

Ack.

> 
>>> @@ -435,6 +497,25 @@ save_cache (const char *cache_name)
>>>        && idx_old < cache_entry_old_count)
>>>      file_entries->libs[idx_old] = file_entries->libs[idx_old - 1];
>>>  
>>> +  /* Compute the location of the extension directory.  This
>>> +     implementation puts the directory after the string table.  The
>>> +     size computation matches the write calls below.  The extension
>>> +     directory does not exist with format 0, so the value does not
>>> +     matter.  */
>>> +  uint32_t extension_offset = 0;
>>> +  if (opt_format != 2)
>>> +    extension_offset += file_entries_size;
>>> +  if (opt_format != 0)
>>> +    {
>>> +      if (opt_format != 2)
>>> +	extension_offset += pad;
>>> +      extension_offset += file_entries_new_size;
>>> +    }
>>
>> Ok, although I think we should be good move the 'opt_format' definition to
>> a proper enumeration.
> 
> Can this wait?  Something for a future patch?

Alright, although it does simplify reading this patchset and the
change should most mechanical. 

> 
>>> +static bool __attribute__ ((unused))
>>
>> Maybe use inline and let the compiler decide? Or the function is
>> really duplicate in a lot of places?
> 
> The compiler already decides for static functions and will inline them
> if they are only used once.  Given the size of the function, I think
> that's the appropriate approach here.

Ack.

> 
>>> +cache_extension_load (const struct cache_file_new *cache,
>>> +		      const void *file_base, size_t file_size,
>>> +		      struct cache_extension_all_loaded *loaded)
>>> +{
>>> +  memset (loaded, 0, sizeof (*loaded));
>>> +  if (cache->extension_offset == 0)
>>> +    /* No extensions present.  This is not a format error.  */
>>> +    return true;
>>> +  if ((cache->extension_offset % 4) != 0)
>>> +    /* Extension offset is misaligned.  */
>>> +    return false;
>>> +  size_t size_tmp;
>>> +  if (__builtin_add_overflow (cache->extension_offset,
>>> +			      sizeof (struct cache_extension), &size_tmp)
>>> +      || size_tmp > file_size)
>>> +    /* Extension extends beyond the end of the file.  */
>>> +    return false;
>>> +  const struct cache_extension *ext = file_base + cache->extension_offset;
>>
>> Maybe we should add an alignment check for 'file_base' as well (to
>> avoid unaligned struct member deference)?
> 
> This pointer comes from mmap, so it's not something that can be wrong as
> the result of the file data.  I could add an assert, but I don't think
> this is likely to go wrong.

Ack.