Patchwork ldconfig: Call fsync on temporary files before renaming them [BZ #20890]

login
register
mail settings
Submitter Florian Weimer
Date Jan. 12, 2018, 11:07 a.m.
Message ID <20180112110757.D4D2440172552@oldenburg.str.redhat.com>
Download mbox | patch
Permalink /patch/25363/
State New
Headers show

Comments

Florian Weimer - Jan. 12, 2018, 11:07 a.m.
If the system crashes before the file data has been written to disk, the
file system recovery upon the next mount may restore a partially
rewritten temporary file under the non-temporary (final) name (after the
rename operation).

Some file systems perform an implicit fsync before renaming a file over
another one, but XFS does not, for example.

2018-01-12  Florian Weimer  <fweimer@redhat.com>

	[BZ #20890]
	* elf/cache.c (save_cache): Call fsync on temporary file before
	renaming it.
	(save_aux_cache): Likewise.
Christoph Hellwig - Jan. 12, 2018, 11:37 a.m.
On Fri, Jan 12, 2018 at 12:07:57PM +0100, Florian Weimer wrote:
> If the system crashes before the file data has been written to disk, the
> file system recovery upon the next mount may restore a partially
> rewritten temporary file under the non-temporary (final) name (after the
> rename operation).
> 
> Some file systems perform an implicit fsync before renaming a file over
> another one, but XFS does not, for example.

At least on Linux no file system performs an actual fsync equivalent.
A few do start a writeout, but don't actually wait on it.
Florian Weimer - Jan. 12, 2018, 11:54 a.m.
On 01/12/2018 12:37 PM, Christoph Hellwig wrote:
> On Fri, Jan 12, 2018 at 12:07:57PM +0100, Florian Weimer wrote:
>> If the system crashes before the file data has been written to disk, the
>> file system recovery upon the next mount may restore a partially
>> rewritten temporary file under the non-temporary (final) name (after the
>> rename operation).
>>
>> Some file systems perform an implicit fsync before renaming a file over
>> another one, but XFS does not, for example.
> 
> At least on Linux no file system performs an actual fsync equivalent.
> A few do start a writeout, but don't actually wait on it.

Oh.  I'll drop the last paragraph from the commit message, then.

Thanks,
Florian
Paul Eggert - Jan. 13, 2018, 2:28 a.m.
Florian Weimer wrote:
>     if (write (fd, strings, total_strlen) != (ssize_t) total_strlen
> -      || close (fd))
> +      || fsync (fd) != 0
> +      || close (fd) != 0)
>       error (EXIT_FAILURE, errno, _("Writing of cache data failed"));

Would fdatasync suffice, instead of fsync?

Patch

diff --git a/elf/cache.c b/elf/cache.c
index 1ec6ab36e7..2165f8d305 100644
--- a/elf/cache.c
+++ b/elf/cache.c
@@ -449,7 +449,8 @@  save_cache (const char *cache_name)
     }
 
   if (write (fd, strings, total_strlen) != (ssize_t) total_strlen
-      || close (fd))
+      || fsync (fd) != 0
+      || close (fd) != 0)
     error (EXIT_FAILURE, errno, _("Writing of cache data failed"));
 
   /* Make sure user can always read cache file */
@@ -812,7 +813,8 @@  save_aux_cache (const char *aux_cache_name)
 
   if (write (fd, file_entries, file_entries_size + total_strlen)
       != (ssize_t) (file_entries_size + total_strlen)
-      || close (fd))
+      || fsync (fd) != 0
+      || close (fd) != 0)
     {
       unlink (temp_name);
       goto out_fail;