[lkml]   [2008]   [Jan]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCH 0/2] yet another attempt to fix the ctime and mtime issue
The POSIX standard requires that the ctime and mtime fields
for memory-mapped files should be updated after a write
reference to the memory region where the file data is mapped.
At least FreeBSD 6.2 and HP-UX 11i implement this properly.
Linux does not, which leads to data loss problems in database
backup applications.

Kernel Bug Tracker contains more information about the problem:

There have been several attempts in the past to address this
issue. Following are a few links to LKML discussions related
to this bug:

All earlier solutions were criticized. Some solutions did not
handle memory-mapped block devices properly. Some led to forcing
applications to explicitly call msync() to update file metadata.
Some contained errors in using kernel synchronization primitives.

In the two patches that follow, I would like to propose a new

This is the third version of my changes. This version takes
into account all feedback I received for the two previous versions.
The overall design remains basically the same as the one that
was acked by Rick van Riel:

To the best of my knowledge, these patches are free of all the
drawbacks found during previous attempts by Peter Staubach,
Miklos Szeredi and myself.

New since the previous version:

1) no need to explicitly call msync() to update file times;
2) changing block device data is visible to all device files
associated with the block device;
3) in the cleanup part, the error checks are separated out as
suggested by Rik van Riel;
4) some small refinements accodring to the LKML comments.

This is how I tested the patches.

1. To test the features mentioned above, I wrote a unit test
available from

I verified that the unit test passed successfully for both
regular files and block device files. For the unit test I
used the following architectures: 32-bit x86, x86_64 and
MIPS32 (cross-compiled from x86_64).

2. I did build tests with allmodconfig and allyesconfig on x86_64.

3. I ran the following test cases from the LTP test suite:


No regressions were found by these test cases.

I think that the bug #2645 is resolved by these patches.

Please apply.

 \ /
  Last update: 2008-01-13 05:43    [W:0.125 / U:1.748 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site