[PATCH v3 0/7] DAX fsync/msync support

From: Ross Zwisler
Date: Tue Dec 08 2015 - 14:19:22 EST


This patch series adds a slimmed down version of fsync/msync support to
DAX. The major change versus v2 of this patch series is that we no longer
remove DAX entries from the radix tree during fsync/msync calls. Instead
the list of DAX entries in the radix tree grows for the lifetime of the
mapping. We reclaim DAX entries from the radix tree via
clear_exceptional_entry() for truncate, when the filesystem is unmounted,
etc.

This change was made because if we try and remove radix tree entries during
writeback operations there are a number of race conditions that exist
between those writeback operations and page faults. In the non-DAX case
these races are dealt with using the page lock, but we don't have a good
replacement lock with the same granularity. These races could leave us in
a place where we have a DAX page that is dirty and writeable from userspace
but no longer in the radix tree. This page would then be skipped during
subsequent writeback operations, which is unacceptable.

I do plan to continue to try and solve these race conditions so that we can
have a more optimal fsync/msync solution for DAX, but I wanted to get this
set out for v4.5 consideration while I continued working. While
suboptimal the solution in this series gives us correct behavior for DAX
fsync/msync and seems like a reasonable short term compromise.

This series is built upon v4.4-rc4 plus the recent ext4 DAX series from Jan
Kara (http://www.spinics.net/lists/linux-ext4/msg49951.html) and a recent
XFS fix from Dave Chinner (https://lkml.org/lkml/2015/12/2/923). The tree
with all this working can be found here:

https://git.kernel.org/cgit/linux/kernel/git/zwisler/linux.git/log/?h=fsync_v3

Other changes versus v2:
- Renamed dax_fsync() to dax_writeback_mapping_range(). (Dave Chinner)
- Removed REQ_FUA/REQ_FLUSH support from the PMEM driver and instead just
make the call to wmb_pmem() in dax_writeback_mapping_range(). (Dan)
- Reworked some BUG_ON() calls to be a WARN_ON() followed by an error
return.
- Moved call to dax_writeback_mapping_range() from the filesystems down
into filemap_write_and_wait_range(). (Dave Chinner)
- Fixed handling of DAX read faults so they create a radix tree entry but
don't mark it as dirty until the follow-up dax_pfn_mkwrite() call.
- Update clear_exceptional_entry() and to dax_writeback_one() so they
validate the DAX radix tree entry before they use it. (Dave Chinner)
- Added a comment to find_get_entries_tag() to explain the restart
condition. (Dave Chinner)

Ross Zwisler (7):
pmem: add wb_cache_pmem() to the PMEM API
dax: support dirty DAX entries in radix tree
mm: add find_get_entries_tag()
dax: add support for fsync/sync
ext2: call dax_pfn_mkwrite() for DAX fsync/msync
ext4: call dax_pfn_mkwrite() for DAX fsync/msync
xfs: call dax_pfn_mkwrite() for DAX fsync/msync

arch/x86/include/asm/pmem.h | 11 ++--
fs/block_dev.c | 3 +-
fs/dax.c | 147 ++++++++++++++++++++++++++++++++++++++++++--
fs/ext2/file.c | 4 +-
fs/ext4/file.c | 4 +-
fs/inode.c | 1 +
fs/xfs/xfs_file.c | 7 ++-
include/linux/dax.h | 7 +++
include/linux/fs.h | 1 +
include/linux/pagemap.h | 3 +
include/linux/pmem.h | 22 ++++++-
include/linux/radix-tree.h | 9 +++
mm/filemap.c | 84 ++++++++++++++++++++++++-
mm/truncate.c | 64 +++++++++++--------
14 files changed, 319 insertions(+), 48 deletions(-)

--
2.5.0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/