From: Mike Marshall <hubcap@xxxxxxxxxxxx>
Until now the Orangefs kernel module has not gone through the pagecache
and has been a de facto O_DIRECT style module.
This patch series adds pagecache support and greatly improves
small IO. For example, xfstest generic/069 used to take several
hours on our test VMs, and now finishes in seconds.
Orangefs' kernel module shares a memory buffer with Orangefs'
userspace parts for IO. Orangefs can best exploit parallelism when
the buffer is significantly filled, and struggles when the buffer
is nearly empty. Buffering reads and writes through the pagecache
helps us to communicate with userspace less often with buffers
that are more full.
Reads can be tuned with "count" from read(2) to simulate page sizes
more in line with Orangefs's buffer size (4 meg by default). Userspace
will deposit "count" bytes in the shared buffer on a read and when
the VFS calls orangefs_readpage, as many pagecache pages as possible
will be filled at once.
Writes are coalesced even across pages when possible up to the size of
the buffer between the OrangeFS client core and the kernel. This is
4 MB by default. This means that an application writing with a block
size less than 4 MB will have its writes upgraded. For sequential
writes, this is a big improvement.
Martin Brandenburg (19):
orangefs: implement xattr cache
orangefs: do not invalidate attributes on inode create
orangefs: simplify orangefs_inode_getattr interface
orangefs: update attributes rather than relying on server
orangefs: hold i_lock during inode_getattr
orangefs: set up and use backing_dev_info
orangefs: let setattr write to cached inode
orangefs: reorganize setattr functions to track attribute changes
orangefs: remove orangefs_readpages
orangefs: service ops done for writeback are not killable
orangefs: migrate to generic_file_read_iter
orangefs: implement writepage
orangefs: do not return successful read when the client-core
disappeared
orangefs: move do_readv_writev to direct_IO
orangefs: skip inode writeout if nothing to write
orangefs: avoid fsync service operation on flush
orangefs: write range tracking
orangefs: implement writepages
orangefs: add orangefs_revalidate_mapping
Mike Marshall (3):
orangefs: remember count when reading.
orangefs: pass slot index back to readpage.
orangefs: copy Orangefs-sized blocks into the pagecache if possible.
fs/orangefs/acl.c | 4 +-
fs/orangefs/file.c | 389 ++++++--------
fs/orangefs/inode.c | 908 +++++++++++++++++++++++++++++----
fs/orangefs/namei.c | 40 +-
fs/orangefs/orangefs-bufmap.c | 13 +
fs/orangefs/orangefs-bufmap.h | 2 +
fs/orangefs/orangefs-debugfs.c | 4 +-
fs/orangefs/orangefs-kernel.h | 56 +-
fs/orangefs/orangefs-mod.c | 1 +
fs/orangefs/orangefs-sysfs.c | 22 +
fs/orangefs/orangefs-utils.c | 179 +++----
fs/orangefs/super.c | 39 +-
fs/orangefs/waitqueue.c | 18 +-
fs/orangefs/xattr.c | 106 +++-
14 files changed, 1293 insertions(+), 488 deletions(-)
--
2.20.1
