Re: [PATCH v1 0/8] VFS: In-kernel copy system call
From: Andreas Dilger <hidden>
Date: 2015-09-04 22:25:27
On Sep 4, 2015, at 2:16 PM, Anna Schumaker [off-list ref] wrote:
Copy system calls came up during Plumbers a couple of weeks ago,
because several filesystems (including NFS and XFS) are currently
working on copy acceleration implementations. We haven't heard from
Zach Brown in a while, so I volunteered to push his patches upstream
so individual filesystems don't need to keep writing their own ioctls.
The first three patches are a simple reposting of Zach's patches
from several months ago, with one minor error code fix. The remaining
patches add in a fallback mechanism when filesystems don't provide a
copy function. This is especially useful when doing a server-side
copy on NFS (using the new COPY operation in NFS v4.2). This fallback
can be disabled by passing the flag COPY_REFLINK to the system call.
The last patch is a man page patch documenting this new system call,
including an example program.
I tested the fallback option by using /dev/urandom to generate files
of varying sizes and copying them. I compared the time to copy
against that of `cp` just to see if there is a noticable difference.
I found that runtimes are roughly the same, but in-kernel copy tends
to use less of the cpu. Values in the tables below are averages
across multiple trials.
/usr/bin/cp | 512 MB | 1024 MB | 1536 MB | 2048 MB
-------------|-----------|-----------|-----------|-----------
user | 0.00s | 0.00s | 0.00s | 0.00s
system | 0.32s | 0.52s | 1.04s | 1.04s
cpu | 73% | 69% | 62% | 62%
total | 0.446 | 0.757 | 1.197 | 1.667
VFS copy | 512 MB | 1024 MB | 1536 MB | 2048 MB
-------------|-----------|-----------|-----------|-----------
user | 0.00s | 0.00s | 0.00s | 0.00s
system | 0.33s | 0.49s | 0.76s | 0.99s
cpu | 77% | 62% | 60% | 59%
total | 0.422 | 0.777 | 1.267 | 1.655
Questions? Comments? Thoughts?
This is a bit of a surprising result, since in my testing in the
past, copy_{to/from}_user() is a major consumer of CPU time (50%
of a CPU core at 1GB/s). What backing filesystem did you test on?
In theory, the VFS copy routines should save at least 50% of the
CPU usage since it only needs to make one copy (src->dest) instead
of two (kernel->user, user->kernel). Ideally it wouldn't make any
data copies at all and just pass page references from the source
to the target.
Cheers, AndreasAnna Anna Schumaker (5): btrfs: Add mountpoint checking during btrfs_copy_file_range vfs: Remove copy_file_range mountpoint checks vfs: Copy should check len after file open mode vfs: Copy should use file_out rather than file_in vfs: Fall back on splice if no copy function defined Zach Brown (3): vfs: add copy_file_range syscall and vfs helper x86: add sys_copy_file_range to syscall tables btrfs: add .copy_file_range file operation arch/x86/entry/syscalls/syscall_32.tbl | 1 + arch/x86/entry/syscalls/syscall_64.tbl | 1 + fs/btrfs/ctree.h | 3 + fs/btrfs/file.c | 1 + fs/btrfs/ioctl.c | 95 ++++++++++++++---------- fs/read_write.c | 132 +++++++++++++++++++++++++++++++++ include/linux/copy.h | 6 ++ include/linux/fs.h | 3 + include/uapi/asm-generic/unistd.h | 4 +- include/uapi/linux/Kbuild | 1 + include/uapi/linux/copy.h | 6 ++ kernel/sys_ni.c | 1 + 12 files changed, 214 insertions(+), 40 deletions(-) create mode 100644 include/linux/copy.h create mode 100644 include/uapi/linux/copy.h -- 2.5.1 -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Cheers, Andreas