| Message ID | 160679383664.447787.14224539520566294960.stgit@magnolia (mailing list archive) |
|---|---|
| State | Deferred, archived |
| Headers | show |
| Series | xfs: faster unaligned copy_file_range | expand |
On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <darrick.wong@oracle.com> > > Add a copy_file_range handler to XFS so that we can accelerate file > copies with reflink when the source and destination ranges are not > block-aligned. We'll use the generic pagecache copy to handle the > unaligned edges and attempt to reflink the middle. Isn't this something we could better handle in the VFS (or a generic helper) so that all file systems that support reflink could benefit?
On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > From: Darrick J. Wong <darrick.wong@oracle.com> > > Add a copy_file_range handler to XFS so that we can accelerate file > copies with reflink when the source and destination ranges are not > block-aligned. We'll use the generic pagecache copy to handle the > unaligned edges and attempt to reflink the middle. > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > --- > fs/xfs/xfs_file.c | 99 +++++++++++++++++++++++++++++++++++++++++++++++++++++ > 1 file changed, 99 insertions(+) > > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c > index 5b0f93f73837..9d1bb0dc30e2 100644 > --- a/fs/xfs/xfs_file.c > +++ b/fs/xfs/xfs_file.c > @@ -1119,6 +1119,104 @@ xfs_file_remap_range( > return remapped > 0 ? remapped : ret; > } > ... > +STATIC ssize_t > +xfs_file_copy_range( > + struct file *src_file, > + loff_t src_off, > + struct file *dst_file, > + loff_t dst_off, > + size_t len, > + unsigned int flags) > +{ > + struct inode *inode_src = file_inode(src_file); > + struct xfs_inode *src = XFS_I(inode_src); > + struct inode *inode_dst = file_inode(dst_file); > + struct xfs_inode *dst = XFS_I(inode_dst); > + struct xfs_mount *mp = src->i_mount; > + loff_t copy_ret; > + loff_t next_block; > + size_t copy_len; > + ssize_t total_copied = 0; > + > + /* Bypass all this if no copy acceleration is possible. */ > + if (!xfs_want_reflink_copy_range(src, src_off, dst, dst_off, len)) > + goto use_generic; > + > + /* Use the regular copy until we're block aligned at the start. */ > + next_block = round_up(src_off + 1, mp->m_sb.sb_blocksize); Why the +1? AFAICT this means we manually copy the first block if src_off does happen to be block aligned. Is this an assumption based on the caller attempting ->remap_file_range() first? BTW, if we do happen to be called in some (theoretical) corner case where remap doesn't work unrelated to alignment, it seems this would unconditionally break the manual copy into multiple parts (first block + the rest). It's not immediately clear to me if that's significant from a performance perspective, but I wonder if it would be nicer here to filter that out more explicitly. For example, run the remap checks on the block aligned offset/len first, or skip the remap if the caller has provided a block aligned start (i.e. hinting that remap failed for other reasons), or perhaps even implement this so it conditionally performs a short manual copy so the next retry would fall into ->remap_file_range() with aligned offsets, etc. Thoughts? > + copy_len = min_t(size_t, len, next_block - src_off); > + if (copy_len > 0) { > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > + dst_off, copy_len, flags); > + if (copy_ret < 0) > + return copy_ret; > + > + src_off += copy_ret; > + dst_off += copy_ret; > + len -= copy_ret; > + total_copied += copy_ret; > + if (copy_ret < copy_len || len == 0) > + return total_copied; > + } > + > + /* > + * Now try to reflink as many full blocks as we can. If the end of the > + * copy request wasn't block-aligned or the reflink fails, we'll just > + * fall into the generic copy to do the rest. > + */ > + copy_len = round_down(len, mp->m_sb.sb_blocksize); > + if (copy_len > 0) { > + copy_ret = xfs_file_remap_range(src_file, src_off, dst_file, > + dst_off, copy_len, REMAP_FILE_CAN_SHORTEN); > + if (copy_ret >= 0) { > + src_off += copy_ret; > + dst_off += copy_ret; > + len -= copy_ret; > + total_copied += copy_ret; > + if (copy_ret < copy_len || len == 0) > + return total_copied; Any reason we return a potential short copy here, but fall into the manual copy if the reflink outright fails? > + } > + } > + > +use_generic: > + /* Use the regular copy to deal with leftover bytes. */ > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > + dst_off, len, flags); > + if (copy_ret < 0) > + return copy_ret; Perhaps this should also check/return total_copied in the event we've already done some work..? Brian > + return total_copied + copy_ret; > +} > + > STATIC int > xfs_file_open( > struct inode *inode, > @@ -1381,6 +1479,7 @@ const struct file_operations xfs_file_operations = { > .get_unmapped_area = thp_get_unmapped_area, > .fallocate = xfs_file_fallocate, > .fadvise = xfs_file_fadvise, > + .copy_file_range = xfs_file_copy_range, > .remap_file_range = xfs_file_remap_range, > }; > >
On Tue, Dec 01, 2020 at 10:02:06AM +0000, Christoph Hellwig wrote: > On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > Add a copy_file_range handler to XFS so that we can accelerate file > > copies with reflink when the source and destination ranges are not > > block-aligned. We'll use the generic pagecache copy to handle the > > unaligned edges and attempt to reflink the middle. > > Isn't this something we could better handle in the VFS (or a generic > helper) so that all file systems that support reflink could benefit? Maybe. I don't know if it's universally true that all filesystems should fall back to reflinking the middle range and pagecache copying the unaligned start/end. The other thing is that xfs can easily support reflink on rtextsize > 1, but that adds the requirement that we set i_blocksize to a larger value than we do now... or find some other way to convey allocation unit size to a generic version of the fallback. OTOH that's pretty easy to do from xfs_copy_file_range. --D
On Tue, Dec 01, 2020 at 10:25:48AM -0500, Brian Foster wrote: > On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > Add a copy_file_range handler to XFS so that we can accelerate file > > copies with reflink when the source and destination ranges are not > > block-aligned. We'll use the generic pagecache copy to handle the > > unaligned edges and attempt to reflink the middle. > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > --- > > fs/xfs/xfs_file.c | 99 +++++++++++++++++++++++++++++++++++++++++++++++++++++ > > 1 file changed, 99 insertions(+) > > > > > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c > > index 5b0f93f73837..9d1bb0dc30e2 100644 > > --- a/fs/xfs/xfs_file.c > > +++ b/fs/xfs/xfs_file.c > > @@ -1119,6 +1119,104 @@ xfs_file_remap_range( > > return remapped > 0 ? remapped : ret; > > } > > > ... > > +STATIC ssize_t > > +xfs_file_copy_range( > > + struct file *src_file, > > + loff_t src_off, > > + struct file *dst_file, > > + loff_t dst_off, > > + size_t len, > > + unsigned int flags) > > +{ > > + struct inode *inode_src = file_inode(src_file); > > + struct xfs_inode *src = XFS_I(inode_src); > > + struct inode *inode_dst = file_inode(dst_file); > > + struct xfs_inode *dst = XFS_I(inode_dst); > > + struct xfs_mount *mp = src->i_mount; > > + loff_t copy_ret; > > + loff_t next_block; > > + size_t copy_len; > > + ssize_t total_copied = 0; > > + > > + /* Bypass all this if no copy acceleration is possible. */ > > + if (!xfs_want_reflink_copy_range(src, src_off, dst, dst_off, len)) > > + goto use_generic; > > + > > + /* Use the regular copy until we're block aligned at the start. */ > > + next_block = round_up(src_off + 1, mp->m_sb.sb_blocksize); > > Why the +1? AFAICT this means we manually copy the first block if > src_off does happen to be block aligned. Is this an assumption based on > the caller attempting ->remap_file_range() first? Yes. The VFS always tries that first. > BTW, if we do happen to be called in some (theoretical) corner case > where remap doesn't work unrelated to alignment, it seems this would > unconditionally break the manual copy into multiple parts (first block + > the rest). It's not immediately clear to me if that's significant from a > performance perspective, I doubt it, since that's usually just copying around the pagecache. > but I wonder if it would be nicer here to > filter that out more explicitly. For example, run the remap checks on > the block aligned offset/len first, or skip the remap if the caller has > provided a block aligned start (i.e. hinting that remap failed for other > reasons), Yes, checking the block alignment is a good suggestion. Will fix. > or perhaps even implement this so it conditionally performs a > short manual copy so the next retry would fall into ->remap_file_range() > with aligned offsets, etc. Hm. That could be a thing too, though my opinion is that we should make as much progress as we can before exiting the kernel. --D > Thoughts? > > > + copy_len = min_t(size_t, len, next_block - src_off); > > + if (copy_len > 0) { > > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > > + dst_off, copy_len, flags); > > + if (copy_ret < 0) > > + return copy_ret; > > + > > + src_off += copy_ret; > > + dst_off += copy_ret; > > + len -= copy_ret; > > + total_copied += copy_ret; > > + if (copy_ret < copy_len || len == 0) > > + return total_copied; > > + } > > + > > + /* > > + * Now try to reflink as many full blocks as we can. If the end of the > > + * copy request wasn't block-aligned or the reflink fails, we'll just > > + * fall into the generic copy to do the rest. > > + */ > > + copy_len = round_down(len, mp->m_sb.sb_blocksize); > > + if (copy_len > 0) { > > + copy_ret = xfs_file_remap_range(src_file, src_off, dst_file, > > + dst_off, copy_len, REMAP_FILE_CAN_SHORTEN); > > + if (copy_ret >= 0) { > > + src_off += copy_ret; > > + dst_off += copy_ret; > > + len -= copy_ret; > > + total_copied += copy_ret; > > + if (copy_ret < copy_len || len == 0) > > + return total_copied; > > Any reason we return a potential short copy here, but fall into the > manual copy if the reflink outright fails? > > > + } > > + } > > + > > +use_generic: > > + /* Use the regular copy to deal with leftover bytes. */ > > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > > + dst_off, len, flags); > > + if (copy_ret < 0) > > + return copy_ret; > > Perhaps this should also check/return total_copied in the event we've > already done some work..? > > Brian > > > + return total_copied + copy_ret; > > +} > > + > > STATIC int > > xfs_file_open( > > struct inode *inode, > > @@ -1381,6 +1479,7 @@ const struct file_operations xfs_file_operations = { > > .get_unmapped_area = thp_get_unmapped_area, > > .fallocate = xfs_file_fallocate, > > .fadvise = xfs_file_fadvise, > > + .copy_file_range = xfs_file_copy_range, > > .remap_file_range = xfs_file_remap_range, > > }; > > > > >
On Sun, Dec 06, 2020 at 03:24:54PM -0800, Darrick J. Wong wrote: > On Tue, Dec 01, 2020 at 10:25:48AM -0500, Brian Foster wrote: > > On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > > > Add a copy_file_range handler to XFS so that we can accelerate file > > > copies with reflink when the source and destination ranges are not > > > block-aligned. We'll use the generic pagecache copy to handle the > > > unaligned edges and attempt to reflink the middle. > > > > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> > > > --- > > > fs/xfs/xfs_file.c | 99 +++++++++++++++++++++++++++++++++++++++++++++++++++++ > > > 1 file changed, 99 insertions(+) > > > > > > > > > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c > > > index 5b0f93f73837..9d1bb0dc30e2 100644 > > > --- a/fs/xfs/xfs_file.c > > > +++ b/fs/xfs/xfs_file.c > > > @@ -1119,6 +1119,104 @@ xfs_file_remap_range( > > > return remapped > 0 ? remapped : ret; > > > } > > > > > ... > > > +STATIC ssize_t > > > +xfs_file_copy_range( > > > + struct file *src_file, > > > + loff_t src_off, > > > + struct file *dst_file, > > > + loff_t dst_off, > > > + size_t len, > > > + unsigned int flags) > > > +{ > > > + struct inode *inode_src = file_inode(src_file); > > > + struct xfs_inode *src = XFS_I(inode_src); > > > + struct inode *inode_dst = file_inode(dst_file); > > > + struct xfs_inode *dst = XFS_I(inode_dst); > > > + struct xfs_mount *mp = src->i_mount; > > > + loff_t copy_ret; > > > + loff_t next_block; > > > + size_t copy_len; > > > + ssize_t total_copied = 0; > > > + > > > + /* Bypass all this if no copy acceleration is possible. */ > > > + if (!xfs_want_reflink_copy_range(src, src_off, dst, dst_off, len)) > > > + goto use_generic; > > > + > > > + /* Use the regular copy until we're block aligned at the start. */ > > > + next_block = round_up(src_off + 1, mp->m_sb.sb_blocksize); > > > > Why the +1? AFAICT this means we manually copy the first block if > > src_off does happen to be block aligned. Is this an assumption based on > > the caller attempting ->remap_file_range() first? > > Yes. The VFS always tries that first. > > > BTW, if we do happen to be called in some (theoretical) corner case > > where remap doesn't work unrelated to alignment, it seems this would > > unconditionally break the manual copy into multiple parts (first block + > > the rest). It's not immediately clear to me if that's significant from a > > performance perspective, > > I doubt it, since that's usually just copying around the pagecache. > Ok, comment please. > > but I wonder if it would be nicer here to > > filter that out more explicitly. For example, run the remap checks on > > the block aligned offset/len first, or skip the remap if the caller has > > provided a block aligned start (i.e. hinting that remap failed for other > > reasons), > > Yes, checking the block alignment is a good suggestion. Will fix. > > > or perhaps even implement this so it conditionally performs a > > short manual copy so the next retry would fall into ->remap_file_range() > > with aligned offsets, etc. > > Hm. That could be a thing too, though my opinion is that we should make > as much progress as we can before exiting the kernel. > Yeah, the more I thought about this the more it seemed like a hack and not really sane for a production system. Brian > --D > > > Thoughts? > > > > > + copy_len = min_t(size_t, len, next_block - src_off); > > > + if (copy_len > 0) { > > > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > > > + dst_off, copy_len, flags); > > > + if (copy_ret < 0) > > > + return copy_ret; > > > + > > > + src_off += copy_ret; > > > + dst_off += copy_ret; > > > + len -= copy_ret; > > > + total_copied += copy_ret; > > > + if (copy_ret < copy_len || len == 0) > > > + return total_copied; > > > + } > > > + > > > + /* > > > + * Now try to reflink as many full blocks as we can. If the end of the > > > + * copy request wasn't block-aligned or the reflink fails, we'll just > > > + * fall into the generic copy to do the rest. > > > + */ > > > + copy_len = round_down(len, mp->m_sb.sb_blocksize); > > > + if (copy_len > 0) { > > > + copy_ret = xfs_file_remap_range(src_file, src_off, dst_file, > > > + dst_off, copy_len, REMAP_FILE_CAN_SHORTEN); > > > + if (copy_ret >= 0) { > > > + src_off += copy_ret; > > > + dst_off += copy_ret; > > > + len -= copy_ret; > > > + total_copied += copy_ret; > > > + if (copy_ret < copy_len || len == 0) > > > + return total_copied; > > > > Any reason we return a potential short copy here, but fall into the > > manual copy if the reflink outright fails? > > > > > + } > > > + } > > > + > > > +use_generic: > > > + /* Use the regular copy to deal with leftover bytes. */ > > > + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, > > > + dst_off, len, flags); > > > + if (copy_ret < 0) > > > + return copy_ret; > > > > Perhaps this should also check/return total_copied in the event we've > > already done some work..? > > > > Brian > > > > > + return total_copied + copy_ret; > > > +} > > > + > > > STATIC int > > > xfs_file_open( > > > struct inode *inode, > > > @@ -1381,6 +1479,7 @@ const struct file_operations xfs_file_operations = { > > > .get_unmapped_area = thp_get_unmapped_area, > > > .fallocate = xfs_file_fallocate, > > > .fadvise = xfs_file_fadvise, > > > + .copy_file_range = xfs_file_copy_range, > > > .remap_file_range = xfs_file_remap_range, > > > }; > > > > > > > > >
On Sun, Dec 06, 2020 at 03:21:54PM -0800, Darrick J. Wong wrote: > On Tue, Dec 01, 2020 at 10:02:06AM +0000, Christoph Hellwig wrote: > > On Mon, Nov 30, 2020 at 07:37:16PM -0800, Darrick J. Wong wrote: > > > From: Darrick J. Wong <darrick.wong@oracle.com> > > > > > > Add a copy_file_range handler to XFS so that we can accelerate file > > > copies with reflink when the source and destination ranges are not > > > block-aligned. We'll use the generic pagecache copy to handle the > > > unaligned edges and attempt to reflink the middle. > > > > Isn't this something we could better handle in the VFS (or a generic > > helper) so that all file systems that support reflink could benefit? > > Maybe. I don't know if it's universally true that all filesystems > should fall back to reflinking the middle range and pagecache copying > the unaligned start/end. > > The other thing is that xfs can easily support reflink on rtextsize > 1, > but that adds the requirement that we set i_blocksize to a larger value > than we do now... or find some other way to convey allocation unit size > to a generic version of the fallback. OTOH that's pretty easy to do > from xfs_copy_file_range. I think you can basically turn xfs_want_reflink_copy_range into a callback supplied by the fs for the generic helper, and to deal with the rtextsize problem just return the relevant block size from the helper.
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c index 5b0f93f73837..9d1bb0dc30e2 100644 --- a/fs/xfs/xfs_file.c +++ b/fs/xfs/xfs_file.c @@ -1119,6 +1119,104 @@ xfs_file_remap_range( return remapped > 0 ? remapped : ret; } +/* + * Decide if we want to use reflink to accelerate a copy_file_range request. + * + * We need to use the generic pagecache copy routine if there's no reflink; if + * the two files are on different filesystems; if the two files are on + * different devices; or if the two offsets are not at the same offset within + * an fs block. Studies on the author's computer show that reflink doesn't + * speed up copies smaller than 32k, so use the page cache for those. + */ +static inline bool +xfs_want_reflink_copy_range( + struct xfs_inode *src, + unsigned int src_off, + struct xfs_inode *dst, + unsigned int dst_off, + size_t len) +{ + struct xfs_mount *mp = src->i_mount; + + if (len < 32768) + return false; + if (mp != dst->i_mount) + return false; + if (!xfs_sb_version_hasreflink(&mp->m_sb)) + return false; + if (XFS_IS_REALTIME_INODE(src) != XFS_IS_REALTIME_INODE(dst)) + return false; + return (src_off & mp->m_blockmask) == (dst_off & mp->m_blockmask); +} + +STATIC ssize_t +xfs_file_copy_range( + struct file *src_file, + loff_t src_off, + struct file *dst_file, + loff_t dst_off, + size_t len, + unsigned int flags) +{ + struct inode *inode_src = file_inode(src_file); + struct xfs_inode *src = XFS_I(inode_src); + struct inode *inode_dst = file_inode(dst_file); + struct xfs_inode *dst = XFS_I(inode_dst); + struct xfs_mount *mp = src->i_mount; + loff_t copy_ret; + loff_t next_block; + size_t copy_len; + ssize_t total_copied = 0; + + /* Bypass all this if no copy acceleration is possible. */ + if (!xfs_want_reflink_copy_range(src, src_off, dst, dst_off, len)) + goto use_generic; + + /* Use the regular copy until we're block aligned at the start. */ + next_block = round_up(src_off + 1, mp->m_sb.sb_blocksize); + copy_len = min_t(size_t, len, next_block - src_off); + if (copy_len > 0) { + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, + dst_off, copy_len, flags); + if (copy_ret < 0) + return copy_ret; + + src_off += copy_ret; + dst_off += copy_ret; + len -= copy_ret; + total_copied += copy_ret; + if (copy_ret < copy_len || len == 0) + return total_copied; + } + + /* + * Now try to reflink as many full blocks as we can. If the end of the + * copy request wasn't block-aligned or the reflink fails, we'll just + * fall into the generic copy to do the rest. + */ + copy_len = round_down(len, mp->m_sb.sb_blocksize); + if (copy_len > 0) { + copy_ret = xfs_file_remap_range(src_file, src_off, dst_file, + dst_off, copy_len, REMAP_FILE_CAN_SHORTEN); + if (copy_ret >= 0) { + src_off += copy_ret; + dst_off += copy_ret; + len -= copy_ret; + total_copied += copy_ret; + if (copy_ret < copy_len || len == 0) + return total_copied; + } + } + +use_generic: + /* Use the regular copy to deal with leftover bytes. */ + copy_ret = generic_copy_file_range(src_file, src_off, dst_file, + dst_off, len, flags); + if (copy_ret < 0) + return copy_ret; + return total_copied + copy_ret; +} + STATIC int xfs_file_open( struct inode *inode, @@ -1381,6 +1479,7 @@ const struct file_operations xfs_file_operations = { .get_unmapped_area = thp_get_unmapped_area, .fallocate = xfs_file_fallocate, .fadvise = xfs_file_fadvise, + .copy_file_range = xfs_file_copy_range, .remap_file_range = xfs_file_remap_range, };