diff mbox series

[5/7] xfs: speed up xfarray sort by sorting xfile page contents directly

Message ID 169049623643.921478.6377149280402650711.stgit@frogsfrogsfrogs (mailing list archive)
State New, archived
Headers show
Series xfs: stage repair information in pageable memory | expand

Commit Message

Darrick J. Wong July 27, 2023, 10:26 p.m. UTC
From: Darrick J. Wong <djwong@kernel.org>

If all the records in an xfarray subset live within the same memory
page, we can short-circuit even more quicksort recursion by mapping that
page into the local CPU and using the kernel's heapsort function to sort
the subset.  On the author's computer, this reduces the runtime by
another 15% on a 500,000 element array.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Kent Overstreet <kent.overstreet@linux.dev>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
---
 fs/xfs/scrub/trace.h   |   20 ++++++++++
 fs/xfs/scrub/xfarray.c |   97 ++++++++++++++++++++++++++++++++++++++++++++++++
 fs/xfs/scrub/xfarray.h |    4 ++
 3 files changed, 121 insertions(+)
diff mbox series

Patch

diff --git a/fs/xfs/scrub/trace.h b/fs/xfs/scrub/trace.h
index f8c814e07587f..e9d7159461428 100644
--- a/fs/xfs/scrub/trace.h
+++ b/fs/xfs/scrub/trace.h
@@ -869,6 +869,26 @@  TRACE_EVENT(xfarray_isort,
 		  __entry->hi - __entry->lo)
 );
 
+TRACE_EVENT(xfarray_pagesort,
+	TP_PROTO(struct xfarray_sortinfo *si, uint64_t lo, uint64_t hi),
+	TP_ARGS(si, lo, hi),
+	TP_STRUCT__entry(
+		__field(unsigned long, ino)
+		__field(unsigned long long, lo)
+		__field(unsigned long long, hi)
+	),
+	TP_fast_assign(
+		__entry->ino = file_inode(si->array->xfile->file)->i_ino;
+		__entry->lo = lo;
+		__entry->hi = hi;
+	),
+	TP_printk("xfino 0x%lx lo %llu hi %llu elts %llu",
+		  __entry->ino,
+		  __entry->lo,
+		  __entry->hi,
+		  __entry->hi - __entry->lo)
+);
+
 TRACE_EVENT(xfarray_qsort,
 	TP_PROTO(struct xfarray_sortinfo *si, uint64_t lo, uint64_t hi),
 	TP_ARGS(si, lo, hi),
diff --git a/fs/xfs/scrub/xfarray.c b/fs/xfs/scrub/xfarray.c
index 2a0599f660d7b..457e56eac5e15 100644
--- a/fs/xfs/scrub/xfarray.c
+++ b/fs/xfs/scrub/xfarray.c
@@ -545,6 +545,87 @@  xfarray_isort(
 	return xfile_obj_store(si->array->xfile, scratch, len, lo_pos);
 }
 
+/* Grab a page for sorting records. */
+static inline int
+xfarray_sort_get_page(
+	struct xfarray_sortinfo	*si,
+	loff_t			pos,
+	uint64_t		len)
+{
+	int			error;
+
+	error = xfile_get_page(si->array->xfile, pos, len, &si->xfpage);
+	if (error)
+		return error;
+
+	/*
+	 * xfile pages must never be mapped into userspace, so we skip the
+	 * dcache flush when mapping the page.
+	 */
+	si->page_kaddr = kmap_local_page(si->xfpage.page);
+	return 0;
+}
+
+/* Release a page we grabbed for sorting records. */
+static inline int
+xfarray_sort_put_page(
+	struct xfarray_sortinfo	*si)
+{
+	if (!si->page_kaddr)
+		return 0;
+
+	kunmap_local(si->page_kaddr);
+	si->page_kaddr = NULL;
+
+	return xfile_put_page(si->array->xfile, &si->xfpage);
+}
+
+/* Decide if these records are eligible for in-page sorting. */
+static inline bool
+xfarray_want_pagesort(
+	struct xfarray_sortinfo	*si,
+	xfarray_idx_t		lo,
+	xfarray_idx_t		hi)
+{
+	pgoff_t			lo_page;
+	pgoff_t			hi_page;
+	loff_t			end_pos;
+
+	/* We can only map one page at a time. */
+	lo_page = xfarray_pos(si->array, lo) >> PAGE_SHIFT;
+	end_pos = xfarray_pos(si->array, hi) + si->array->obj_size - 1;
+	hi_page = end_pos >> PAGE_SHIFT;
+
+	return lo_page == hi_page;
+}
+
+/* Sort a bunch of records that all live in the same memory page. */
+STATIC int
+xfarray_pagesort(
+	struct xfarray_sortinfo	*si,
+	xfarray_idx_t		lo,
+	xfarray_idx_t		hi)
+{
+	void			*startp;
+	loff_t			lo_pos = xfarray_pos(si->array, lo);
+	uint64_t		len = xfarray_pos(si->array, hi - lo);
+	int			error = 0;
+
+	trace_xfarray_pagesort(si, lo, hi);
+
+	xfarray_sort_bump_loads(si);
+	error = xfarray_sort_get_page(si, lo_pos, len);
+	if (error)
+		return error;
+
+	xfarray_sort_bump_heapsorts(si);
+	startp = si->page_kaddr + offset_in_page(lo_pos);
+	sort(startp, hi - lo + 1, si->array->obj_size, si->cmp_fn, NULL);
+
+	xfarray_sort_bump_stores(si);
+	return xfarray_sort_put_page(si);
+}
+
 /* Return a pointer to the xfarray pivot record within the sortinfo struct. */
 static inline void *xfarray_sortinfo_pivot(struct xfarray_sortinfo *si)
 {
@@ -699,6 +780,10 @@  xfarray_qsort_push(
  * 4. For small sets, load the records into the scratchpad and run heapsort on
  *    them because that is very fast.  In the author's experience, this yields
  *    a ~10% reduction in runtime.
+ *
+ *    If a small set is contained entirely within a single xfile memory page,
+ *    map the page directly and run heap sort directly on the xfile page
+ *    instead of using the load/store interface.  This halves the runtime.
  */
 
 /*
@@ -744,6 +829,18 @@  xfarray_sort(
 			continue;
 		}
 
+		/*
+		 * If directly mapping the page and sorting can solve our
+		 * problems, we're done.
+		 */
+		if (xfarray_want_pagesort(si, lo, hi)) {
+			error = xfarray_pagesort(si, lo, hi);
+			if (error)
+				goto out_free;
+			si->stack_depth--;
+			continue;
+		}
+
 		/* If insertion sort can solve our problems, we're done. */
 		if (xfarray_want_isort(si, lo, hi)) {
 			error = xfarray_isort(si, lo, hi);
diff --git a/fs/xfs/scrub/xfarray.h b/fs/xfs/scrub/xfarray.h
index 3661c98272cd5..091614e7f6836 100644
--- a/fs/xfs/scrub/xfarray.h
+++ b/fs/xfs/scrub/xfarray.h
@@ -80,6 +80,10 @@  struct xfarray_sortinfo {
 	/* XFARRAY_SORT_* flags; see below. */
 	unsigned int		flags;
 
+	/* Cache a page here for faster access. */
+	struct xfile_page	xfpage;
+	void			*page_kaddr;
+
 #ifdef DEBUG
 	/* Performance statistics. */
 	uint64_t		loads;