xfs: speed up xfarray sort by sorting xfile page contents directly

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>

3 files changed, 121 insertions, 0 deletions

diff --git a/fs/xfs/scrub/trace.h b/fs/xfs/scrub/trace.h
index c2e611e2af49..5100745086c8 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 2a0599f660d7..457e56eac5e1 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 3661c98272cd..091614e7f683 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;