[PATCH 1/3] erofs: avoid redundant page copy for aligned shifted/plain pclusters

[Sarthak Kukreti]

When a Z_EROFS_COMPRESSION_SHIFTED pcluster is read in-place into its
target page cache folio(s) with page-aligned input and output offsets
(pageofs_in == 0 && pageofs_out == 0), z_erofs_transform_plain() still
maps each page via kmap_local_page() and calls memmove(kin + po,
kin + rq->pageofs_in + pi, cnt) where dst == src. On x86_64,
arch/x86/lib/memmove_64.S does not early-exit when dst == src, executing
rep movsb over the entire 4 KiB page and dirtying all 64 cachelines.

Furthermore, z_erofs_attach_page() restricts uncompressed pclusters to
a single inplace I/O page (fe->icur <= 1), forcing multi-page aligned
Z_EROFS_COMPRESSION_SHIFTED extents to allocate temporary shortlived
pages and copy every page beyond the first via memcpy_to_page().

1. Allow multi-page inplace I/O in z_erofs_attach_page() when the
   pcluster uses Z_EROFS_COMPRESSION_SHIFTED with page-aligned input
   and output offsets (!pcl->pageofs_in && !pcl->pageofs_out) and full
   page bvecs.
2. In z_erofs_transform_plain(), return immediately when a page-aligned
   Z_EROFS_COMPRESSION_SHIFTED request has identical input and output
   page pointers (rq->in[ni] == rq->out[ni]), and skip memmove() when
   po == rq->pageofs_in + pi.

Signed-off-by: Sarthak Kukreti <sarthakkukreti@xxxxxxxxxx>
---
 fs/erofs/decompressor.c | 24 ++++++++++++++++++++----
 fs/erofs/zdata.c        | 14 ++++++++++++--
 2 files changed, 32 insertions(+), 6 deletions(-)

diff --git a/fs/erofs/decompressor.c b/fs/erofs/decompressor.c
index d387b27c4ee2..0f6517f329ee 100644
--- a/fs/erofs/decompressor.c
+++ b/fs/erofs/decompressor.c
@@ -283,6 +283,21 @@ static const char *z_erofs_transform_plain(struct z_erofs_decompress_req *rq,
 		rq->outputsize -= cur;
 	}
 
+	if (rq->alg == Z_EROFS_COMPRESSION_SHIFTED && !rq->pageofs_in &&
+	    !rq->pageofs_out && nrpages_in == nrpages_out) {
+		bool same = true;
+
+		for (ni = 0; ni < nrpages_in; ++ni) {
+			if (!rq->in[ni] || rq->in[ni] != rq->out[ni]) {
+				same = false;
+				break;
+			}
+		}
+		if (same)
+			return NULL;
+		ni = 0;
+	}
+
 	for (; rq->outputsize; rq->pageofs_in = 0, cur += insz, ni++) {
 		insz = min(PAGE_SIZE - rq->pageofs_in, rq->outputsize);
 		rq->outputsize -= insz;
@@ -295,10 +310,11 @@ static const char *z_erofs_transform_plain(struct z_erofs_decompress_req *rq,
 			po = (rq->pageofs_out + cur + pi) & ~PAGE_MASK;
 			DBG_BUGON(no >= nrpages_out);
 			cnt = min(insz - pi, PAGE_SIZE - po);
-			if (rq->out[no] == rq->in[ni])
-				memmove(kin + po,
-					kin + rq->pageofs_in + pi, cnt);
-			else if (rq->out[no])
+			if (rq->out[no] == rq->in[ni]) {
+				if (po != rq->pageofs_in + pi)
+					memmove(kin + po,
+						kin + rq->pageofs_in + pi, cnt);
+			} else if (rq->out[no])
 				memcpy_to_page(rq->out[no], po,
 					       kin + rq->pageofs_in + pi, cnt);
 			pi += cnt;
diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
index 6b07e73ee2aa..0517caf61959 100644
--- a/fs/erofs/zdata.c
+++ b/fs/erofs/zdata.c
@@ -698,9 +698,19 @@ static int z_erofs_attach_page(struct z_erofs_frontend *fe,
 	int ret;
 
 	if (exclusive) {
-		/* Inplace I/O is limited to one page for uncompressed data */
+		/*
+		 * Allow multi-page inplace I/O for page-aligned SHIFTED
+		 * pclusters so that uncompressed multi-block extents read
+		 * directly into the target page cache folios without allocating
+		 * intermediate shortlived pages or copying in
+		 * z_erofs_transform_plain().
+		 */
 		if (pcl->algorithmformat < Z_EROFS_COMPRESSION_MAX ||
-		    fe->icur <= 1) {
+		    fe->icur <= 1 ||
+		    (pcl->algorithmformat == Z_EROFS_COMPRESSION_SHIFTED &&
+		     !pcl->pageofs_in && !pcl->pageofs_out &&
+		     !(bvec->offset & ~PAGE_MASK) &&
+		     bvec->end == PAGE_SIZE)) {
 			/* Try to prioritize inplace I/O here */
 			spin_lock(&pcl->lockref.lock);
 			while (fe->icur > 0) {
-- 
2.56.0.rc1.315.gc6ed9934b7-goog