[Devel] [PATCH VZ9] fs/fuse: hashed write buckets

Alexey Kuznetsov kuznet at virtuozzo.com
Mon Feb 5 20:34:44 MSK 2024


Previous write record was 3.5G/sec and this ceiling
could not be penetrated even though eventloop had lots
of spare cpu. The bottlneck is diagnosed as saturation
of thread copying requests from kernel. So, we have
to switch to spreading it over multiple threads,
similar to scheme used for reads. So, for writes we
introduce two-level table, keyed by request size
and by inode hash.

New record is 4.8G/sec. Further progress is not going to be
easy, now we simultaneously:
- saturate eventloop
- saturate ec_offload thread, increasing offload threads does
  not help, we get 2 threads each 50% busy
- we are near to 100Gib ethernet limit, 9.6G/sec of network load

https://pmc.acronis.work/browse/VSTOR-79527

Signed-off-by: Alexey Kuznetsov <kuznet at acronis.com>
---
 fs/fuse/file.c   | 4 ++--
 fs/fuse/fuse_i.h | 1 +
 fs/fuse/inode.c  | 9 +++++++--
 3 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/fs/fuse/file.c b/fs/fuse/file.c
index 071225151a5f..b5abc7b260b6 100644
--- a/fs/fuse/file.c
+++ b/fs/fuse/file.c
@@ -791,8 +791,8 @@ struct fuse_iqueue *fuse_route_io(struct fuse_conn *fc, struct fuse_rtable *rt,
 		if (iosize == 0)
 			return NULL;
 
-		for (i = 0; i < rt->rt_size; i++) {
-			fiq = rt->iqs_table + i;
+		for (i = 0; i < rt->rt_size; i += rt->divisor) {
+			fiq = rt->iqs_table + i + (jhash_1word((u32)inode->i_ino, 0) % rt->divisor);
 			if (iosize <= fiq->size && fiq->handled_by_fud)
 				return fiq;
 		}
diff --git a/fs/fuse/fuse_i.h b/fs/fuse/fuse_i.h
index c8bc1f75f4a7..9a5a52afaaa6 100644
--- a/fs/fuse/fuse_i.h
+++ b/fs/fuse/fuse_i.h
@@ -673,6 +673,7 @@ static inline unsigned int fuse_qhash_bucket(void)
 struct fuse_rtable {
 	int	type;
 	int	rt_size;
+	int	divisor;
 	union {
 		void				*iqs;
 		struct fuse_iqueue __percpu	*iqs_cpu;
diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
index d5c3f7ef7da7..53ee80addbe2 100644
--- a/fs/fuse/inode.c
+++ b/fs/fuse/inode.c
@@ -1005,6 +1005,8 @@ static int alloc_rt_table(struct fuse_dev *fud, struct fuse_rtable *rt,
 	int res = -EINVAL;
 	int idx;
 
+	rt->divisor = 1;
+
 	switch (req->type) {
 	case FUSE_ROUTING_CPU:
 		if (req->index >= NR_CPUS || !cpu_possible(req->index))
@@ -1020,7 +1022,9 @@ static int alloc_rt_table(struct fuse_dev *fud, struct fuse_rtable *rt,
 		res = 0;
 		break;
 	case FUSE_ROUTING_SIZE:
-		if (req->key > FUSE_MAX_MAX_PAGES*PAGE_SIZE || (req->key % PAGE_SIZE))
+		rt->divisor = 1 + (req->key & (PAGE_SIZE - 1));
+		if (rt->divisor > req->table_size ||
+		   (req->key & ~(PAGE_SIZE - 1)) > FUSE_MAX_MAX_PAGES*PAGE_SIZE)
 			break;
 		fallthrough;
 	case FUSE_ROUTING_HASH:
@@ -1044,7 +1048,7 @@ static int alloc_rt_table(struct fuse_dev *fud, struct fuse_rtable *rt,
 static void adjust_rt_table(struct fuse_dev *fud, struct fuse_iqueue *fiq,
 			    struct fuse_iq_routing *req)
 {
-	u32 size = req->key;
+	u32 size = req->key & ~(PAGE_SIZE - 1);
 
 	fiq->size = size;
 
@@ -1090,6 +1094,7 @@ int fuse_install_iq_route(struct fuse_dev *fud, struct fuse_iq_routing *req)
 
 	if (rt->iqs == NULL) {
 		rt->iqs = rtl.iqs;
+		rt->divisor = rtl.divisor;
 		rt->type = rtl.type;
 		rt->rt_size = rtl.rt_size;
 	} else if (rt->iqs != rtl.iqs) {
-- 
2.39.3



More information about the Devel mailing list