Skip to contentSkip to conversation

xprtrdma: Use xprt_pin_rqst in rpcrdma_reply_handler

PATCH In review Chuck Lever 4 participants 8 messages Aug 23, 2017 – Sep 6, 2017

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xprtrdma: Use xprt_pin_rqst in rpcrdma_reply_handler

Adopt the use of xprt_pin_rqst to eliminate contention between Call-side users of rb_lock and the use of rb_lock in rpcrdma_reply_handler.

This replaces the mechanism introduced in 431af645cf66 ("xprtrdma:
Fix client lock-up after application signal fires").

Use recv_lock to quickly find the completing rqst, pin it, then drop the lock. At that point invalidation and pull-up of the Reply XDR can be done. Both are often expensive operations.

Finally, take recv_lock again to signal completion to the RPC layer. It also protects adjustment of "cwnd".

This greatly reduces the amount of time a lock is held by the reply handler. Comparing lock_stat results shows a marked decrease in contention on rb_lock and recv_lock.

---
Hi-

If Trond's lock contention series is going into v4.14, I'd like you to consider this one (applied at the end of that series) as well. Without it, NFS/RDMA performance regresses a bit after the first xprt_pin_rqst patch is applied. Thanks!

 net/sunrpc/xprt.c               |    2 +
 net/sunrpc/xprtrdma/rpc_rdma.c  |   57 ++++++++++-----------------------------
 net/sunrpc/xprtrdma/transport.c |    1 -
 net/sunrpc/xprtrdma/verbs.c     |    1 -
 net/sunrpc/xprtrdma/xprt_rdma.h |   30 ---------------------
 5 files changed, 16 insertions(+), 75 deletions(-)
net/sunrpc/xprt.c +2 -0
index 2af189c..e741ec2 100644
@@ -855,6 +855,7 @@ void xprt_pin_rqst(struct rpc_rqst *req)
 {
 	set_bit(RPC_TASK_MSG_RECV, &req->rq_task->tk_runstate);
 }
+EXPORT_SYMBOL_GPL(xprt_pin_rqst);
 
 /**
  * xprt_unpin_rqst - Unpin a request on the transport receive list
@@ -870,6 +871,7 @@ void xprt_unpin_rqst(struct rpc_rqst *req)
 	if (test_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate))
 		wake_up_bit(&task->tk_runstate, RPC_TASK_MSG_RECV);
 }
+EXPORT_SYMBOL_GPL(xprt_unpin_rqst);
 
 static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req)
 __must_hold(&req->rq_xprt->recv_lock)
net/sunrpc/xprtrdma/rpc_rdma.c +14 -43

107 lines not on this page · raw message

net/sunrpc/xprtrdma/transport.c +0 -1
index d1c458e..1cb1a07 100644
@@ -684,7 +684,6 @@
 
 	dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 
-	rpcrdma_remove_req(&r_xprt->rx_buf, req);
 	if (!list_empty(&req->rl_registered))
 		ia->ri_ops->ro_unmap_safe(r_xprt, req, !RPC_IS_ASYNC(task));
 	rpcrdma_unmap_sges(ia, req);
net/sunrpc/xprtrdma/verbs.c +0 -1
index e4171f2..23ec6ed 100644
@@ -1004,7 +1004,6 @@ struct rpcrdma_rep *
 	spin_lock_init(&buf->rb_recovery_lock);
 	INIT_LIST_HEAD(&buf->rb_mws);
 	INIT_LIST_HEAD(&buf->rb_all);
-	INIT_LIST_HEAD(&buf->rb_pending);
 	INIT_LIST_HEAD(&buf->rb_stale_mrs);
 	INIT_DELAYED_WORK(&buf->rb_refresh_worker,
 			  rpcrdma_mr_refresh_worker);
net/sunrpc/xprtrdma/xprt_rdma.h +0 -30
index b282d3f..ad918c8 100644
@@ -341,7 +341,6 @@ enum {
 struct rpcrdma_buffer;
 struct rpcrdma_req {
 	struct list_head	rl_list;
-	__be32			rl_xid;
 	unsigned int		rl_mapped_sges;
 	unsigned int		rl_connect_cookie;
 	struct rpcrdma_buffer	*rl_buffer;
@@ -403,7 +402,6 @@ struct rpcrdma_buffer {
 	int			rb_send_count, rb_recv_count;
 	struct list_head	rb_send_bufs;
 	struct list_head	rb_recv_bufs;
-	struct list_head	rb_pending;
 	u32			rb_max_requests;
 	atomic_t		rb_credits;	/* most recent credit grant */
 
@@ -552,34 +550,6 @@ int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
 
-static inline void
-rpcrdma_insert_req(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
-{
-	spin_lock(&buffers->rb_lock);
-	if (list_empty(&req->rl_list))
-		list_add_tail(&req->rl_list, &buffers->rb_pending);
-	spin_unlock(&buffers->rb_lock);
-}
-
-static inline struct rpcrdma_req *
-rpcrdma_lookup_req_locked(struct rpcrdma_buffer *buffers, __be32 xid)
-{
-	struct rpcrdma_req *pos;
-
-	list_for_each_entry(pos, &buffers->rb_pending, rl_list)
-		if (pos->rl_xid == xid)
-			return pos;
-	return NULL;
-}
-
-static inline void
-rpcrdma_remove_req(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
-{
-	spin_lock(&buffers->rb_lock);
-	list_del(&req->rl_list);
-	spin_unlock(&buffers->rb_lock);
-}
-
 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
#1 lore raw

T24gV2VkLCAyMDE3LTA4LTIzIGF0IDE3OjA1IC0wNDAwLCBDaHVjayBMZXZlciB3cm90ZToNCj4g QWRvcHQgdGhlIHVzZSBvZiB4cHJ0X3Bpbl9ycXN0IHRvIGVsaW1pbmF0ZSBjb250ZW50aW9uIGJl dHdlZW4NCj4gQ2FsbC1zaWRlIHVzZXJzIG9mIHJiX2xvY2sgYW5kIHRoZSB1c2Ugb2YgcmJfbG9j ayBpbg0KPiBycGNyZG1hX3JlcGx5X2hhbmRsZXIuDQo+IA0KPiBUaGlzIHJlcGxhY2VzIHRoZSBt ZWNoYW5pc20gaW50cm9kdWNlZCBpbiA0MzFhZjY0NWNmNjYgKCJ4cHJ0cmRtYToNCj4gRml4IGNs aWVudCBsb2NrLXVwIGFmdGVyIGFwcGxpY2F0aW9uIHNpZ25hbCBmaXJlcyIpLg0KPiANCj4gVXNl IHJlY3ZfbG9jayB0byBxdWlja2x5IGZpbmQgdGhlIGNvbXBsZXRpbmcgcnFzdCwgcGluIGl0LCB0 aGVuDQo+IGRyb3AgdGhlIGxvY2suIEF0IHRoYXQgcG9pbnQgaW52YWxpZGF0aW9uIGFuZCBwdWxs LXVwIG9mIHRoZSBSZXBseQ0KPiBYRFIgY2FuIGJlIGRvbmUuIEJvdGggYXJlIG9mdGVuIGV4cGVu c2l2ZSBvcGVyYXRpb25zLg0KPiANCj4gRmluYWxseSwgdGFrZSByZWN2X2xvY2sgYWdhaW4gdG8g c2lnbmFsIGNvbXBsZXRpb24gdG8gdGhlIFJQQw0KPiBsYXllci4gSXQgYWxzbyBwcm90ZWN0cyBh ZGp1c3RtZW50IG9mICJjd25kIi4NCj4gDQo+IFRoaXMgZ3JlYXRseSByZWR1Y2VzIHRoZSBhbW91 bnQgb2YgdGltZSBhIGxvY2sgaXMgaGVsZCBieSB0aGUNCj4gcmVwbHkgaGFuZGxlci4gQ29tcGFy aW5nIGxvY2tfc3RhdCByZXN1bHRzIHNob3dzIGEgbWFya2VkIGRlY3JlYXNlDQo+IGluIGNvbnRl bnRpb24gb24gcmJfbG9jayBhbmQgcmVjdl9sb2NrLg0KPiANCj4gU2lnbmVkLW9mZi1ieTogQ2h1 Y2sgTGV2ZXIgPGNodWNrLmxldmVyQG9yYWNsZS5jb20+DQo+IC0tLQ0KPiBIaS0NCj4gDQo+IElm IFRyb25kJ3MgbG9jayBjb250ZW50aW9uIHNlcmllcyBpcyBnb2luZyBpbnRvIHY0LjE0LCBJJ2Qg bGlrZSB5b3UNCj4gdG8gY29uc2lkZXIgdGhpcyBvbmUgKGFwcGxpZWQgYXQgdGhlIGVuZCBvZiB0 aGF0IHNlcmllcykgYXMgd2VsbC4NCj4gV2l0aG91dCBpdCwgTkZTL1JETUEgcGVyZm9ybWFuY2Ug cmVncmVzc2VzIGEgYml0IGFmdGVyIHRoZQ0KPiBmaXJzdCB4cHJ0X3Bpbl9ycXN0IHBhdGNoIGlz IGFwcGxpZWQuIFRoYW5rcyENCj4gDQoNCklmIEFubmEgaXMgT0sgd2l0aCBpdCwgdGhlbiBJIGNh biBhcHBseSB0aGlzIHBhdGNoIG9uY2Ugc2hlJ3Mgc2VudCBtZQ0KdGhlIHB1bGwgcmVxdWVzdCBm b3IgdGhlIG90aGVyIFJETUEgY2xpZW50IHBhdGNoZXMuDQoNCi0tIA0KVHJvbmQgTXlrbGVidXN0 DQpMaW51eCBORlMgY2xpZW50IG1haW50YWluZXIsIFByaW1hcnlEYXRhDQp0cm9uZC5teWtsZWJ1 c3RAcHJpbWFyeWRhdGEuY29tDQo=

#2 lore raw
Quoted 23 lines · Trond Myklebust
Quoted 22 lines · Chuck Lever

Adopt the use of xprt_pin_rqst to eliminate contention between Call-side users of rb_lock and the use of rb_lock in rpcrdma_reply_handler.

This replaces the mechanism introduced in 431af645cf66 ("xprtrdma:
Fix client lock-up after application signal fires").

Use recv_lock to quickly find the completing rqst, pin it, then drop the lock. At that point invalidation and pull-up of the Reply XDR can be done. Both are often expensive operations.

Finally, take recv_lock again to signal completion to the RPC layer. It also protects adjustment of "cwnd".

This greatly reduces the amount of time a lock is held by the reply handler. Comparing lock_stat results shows a marked decrease in contention on rb_lock and recv_lock.

---
Hi-

Quoted 8 lines
Quoted 5 lines

If Trond's lock contention series is going into v4.14, I'd like you to consider this one (applied at the end of that series) as well. Without it, NFS/RDMA performance regresses a bit after the first xprt_pin_rqst patch is applied. Thanks!

If Anna is OK with it, then I can apply this patch once she's sent me the pull request for the other RDMA client patches.

Works for me. Thanks!

Quoted 1 line
#3 lore raw

T24gVGh1LCAyMDE3LTA4LTI0IGF0IDE0OjE4IC0wNDAwLCBBbm5hIFNjaHVtYWtlciB3cm90ZToN Cj4gDQo+IE9uIDA4LzIzLzIwMTcgMDU6NDAgUE0sIFRyb25kIE15a2xlYnVzdCB3cm90ZToNCj4g PiBPbiBXZWQsIDIwMTctMDgtMjMgYXQgMTc6MDUgLTA0MDAsIENodWNrIExldmVyIHdyb3RlOg0K PiA+ID4gQWRvcHQgdGhlIHVzZSBvZiB4cHJ0X3Bpbl9ycXN0IHRvIGVsaW1pbmF0ZSBjb250ZW50 aW9uIGJldHdlZW4NCj4gPiA+IENhbGwtc2lkZSB1c2VycyBvZiByYl9sb2NrIGFuZCB0aGUgdXNl IG9mIHJiX2xvY2sgaW4NCj4gPiA+IHJwY3JkbWFfcmVwbHlfaGFuZGxlci4NCj4gPiA+IA0KPiA+ ID4gVGhpcyByZXBsYWNlcyB0aGUgbWVjaGFuaXNtIGludHJvZHVjZWQgaW4gNDMxYWY2NDVjZjY2 DQo+ID4gPiAoInhwcnRyZG1hOg0KPiA+ID4gRml4IGNsaWVudCBsb2NrLXVwIGFmdGVyIGFwcGxp Y2F0aW9uIHNpZ25hbCBmaXJlcyIpLg0KPiA+ID4gDQo+ID4gPiBVc2UgcmVjdl9sb2NrIHRvIHF1 aWNrbHkgZmluZCB0aGUgY29tcGxldGluZyBycXN0LCBwaW4gaXQsIHRoZW4NCj4gPiA+IGRyb3Ag dGhlIGxvY2suIEF0IHRoYXQgcG9pbnQgaW52YWxpZGF0aW9uIGFuZCBwdWxsLXVwIG9mIHRoZQ0K PiA+ID4gUmVwbHkNCj4gPiA+IFhEUiBjYW4gYmUgZG9uZS4gQm90aCBhcmUgb2Z0ZW4gZXhwZW5z aXZlIG9wZXJhdGlvbnMuDQo+ID4gPiANCj4gPiA+IEZpbmFsbHksIHRha2UgcmVjdl9sb2NrIGFn YWluIHRvIHNpZ25hbCBjb21wbGV0aW9uIHRvIHRoZSBSUEMNCj4gPiA+IGxheWVyLiBJdCBhbHNv IHByb3RlY3RzIGFkanVzdG1lbnQgb2YgImN3bmQiLg0KPiA+ID4gDQo+ID4gPiBUaGlzIGdyZWF0 bHkgcmVkdWNlcyB0aGUgYW1vdW50IG9mIHRpbWUgYSBsb2NrIGlzIGhlbGQgYnkgdGhlDQo+ID4g PiByZXBseSBoYW5kbGVyLiBDb21wYXJpbmcgbG9ja19zdGF0IHJlc3VsdHMgc2hvd3MgYSBtYXJr ZWQNCj4gPiA+IGRlY3JlYXNlDQo+ID4gPiBpbiBjb250ZW50aW9uIG9uIHJiX2xvY2sgYW5kIHJl Y3ZfbG9jay4NCj4gPiA+IA0KPiA+ID4gU2lnbmVkLW9mZi1ieTogQ2h1Y2sgTGV2ZXIgPGNodWNr LmxldmVyQG9yYWNsZS5jb20+DQo+ID4gPiAtLS0NCj4gPiA+IEhpLQ0KPiA+ID4gDQo+ID4gPiBJ ZiBUcm9uZCdzIGxvY2sgY29udGVudGlvbiBzZXJpZXMgaXMgZ29pbmcgaW50byB2NC4xNCwgSSdk IGxpa2UNCj4gPiA+IHlvdQ0KPiA+ID4gdG8gY29uc2lkZXIgdGhpcyBvbmUgKGFwcGxpZWQgYXQg dGhlIGVuZCBvZiB0aGF0IHNlcmllcykgYXMgd2VsbC4NCj4gPiA+IFdpdGhvdXQgaXQsIE5GUy9S RE1BIHBlcmZvcm1hbmNlIHJlZ3Jlc3NlcyBhIGJpdCBhZnRlciB0aGUNCj4gPiA+IGZpcnN0IHhw cnRfcGluX3Jxc3QgcGF0Y2ggaXMgYXBwbGllZC4gVGhhbmtzIQ0KPiA+ID4gDQo+ID4gDQo+ID4g SWYgQW5uYSBpcyBPSyB3aXRoIGl0LCB0aGVuIEkgY2FuIGFwcGx5IHRoaXMgcGF0Y2ggb25jZSBz aGUncyBzZW50DQo+ID4gbWUNCj4gPiB0aGUgcHVsbCByZXF1ZXN0IGZvciB0aGUgb3RoZXIgUkRN QSBjbGllbnQgcGF0Y2hlcy4NCj4gDQo+IFdvcmtzIGZvciBtZS4gIFRoYW5rcyENCg0KUGxlYXNl IGNhbiB5b3UgYm90aCBjaGVjayBodHRwOi8vZ2l0LmxpbnV4LW5mcy5vcmcvP3A9dHJvbmRteS9s aW51eC1uZnMNCi5naXQ7YT1jb21taXQ7aD1iMWRhNWQ5MDg1N2VlNGI4ZjVhY2VlMTE0MzcwNTQx MmIxNmZjZTU2IGFuZCBzZWUgaWYgYWxsDQppcyB3ZWxsPw0KDQpEbyBub3RlIHRoYXQgSSBkaWQg cmVtb3ZlIHRoZSBjYWxsIHRvIHJwY3JkbWFfYnVmZmVyX3B1dCgpIHdoaWNoIHdhcw0KYmVpbmcg Y2FsbGVkIHdpdGggYW4gdW5pbml0aWFsaXNlZCBhcmd1bWVudC4NCg0KQ2hlZXJzDQogIFRyb25k DQoNCi0tIA0KVHJvbmQgTXlrbGVidXN0DQpMaW51eCBORlMgY2xpZW50IG1haW50YWluZXIsIFBy aW1hcnlEYXRhDQp0cm9uZC5teWtsZWJ1c3RAcHJpbWFyeWRhdGEuY29tDQo=

#4 lore raw
Quoted 40 lines · Trond Myklebust
Quoted 39 lines · Anna Schumaker
Quoted 36 lines · Trond Myklebust
Quoted 22 lines · Chuck Lever

Adopt the use of xprt_pin_rqst to eliminate contention between Call-side users of rb_lock and the use of rb_lock in rpcrdma_reply_handler.

This replaces the mechanism introduced in 431af645cf66 ("xprtrdma:
Fix client lock-up after application signal fires").

Use recv_lock to quickly find the completing rqst, pin it, then drop the lock. At that point invalidation and pull-up of the Reply
XDR can be done. Both are often expensive operations.

Finally, take recv_lock again to signal completion to the RPC layer. It also protects adjustment of "cwnd".

This greatly reduces the amount of time a lock is held by the reply handler. Comparing lock_stat results shows a marked decrease
in contention on rb_lock and recv_lock.

Quoted 9 lines

---
Hi-

If Trond's lock contention series is going into v4.14, I'd like you
to consider this one (applied at the end of that series) as well. Without it, NFS/RDMA performance regresses a bit after the first xprt_pin_rqst patch is applied. Thanks!

If Anna is OK with it, then I can apply this patch once she's sent me
the pull request for the other RDMA client patches.

Quoted 5 lines
Quoted 1 line

Works for me. Thanks!

Please can you both check http://git.linux-nfs.org/?p=trondmy/linux-nfs
.git;a=commit;h=b1da5d90857ee4b8f5acee1143705412b16fce56 and see if all is well?

Looks okay to me.

Anna

Quoted 7 lines

Do note that I did remove the call to rpcrdma_buffer_put() which was being called with an uninitialised argument.

Cheers
Trond

#5 lore raw
In reply to Trond Myklebust
Quoted 43 lines · Trond Myklebust
Quoted 39 lines · Anna Schumaker
Quoted 36 lines · Trond Myklebust
Quoted 22 lines · Chuck Lever

Adopt the use of xprt_pin_rqst to eliminate contention between Call-side users of rb_lock and the use of rb_lock in rpcrdma_reply_handler.

This replaces the mechanism introduced in 431af645cf66 ("xprtrdma:
Fix client lock-up after application signal fires").

Use recv_lock to quickly find the completing rqst, pin it, then drop the lock. At that point invalidation and pull-up of the Reply
XDR can be done. Both are often expensive operations.

Finally, take recv_lock again to signal completion to the RPC layer. It also protects adjustment of "cwnd".

This greatly reduces the amount of time a lock is held by the reply handler. Comparing lock_stat results shows a marked decrease
in contention on rb_lock and recv_lock.

Quoted 9 lines

---
Hi-

If Trond's lock contention series is going into v4.14, I'd like you
to consider this one (applied at the end of that series) as well. Without it, NFS/RDMA performance regresses a bit after the first xprt_pin_rqst patch is applied. Thanks!

If Anna is OK with it, then I can apply this patch once she's sent me
the pull request for the other RDMA client patches.

Quoted 8 lines
Quoted 1 line

Works for me. Thanks!

Please can you both check http://git.linux-nfs.org/?p=trondmy/linux-nfs
.git;a=commit;h=b1da5d90857ee4b8f5acee1143705412b16fce56 and see if all is well?

Do note that I did remove the call to rpcrdma_buffer_put() which was being called with an uninitialised argument.

Looking at that again. rpcrdma_buffer_put() does look unnecessary, since out_norqst is being called now before "req" is initialized.

On further inspection, "return;" should be replaced with "goto repost;". This is similar to "out_nomatch" and "out_duplicate", which are both being replaced in this patch.

So, out_norqst should look something like this:

out_norqst:

        spin_unlock(&xprt->recv_lock);
        dprintk("RPC:       %s: no match for incoming xid 0x%08x\n",
                __func__, be32_to_cpu(xid));
        goto repost;

Otherwise, I don't see any obvious problems.

--
Chuck Lever

#6 lore raw

T24gVHVlLCAyMDE3LTA5LTA1IGF0IDE3OjA2IC0wNDAwLCBDaHVjayBMZXZlciB3cm90ZToNCj4g PiBPbiBTZXAgNSwgMjAxNywgYXQgNDozMyBQTSwgVHJvbmQgTXlrbGVidXN0IDx0cm9uZG15QHBy aW1hcnlkYXRhLmNvDQo+ID4gbT4gd3JvdGU6DQo+ID4gDQo+ID4gT24gVGh1LCAyMDE3LTA4LTI0 IGF0IDE0OjE4IC0wNDAwLCBBbm5hIFNjaHVtYWtlciB3cm90ZToNCj4gPiA+IA0KPiA+ID4gT24g MDgvMjMvMjAxNyAwNTo0MCBQTSwgVHJvbmQgTXlrbGVidXN0IHdyb3RlOg0KPiA+ID4gPiBPbiBX ZWQsIDIwMTctMDgtMjMgYXQgMTc6MDUgLTA0MDAsIENodWNrIExldmVyIHdyb3RlOg0KPiA+ID4g PiA+IEFkb3B0IHRoZSB1c2Ugb2YgeHBydF9waW5fcnFzdCB0byBlbGltaW5hdGUgY29udGVudGlv bg0KPiA+ID4gPiA+IGJldHdlZW4NCj4gPiA+ID4gPiBDYWxsLXNpZGUgdXNlcnMgb2YgcmJfbG9j ayBhbmQgdGhlIHVzZSBvZiByYl9sb2NrIGluDQo+ID4gPiA+ID4gcnBjcmRtYV9yZXBseV9oYW5k bGVyLg0KPiA+ID4gPiA+IA0KPiA+ID4gPiA+IFRoaXMgcmVwbGFjZXMgdGhlIG1lY2hhbmlzbSBp bnRyb2R1Y2VkIGluIDQzMWFmNjQ1Y2Y2Ng0KPiA+ID4gPiA+ICgieHBydHJkbWE6DQo+ID4gPiA+ ID4gRml4IGNsaWVudCBsb2NrLXVwIGFmdGVyIGFwcGxpY2F0aW9uIHNpZ25hbCBmaXJlcyIpLg0K PiA+ID4gPiA+IA0KPiA+ID4gPiA+IFVzZSByZWN2X2xvY2sgdG8gcXVpY2tseSBmaW5kIHRoZSBj b21wbGV0aW5nIHJxc3QsIHBpbiBpdCwNCj4gPiA+ID4gPiB0aGVuDQo+ID4gPiA+ID4gZHJvcCB0 aGUgbG9jay4gQXQgdGhhdCBwb2ludCBpbnZhbGlkYXRpb24gYW5kIHB1bGwtdXAgb2YgdGhlDQo+ ID4gPiA+ID4gUmVwbHkNCj4gPiA+ID4gPiBYRFIgY2FuIGJlIGRvbmUuIEJvdGggYXJlIG9mdGVu IGV4cGVuc2l2ZSBvcGVyYXRpb25zLg0KPiA+ID4gPiA+IA0KPiA+ID4gPiA+IEZpbmFsbHksIHRh a2UgcmVjdl9sb2NrIGFnYWluIHRvIHNpZ25hbCBjb21wbGV0aW9uIHRvIHRoZSBSUEMNCj4gPiA+ ID4gPiBsYXllci4gSXQgYWxzbyBwcm90ZWN0cyBhZGp1c3RtZW50IG9mICJjd25kIi4NCj4gPiA+ ID4gPiANCj4gPiA+ID4gPiBUaGlzIGdyZWF0bHkgcmVkdWNlcyB0aGUgYW1vdW50IG9mIHRpbWUg YSBsb2NrIGlzIGhlbGQgYnkgdGhlDQo+ID4gPiA+ID4gcmVwbHkgaGFuZGxlci4gQ29tcGFyaW5n IGxvY2tfc3RhdCByZXN1bHRzIHNob3dzIGEgbWFya2VkDQo+ID4gPiA+ID4gZGVjcmVhc2UNCj4g PiA+ID4gPiBpbiBjb250ZW50aW9uIG9uIHJiX2xvY2sgYW5kIHJlY3ZfbG9jay4NCj4gPiA+ID4g PiANCj4gPiA+ID4gPiBTaWduZWQtb2ZmLWJ5OiBDaHVjayBMZXZlciA8Y2h1Y2subGV2ZXJAb3Jh Y2xlLmNvbT4NCj4gPiA+ID4gPiAtLS0NCj4gPiA+ID4gPiBIaS0NCj4gPiA+ID4gPiANCj4gPiA+ ID4gPiBJZiBUcm9uZCdzIGxvY2sgY29udGVudGlvbiBzZXJpZXMgaXMgZ29pbmcgaW50byB2NC4x NCwgSSdkDQo+ID4gPiA+ID4gbGlrZQ0KPiA+ID4gPiA+IHlvdQ0KPiA+ID4gPiA+IHRvIGNvbnNp ZGVyIHRoaXMgb25lIChhcHBsaWVkIGF0IHRoZSBlbmQgb2YgdGhhdCBzZXJpZXMpIGFzDQo+ID4g PiA+ID4gd2VsbC4NCj4gPiA+ID4gPiBXaXRob3V0IGl0LCBORlMvUkRNQSBwZXJmb3JtYW5jZSBy ZWdyZXNzZXMgYSBiaXQgYWZ0ZXIgdGhlDQo+ID4gPiA+ID4gZmlyc3QgeHBydF9waW5fcnFzdCBw YXRjaCBpcyBhcHBsaWVkLiBUaGFua3MhDQo+ID4gPiA+ID4gDQo+ID4gPiA+IA0KPiA+ID4gPiBJ ZiBBbm5hIGlzIE9LIHdpdGggaXQsIHRoZW4gSSBjYW4gYXBwbHkgdGhpcyBwYXRjaCBvbmNlIHNo ZSdzDQo+ID4gPiA+IHNlbnQNCj4gPiA+ID4gbWUNCj4gPiA+ID4gdGhlIHB1bGwgcmVxdWVzdCBm b3IgdGhlIG90aGVyIFJETUEgY2xpZW50IHBhdGNoZXMuDQo+ID4gPiANCj4gPiA+IFdvcmtzIGZv ciBtZS4gIFRoYW5rcyENCj4gPiANCj4gPiBQbGVhc2UgY2FuIHlvdSBib3RoIGNoZWNrIGh0dHA6 Ly9naXQubGludXgtbmZzLm9yZy8/cD10cm9uZG15L2xpbnV4DQo+ID4gLW5mcw0KPiA+IC5naXQ7 YT1jb21taXQ7aD1iMWRhNWQ5MDg1N2VlNGI4ZjVhY2VlMTE0MzcwNTQxMmIxNmZjZTU2IGFuZCBz ZWUgaWYNCj4gPiBhbGwNCj4gPiBpcyB3ZWxsPw0KPiA+IA0KPiA+IERvIG5vdGUgdGhhdCBJIGRp ZCByZW1vdmUgdGhlIGNhbGwgdG8gcnBjcmRtYV9idWZmZXJfcHV0KCkgd2hpY2gNCj4gPiB3YXMN Cj4gPiBiZWluZyBjYWxsZWQgd2l0aCBhbiB1bmluaXRpYWxpc2VkIGFyZ3VtZW50Lg0KPiANCj4g TG9va2luZyBhdCB0aGF0IGFnYWluLiBycGNyZG1hX2J1ZmZlcl9wdXQoKSBkb2VzIGxvb2sgdW5u ZWNlc3NhcnksDQo+IHNpbmNlIG91dF9ub3Jxc3QgaXMgYmVpbmcgY2FsbGVkIG5vdyBiZWZvcmUg InJlcSIgaXMgaW5pdGlhbGl6ZWQuDQo+IA0KPiBPbiBmdXJ0aGVyIGluc3BlY3Rpb24sICJyZXR1 cm47IiBzaG91bGQgYmUgcmVwbGFjZWQgd2l0aA0KPiAiZ290byByZXBvc3Q7Ii4gVGhpcyBpcyBz aW1pbGFyIHRvICJvdXRfbm9tYXRjaCIgYW5kDQo+ICJvdXRfZHVwbGljYXRlIiwgd2hpY2ggYXJl IGJvdGggYmVpbmcgcmVwbGFjZWQgaW4gdGhpcyBwYXRjaC4NCj4gDQo+IFNvLCBvdXRfbm9ycXN0 IHNob3VsZCBsb29rIHNvbWV0aGluZyBsaWtlIHRoaXM6DQo+IA0KPiANCj4gb3V0X25vcnFzdDoN Cj4gICAgICAgICBzcGluX3VubG9jaygmeHBydC0+cmVjdl9sb2NrKTsNCj4gICAgICAgICBkcHJp bnRrKCJSUEM6ICAgICAgICVzOiBubyBtYXRjaCBmb3IgaW5jb21pbmcgeGlkIDB4JTA4eFxuIiwN Cj4gICAgICAgICAgICAgICAgIF9fZnVuY19fLCBiZTMyX3RvX2NwdSh4aWQpKTsNCj4gICAgICAg ICBnb3RvIHJlcG9zdDsNCj4gDQo+IA0KPiBPdGhlcndpc2UsIEkgZG9uJ3Qgc2VlIGFueSBvYnZp b3VzIHByb2JsZW1zLg0KPiANCg0KUGF0Y2ggdXBkYXRlZC4gUGxlYXNlIHNlZSBodHRwOi8vZ2l0 LmxpbnV4LW5mcy5vcmcvP3A9dHJvbmRteS9saW51eC1uZnMNCi5naXQ7YT1jb21taXQ7aD05NTkw ZDA4M2MxYmIxNDE5Yjc5OTI2MDlkMWEwYTNlMzUxN2QzODkzDQoNCi0tIA0KVHJvbmQgTXlrbGVi dXN0DQpMaW51eCBORlMgY2xpZW50IG1haW50YWluZXIsIFByaW1hcnlEYXRhDQp0cm9uZC5teWts ZWJ1c3RAcHJpbWFyeWRhdGEuY29tDQo=

#7 lore raw
Quoted 81 lines · Trond Myklebust
Quoted 77 lines · Chuck Lever
Quoted 56 lines

On Sep 5, 2017, at 4:33 PM, Trond Myklebust <[email protected] m> wrote:

Quoted 39 lines · Anna Schumaker
Quoted 37 lines · Trond Myklebust
Quoted 35 lines · Chuck Lever

Adopt the use of xprt_pin_rqst to eliminate contention between
Call-side users of rb_lock and the use of rb_lock in rpcrdma_reply_handler.

This replaces the mechanism introduced in 431af645cf66 ("xprtrdma:
Fix client lock-up after application signal fires").

Use recv_lock to quickly find the completing rqst, pin it, then
drop the lock. At that point invalidation and pull-up of the Reply
XDR can be done. Both are often expensive operations.

Finally, take recv_lock again to signal completion to the RPC layer. It also protects adjustment of "cwnd".

This greatly reduces the amount of time a lock is held by the reply handler. Comparing lock_stat results shows a marked decrease
in contention on rb_lock and recv_lock.

---
Hi-

If Trond's lock contention series is going into v4.14, I'd like
you
to consider this one (applied at the end of that series) as well.
Without it, NFS/RDMA performance regresses a bit after the first xprt_pin_rqst patch is applied. Thanks!

Quoted 6 lines
Quoted 4 lines

If Anna is OK with it, then I can apply this patch once she's sent
me
the pull request for the other RDMA client patches.

Works for me. Thanks!

Please can you both check http://git.linux-nfs.org/?p=trondmy/linux
-nfs .git;a=commit;h=b1da5d90857ee4b8f5acee1143705412b16fce56 and see if all
is well?

Quoted 3 lines

Do note that I did remove the call to rpcrdma_buffer_put() which was
being called with an uninitialised argument.

Looking at that again. rpcrdma_buffer_put() does look unnecessary, since out_norqst is being called now before "req" is initialized.

On further inspection, "return;" should be replaced with "goto repost;". This is similar to "out_nomatch" and "out_duplicate", which are both being replaced in this patch.

So, out_norqst should look something like this:

out_norqst:

       spin_unlock(&xprt->recv_lock);
       dprintk("RPC:       %s: no match for incoming xid 0x%08x\n",
               __func__, be32_to_cpu(xid));
       goto repost;
Quoted 5 lines
Quoted 2 lines

Otherwise, I don't see any obvious problems.

Patch updated. Please see http://git.linux-nfs.org/?p=trondmy/linux-nfs
.git;a=commit;h=9590d083c1bb1419b7992609d1a0a3e3517d3893

out_norqst looks correct.

Note that for out_shortreply:

+out_shortreply:
+ dprintk("RPC: %s: short/invalid reply\n", __func__);
+ goto repost;

The "goto repost;" can be dropped, since this code drops right into "repost:". That's harmless, though.

--
Chuck Lever