net/9p: convert to new CQ API

net/9p: convert to new CQ API

[Christoph Hellwig]

Hi all,

who is maintaining the "RDMA transport" (1) for 9p?  Below patch converts
it to your new CQ API.  It's fairly trivial, but untested as I can't figure
out how to actually test this code.

[1] RDMA seems a bit of a misowner as it's never doing RDMA data transfers,
but that's a separate story :)

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH] net/9p: convert to new CQ API

[Christoph Hellwig]

Trivial conversion to the new RDMA CQ API.

Signed-off-by: Christoph Hellwig <hch-jcswGhMUV9g@public.gmane.org>
---
 net/9p/trans_rdma.c | 87 +++++++++++++++++++----------------------------------
 1 file changed, 31 insertions(+), 56 deletions(-)

diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c
index 52b4a2f..668c3be 100644
--- a/net/9p/trans_rdma.c
+++ b/net/9p/trans_rdma.c
@@ -109,14 +109,13 @@ struct p9_trans_rdma {
 /**
  * p9_rdma_context - Keeps track of in-process WR
  *
- * @wc_op: The original WR op for when the CQE completes in error.
  * @busa: Bus address to unmap when the WR completes
  * @req: Keeps track of requests (send)
  * @rc: Keepts track of replies (receive)
  */
 struct p9_rdma_req;
-struct p9_rdma_context {
-	enum ib_wc_opcode wc_op;
+struct p9_rdma_context {	
+	struct ib_cqe cqe;
 	dma_addr_t busa;
 	union {
 		struct p9_req_t *req;
@@ -284,9 +283,12 @@ p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 }
 
 static void
-handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
-	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+recv_done(struct ib_cq *cq, struct ib_wc *wc)
 {
+	struct p9_client *client = cq->cq_context;
+	struct p9_trans_rdma *rdma = client->trans;
+	struct p9_rdma_context *c =
+		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
 	struct p9_req_t *req;
 	int err = 0;
 	int16_t tag;
@@ -295,7 +297,7 @@ handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 							 DMA_FROM_DEVICE);
 
-	if (status != IB_WC_SUCCESS)
+	if (wc->status != IB_WC_SUCCESS)
 		goto err_out;
 
 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
@@ -316,21 +318,31 @@ handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 	req->rc = c->rc;
 	p9_client_cb(client, req, REQ_STATUS_RCVD);
 
+ out:
+	up(&rdma->rq_sem);
+	kfree(c);
 	return;
 
  err_out:
-	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
+	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, wc->status);
 	rdma->state = P9_RDMA_FLUSHING;
 	client->status = Disconnected;
+	goto out;
 }
 
 static void
-handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
-	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+send_done(struct ib_cq *cq, struct ib_wc *wc)
 {
+	struct p9_client *client = cq->cq_context;
+	struct p9_trans_rdma *rdma = client->trans;
+	struct p9_rdma_context *c =
+		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
+
 	ib_dma_unmap_single(rdma->cm_id->device,
 			    c->busa, c->req->tc->size,
 			    DMA_TO_DEVICE);
+	up(&rdma->sq_sem);
+	kfree(c);
 }
 
 static void qp_event_handler(struct ib_event *event, void *context)
@@ -339,42 +351,6 @@ static void qp_event_handler(struct ib_event *event, void *context)
 		 event->event, context);
 }
 
-static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
-{
-	struct p9_client *client = cq_context;
-	struct p9_trans_rdma *rdma = client->trans;
-	int ret;
-	struct ib_wc wc;
-
-	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
-	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
-		struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
-
-		switch (c->wc_op) {
-		case IB_WC_RECV:
-			handle_recv(client, rdma, c, wc.status, wc.byte_len);
-			up(&rdma->rq_sem);
-			break;
-
-		case IB_WC_SEND:
-			handle_send(client, rdma, c, wc.status, wc.byte_len);
-			up(&rdma->sq_sem);
-			break;
-
-		default:
-			pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
-			       c->wc_op, wc.opcode, wc.status);
-			break;
-		}
-		kfree(c);
-	}
-}
-
-static void cq_event_handler(struct ib_event *e, void *v)
-{
-	p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
-}
-
 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 {
 	if (!rdma)
@@ -387,7 +363,7 @@ static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 		ib_dealloc_pd(rdma->pd);
 
 	if (rdma->cq && !IS_ERR(rdma->cq))
-		ib_destroy_cq(rdma->cq);
+		ib_free_cq(rdma->cq);
 
 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 		rdma_destroy_id(rdma->cm_id);
@@ -408,13 +384,14 @@ post_recv(struct p9_client *client, struct p9_rdma_context *c)
 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 		goto error;
 
+	c->cqe.done = recv_done;
+
 	sge.addr = c->busa;
 	sge.length = client->msize;
 	sge.lkey = rdma->pd->local_dma_lkey;
 
 	wr.next = NULL;
-	c->wc_op = IB_WC_RECV;
-	wr.wr_id = (unsigned long) c;
+	wr.wr_cqe = &c->cqe;
 	wr.sg_list = &sge;
 	wr.num_sge = 1;
 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
@@ -499,13 +476,14 @@ dont_need_post_recv:
 		goto send_error;
 	}
 
+	c->cqe.done = send_done;
+
 	sge.addr = c->busa;
 	sge.length = c->req->tc->size;
 	sge.lkey = rdma->pd->local_dma_lkey;
 
 	wr.next = NULL;
-	c->wc_op = IB_WC_SEND;
-	wr.wr_id = (unsigned long) c;
+	wr.wr_cqe = &c->cqe;
 	wr.opcode = IB_WR_SEND;
 	wr.send_flags = IB_SEND_SIGNALED;
 	wr.sg_list = &sge;
@@ -642,7 +620,6 @@ rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 	struct p9_trans_rdma *rdma;
 	struct rdma_conn_param conn_param;
 	struct ib_qp_init_attr qp_attr;
-	struct ib_cq_init_attr cq_attr = {};
 
 	/* Parse the transport specific mount options */
 	err = parse_opts(args, &opts);
@@ -695,13 +672,11 @@ rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 		goto error;
 
 	/* Create the Completion Queue */
-	cq_attr.cqe = opts.sq_depth + opts.rq_depth + 1;
-	rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
-				cq_event_handler, client,
-				&cq_attr);
+	rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
+			opts.sq_depth + opts.rq_depth + 1,
+			0, IB_POLL_SOFTIRQ);
 	if (IS_ERR(rdma->cq))
 		goto error;
-	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 
 	/* Create the Protection Domain */
 	rdma->pd = ib_alloc_pd(rdma->cm_id->device);
-- 
2.1.4

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [V9fs-developer] [PATCH] net/9p: convert to new CQ API

[Dominique Martinet]

Hi,

Couple of checkpatch complains:

Christoph Hellwig wrote on Sat, Feb 27, 2016:
> -struct p9_rdma_context {
> -	enum ib_wc_opcode wc_op;
> +struct p9_rdma_context {	

trailing tab

> -	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
> +	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, wc->status);

line over 80 chars


That aside it looks good ; I need to check on the new API (hadn't
noticed the change) but it looks nice.

Will do the actual testing likely only next week only though;
Eric has been taking my patches for 9p/RDMA so I suspect he'll take
your's as well eventually (get_maintainer.pl has a long-ish list of CC
for us usually)


BTW I think it's easy enough to do the testing if you have a server that
can dish it out. diod[1] and nfs-ganesha[2] are the only two I'm aware
of but there might be more (using ganesha myself; happy to help you set
it up in private if you need)

[1] https://github.com/chaos/diod
[2] https://github.com/nfs-ganesha/nfs-ganesha

-- 
Dominique Martinet

Re: [PATCH] net/9p: convert to new CQ API

[Sagi Grimberg]

> Trivial conversion to the new RDMA CQ API.

Looks nice and simple :)

But I think that the fact that CQ processing is now
done in soft-IRQ (which is an improvement!) needs to
be documented.

Other than that, looks great

Reviewed-by: Sagi Grimberg <sagig@mellanox.com>

P.S.
I was also confused in the past about 9p and if anyone
actually uses it...

Re: [PATCH] net/9p: convert to new CQ API

[Christoph Hellwig]

On Sun, Feb 28, 2016 at 01:03:47PM +0200, Sagi Grimberg wrote:
>
>> Trivial conversion to the new RDMA CQ API.
>
> Looks nice and simple :)
>
> But I think that the fact that CQ processing is now
> done in soft-IRQ (which is an improvement!) needs to
> be documented.

Yeah, I meant to do that, but ended up beeing lazy as usual :)
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [V9fs-developer] [PATCH] net/9p: convert to new CQ API

[Christoph Hellwig]

New version with the nits fixed below.  Now that checkpath started
a stupid warning about not using tabs for indentation which I've
ignored here and will take up in my usual fights against Joes
idicotic opinions separately..

--
>From 0055e31f24c7b3642ece5ebb998fa63d772dcb1a Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch-jcswGhMUV9g@public.gmane.org>
Date: Sat, 27 Feb 2016 10:22:40 +0100
Subject: net/9p: convert to new CQ API

Trivial conversion to the new RDMA CQ API.

Signed-off-by: Christoph Hellwig <hch-jcswGhMUV9g@public.gmane.org>
---
 net/9p/trans_rdma.c | 86 +++++++++++++++++++----------------------------------
 1 file changed, 31 insertions(+), 55 deletions(-)

diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c
index 52b4a2f..1852e38 100644
--- a/net/9p/trans_rdma.c
+++ b/net/9p/trans_rdma.c
@@ -109,14 +109,13 @@ struct p9_trans_rdma {
 /**
  * p9_rdma_context - Keeps track of in-process WR
  *
- * @wc_op: The original WR op for when the CQE completes in error.
  * @busa: Bus address to unmap when the WR completes
  * @req: Keeps track of requests (send)
  * @rc: Keepts track of replies (receive)
  */
 struct p9_rdma_req;
 struct p9_rdma_context {
-	enum ib_wc_opcode wc_op;
+	struct ib_cqe cqe;
 	dma_addr_t busa;
 	union {
 		struct p9_req_t *req;
@@ -284,9 +283,12 @@ p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 }
 
 static void
-handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
-	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+recv_done(struct ib_cq *cq, struct ib_wc *wc)
 {
+	struct p9_client *client = cq->cq_context;
+	struct p9_trans_rdma *rdma = client->trans;
+	struct p9_rdma_context *c =
+		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
 	struct p9_req_t *req;
 	int err = 0;
 	int16_t tag;
@@ -295,7 +297,7 @@ handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 							 DMA_FROM_DEVICE);
 
-	if (status != IB_WC_SUCCESS)
+	if (wc->status != IB_WC_SUCCESS)
 		goto err_out;
 
 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
@@ -316,21 +318,32 @@ handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 	req->rc = c->rc;
 	p9_client_cb(client, req, REQ_STATUS_RCVD);
 
+ out:
+	up(&rdma->rq_sem);
+	kfree(c);
 	return;
 
  err_out:
-	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
+	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
+			req, err, wc->status);
 	rdma->state = P9_RDMA_FLUSHING;
 	client->status = Disconnected;
+	goto out;
 }
 
 static void
-handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
-	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+send_done(struct ib_cq *cq, struct ib_wc *wc)
 {
+	struct p9_client *client = cq->cq_context;
+	struct p9_trans_rdma *rdma = client->trans;
+	struct p9_rdma_context *c =
+		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
+
 	ib_dma_unmap_single(rdma->cm_id->device,
 			    c->busa, c->req->tc->size,
 			    DMA_TO_DEVICE);
+	up(&rdma->sq_sem);
+	kfree(c);
 }
 
 static void qp_event_handler(struct ib_event *event, void *context)
@@ -339,42 +352,6 @@ static void qp_event_handler(struct ib_event *event, void *context)
 		 event->event, context);
 }
 
-static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
-{
-	struct p9_client *client = cq_context;
-	struct p9_trans_rdma *rdma = client->trans;
-	int ret;
-	struct ib_wc wc;
-
-	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
-	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
-		struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
-
-		switch (c->wc_op) {
-		case IB_WC_RECV:
-			handle_recv(client, rdma, c, wc.status, wc.byte_len);
-			up(&rdma->rq_sem);
-			break;
-
-		case IB_WC_SEND:
-			handle_send(client, rdma, c, wc.status, wc.byte_len);
-			up(&rdma->sq_sem);
-			break;
-
-		default:
-			pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
-			       c->wc_op, wc.opcode, wc.status);
-			break;
-		}
-		kfree(c);
-	}
-}
-
-static void cq_event_handler(struct ib_event *e, void *v)
-{
-	p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
-}
-
 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 {
 	if (!rdma)
@@ -387,7 +364,7 @@ static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 		ib_dealloc_pd(rdma->pd);
 
 	if (rdma->cq && !IS_ERR(rdma->cq))
-		ib_destroy_cq(rdma->cq);
+		ib_free_cq(rdma->cq);
 
 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 		rdma_destroy_id(rdma->cm_id);
@@ -408,13 +385,14 @@ post_recv(struct p9_client *client, struct p9_rdma_context *c)
 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 		goto error;
 
+	c->cqe.done = recv_done;
+
 	sge.addr = c->busa;
 	sge.length = client->msize;
 	sge.lkey = rdma->pd->local_dma_lkey;
 
 	wr.next = NULL;
-	c->wc_op = IB_WC_RECV;
-	wr.wr_id = (unsigned long) c;
+	wr.wr_cqe = &c->cqe;
 	wr.sg_list = &sge;
 	wr.num_sge = 1;
 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
@@ -499,13 +477,14 @@ dont_need_post_recv:
 		goto send_error;
 	}
 
+	c->cqe.done = send_done;
+
 	sge.addr = c->busa;
 	sge.length = c->req->tc->size;
 	sge.lkey = rdma->pd->local_dma_lkey;
 
 	wr.next = NULL;
-	c->wc_op = IB_WC_SEND;
-	wr.wr_id = (unsigned long) c;
+	wr.wr_cqe = &c->cqe;
 	wr.opcode = IB_WR_SEND;
 	wr.send_flags = IB_SEND_SIGNALED;
 	wr.sg_list = &sge;
@@ -642,7 +621,6 @@ rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 	struct p9_trans_rdma *rdma;
 	struct rdma_conn_param conn_param;
 	struct ib_qp_init_attr qp_attr;
-	struct ib_cq_init_attr cq_attr = {};
 
 	/* Parse the transport specific mount options */
 	err = parse_opts(args, &opts);
@@ -695,13 +673,11 @@ rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 		goto error;
 
 	/* Create the Completion Queue */
-	cq_attr.cqe = opts.sq_depth + opts.rq_depth + 1;
-	rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
-				cq_event_handler, client,
-				&cq_attr);
+	rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
+			opts.sq_depth + opts.rq_depth + 1,
+			0, IB_POLL_SOFTIRQ);
 	if (IS_ERR(rdma->cq))
 		goto error;
-	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 
 	/* Create the Protection Domain */
 	rdma->pd = ib_alloc_pd(rdma->cm_id->device);
-- 
2.1.4

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [V9fs-developer] [PATCH] net/9p: convert to new CQ API

[Dominique Martinet]

Christoph Hellwig wrote on Thu, Mar 03, 2016:
> New version with the nits fixed below.  Now that checkpath started
> a stupid warning about not using tabs for indentation which I've
> ignored here and will take up in my usual fights against Joes
> idicotic opinions separately..

Thanks for the nitpicks, I can confirm it works as expected as well so
all good with me.
I like the new CQ interface :)

(if someone adds an Acked-by please use dominique.martinet@cea.fr for my
mail; sorry for the split personality)

-- 
Dominique

Re: [V9fs-developer] [PATCH] net/9p: convert to new CQ API

[Doug Ledford]

[-- Attachment #1: Type: text/plain, Size: 835 bytes --]

On 03/08/2016 09:38 AM, Dominique Martinet wrote:
> Christoph Hellwig wrote on Thu, Mar 03, 2016:
>> New version with the nits fixed below.  Now that checkpath started
>> a stupid warning about not using tabs for indentation which I've
>> ignored here and will take up in my usual fights against Joes
>> idicotic opinions separately..
> 
> Thanks for the nitpicks, I can confirm it works as expected as well so
> all good with me.
> I like the new CQ interface :)
> 
> (if someone adds an Acked-by please use dominique.martinet-KCE40YydGKI@public.gmane.org for my
> mail; sorry for the split personality)
> 

Since I haven't heard anyone else say they are picking this up, I've
grabbed it for 4.6.  Thanks.

-- 
Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
              GPG KeyID: 0E572FDD



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 884 bytes --]