[PATCH 6.17 029/563] smb: client: fix sending the iwrap custom IRD/ORD negotiation messages

[PATCH 6.17 029/563] smb: client: fix sending the iwrap custom IRD/ORD negotiation messages

[Greg Kroah-Hartman]

6.17-stable review patch.  If anyone has any objections, please let me know.

------------------

From: Stefan Metzmacher <metze@samba.org>

[ Upstream commit ef71f1e046489c77a2f7d012edc762fba0a7aadc ]

Do a real negotiation and check the servers initiator_depth and
responder_resources.

This should use big endian in order to be useful.
I have captures of windows clients showing this.

The fact that we used little endian up to now
means that we sent very large numbers and the
negotiation with the server truncated them to the
server limits.

Note the reason why this uses u8 for
initiator_depth and responder_resources is
that the rdma layer also uses it.

The inconsitency regarding the initiator_depth
and responder_resources values being reversed
for iwarp devices in RDMA_CM_EVENT_ESTABLISHED
should also be fixed later, but for now we should
fix it.

Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Cc: linux-rdma@vger.kernel.org
Fixes: c7398583340a ("CIFS: SMBD: Implement RDMA memory registration")
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 fs/smb/client/smbdirect.c | 110 ++++++++++++++++++++++++++++++++++----
 fs/smb/client/smbdirect.h |   4 +-
 2 files changed, 103 insertions(+), 11 deletions(-)

diff --git a/fs/smb/client/smbdirect.c b/fs/smb/client/smbdirect.c
index e0fce5033004c..6480945c24592 100644
--- a/fs/smb/client/smbdirect.c
+++ b/fs/smb/client/smbdirect.c
@@ -179,6 +179,8 @@ static int smbd_conn_upcall(
 	struct smbd_connection *info = id->context;
 	struct smbdirect_socket *sc = &info->socket;
 	const char *event_name = rdma_event_msg(event->event);
+	u8 peer_initiator_depth;
+	u8 peer_responder_resources;
 
 	log_rdma_event(INFO, "event=%s status=%d\n",
 		event_name, event->status);
@@ -204,6 +206,85 @@ static int smbd_conn_upcall(
 
 	case RDMA_CM_EVENT_ESTABLISHED:
 		log_rdma_event(INFO, "connected event=%s\n", event_name);
+
+		/*
+		 * Here we work around an inconsistency between
+		 * iWarp and other devices (at least rxe and irdma using RoCEv2)
+		 */
+		if (rdma_protocol_iwarp(id->device, id->port_num)) {
+			/*
+			 * iWarp devices report the peer's values
+			 * with the perspective of the peer here.
+			 * Tested with siw and irdma (in iwarp mode)
+			 * We need to change to our perspective here,
+			 * so we need to switch the values.
+			 */
+			peer_initiator_depth = event->param.conn.responder_resources;
+			peer_responder_resources = event->param.conn.initiator_depth;
+		} else {
+			/*
+			 * Non iWarp devices report the peer's values
+			 * already changed to our perspective here.
+			 * Tested with rxe and irdma (in roce mode).
+			 */
+			peer_initiator_depth = event->param.conn.initiator_depth;
+			peer_responder_resources = event->param.conn.responder_resources;
+		}
+		if (rdma_protocol_iwarp(id->device, id->port_num) &&
+		    event->param.conn.private_data_len == 8) {
+			/*
+			 * Legacy clients with only iWarp MPA v1 support
+			 * need a private blob in order to negotiate
+			 * the IRD/ORD values.
+			 */
+			const __be32 *ird_ord_hdr = event->param.conn.private_data;
+			u32 ird32 = be32_to_cpu(ird_ord_hdr[0]);
+			u32 ord32 = be32_to_cpu(ird_ord_hdr[1]);
+
+			/*
+			 * cifs.ko sends the legacy IRD/ORD negotiation
+			 * event if iWarp MPA v2 was used.
+			 *
+			 * Here we check that the values match and only
+			 * mark the client as legacy if they don't match.
+			 */
+			if ((u32)event->param.conn.initiator_depth != ird32 ||
+			    (u32)event->param.conn.responder_resources != ord32) {
+				/*
+				 * There are broken clients (old cifs.ko)
+				 * using little endian and also
+				 * struct rdma_conn_param only uses u8
+				 * for initiator_depth and responder_resources,
+				 * so we truncate the value to U8_MAX.
+				 *
+				 * smb_direct_accept_client() will then
+				 * do the real negotiation in order to
+				 * select the minimum between client and
+				 * server.
+				 */
+				ird32 = min_t(u32, ird32, U8_MAX);
+				ord32 = min_t(u32, ord32, U8_MAX);
+
+				info->legacy_iwarp = true;
+				peer_initiator_depth = (u8)ird32;
+				peer_responder_resources = (u8)ord32;
+			}
+		}
+
+		/*
+		 * negotiate the value by using the minimum
+		 * between client and server if the client provided
+		 * non 0 values.
+		 */
+		if (peer_initiator_depth != 0)
+			info->initiator_depth =
+					min_t(u8, info->initiator_depth,
+					      peer_initiator_depth);
+		if (peer_responder_resources != 0)
+			info->responder_resources =
+					min_t(u8, info->responder_resources,
+					      peer_responder_resources);
+
 		sc->status = SMBDIRECT_SOCKET_CONNECTED;
 		wake_up_interruptible(&info->status_wait);
 		break;
@@ -1551,7 +1632,7 @@ static struct smbd_connection *_smbd_get_connection(
 	struct ib_qp_init_attr qp_attr;
 	struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
 	struct ib_port_immutable port_immutable;
-	u32 ird_ord_hdr[2];
+	__be32 ird_ord_hdr[2];
 
 	info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
 	if (!info)
@@ -1559,6 +1640,9 @@ static struct smbd_connection *_smbd_get_connection(
 	sc = &info->socket;
 	sp = &sc->parameters;
 
+	info->initiator_depth = 1;
+	info->responder_resources = SMBD_CM_RESPONDER_RESOURCES;
+
 	sc->status = SMBDIRECT_SOCKET_CONNECTING;
 	rc = smbd_ia_open(info, dstaddr, port);
 	if (rc) {
@@ -1639,22 +1723,22 @@ static struct smbd_connection *_smbd_get_connection(
 	}
 	sc->ib.qp = sc->rdma.cm_id->qp;
 
-	memset(&conn_param, 0, sizeof(conn_param));
-	conn_param.initiator_depth = 0;
-
-	conn_param.responder_resources =
-		min(sc->ib.dev->attrs.max_qp_rd_atom,
-		    SMBD_CM_RESPONDER_RESOURCES);
-	info->responder_resources = conn_param.responder_resources;
+	info->responder_resources =
+		min_t(u8, info->responder_resources,
+		      sc->ib.dev->attrs.max_qp_rd_atom);
 	log_rdma_mr(INFO, "responder_resources=%d\n",
 		info->responder_resources);
 
+	memset(&conn_param, 0, sizeof(conn_param));
+	conn_param.initiator_depth = info->initiator_depth;
+	conn_param.responder_resources = info->responder_resources;
+
 	/* Need to send IRD/ORD in private data for iWARP */
 	sc->ib.dev->ops.get_port_immutable(
 		sc->ib.dev, sc->rdma.cm_id->port_num, &port_immutable);
 	if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
-		ird_ord_hdr[0] = info->responder_resources;
-		ird_ord_hdr[1] = 1;
+		ird_ord_hdr[0] = cpu_to_be32(conn_param.responder_resources);
+		ird_ord_hdr[1] = cpu_to_be32(conn_param.initiator_depth);
 		conn_param.private_data = ird_ord_hdr;
 		conn_param.private_data_len = sizeof(ird_ord_hdr);
 	} else {
@@ -2121,6 +2205,12 @@ static int allocate_mr_list(struct smbd_connection *info)
 	atomic_set(&info->mr_used_count, 0);
 	init_waitqueue_head(&info->wait_for_mr_cleanup);
 	INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
+
+	if (info->responder_resources == 0) {
+		log_rdma_mr(ERR, "responder_resources negotiated as 0\n");
+		return -EINVAL;
+	}
+
 	/* Allocate more MRs (2x) than hardware responder_resources */
 	for (i = 0; i < info->responder_resources * 2; i++) {
 		smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
diff --git a/fs/smb/client/smbdirect.h b/fs/smb/client/smbdirect.h
index e45aa9ddd71da..4ca9b2b2c57f9 100644
--- a/fs/smb/client/smbdirect.h
+++ b/fs/smb/client/smbdirect.h
@@ -67,7 +67,9 @@ struct smbd_connection {
 
 	/* Memory registrations */
 	/* Maximum number of RDMA read/write outstanding on this connection */
-	int responder_resources;
+	bool legacy_iwarp;
+	u8 initiator_depth;
+	u8 responder_resources;
 	/* Maximum number of pages in a single RDMA write/read on this connection */
 	int max_frmr_depth;
 	/*
-- 
2.51.0

[PATCH 6.17 030/563] smb: server: fix IRD/ORD negotiation with the client

[Greg Kroah-Hartman]

6.17-stable review patch.  If anyone has any objections, please let me know.

------------------

From: Stefan Metzmacher <metze@samba.org>

[ Upstream commit fad988a2158d743da7971884b93482a73735b25e ]

Already do real negotiation in smb_direct_handle_connect_request()
where we see the requested initiator_depth and responder_resources
from the client.

We should detect legacy iwarp clients using MPA v1
with the custom IRD/ORD negotiation.

We need to send the custom IRD/ORD in big endian,
but we need to try to let clients with broken requests
using little endian (older cifs.ko) to work.

Note the reason why this uses u8 for
initiator_depth and responder_resources is
that the rdma layer also uses it.

Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Cc: linux-rdma@vger.kernel.org
Fixes: 0626e6641f6b ("cifsd: add server handler for central processing and tranport layers")
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 fs/smb/server/transport_rdma.c | 99 +++++++++++++++++++++++++++++-----
 1 file changed, 85 insertions(+), 14 deletions(-)

diff --git a/fs/smb/server/transport_rdma.c b/fs/smb/server/transport_rdma.c
index 74dfb6496095d..e1f659d3b4cf5 100644
--- a/fs/smb/server/transport_rdma.c
+++ b/fs/smb/server/transport_rdma.c
@@ -153,6 +153,10 @@ struct smb_direct_transport {
 	struct work_struct	disconnect_work;
 
 	bool			negotiation_requested;
+
+	bool			legacy_iwarp;
+	u8			initiator_depth;
+	u8			responder_resources;
 };
 
 #define KSMBD_TRANS(t) ((struct ksmbd_transport *)&((t)->transport))
@@ -347,6 +351,9 @@ static struct smb_direct_transport *alloc_transport(struct rdma_cm_id *cm_id)
 	t->cm_id = cm_id;
 	cm_id->context = t;
 
+	t->initiator_depth = SMB_DIRECT_CM_INITIATOR_DEPTH;
+	t->responder_resources = 1;
+
 	t->status = SMB_DIRECT_CS_NEW;
 	init_waitqueue_head(&t->wait_status);
 
@@ -1676,21 +1683,21 @@ static int smb_direct_send_negotiate_response(struct smb_direct_transport *t,
 static int smb_direct_accept_client(struct smb_direct_transport *t)
 {
 	struct rdma_conn_param conn_param;
-	struct ib_port_immutable port_immutable;
-	u32 ird_ord_hdr[2];
+	__be32 ird_ord_hdr[2];
 	int ret;
 
+	/*
+	 * smb_direct_handle_connect_request()
+	 * already negotiated t->initiator_depth
+	 * and t->responder_resources
+	 */
 	memset(&conn_param, 0, sizeof(conn_param));
-	conn_param.initiator_depth = min_t(u8, t->cm_id->device->attrs.max_qp_rd_atom,
-					   SMB_DIRECT_CM_INITIATOR_DEPTH);
-	conn_param.responder_resources = 0;
-
-	t->cm_id->device->ops.get_port_immutable(t->cm_id->device,
-						 t->cm_id->port_num,
-						 &port_immutable);
-	if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
-		ird_ord_hdr[0] = conn_param.responder_resources;
-		ird_ord_hdr[1] = 1;
+	conn_param.initiator_depth = t->initiator_depth;
+	conn_param.responder_resources = t->responder_resources;
+
+	if (t->legacy_iwarp) {
+		ird_ord_hdr[0] = cpu_to_be32(conn_param.responder_resources);
+		ird_ord_hdr[1] = cpu_to_be32(conn_param.initiator_depth);
 		conn_param.private_data = ird_ord_hdr;
 		conn_param.private_data_len = sizeof(ird_ord_hdr);
 	} else {
@@ -2081,10 +2088,13 @@ static bool rdma_frwr_is_supported(struct ib_device_attr *attrs)
 	return true;
 }
 
-static int smb_direct_handle_connect_request(struct rdma_cm_id *new_cm_id)
+static int smb_direct_handle_connect_request(struct rdma_cm_id *new_cm_id,
+					     struct rdma_cm_event *event)
 {
 	struct smb_direct_transport *t;
 	struct task_struct *handler;
+	u8 peer_initiator_depth;
+	u8 peer_responder_resources;
 	int ret;
 
 	if (!rdma_frwr_is_supported(&new_cm_id->device->attrs)) {
@@ -2098,6 +2108,67 @@ static int smb_direct_handle_connect_request(struct rdma_cm_id *new_cm_id)
 	if (!t)
 		return -ENOMEM;
 
+	peer_initiator_depth = event->param.conn.initiator_depth;
+	peer_responder_resources = event->param.conn.responder_resources;
+	if (rdma_protocol_iwarp(new_cm_id->device, new_cm_id->port_num) &&
+	    event->param.conn.private_data_len == 8) {
+		/*
+		 * Legacy clients with only iWarp MPA v1 support
+		 * need a private blob in order to negotiate
+		 * the IRD/ORD values.
+		 */
+		const __be32 *ird_ord_hdr = event->param.conn.private_data;
+		u32 ird32 = be32_to_cpu(ird_ord_hdr[0]);
+		u32 ord32 = be32_to_cpu(ird_ord_hdr[1]);
+
+		/*
+		 * cifs.ko sends the legacy IRD/ORD negotiation
+		 * event if iWarp MPA v2 was used.
+		 *
+		 * Here we check that the values match and only
+		 * mark the client as legacy if they don't match.
+		 */
+		if ((u32)event->param.conn.initiator_depth != ird32 ||
+		    (u32)event->param.conn.responder_resources != ord32) {
+			/*
+			 * There are broken clients (old cifs.ko)
+			 * using little endian and also
+			 * struct rdma_conn_param only uses u8
+			 * for initiator_depth and responder_resources,
+			 * so we truncate the value to U8_MAX.
+			 *
+			 * smb_direct_accept_client() will then
+			 * do the real negotiation in order to
+			 * select the minimum between client and
+			 * server.
+			 */
+			ird32 = min_t(u32, ird32, U8_MAX);
+			ord32 = min_t(u32, ord32, U8_MAX);
+
+			t->legacy_iwarp = true;
+			peer_initiator_depth = (u8)ird32;
+			peer_responder_resources = (u8)ord32;
+		}
+	}
+
+	/*
+	 * First set what the we as server are able to support
+	 */
+	t->initiator_depth = min_t(u8, t->initiator_depth,
+				   new_cm_id->device->attrs.max_qp_rd_atom);
+
+	/*
+	 * negotiate the value by using the minimum
+	 * between client and server if the client provided
+	 * non 0 values.
+	 */
+	if (peer_initiator_depth != 0)
+		t->initiator_depth = min_t(u8, t->initiator_depth,
+					   peer_initiator_depth);
+	if (peer_responder_resources != 0)
+		t->responder_resources = min_t(u8, t->responder_resources,
+					       peer_responder_resources);
+
 	ret = smb_direct_connect(t);
 	if (ret)
 		goto out_err;
@@ -2122,7 +2193,7 @@ static int smb_direct_listen_handler(struct rdma_cm_id *cm_id,
 {
 	switch (event->event) {
 	case RDMA_CM_EVENT_CONNECT_REQUEST: {
-		int ret = smb_direct_handle_connect_request(cm_id);
+		int ret = smb_direct_handle_connect_request(cm_id, event);
 
 		if (ret) {
 			pr_err("Can't create transport: %d\n", ret);
-- 
2.51.0