rdmaio: enable fio-native io-direction semantics

Currently rdmaio configuration api is far from ideal.
1) It is too rdma speciffic
2) since client always use in rw=write mode so all stats accounted to write,

Let's make it more fio friendly 
 - odirect={0,1}:  represent  whenever we use zerocopy IO FIO_RDMA_MEM_XXX or 
 - rw={read,write}: represend data direction
 - listen={0,1} use same semantics as net and netsplice engines
So new configuration matrix look like follows:
if (td->o.odirect) {
       td_read(td)  ==> FIO_RDMA_MEM_READ
       td_write(td) ==> FIO_RDMA_MEM_WRITE
   } else {
       td_read(td)  ==> FIO_RDMA_CHA_RECV
       td_write(td) ==> FIO_RDMA_CHA_SEND
   }
   
This definitely looks much cleaner, and represend actual difference between modes
Also this allow to have sane read/write statistics
- remove option: verb

This breaks old configs, but make it explicit and visiable.
IMHO this is not a problem since rdmaio was completey broken for too long.

HOWTO

@@ -1900,8 +1900,15 @@ I/O engine
 		**rdma**
 			The RDMA I/O engine supports both RDMA memory semantics
 			(RDMA_WRITE/RDMA_READ) and channel semantics (Send/Recv) for the
-			InfiniBand, RoCE and iWARP protocols. This engine defines engine
-			specific options.
+			InfiniBand, RoCE and iWARP protocols. This engine defines specific
+			io derection transformation matrix:
+			With :option: `direct` enbled:
+				`DDIR_READ`  :  `IBV_WC_RDMA_READ`
+				`DDIR_WRITE` :  `IBV_WC_RDMA_WRITE`
+			With options.direct disabled:
+				`DDIR_READ`  :  `IBV_WC_RECV`
+				`DDIR_WRITE` :  `IBV_WC_SEND`
+			This engine defines engine specific options.
 
 		**falloc**
 			I/O engine that does regular fallocate to simulate data transfer as
@@ -2169,11 +2176,13 @@ with the caveat that when used on the command line, they must come after the
 	hostname if the job is a TCP listener or UDP reader. For unix sockets, the
 	normal :option:`filename` option should be used and the port is invalid.
 
-.. option:: listen : [netsplice] [net]
+.. option:: listen : [netsplice] [net] [rdma]
 
 	For TCP network connections, tell fio to listen for incoming connections
 	rather than initiating an outgoing connection. The :option:`hostname` must
 	be omitted if this option is used.
+	For RDMA engine, tell fio to listen for incoming RDMA-CM connection,
+	rather than initiating an outgoing connection.
 
 .. option:: pingpong : [netsplice] [net]
 
@@ -2245,13 +2254,6 @@ with the caveat that when used on the command line, they must come after the
 
 	The size of the chunk to use for each file.
 
-.. option:: verb=str : [rdma]
-
-	The RDMA verb to use on this side of the RDMA ioengine connection. Valid
-	values are write, read, send and recv. These correspond to the equivalent
-	RDMA verbs (e.g. write = rdma_write etc.). Note that this only needs to be
-	specified on the client side of the connection. See the examples folder.
-
 .. option:: bindname=str : [rdma]
 
 	The name to use to bind the local RDMA-CM connection to a local RDMA device.

engines/rdma.c

@@ -57,6 +57,7 @@ enum rdma_io_mode {
 
 struct rdmaio_options {
 	struct thread_data *td;
+	bool listen;
 	unsigned int port;
 	enum rdma_io_mode verb;
 	char *bindname;
@@ -104,33 +105,14 @@ static struct fio_option options[] = {
 		.group	= FIO_OPT_G_RDMA,
 	},
 	{
-		.name	= "verb",
-		.lname	= "RDMA engine verb",
-		.alias	= "proto",
-		.type	= FIO_OPT_STR,
-		.off1	= offsetof(struct rdmaio_options, verb),
-		.help	= "RDMA engine verb",
-		.def	= "write",
-		.posval = {
-			  { .ival = "write",
-			    .oval = FIO_RDMA_MEM_WRITE,
-			    .help = "Memory Write",
-			  },
-			  { .ival = "read",
-			    .oval = FIO_RDMA_MEM_READ,
-			    .help = "Memory Read",
-			  },
-			  { .ival = "send",
-			    .oval = FIO_RDMA_CHA_SEND,
-			    .help = "Posted Send",
-			  },
-			  { .ival = "recv",
-			    .oval = FIO_RDMA_CHA_RECV,
-			    .help = "Posted Receive",
-			  },
-		},
+		.name     = "listen",
+		.lname    = "rdma engine listen",
+		.help     = "Listen for incoming RDMA-CM connections",
+		.type     = FIO_OPT_BOOL,
+		.off1     = offsetof(struct rdmaio_options, listen),
+		.def	  = "0",
 		.category = FIO_OPT_C_ENGINE,
-		.group	= FIO_OPT_G_RDMA,
+		.group    = FIO_OPT_G_RDMA,
 	},
 	{
 		.name	= NULL,
@@ -896,6 +878,18 @@ static int fio_rdmaio_connect(struct thread_data *td, struct fio_file *f)
 		return 1;
 	}
 
+	if (td->o.odirect) {
+		if (td_read(td))
+			rd->rdma_protocol = FIO_RDMA_MEM_READ;
+		else
+			rd->rdma_protocol = FIO_RDMA_MEM_WRITE;
+	} else {
+		if (td_read(td))
+			rd->rdma_protocol = FIO_RDMA_CHA_RECV;
+		else
+			rd->rdma_protocol = FIO_RDMA_CHA_SEND;
+
+	}
 	/* send task request */
 	strncpy(rd->send_buf.name, td->o.name, FIO_RDMA_NAME_MAX - 1);
 	rd->send_buf.name[FIO_RDMA_NAME_MAX - 1] = '\0';
@@ -960,7 +954,9 @@ static int fio_rdmaio_accept(struct thread_data *td, struct fio_file *f)
 
 static int fio_rdmaio_open_file(struct thread_data *td, struct fio_file *f)
 {
-	if (td_read(td))
+	struct rdmaio_options *o = td->eo;
+
+	if (o->listen)
 		return fio_rdmaio_accept(td, f);
 	else
 		return fio_rdmaio_connect(td, f);
@@ -1299,16 +1295,13 @@ static int fio_rdmaio_init(struct thread_data *td)
 		log_err("fio: rdma_create_id fail: %m\n");
 		return 1;
 	}
-
-	if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
-	    (rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
+	if (td->o.odirect) {
 		rd->rmt_us =
 			malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
 		memset(rd->rmt_us, 0,
-			FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
+		       FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
 		rd->rmt_nr = 0;
 	}
-
 	rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
 	memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
 	rd->io_u_queued_nr = 0;
@@ -1321,7 +1314,7 @@ static int fio_rdmaio_init(struct thread_data *td)
 	memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
 	rd->io_u_completed_nr = 0;
 
-	if (td_read(td)) {	/* READ as the server */
+	if (o->listen) {	/* READ as the server */
 		rd->is_client = 0;
 		td->flags |= TD_F_NO_PROGRESS;
 		/* server rd->rdma_buf_len will be setup after got request */

examples/gpudirect-rdmaio-client.fio

@@ -1,15 +1,23 @@
 # Example gpudirect rdma client job
 [global]
 ioengine=rdma
-hostname=[hostname]
-port=[port]
-verb=[read/write/send/recv]
-mem=cudamalloc
-gpu_dev_id=0
+hostname=${RDMA_SERVER}
+port=8888
 bs=1m
 size=100g
+mem=cudamalloc
+gpu_dev_id=0
 
-[sender]
+[cudadirect-client-rdma-write]
+# use IBV_WC_RDMA_WRITE
 rw=write
-iodepth=1
-iodepth_batch_complete=1
+direct=1
+numjobs=2
+iodepth=16
+
+[cudadirect-client-rdma-read]
+# use IBV_WC_RDMA_READ
+rw=read
+direct=1
+numjobs=2
+iodepth=16

examples/gpudirect-rdmaio-server.fio

@@ -1,12 +1,13 @@
-# Example rdma server job
+# Example gpudirect rdma server job
 [global]
 ioengine=rdma
-port=[port]
-mem=cudamalloc
-gpu_dev_id=0
+listen=1
+port=8888
 bs=1m
 size=100g
+mem=cudamalloc
+gpu_dev_id=0
 
-[receiver]
-rw=read
+[server]
+numjobs=4
 iodepth=16

examples/rdmaio-client.fio

@@ -3,14 +3,19 @@
 ioengine=rdma
 hostname=${RDMA_SERVER}
 port=8888
-# Data exchange method
-# verb=[read/write/send/recv]
-verb=write
 bs=1m
 size=100g
-# IO direction:  [read] as the server, [write] as for client
-rw=write
 
 [client-rdma-write]
-numjobs=4
+# use IBV_WC_RDMA_WRITE
+rw=write
+direct=1
+numjobs=2
+iodepth=16
+
+[client-rdma-read]
+# use IBV_WC_RDMA_READ
+rw=read
+direct=1
+numjobs=2
 iodepth=16

examples/rdmaio-server.fio

@@ -1,11 +1,10 @@
 # Example rdma server job
 [global]
 ioengine=rdma
+listen=1
 port=8888
 bs=1m
 size=100g
-# IO direction:  [read] as the server, [write] as for client
-rw=read
 
 [server]
 numjobs=4