Merge pull request #55 from plieven/tcp_nodelay

SOCKET set TCP_NODELAY on iscsi->fd

include/iscsi-private.h

@@ -290,9 +290,8 @@ int iscsi_send_target_nop_out(struct iscsi_context *iscsi, uint32_t ttt);
 void iscsi_set_error(struct iscsi_context *iscsi, const char *error_string,
 		     ...) __attribute__((format(printf, 2, 3)));
 
-unsigned char *iscsi_get_user_in_buffer(struct iscsi_context *iscsi, struct iscsi_in_pdu *in, uint32_t pos, ssize_t *count);
+struct scsi_iovector *iscsi_get_scsi_task_iovector_in(struct iscsi_context *iscsi, struct iscsi_in_pdu *in);
-unsigned char *iscsi_get_user_out_buffer(struct iscsi_context *iscsi, struct iscsi_pdu *pdu, uint32_t pos, ssize_t *count);
+struct scsi_iovector *iscsi_get_scsi_task_iovector_out(struct iscsi_context *iscsi, struct iscsi_pdu *pdu);
-
 
 inline void* iscsi_malloc(struct iscsi_context *iscsi, size_t size);
 inline void* iscsi_zmalloc(struct iscsi_context *iscsi, size_t size);

lib/iscsi-command.c

@@ -436,7 +436,7 @@ iscsi_process_scsi_data_in(struct iscsi_context *iscsi, struct iscsi_pdu *pdu,
 	dsl = scsi_get_uint32(&in->hdr[4]) & 0x00ffffff;
 
 	/* Dont add to reassembly buffer if we already have a user buffer */
-	if (scsi_task_get_data_in_buffer(task, 0, NULL) == NULL) {
+	if (task->iovector_in.iov == NULL) {
 		if (iscsi_add_data(iscsi, &pdu->indata, in->data, dsl, 0) != 0) {
 		    iscsi_set_error(iscsi, "Out-of-memory: failed to add data "
 				"to pdu in buffer.");
@@ -1422,30 +1422,42 @@ iscsi_unmap_task(struct iscsi_context *iscsi, int lun, int anchor, int group,
 	return task;
 }
 
-unsigned char *
+struct scsi_iovector *
-iscsi_get_user_in_buffer(struct iscsi_context *iscsi, struct iscsi_in_pdu *in, uint32_t pos, ssize_t *count)
+iscsi_get_scsi_task_iovector_in(struct iscsi_context *iscsi, struct iscsi_in_pdu *in)
 {
 	struct iscsi_pdu *pdu;
-	uint32_t offset;
 	uint32_t itt;
 
 	if ((in->hdr[0] & 0x3f) != ISCSI_PDU_DATA_IN) {
 		return NULL;
 	}
 
-	offset = scsi_get_uint32(&in->hdr[40]);
-
 	itt = scsi_get_uint32(&in->hdr[16]);
 	for (pdu = iscsi->waitpdu; pdu; pdu = pdu->next) {
 		if (pdu->itt == itt) {
 			break;
 		}
 	}
+
 	if (pdu == NULL) {
 		return NULL;
 	}
 
-	return scsi_task_get_data_in_buffer(pdu->scsi_cbdata.task, offset + pos, count);
+	if (pdu->scsi_cbdata.task->iovector_in.iov == NULL) {
+		return NULL;
+	}
+
+	return &pdu->scsi_cbdata.task->iovector_in;
+}
+
+struct scsi_iovector *
+iscsi_get_scsi_task_iovector_out(struct iscsi_context *iscsi _U_, struct iscsi_pdu *pdu)
+{
+	if (pdu->scsi_cbdata.task->iovector_out.iov == NULL) {
+		return NULL;
+	}
+
+	return &pdu->scsi_cbdata.task->iovector_out;
 }
 
 struct scsi_task *
@@ -1593,10 +1605,3 @@ iscsi_scsi_cancel_all_tasks(struct iscsi_context *iscsi)
 		iscsi_free_pdu(iscsi, pdu);
 	}
 }
-
-unsigned char *
-iscsi_get_user_out_buffer(struct iscsi_context *iscsi _U_, struct iscsi_pdu *pdu, uint32_t pos, ssize_t *count)
-{
-	return scsi_task_get_data_out_buffer(pdu->scsi_cbdata.task, pos, count);
-}
-

lib/scsi-lowlevel.c

@@ -34,6 +34,7 @@
 #include <stddef.h>
 #include <string.h>
 #include <stdint.h>
+#include <errno.h>
 #include "slist.h"
 #include "scsi-lowlevel.h"
 
@@ -2593,69 +2594,14 @@ scsi_iovector_add(struct scsi_task *task, struct scsi_iovector *iovector, int le
 	return 0;
 }
 
-unsigned char *
-scsi_iovector_get_buffer(struct scsi_iovector *iovector, uint32_t pos, ssize_t *count)
-{
-	if (iovector->iov == NULL) {
-		return NULL;
-	}
-
-	if (pos == 0 && count == NULL) return iovector->iov[0].iov_base;
-
-	if (pos < iovector->offset) {
-		/* start over in case we are going backwards */
-		iovector->offset = 0;
-		iovector->consumed = 0;
-	}
-
-	if (iovector->niov <= iovector->consumed) {
-		/* someone issued a read but did not provide enough user buffers for all the data.
-		 * maybe someone tried to read just 512 bytes off a MMC device?
-		 */
-		return NULL;
-	}
-
-	struct scsi_iovec *iov = &iovector->iov[iovector->consumed];
-
-	pos-= iovector->offset;
-
-	while (pos >= iov->iov_len) {
-		iovector->offset += iov->iov_len;
-		iovector->consumed++;
-		pos -= iov->iov_len;
-		if (iovector->niov <= iovector->consumed) {
-			return NULL;
-		}
-		iov = &iovector->iov[iovector->consumed];
-	}
-
-	if (count && *count >= (ssize_t)(iov->iov_len - pos)) {
-		*count = iov->iov_len - pos;
-	}
-
-	return (unsigned char *) iov->iov_base + pos;
-}
-
 int
 scsi_task_add_data_in_buffer(struct scsi_task *task, int len, unsigned char *buf)
 {
 	return scsi_iovector_add(task, &task->iovector_in, len, buf);
 }
 
-unsigned char *
-scsi_task_get_data_in_buffer(struct scsi_task *task, uint32_t pos, ssize_t *count)
-{
-	return scsi_iovector_get_buffer(&task->iovector_in, pos, count);
-}
-
 int
 scsi_task_add_data_out_buffer(struct scsi_task *task, int len, unsigned char *buf)
 {
 	return scsi_iovector_add(task, &task->iovector_out, len, buf);
 }
-
-unsigned char *
-scsi_task_get_data_out_buffer(struct scsi_task *task, uint32_t pos, ssize_t *count)
-{
-	return scsi_iovector_get_buffer(&task->iovector_out, pos, count);
-}

lib/socket.c

@@ -42,6 +42,7 @@
 #include <string.h>
 #include <errno.h>
 #include <fcntl.h>
+#include "scsi-lowlevel.h"
 #include "iscsi.h"
 #include "iscsi-private.h"
 #include "slist.h"
@@ -283,6 +284,12 @@ iscsi_connect_async(struct iscsi_context *iscsi, const char *portal,
 	}
 #endif
 
+	if (set_tcp_sockopt(iscsi->fd, TCP_NODELAY, 1) != 0) {
+		ISCSI_LOG(iscsi,1,"failed to set TCP_NODELAY sockopt: %s",strerror(errno));
+	} else {
+		ISCSI_LOG(iscsi,3,"TCP_NODELAY set to 1");
+	}
+
 	if (connect(iscsi->fd, ai->ai_addr, socksize) != 0
 	    && errno != EINPROGRESS) {
 		iscsi_set_error(iscsi, "Connect failed with errno : "
@@ -357,6 +364,91 @@ iscsi_queue_length(struct iscsi_context *iscsi)
 	return i;
 }
 
+ssize_t
+iscsi_iovector_readv_writev(struct iscsi_context *iscsi, struct scsi_iovector *iovector, uint32_t pos, ssize_t max_read, int do_write)
+{
+	if (iovector->iov == NULL) {
+		errno = EINVAL;
+		return -1;
+	}
+
+	if (pos < iovector->offset) {
+		/* start over in case we are going backwards */
+		iovector->offset = 0;
+		iovector->consumed = 0;
+	}
+
+	if (iovector->niov <= iovector->consumed) {
+		/* someone issued a read/write but did not provide enough user buffers for all the data.
+		 * maybe someone tried to read just 512 bytes off a MMC device?
+		 */
+		errno = EINVAL;
+		return -1;
+	}
+
+	/* iov is a pointer to the first iovec to pass */
+	struct scsi_iovec *iov = &iovector->iov[iovector->consumed];
+	pos -= iovector->offset;
+
+	/* forward until iov points to the first iov to pass */
+	while (pos >= iov->iov_len) {
+		iovector->offset += iov->iov_len;
+		iovector->consumed++;
+		pos -= iov->iov_len;
+		if (iovector->niov <= iovector->consumed) {
+			errno = EINVAL;
+			return -1;
+		}
+		iov = &iovector->iov[iovector->consumed];
+	}
+
+	/* iov2 is a pointer to the last iovec to pass */
+	struct scsi_iovec *iov2 = iov;
+
+	int niov=1; /* number of iovectors to pass */
+	uint32_t len2 = pos + max_read; /* adjust length of iov2 */
+	
+	/* forward until iov2 points to the last iovec we pass later. it might
+	   happen that we have a lot of iovectors but are limited by max_read */
+	while (len2 > iov2->iov_len) {
+		if (iovector->niov <= iovector->consumed+niov-1) {
+			errno = EINVAL;
+			return -1;
+		}
+		niov++;
+		len2 -= iov2->iov_len;
+		iov2 = &iovector->iov[iovector->consumed+niov-1];
+	}
+
+	/* we might limit the length of the last iovec we pass to readv/writev
+	   store its orignal length to restore it later */
+	size_t _len2 = iov2->iov_len;
+
+	/* adjust base+len of start iovec and len of last iovec */
+	iov2->iov_len = len2;
+	iov->iov_base = (void*) ((uintptr_t)iov->iov_base + pos);
+	iov->iov_len -= pos;
+
+	ssize_t n;
+	if (do_write) {
+		n = writev(iscsi->fd, (struct iovec*) iov, niov);
+	} else {
+		n = readv(iscsi->fd, (struct iovec*) iov, niov);
+	}
+
+	/* restore original values */
+	iov->iov_base = (void*) ((uintptr_t)iov->iov_base - pos);
+	iov->iov_len += pos;
+	iov2->iov_len = _len2;
+
+	if (n > max_read) {
+		/* we read/write more bytes than expected, this MUST not happen */
+		errno = EINVAL;
+		return -1;
+	}
+	return n;
+}
+
 static int
 iscsi_read_from_socket(struct iscsi_context *iscsi)
 {
@@ -405,13 +497,16 @@ iscsi_read_from_socket(struct iscsi_context *iscsi)
 	}
 	if (data_size != 0) {
 		unsigned char *buf = NULL;
+		struct scsi_iovector * iovector_in;
 
 		count = data_size - in->data_pos;
 
 		/* first try to see if we already have a user buffer */
-		buf = iscsi_get_user_in_buffer(iscsi, in, in->data_pos, &count);
+		iovector_in = iscsi_get_scsi_task_iovector_in(iscsi, in);
-		/* if not, allocate one */
+		if (iovector_in != NULL) {
-		if (buf == NULL) {
+			uint32_t offset = scsi_get_uint32(&in->hdr[40]);
+			count = iscsi_iovector_readv_writev(iscsi, iovector_in, in->data_pos + offset, count, 0);
+		} else {
 			if (in->data == NULL) {
 				in->data = iscsi_malloc(iscsi, data_size);
 				if (in->data == NULL) {
@@ -420,9 +515,9 @@ iscsi_read_from_socket(struct iscsi_context *iscsi)
 				}
 			}
 			buf = &in->data[in->data_pos];
+			count = recv(iscsi->fd, buf, count, 0);
 		}
 		
-		count = recv(iscsi->fd, buf, count, 0);
 		if (count == 0) {
 			return -1;
 		}
@@ -434,6 +529,7 @@ iscsi_read_from_socket(struct iscsi_context *iscsi)
 				"errno:%d", errno);
 			return -1;
 		}
+
 		in->data_pos += count;
 	}
 
@@ -518,19 +614,17 @@ iscsi_write_to_socket(struct iscsi_context *iscsi)
 
 		/* Write any iovectors that might have been passed to us */
 		while (pdu->out_written < pdu->out_len) {
-			unsigned char *buf;
+			struct scsi_iovector* iovector_out;
 
-			count = pdu->out_len - pdu->out_written;
+			iovector_out = iscsi_get_scsi_task_iovector_out(iscsi, pdu);
-			buf = iscsi_get_user_out_buffer(iscsi, pdu, pdu->out_offset + pdu->out_written, &count);
+
-			if (buf == NULL) {
+			if (iovector_out == NULL) {
 				iscsi_set_error(iscsi, "Can't find iovector data for DATA-OUT");
 				return -1;
 			}
 
-			count = send(iscsi->fd,
+			count = iscsi_iovector_readv_writev(iscsi, iovector_out, pdu->out_offset + pdu->out_written, pdu->out_len - pdu->out_written, 1);
-				     buf,
+
-				     count,
-				     0);
 			if (count == -1) {
 				if (errno == EAGAIN || errno == EWOULDBLOCK) {
 					return 0;
@@ -539,6 +633,7 @@ iscsi_write_to_socket(struct iscsi_context *iscsi)
 						"socket :%d", errno);
 				return -1;
 			}
+
 			pdu->out_written += count;
 		}