4496 ndmpd handles wrongly EOM and EOF conditions

Reviewed by: Albert Lee <trisk@nexenta.com>
Approved by: Robert Mustacchi <rm@joyent.com>

4 files changed, 255 insertions, 428 deletions

diff --git a/usr/src/cmd/ndmpd/ndmp/ndmpd.h b/usr/src/cmd/ndmpd/ndmp/ndmpd.h
index b24c1a58a2..efee818c14 100644
--- a/usr/src/cmd/ndmpd/ndmp/ndmpd.h
+++ b/usr/src/cmd/ndmpd/ndmp/ndmpd.h
@@ -37,6 +37,7 @@
  */
 /* Copyright (c) 2007, The Storage Networking Industry Association. */
 /* Copyright (c) 1996, 1997 PDC, Network Appliance. All Rights Reserved */
+/* Copyright 2014 Nexenta Systems, Inc.  All rights reserved.  */
 
 #ifndef _NDMPD_H
 #define	_NDMPD_H
@@ -310,9 +311,6 @@ typedef struct ndmpd_session_tape_desc {
 	int td_sid;
 	int td_lun;
 	char td_adapter_name[SCSI_MAX_NAME];
-	ulong_t td_eom_seen:1,
-		td_io_err:1,
-		td_write:1;
 } ndmpd_session_tape_desc_t;
 
 typedef struct ndmpd_session_mover_desc {
@@ -798,7 +796,6 @@ extern int ndmpd_remote_read_v3(ndmpd_session_t *,
     char *,
     ulong_t);
 extern int ndmpd_mover_wait_v3(ndmpd_session_t *);
-extern void ndmpd_write_eom(int);
 
 
 /*
@@ -1022,6 +1019,9 @@ extern int ndmp_restore_extract_params(ndmpd_session_t *,
 extern int ndmp_tar_reader(ndmp_tar_reader_arg_t *);
 
 extern int tape_open(char *, int);
+extern int tape_is_at_bot(ndmpd_session_t *);
+extern int tape_is_at_bof(ndmpd_session_t *);
+extern void fm_dance(ndmpd_session_t *);
 
 extern void ndmp_session_ref(ndmpd_session_t *);
 extern void ndmp_session_unref(ndmpd_session_t *);
diff --git a/usr/src/cmd/ndmpd/ndmp/ndmpd_handler.c b/usr/src/cmd/ndmpd/ndmp/ndmpd_handler.c
index 30bafa0f73..2658bb2cef 100644
--- a/usr/src/cmd/ndmpd/ndmp/ndmpd_handler.c
+++ b/usr/src/cmd/ndmpd/ndmp/ndmpd_handler.c
@@ -37,6 +37,9 @@
  * POSSIBILITY OF SUCH DAMAGE.
  */
 /* Copyright (c) 1996, 1997 PDC, Network Appliance. All Rights Reserved */
+/*
+ * Copyright 2014 Nexenta Systems, Inc.  All rights reserved.
+ */
 
 #include <sys/note.h>
 #include "ndmpd.h"
@@ -348,7 +351,7 @@ ndmp_handler_t ndmp_msghdl_tab[] = {
 				NDMP_TAPE_WRITE,
 				AUTH_REQUIRED,
 				{
-					HANDL(tape_write, 2, 2),
+					HANDL(tape_write, 3, 2),
 					HANDL(tape_write, 3, 3),
 					HANDL(tape_write, 3, 4),
 				}
diff --git a/usr/src/cmd/ndmpd/ndmp/ndmpd_mover.c b/usr/src/cmd/ndmpd/ndmp/ndmpd_mover.c
index 6bf0a8808b..8772edc199 100644
--- a/usr/src/cmd/ndmpd/ndmp/ndmpd_mover.c
+++ b/usr/src/cmd/ndmpd/ndmp/ndmpd_mover.c
@@ -37,6 +37,7 @@
  */
 /* Copyright (c) 2007, The Storage Networking Industry Association. */
 /* Copyright (c) 1996, 1997 PDC, Network Appliance. All Rights Reserved */
+/* Copyright 2014 Nexenta Systems, Inc.  All rights reserved. */
 
 #include <sys/ioctl.h>
 #include <sys/types.h>
@@ -62,7 +63,6 @@
 
 static int create_listen_socket_v2(ndmpd_session_t *session, ulong_t *addr,
     ushort_t *port);
-static int tape_write(ndmpd_session_t *session, char *data, ssize_t length);
 static int tape_read(ndmpd_session_t *session, char *data);
 static int change_tape(ndmpd_session_t *session);
 static int discard_data(ndmpd_session_t *session, ulong_t length);
@@ -1354,30 +1354,6 @@ ndmpd_mover_connect_v4(ndmp_connection_t *connection, void *body)
  */
 
 /*
- * ndmpd_write_eom
- *
- * Write end-of-media magic string.  This is called after hitting the LEOT.
- */
-void
-ndmpd_write_eom(int fd)
-{
-	int n;
-
-	(void) ndmp_mtioctl(fd, MTWEOF, 1);
-	n = write(fd, NDMP_EOM_MAGIC, strlen(NDMP_EOM_MAGIC));
-
-	NDMP_LOG(LOG_DEBUG, "%d EOM bytes wrote", n);
-	(void) ndmp_mtioctl(fd, MTWEOF, 1);
-
-	/*
-	 * Rewind to the previous file since the last two files are used
-	 * as the indicator for logical EOM.
-	 */
-	(void) ndmp_mtioctl(fd, MTBSF, 2);
-}
-
-
-/*
  * ndmpd_local_write
  *
  * Writes data to the mover.
@@ -1413,7 +1389,7 @@ ndmpd_local_write(ndmpd_session_t *session, char *data, ulong_t length)
 		    0, session->ns_mover.md_record_size -
 		    session->ns_mover.md_w_index);
 
-		n = tape_write(session, session->ns_mover.md_buf,
+		n = mover_tape_write_v3(session, session->ns_mover.md_buf,
 		    session->ns_mover.md_record_size);
 		if (n <= 0) {
 			ndmpd_mover_error(session,
@@ -1439,7 +1415,7 @@ ndmpd_local_write(ndmpd_session_t *session, char *data, ulong_t length)
 		 */
 		if (session->ns_mover.md_w_index == 0 &&
 		    length - count >= session->ns_mover.md_record_size) {
-			n = tape_write(session, &data[count],
+			n = mover_tape_write_v3(session, &data[count],
 			    session->ns_mover.md_record_size);
 			if (n <= 0) {
 				ndmpd_mover_error(session,
@@ -1469,9 +1445,10 @@ ndmpd_local_write(ndmpd_session_t *session, char *data, ulong_t length)
 		/* Write the buffer if its full */
 		if (session->ns_mover.md_w_index ==
 		    session->ns_mover.md_record_size) {
-			n = tape_write(session, session->ns_mover.md_buf,
+			n = mover_tape_write_v3(session,
+			    session->ns_mover.md_buf,
 			    session->ns_mover.md_record_size);
-			if (n < 0) {
+			if (n <= 0) {
 				ndmpd_mover_error(session,
 				    (n == 0 ? NDMP_MOVER_HALT_ABORTED :
 				    NDMP_MOVER_HALT_INTERNAL_ERROR));
@@ -2279,99 +2256,6 @@ accept_connection(void *cookie, int fd, ulong_t mode)
 }
 
 /*
- * tape_write
- *
- * Writes a data record to tape. Detects and handles EOT conditions.
- *
- * Parameters:
- *   session (input) - session pointer.
- *   data    (input) - data to be written.
- *   length  (input) - length of data to be written.
- *
- * Returns:
- *    0 - operation aborted by client.
- *   -1 - error.
- *   otherwise - number of bytes written.
- */
-static int
-tape_write(ndmpd_session_t *session, char *data, ssize_t length)
-{
-	ssize_t n;
-	int err;
-
-	for (; ; ) {
-		/*
-		 * Refer to the comment at the top of ndmpd_tape.c file for
-		 * Mammoth2 tape drives.
-		 */
-		if (session->ns_tape.td_eom_seen) {
-			NDMP_LOG(LOG_DEBUG, "eom_seen");
-			session->ns_tape.td_eom_seen = FALSE;
-			/*
-			 * End of media reached.
-			 * Notify client and wait for the client to
-			 * either abort the operation or continue the
-			 * operation after changing the tape.
-			 */
-			NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
-			    ++ndmp_log_msg_id,
-			    "End of tape reached. Load next tape.\n");
-
-			err = change_tape(session);
-
-			/* Operation aborted or connection terminated? */
-			if (err < 0)
-				return (-1);
-
-			continue;
-		}
-
-		n = write(session->ns_tape.td_fd, data, length);
-		if (n < 0) {
-			NDMP_LOG(LOG_ERR, "Tape write error: %m.");
-			return (-1);
-		}
-		NS_ADD(wtape, n);
-
-		if (n == 0 || n != length) {
-			if (n != 0) {
-				NDMP_LOG(LOG_DEBUG, "LEOT n: %d", n);
-
-				NDMP_LOG(LOG_DEBUG, "Backup one record");
-				(void) ndmp_mtioctl(session->ns_tape.td_fd,
-				    MTBSR, 1);
-
-				/* setting logical EOM */
-				ndmpd_write_eom(session->ns_tape.td_fd);
-			}
-
-			/*
-			 * End of media reached.
-			 * Notify client and wait for the client to
-			 * either abort the operation or continue the
-			 * operation after changing the tape.
-			 */
-			NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
-			    ++ndmp_log_msg_id,
-			    "End of tape reached. Load next tape.\n");
-
-			err = change_tape(session);
-
-			/* Operation aborted or connection terminated? */
-			if (err < 0)
-				return (-1);
-
-			/* Retry the write to the new tape. */
-			continue;
-		}
-
-		session->ns_tape.td_record_count++;
-		return (n);
-	}
-}
-
-
-/*
  * tape_read
  *
  * Reads a data record from tape. Detects and handles EOT conditions.
@@ -3143,7 +3027,8 @@ mover_tape_write_one_buf(ndmpd_session_t *session, tlm_buffer_t *buf)
 	    buf->tb_full, buf->tb_eot, buf->tb_eof, buf->tb_errno,
 	    buf->tb_buffer_size, buf->tb_buffer_data);
 
-	n = tape_write(session, buf->tb_buffer_data, buf->tb_buffer_size);
+	n = mover_tape_write_v3(session, buf->tb_buffer_data,
+	    buf->tb_buffer_size);
 
 	NDMP_LOG(LOG_DEBUG, "n: %d", n);
 
@@ -3668,37 +3553,9 @@ static int
 mover_tape_write_v3(ndmpd_session_t *session, char *data, ssize_t length)
 {
 	ssize_t n;
-	int err;
-
-	for (; ; ) {
-		/*
-		 * Refer to the comment at the top of ndmpd_tape.c file for
-		 * Mammoth2 tape drives.
-		 */
-		if (session->ns_tape.td_eom_seen) {
-			NDMP_LOG(LOG_DEBUG, "eom_seen");
-
-			session->ns_tape.td_eom_seen = FALSE;
-			/*
-			 * End of media reached.
-			 * Notify client and wait for the client to
-			 * either abort the operation or continue the
-			 * operation after changing the tape.
-			 */
-			NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
-			    ++ndmp_log_msg_id,
-			    "End of tape reached. Load next tape");
-
-			err = mover_pause_v3(session, NDMP_MOVER_PAUSE_EOM);
-
-			/* Operation aborted or connection terminated? */
-			if (err < 0)
-				return (-1);
-
-			/* Retry the write to the new tape. */
-			continue;
-		}
+	ssize_t count = length;
 
+	while (count > 0) {
 		/*
 		 * Enforce mover window on write.
 		 */
@@ -3707,57 +3564,49 @@ mover_tape_write_v3(ndmpd_session_t *session, char *data, ssize_t length)
 		    session->ns_mover.md_window_length) {
 			NDMP_LOG(LOG_DEBUG, "MOVER_PAUSE_EOW");
 
-			err = mover_pause_v3(session, NDMP_MOVER_PAUSE_EOW);
-			/* Operation aborted or connection terminated? */
-			if (err < 0)
+			if (mover_pause_v3(session, NDMP_MOVER_PAUSE_EOW) < 0)
+				/* Operation aborted or connection terminated */
 				return (-1);
 
 		}
 
-		n = write(session->ns_tape.td_fd, data, length);
+		n = write(session->ns_tape.td_fd, data, count);
 		if (n < 0) {
 			NDMP_LOG(LOG_ERR, "Tape write error: %m.");
 			return (-1);
+		} else if (n > 0) {
+			NS_ADD(wtape, n);
+			count -= n;
+			data += n;
+			session->ns_tape.td_record_count++;
 		}
-		NS_ADD(wtape, n);
 
-		if (n == 0 || n != length) {
-			if (n != 0) {
-				/*
-				 * Backup one record since the record
-				 * hits the EOM.
-				 */
-				NDMP_LOG(LOG_DEBUG, "Back up one record");
-				(void) ndmp_mtioctl(session->ns_tape.td_fd,
-				    MTBSR, 1);
+		/* EOM handling */
+		if (count > 0) {
+			struct mtget mtstatus;
 
-				/* setting logical EOM */
-				ndmpd_write_eom(session->ns_tape.td_fd);
-			}
+			(void) ioctl(session->ns_tape.td_fd, MTIOCGET,
+			    &mtstatus);
+			NDMP_LOG(LOG_DEBUG, "EOM detected (%d written bytes, "
+			    "mover record %d, file #%d, block #%d)", n,
+			    session->ns_tape.td_record_count,
+			    mtstatus.mt_fileno, mtstatus.mt_blkno);
 
 			/*
-			 * End of media reached.
-			 * Notify client and wait for the client to
-			 * either abort the operation or continue the
-			 * operation after changing the tape.
+			 * Notify the client to either abort the operation
+			 * or change the tape.
 			 */
 			NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
 			    ++ndmp_log_msg_id,
 			    "End of tape reached. Load next tape");
 
-			err = mover_pause_v3(session, NDMP_MOVER_PAUSE_EOM);
-
-			/* Operation aborted or connection terminated? */
-			if (err < 0)
+			if (mover_pause_v3(session, NDMP_MOVER_PAUSE_EOM) < 0)
+				/* Operation aborted or connection terminated */
 				return (-1);
-
-			/* Retry the write to the new tape. */
-			continue;
 		}
-
-		session->ns_tape.td_record_count++;
-		return (n);
 	}
+
+	return (length);
 }
 
 
@@ -3904,7 +3753,7 @@ ndmpd_local_write_v3(ndmpd_session_t *session, char *data, ulong_t length)
 			n = mover_tape_write_v3(session,
 			    session->ns_mover.md_buf,
 			    session->ns_mover.md_record_size);
-			if (n < 0) {
+			if (n <= 0) {
 				ndmpd_mover_error(session,
 				    (n == 0 ?  NDMP_MOVER_HALT_ABORTED :
 				    NDMP_MOVER_HALT_MEDIA_ERROR));
@@ -4017,59 +3866,84 @@ mover_data_read_v3(void *cookie, int fd, ulong_t mode)
 static int
 mover_tape_read_v3(ndmpd_session_t *session, char *data)
 {
+	int pause_reason;
 	ssize_t	 n;
 	int err;
 	int count;
 
 	count = session->ns_mover.md_record_size;
-	for (; ; ) {
+	while (count > 0) {
+		pause_reason = NDMP_MOVER_PAUSE_NA;
+
 		n = read(session->ns_tape.td_fd, data, count);
 		if (n < 0) {
-			NDMP_LOG(LOG_ERR, "Tape read error: %m.");
-			return (TAPE_READ_ERR);
-		}
-		NS_ADD(rtape, n);
-
-		if (n == 0) {
+			/*
+			 * If at beginning of file and read fails with EIO,
+			 * then it's repeated attempt to read at EOT.
+			 */
+			if (errno == EIO && tape_is_at_bof(session)) {
+				NDMP_LOG(LOG_DEBUG, "Repeated read at EOT");
+				pause_reason = NDMP_MOVER_PAUSE_EOM;
+				NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
+				    ++ndmp_log_msg_id,
+				    "End of tape reached. Load next tape");
+			}
+			/*
+			 * According to NDMPv4 spec preferred error code when
+			 * trying to read from blank tape is NDMP_EOM_ERR.
+			 */
+			else if (errno == EIO && tape_is_at_bot(session)) {
+				NDMP_LOG(LOG_ERR,
+				    "Blank tape detected, returning EOM");
+				NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
+				    ++ndmp_log_msg_id,
+				    "Blank tape. Load another tape");
+				pause_reason = NDMP_MOVER_PAUSE_EOM;
+			} else {
+				NDMP_LOG(LOG_ERR, "Tape read error: %m.");
+				return (TAPE_READ_ERR);
+			}
+		} else if (n > 0) {
+			NS_ADD(rtape, n);
+			data += n;
+			count -= n;
+			session->ns_tape.td_record_count++;
+		} else {
 			if (!is_writer_running_v3(session))
 				return (TAPE_NO_WRITER_ERR);
 
 			/*
-			 * End of media reached.
-			 * Notify client and wait for the client to
-			 * either abort the data operation or continue the
-			 * operation after changing the tape.
+			 * End of file or media reached. Notify client and
+			 * wait for the client to either abort the data
+			 * operation or continue the operation after changing
+			 * the tape.
 			 */
-			NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
-			    ++ndmp_log_msg_id,
-			    "End of tape reached. Load next tape");
+			if (tape_is_at_bof(session)) {
+				NDMP_LOG(LOG_DEBUG, "EOT detected");
+				pause_reason = NDMP_MOVER_PAUSE_EOM;
+				NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
+				    ++ndmp_log_msg_id, "End of medium reached");
+			} else {
+				NDMP_LOG(LOG_DEBUG, "EOF detected");
+				/* reposition the tape to BOT side of FM */
+				fm_dance(session);
+				pause_reason = NDMP_MOVER_PAUSE_EOF;
+				NDMP_APILOG((void*)session, NDMP_LOG_NORMAL,
+				    ++ndmp_log_msg_id, "End of file reached.");
+			}
+		}
 
-			err = mover_pause_v3(session, NDMP_MOVER_PAUSE_EOF);
+		if (pause_reason != NDMP_MOVER_PAUSE_NA) {
+			err = mover_pause_v3(session, pause_reason);
 
 			/* Operation aborted or connection terminated? */
 			if (err < 0) {
-				/*
-				 * Back up one record if it's read but not
-				 * used.
-				 */
-				if (count != session->ns_mover.md_record_size)
-					(void) ndmp_mtioctl(
-					    session->ns_tape.td_fd, MTBSR, 1);
 				return (0);
 			}
-
-			/* Retry the read from the new tape. */
-			continue;
-		}
-
-		data += n;
-		count -= n;
-		if (count <= 0) {
-			session->ns_mover.md_record_num++;
-			session->ns_tape.td_record_count++;
-			return (n);
+			/* Retry the read from new location */
 		}
 	}
+	return (session->ns_mover.md_record_size);
 }
 
 
@@ -4168,6 +4042,7 @@ mover_data_write_v3(void *cookie, int fd, ulong_t mode)
 		}
 
 		session->ns_mover.md_w_index = n;
+		session->ns_mover.md_record_num++;
 	}
 
 	/*
@@ -4540,6 +4415,7 @@ ndmpd_local_read_v3(ndmpd_session_t *session, char *data, ulong_t length)
 			count += n;
 			session->ns_mover.md_bytes_left_to_read -= n;
 			session->ns_mover.md_position += n;
+			session->ns_mover.md_record_num++;
 			continue;
 		}
 
@@ -4557,6 +4433,7 @@ ndmpd_local_read_v3(ndmpd_session_t *session, char *data, ulong_t length)
 
 		session->ns_mover.md_w_index = n;
 		session->ns_mover.md_r_index = 0;
+		session->ns_mover.md_record_num++;
 
 		NDMP_LOG(LOG_DEBUG, "n: %d", n);
 
diff --git a/usr/src/cmd/ndmpd/ndmp/ndmpd_tape.c b/usr/src/cmd/ndmpd/ndmp/ndmpd_tape.c
index 8dd46411f9..0d5194c27e 100644
--- a/usr/src/cmd/ndmpd/ndmp/ndmpd_tape.c
+++ b/usr/src/cmd/ndmpd/ndmp/ndmpd_tape.c
@@ -2,7 +2,6 @@
  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
  * Use is subject to license terms.
  */
-
 /*
  * BSD 3 Clause License
  *
@@ -38,6 +37,7 @@
  */
 /* Copyright (c) 2007, The Storage Networking Industry Association. */
 /* Copyright (c) 1996, 1997 PDC, Network Appliance. All Rights Reserved */
+/* Copyright 2014 Nexenta Systems, Inc.  All rights reserved. */
 
 #include <sys/param.h>
 #include <fcntl.h>
@@ -65,6 +65,29 @@ int ndmp_tape_open_retries = 5;
 int ndmp_tape_open_delay = 1000;
 
 /*
+ * A few words about EOT (end-of-tape) and EOM handling on tapes with SVR4
+ * semantic:
+ *
+ * We adhere to terminology as used in st driver.  EOT means end of recorded
+ * data on a tape. This is different from EOM (somewhere referred to as LEOT)
+ * which is the end of tape medium. EOT is meaningful only for reads while EOM
+ * is meaningful only for writes. It's not possible to read after EOT (fails
+ * with EIO), but it's possible to write data after EOM. EOM returned by st
+ * driver on modern tape drives is just indication that the physical end of
+ * tape medium is nearing and that writer should write just the necessary
+ * minimum and stop writing. When physical end of tape is reached all writes
+ * return EIO. If EOM is crossed during read operation then st driver doesn't
+ * bother to report it to client and that's alright because reads don't care
+ * where medium physically ends but they care about meaningful data recorded on
+ * the tape and as long as there are such data reads should continue to work.
+ *
+ * When reading EOT is signalled by st driver by two empty consecutive reads
+ * (with FSF done between them).  When writing EOM is signalled by empty write
+ * (a write which writes zero bytes). Following writes succeed until physical
+ * end of tape is reached in which case EIO is returned.
+ */
+
+/*
  * ************************************************************************
  * NDMP V2 HANDLERS
  * ************************************************************************
@@ -196,7 +219,6 @@ ndmpd_tape_open_v2(ndmp_connection_t *connection, void *body)
 	session->ns_tape.td_lun = lun;
 	(void) strlcpy(session->ns_tape.td_adapter_name, adptnm, SCSI_MAX_NAME);
 	session->ns_tape.td_record_count = 0;
-	session->ns_tape.td_eom_seen = FALSE;
 
 	NDMP_LOG(LOG_DEBUG, "Tape is opened fd: %d", session->ns_tape.td_fd);
 
@@ -389,6 +411,7 @@ ndmpd_tape_mtio_v2(ndmp_connection_t *connection, void *body)
 
 		do {
 			NS_UPD(twait, trun);
+			errno = 0;
 			rc = ioctl(session->ns_tape.td_fd, MTIOCTOP, &tapeop);
 			NS_UPD(trun, twait);
 			NDMP_LOG(LOG_DEBUG,
@@ -445,104 +468,6 @@ ndmpd_tape_mtio_v2(ndmp_connection_t *connection, void *body)
 
 
 /*
- * ndmpd_tape_write_v2
- *
- * This handler handles tape_write requests.
- * This interface is a non-buffered interface. Each write request
- * maps directly to a write to the tape device. It is the responsibility
- * of the NDMP client to pad the data to the desired record size.
- * It is the responsibility of the NDMP client to ensure that the
- * length is a multiple of the tape block size if the tape device
- * is in fixed block mode.
- *
- * Parameters:
- *   connection (input) - connection handle.
- *   body       (input) - request message body.
- *
- * Returns:
- *   void
- */
-void
-ndmpd_tape_write_v2(ndmp_connection_t *connection, void *body)
-{
-	ndmp_tape_write_request *request = (ndmp_tape_write_request *) body;
-	ndmp_tape_write_reply reply;
-	ndmpd_session_t *session = ndmp_get_client_data(connection);
-	ssize_t n;
-
-	reply.count = 0;
-
-	if (session->ns_tape.td_fd == -1) {
-		NDMP_LOG(LOG_ERR, "Tape device is not open.");
-		reply.error = NDMP_DEV_NOT_OPEN_ERR;
-		ndmp_send_reply(connection, (void *) &reply,
-		    "sending tape_write reply");
-		return;
-	}
-	if (session->ns_tape.td_mode == NDMP_TAPE_READ_MODE) {
-		NDMP_LOG(LOG_INFO, "Tape device opened in read-only mode");
-		reply.error = NDMP_PERMISSION_ERR;
-		ndmp_send_reply(connection, (void *) &reply,
-		    "sending tape_write reply");
-		return;
-	}
-	if (request->data_out.data_out_len == 0) {
-		reply.error = NDMP_NO_ERR;
-		ndmp_send_reply(connection, (void *) &reply,
-		    "sending tape_write reply");
-		return;
-	}
-
-	if (session->ns_tape.td_eom_seen) {
-		/*
-		 * Refer to the comment at the top of this file for
-		 * Mammoth2 tape drives.
-		 */
-		NDMP_LOG(LOG_DEBUG, "eom_seen");
-		ndmpd_write_eom(session->ns_tape.td_fd);
-
-		session->ns_tape.td_eom_seen = FALSE;
-		reply.error = NDMP_EOM_ERR;
-		ndmp_send_reply(connection, (void *) &reply,
-		    "sending tape_write reply");
-		return;
-	}
-
-	n = write(session->ns_tape.td_fd, request->data_out.data_out_val,
-	    request->data_out.data_out_len);
-	if (n >= 0) {
-		session->ns_tape.td_write = 1;
-		NS_ADD(wtape, n);
-	}
-	if (n == 0) {
-		NDMP_LOG(LOG_DEBUG, "n == 0");
-		reply.error = NDMP_EOM_ERR;
-		session->ns_tape.td_eom_seen = FALSE;
-	} else if (n < 0) {
-		NDMP_LOG(LOG_ERR, "Tape write error: %m.");
-		reply.error = NDMP_IO_ERR;
-	} else {
-		reply.count = n;
-		reply.error = NDMP_NO_ERR;
-
-		/*
-		 * a logical end of tape will return number of bytes written
-		 * less than rquested, and one more request to write will
-		 * give 0, and then no-space
-		 */
-		if (n < request->data_out.data_out_len) {
-			NDMP_LOG(LOG_DEBUG, "LEOT: n: %d", n);
-			session->ns_tape.td_eom_seen = TRUE;
-		} else {
-			session->ns_tape.td_eom_seen = FALSE;
-		}
-	}
-	ndmp_send_reply(connection, &reply,
-	    "sending tape_write reply");
-}
-
-
-/*
  * ndmpd_tape_read_v2
  *
  * This handler handles tape_read requests.
@@ -593,8 +518,6 @@ ndmpd_tape_read_v2(ndmp_connection_t *connection, void *body)
 		return;
 	}
 
-	session->ns_tape.td_eom_seen = FALSE;
-
 	unbuffered_read(session, buf, request->count, &reply);
 
 	ndmp_send_reply(connection, (void *) &reply, "sending tape_read reply");
@@ -631,7 +554,6 @@ ndmpd_tape_execute_cdb_v2(ndmp_connection_t *connection, void *body)
 		ndmp_send_reply(connection, (void *) &reply,
 		    "sending tape_execute_cdb reply");
 	} else {
-		session->ns_tape.td_eom_seen = FALSE;
 		ndmp_execute_cdb(session, session->ns_tape.td_adapter_name,
 		    session->ns_tape.td_sid, session->ns_tape.td_lun,
 		    (ndmp_execute_cdb_request *)request);
@@ -752,32 +674,92 @@ ndmpd_tape_get_state_v3(ndmp_connection_t *connection, void *body)
 	    "sending tape_get_state reply");
 }
 
+/*
+ * tape_is_at_bot
+ *
+ * Returns 1 if tape is at BOT, 0 on error or not at BOT.
+ *
+ */
+int
+tape_is_at_bot(ndmpd_session_t *session)
+{
+	struct mtget mtstatus;
+
+	if (ioctl(session->ns_tape.td_fd, MTIOCGET, &mtstatus) == 0 &&
+	    mtstatus.mt_fileno == 0 && mtstatus.mt_blkno == 0)
+		return (1);
+
+	return (0);
+}
+
+/*
+ * If we are at the beginning of a file (block # is zero) and read returns
+ * zero bytes then this has to be end of recorded data on the tape. Repeated
+ * reads at EOT return EIO. In both cases (zero read and EIO read) this
+ * function should be used to test if we are at EOT.
+ *
+ * Returns 1 if tape is at BOF, 0 on error or not at BOF.
+ */
+int
+tape_is_at_bof(ndmpd_session_t *session)
+{
+	struct mtget mtstatus;
+
+	if ((ioctl(session->ns_tape.td_fd, MTIOCGET, &mtstatus) == 0) &&
+	    (mtstatus.mt_fileno > 0) && (mtstatus.mt_blkno == 0))
+		return (1);
+
+	return (0);
+}
+
+/*
+ * Skips forward over a file mark and then back before the file mark. Why is
+ * this needed? There are two reasons for it:
+ *
+ * 1) Because NDMPv4 spec requires that when EOF is encountered, the tape
+ * position should remain on BOT side of the file mark. When st driver reaches
+ * end of file get-position mtioctl reports position before file mark, however
+ * the file mark has already been read and the real position is thus after the
+ * file mark (real position as reported for example by uscsi commands). Thus we
+ * need to do FSF, which does nothing but only updates file & block counter in
+ * st driver and then BSF, which sets the position before the file mark. Thus
+ * current position as reported by scsi and mtioctl will be in sync.
+ *
+ * 2) st driver returns EIO for repeated reads at EOF while according to NDMP
+ * spec we should continue to return zero bytes until FSF is done. By skipping
+ * forward and backward, st driver will return zero bytes for the next read
+ * again and we don't need to specifically handle this case.
+ */
+void
+fm_dance(ndmpd_session_t *session)
+{
+	(void) ndmp_mtioctl(session->ns_tape.td_fd, MTFSF, 1);
+	(void) ndmp_mtioctl(session->ns_tape.td_fd, MTBSF, 1);
+}
 
 /*
  * ndmpd_tape_write_v3
  *
- * This handler handles tape_write requests.
- * This interface is a non-buffered interface. Each write request
- * maps directly to a write to the tape device. It is the responsibility
- * of the NDMP client to pad the data to the desired record size.
- * It is the responsibility of the NDMP client to ensure that the
- * length is a multiple of the tape block size if the tape device
- * is in fixed block mode.
+ * This handler handles tape_write requests.  This interface is a non-buffered
+ * interface. Each write request maps directly to a write to the tape device.
+ * It is the responsibility of the NDMP client to pad the data to the desired
+ * record size.  It is the responsibility of the NDMP client to ensure that the
+ * length is a multiple of the tape block size if the tape device is in fixed
+ * block mode.
+ *
+ * A logical end of tape will return number of bytes written less than
+ * requested, and one more request to write will give 0 and NDMP_EOM_ERR,
+ * followed by NDMP_NO_ERR until NDMP_IO_ERR when physical end of tape is
+ * reached.
  *
  * Parameters:
  *   connection (input) - connection handle.
  *   body       (input) - request message body.
- *
- * Returns:
- *   void
  */
-void
-ndmpd_tape_write_v3(ndmp_connection_t *connection, void *body)
-{
-	ndmp_tape_write_request *request = (ndmp_tape_write_request *) body;
-	ndmp_tape_write_reply reply;
-	ndmpd_session_t *session = ndmp_get_client_data(connection);
-	ssize_t n;
+void ndmpd_tape_write_v3(ndmp_connection_t *connection, void *body) {
+	ndmp_tape_write_request *request = (ndmp_tape_write_request *)body;
+	ndmp_tape_write_reply reply; ndmpd_session_t *session =
+		ndmp_get_client_data(connection); ssize_t n;
 
 	reply.count = 0;
 
@@ -816,65 +798,48 @@ ndmpd_tape_write_v3(ndmp_connection_t *connection, void *body)
 		return;
 	}
 
-	/*
-	 * Refer to the comment at the top of this file for
-	 * Mammoth2 tape drives.
-	 */
-	if (session->ns_tape.td_eom_seen) {
-		NDMP_LOG(LOG_DEBUG, "eom_seen");
-		ndmpd_write_eom(session->ns_tape.td_fd);
-		session->ns_tape.td_eom_seen = FALSE;
-		reply.error = NDMP_EOM_ERR;
-		ndmp_send_reply(connection, (void *) &reply,
-		    "sending tape_write reply");
-		return;
-	}
-
 	n = write(session->ns_tape.td_fd, request->data_out.data_out_val,
 	    request->data_out.data_out_len);
 
-	session->ns_tape.td_eom_seen = FALSE;
-	if (n >= 0) {
-		session->ns_tape.td_write = 1;
-		NS_ADD(wtape, n);
-	}
-	if (n == 0) {
-		NDMP_LOG(LOG_INFO, "EOM detected");
-		reply.error = NDMP_EOM_ERR;
-		session->ns_tape.td_eom_seen = TRUE;
-	} else if (n < 0) {
+	if (n < 0) {
 		NDMP_LOG(LOG_ERR, "Tape write error: %m.");
 		reply.error = NDMP_IO_ERR;
+	} else if (n == 0) {
+		NDMP_LOG(LOG_INFO, "EOM detected");
+		reply.error = NDMP_EOM_ERR;
 	} else {
+		NS_ADD(wtape, n);
 		reply.count = n;
 		reply.error = NDMP_NO_ERR;
+
+		if (n < request->data_out.data_out_len)
+			NDMP_LOG(LOG_DEBUG,
+				"EOM is coming (partial write of %d bytes)", n);
 	}
 
 	ndmp_send_reply(connection, (void *) &reply,
 	    "sending tape_write reply");
 }
 
-
 /*
  * ndmpd_tape_read_v3
  *
- * This handler handles tape_read requests.
- * This interface is a non-buffered interface. Each read request
- * maps directly to a read to the tape device. It is the responsibility
- * of the NDMP client to issue read requests with a length that is at
- * least as large as the record size used write the tape. The tape driver
- * always reads a full record. Data is discarded if the read request is
- * smaller than the record size.
- * It is the responsibility of the NDMP client to ensure that the
- * length is a multiple of the tape block size if the tape device
- * is in fixed block mode.
+ * This handler handles tape_read requests.  This interface is a non-buffered
+ * interface. Each read request maps directly to a read to the tape device. It
+ * is the responsibility of the NDMP client to issue read requests with a
+ * length that is at least as large as the record size used write the tape. The
+ * tape driver always reads a full record. Data is discarded if the read
+ * request is smaller than the record size.  It is the responsibility of the
+ * NDMP client to ensure that the length is a multiple of the tape block size
+ * if the tape device is in fixed block mode.
+ *
+ * A logical end of tape will return less bytes than requested, and one more
+ * request to read will give 0 and NDMP_EOM_ERR.  All subsequent reads will
+ * return NDMP_EOM_ERR until the tape is repositioned.
  *
  * Parameters:
  *   connection (input) - connection handle.
  *   body       (input) - request message body.
- *
- * Returns:
- *   void
  */
 void
 ndmpd_tape_read_v3(ndmp_connection_t *connection, void *body)
@@ -883,7 +848,7 @@ ndmpd_tape_read_v3(ndmp_connection_t *connection, void *body)
 	ndmp_tape_read_reply reply;
 	ndmpd_session_t *session = ndmp_get_client_data(connection);
 	char *buf;
-	int n, len;
+	int n;
 
 	reply.data_in.data_in_len = 0;
 
@@ -921,7 +886,6 @@ ndmpd_tape_read_v3(ndmp_connection_t *connection, void *body)
 		    "sending tape_read reply");
 		return;
 	}
-	session->ns_tape.td_eom_seen = FALSE;
 
 	n = read(session->ns_tape.td_fd, buf, request->count);
 	if (n < 0) {
@@ -931,51 +895,41 @@ ndmpd_tape_read_v3(ndmp_connection_t *connection, void *body)
 		 */
 		if (errno == ENOSPC) {
 			reply.error = NDMP_EOF_ERR;
+		}
+		/*
+		 * If at beginning of file and read fails with EIO, then it's
+		 * repeated attempt to read at EOT.
+		 */
+		else if (errno == EIO && tape_is_at_bof(session)) {
+			NDMP_LOG(LOG_DEBUG, "Repeated read at EOT");
+			reply.error = NDMP_EOM_ERR;
+		}
+		/*
+		 * According to NDMPv4 spec preferred error code when
+		 * trying to read from blank tape is NDMP_EOM_ERR.
+		 */
+		else if (errno == EIO && tape_is_at_bot(session)) {
+			NDMP_LOG(LOG_ERR, "Blank tape detected, returning EOM");
+			reply.error = NDMP_EOM_ERR;
 		} else {
 			NDMP_LOG(LOG_ERR, "Tape read error: %m.");
 			reply.error = NDMP_IO_ERR;
 		}
 	} else if (n == 0) {
-		(void) ndmp_mtioctl(session->ns_tape.td_fd, MTFSF, 1);
-
-		len = strlen(NDMP_EOM_MAGIC);
-		(void) memset(buf, 0, len);
-		n = read(session->ns_tape.td_fd, buf, len);
-		buf[len] = '\0';
-
-		NDMP_LOG(LOG_DEBUG, "Checking EOM: nread %d [%s]", n, buf);
-
-		if (strncmp(buf, NDMP_EOM_MAGIC, len) == 0) {
+		if (tape_is_at_bof(session)) {
+			NDMP_LOG(LOG_DEBUG, "EOT detected");
 			reply.error = NDMP_EOM_ERR;
-			NDMP_LOG(LOG_DEBUG, "NDMP_EOM_ERR");
 		} else {
+			/* reposition the tape to BOT side of FM */
+			fm_dance(session);
+			NDMP_LOG(LOG_DEBUG, "EOF detected");
 			reply.error = NDMP_EOF_ERR;
-			NDMP_LOG(LOG_DEBUG, "NDMP_EOF_ERR");
 		}
-		if (n > 0)
-			(void) ndmp_mtioctl(session->ns_tape.td_fd, MTBSR, 1);
 	} else {
-		/*
-		 * Symantec fix for import phase
-		 *
-		 * As import process from symantec skips filemarks
-		 * they can come across to NDMP_EOM_MAGIC and treat
-		 * it as data. This fix prevents the magic to be
-		 * sent to the client and the read will return zero bytes
-		 * and set the NDMP_EOM_ERR error. The tape should
-		 * be positioned at the EOT side of the file mark.
-		 */
-		len = strlen(NDMP_EOM_MAGIC);
-		if (n == len && strncmp(buf, NDMP_EOM_MAGIC, len) == 0) {
-			reply.error = NDMP_EOM_ERR;
-			(void) ndmp_mtioctl(session->ns_tape.td_fd, MTFSF, 1);
-			NDMP_LOG(LOG_DEBUG, "NDMP_EOM_ERR");
-		} else {
-			session->ns_tape.td_pos += n;
-			reply.data_in.data_in_len = n;
-			reply.data_in.data_in_val = buf;
-			reply.error = NDMP_NO_ERR;
-		}
+		session->ns_tape.td_pos += n;
+		reply.data_in.data_in_len = n;
+		reply.data_in.data_in_val = buf;
+		reply.error = NDMP_NO_ERR;
 		NS_ADD(rtape, n);
 	}
 
@@ -1058,21 +1012,17 @@ ndmpd_tape_get_state_v4(ndmp_connection_t *connection, void *body)
 	if (dtp.bsize == 0)
 		reply.blockno = mtstatus.mt_blkno;
 	else
-		reply.blockno = mtstatus.mt_blkno *
+		reply.blockno = mtstatus.mt_blkno /
 		    (session->ns_mover.md_record_size / dtp.bsize);
 
-	reply.total_space = long_long_to_quad(0); /* not supported */
-	reply.space_remain = long_long_to_quad(0); /* not supported */
-
+	reply.total_space = long_long_to_quad(0LL); /* not supported */
+	reply.space_remain = long_long_to_quad(0LL); /* not supported */
 	reply.soft_errors = 0;
-	reply.total_space = long_long_to_quad(0LL);
-	reply.space_remain = long_long_to_quad(0LL);
 	reply.unsupported = NDMP_TAPE_STATE_SOFT_ERRORS_INVALID |
 	    NDMP_TAPE_STATE_TOTAL_SPACE_INVALID |
 	    NDMP_TAPE_STATE_SPACE_REMAIN_INVALID |
 	    NDMP_TAPE_STATE_PARTITION_INVALID;
 
-
 	NDMP_LOG(LOG_DEBUG, "f 0x%x, fnum %d, bsize %d, bno: %d",
 	    reply.flags, reply.file_num, reply.block_size, reply.blockno);
 
@@ -1366,7 +1316,6 @@ common_tape_open(ndmp_connection_t *connection, char *devname, int ndmpmode)
 	session->ns_tape.td_lun = lun;
 	(void) strlcpy(session->ns_tape.td_adapter_name, adptnm, SCSI_MAX_NAME);
 	session->ns_tape.td_record_count = 0;
-	session->ns_tape.td_eom_seen = FALSE;
 
 	NDMP_LOG(LOG_DEBUG, "Tape is opened fd: %d", session->ns_tape.td_fd);
 
@@ -1397,11 +1346,9 @@ common_tape_close(ndmp_connection_t *connection)
 	session->ns_tape.td_fd = -1;
 	session->ns_tape.td_sid = 0;
 	session->ns_tape.td_lun = 0;
-	session->ns_tape.td_write = 0;
 	(void) memset(session->ns_tape.td_adapter_name, 0,
 	    sizeof (session->ns_tape.td_adapter_name));
 	session->ns_tape.td_record_count = 0;
-	session->ns_tape.td_eom_seen = FALSE;
 
 	reply.error = NDMP_NO_ERR;
 	ndmp_send_reply(connection, (void *) &reply,