MFC r311804: Rewrite CTL statistics in more simple and scalable way.

Instead of collecting statistics for each combination of ports and logical
units, that consumed ~45KB per LU with present number of ports, collect
separate statistics for every port and every logical unit separately, that
consume only 176 bytes per each single LU/port.  This reduces struct
ctl_lun size down to just 6KB.

Also new IOCTL API/ABI does not hardcode number of LUs/ports, and should
allow handling of very large quantities.

Old API is still enabled in stable branches for compatibility reasons.

8 files changed, 456 insertions, 367 deletions

diff --git a/sys/cam/ctl/ctl.c b/sys/cam/ctl/ctl.c
index 8962f03..30082cb 100644
--- a/sys/cam/ctl/ctl.c
+++ b/sys/cam/ctl/ctl.c
@@ -1,7 +1,7 @@
 /*-
  * Copyright (c) 2003-2009 Silicon Graphics International Corp.
  * Copyright (c) 2012 The FreeBSD Foundation
- * Copyright (c) 2015 Alexander Motin <mav@FreeBSD.org>
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Portions of this software were developed by Edward Tomasz Napierala
@@ -2558,6 +2558,7 @@ ctl_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag,
 	  struct thread *td)
 {
 	struct ctl_softc *softc = dev->si_drv1;
+	struct ctl_port *port;
 	struct ctl_lun *lun;
 	int retval;
 
@@ -2769,6 +2770,7 @@ ctl_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag,
 #endif /* CTL_IO_DELAY */
 		break;
 	}
+#ifdef CTL_LEGACY_STATS
 	case CTL_GETSTATS: {
 		struct ctl_stats *stats = (struct ctl_stats *)addr;
 		int i;
@@ -2781,26 +2783,26 @@ ctl_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag,
 		stats->status = CTL_SS_OK;
 		stats->fill_len = 0;
 		STAILQ_FOREACH(lun, &softc->lun_list, links) {
-			if (stats->fill_len + sizeof(lun->stats) >
+			if (stats->fill_len + sizeof(lun->legacy_stats) >
 			    stats->alloc_len) {
 				stats->status = CTL_SS_NEED_MORE_SPACE;
 				break;
 			}
-			retval = copyout(&lun->stats, &stats->lun_stats[i++],
-					 sizeof(lun->stats));
+			retval = copyout(&lun->legacy_stats, &stats->lun_stats[i++],
+					 sizeof(lun->legacy_stats));
 			if (retval != 0)
 				break;
-			stats->fill_len += sizeof(lun->stats);
+			stats->fill_len += sizeof(lun->legacy_stats);
 		}
 		stats->num_luns = softc->num_luns;
-#ifdef CTL_TIME_IO
-		stats->flags = CTL_STATS_FLAG_TIME_VALID;
-#else
 		stats->flags = CTL_STATS_FLAG_NONE;
+#ifdef CTL_TIME_IO
+		stats->flags |= CTL_STATS_FLAG_TIME_VALID;
 #endif
 		getnanouptime(&stats->timestamp);
 		break;
 	}
+#endif /* CTL_LEGACY_STATS */
 	case CTL_ERROR_INJECT: {
 		struct ctl_error_desc *err_desc, *new_err_desc;
 
@@ -3388,6 +3390,72 @@ ctl_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag,
 			ctl_isc_announce_port(port);
 		break;
 	}
+	case CTL_GET_LUN_STATS: {
+		struct ctl_get_io_stats *stats = (struct ctl_get_io_stats *)addr;
+		int i;
+
+		/*
+		 * XXX KDM no locking here.  If the LUN list changes,
+		 * things can blow up.
+		 */
+		i = 0;
+		stats->status = CTL_SS_OK;
+		stats->fill_len = 0;
+		STAILQ_FOREACH(lun, &softc->lun_list, links) {
+			if (lun->lun < stats->first_item)
+				continue;
+			if (stats->fill_len + sizeof(lun->stats) >
+			    stats->alloc_len) {
+				stats->status = CTL_SS_NEED_MORE_SPACE;
+				break;
+			}
+			retval = copyout(&lun->stats, &stats->stats[i++],
+					 sizeof(lun->stats));
+			if (retval != 0)
+				break;
+			stats->fill_len += sizeof(lun->stats);
+		}
+		stats->num_items = softc->num_luns;
+		stats->flags = CTL_STATS_FLAG_NONE;
+#ifdef CTL_TIME_IO
+		stats->flags |= CTL_STATS_FLAG_TIME_VALID;
+#endif
+		getnanouptime(&stats->timestamp);
+		break;
+	}
+	case CTL_GET_PORT_STATS: {
+		struct ctl_get_io_stats *stats = (struct ctl_get_io_stats *)addr;
+		int i;
+
+		/*
+		 * XXX KDM no locking here.  If the LUN list changes,
+		 * things can blow up.
+		 */
+		i = 0;
+		stats->status = CTL_SS_OK;
+		stats->fill_len = 0;
+		STAILQ_FOREACH(port, &softc->port_list, links) {
+			if (port->targ_port < stats->first_item)
+				continue;
+			if (stats->fill_len + sizeof(port->stats) >
+			    stats->alloc_len) {
+				stats->status = CTL_SS_NEED_MORE_SPACE;
+				break;
+			}
+			retval = copyout(&port->stats, &stats->stats[i++],
+					 sizeof(port->stats));
+			if (retval != 0)
+				break;
+			stats->fill_len += sizeof(port->stats);
+		}
+		stats->num_items = softc->num_ports;
+		stats->flags = CTL_STATS_FLAG_NONE;
+#ifdef CTL_TIME_IO
+		stats->flags |= CTL_STATS_FLAG_TIME_VALID;
+#endif
+		getnanouptime(&stats->timestamp);
+		break;
+	}
 	default: {
 		/* XXX KDM should we fix this? */
 #if 0
@@ -4382,7 +4450,7 @@ ctl_alloc_lun(struct ctl_softc *ctl_softc, struct ctl_lun *ctl_lun,
 	struct scsi_vpd_id_descriptor *desc;
 	struct scsi_vpd_id_t10 *t10id;
 	const char *eui, *naa, *scsiname, *uuid, *vendor, *value;
-	int lun_number, i, lun_malloced;
+	int lun_number, lun_malloced;
 	int devidlen, idlen1, idlen2 = 0, len;
 
 	if (be_lun == NULL)
@@ -4604,13 +4672,16 @@ ctl_alloc_lun(struct ctl_softc *ctl_softc, struct ctl_lun *ctl_lun,
 	ctl_softc->num_luns++;
 
 	/* Setup statistics gathering */
-	lun->stats.device_type = be_lun->lun_type;
-	lun->stats.lun_number = lun_number;
-	lun->stats.blocksize = be_lun->blocksize;
+#ifdef CTL_LEGACY_STATS
+	lun->legacy_stats.device_type = be_lun->lun_type;
+	lun->legacy_stats.lun_number = lun_number;
+	lun->legacy_stats.blocksize = be_lun->blocksize;
 	if (be_lun->blocksize == 0)
-		lun->stats.flags = CTL_LUN_STATS_NO_BLOCKSIZE;
-	for (i = 0;i < CTL_MAX_PORTS;i++)
-		lun->stats.ports[i].targ_port = i;
+		lun->legacy_stats.flags = CTL_LUN_STATS_NO_BLOCKSIZE;
+	for (len = 0; len < CTL_MAX_PORTS; len++)
+		lun->legacy_stats.ports[len].targ_port = len;
+#endif /* CTL_LEGACY_STATS */
+	lun->stats.item = lun_number;
 
 	mtx_unlock(&ctl_softc->ctl_lock);
 
@@ -6676,9 +6747,7 @@ ctl_sap_log_sense_handler(struct ctl_scsiio *ctsio,
 {
 	struct ctl_lun *lun = CTL_LUN(ctsio);
 	struct stat_page *data;
-	uint64_t rn, wn, rb, wb;
-	struct bintime rt, wt;
-	int i;
+	struct bintime *t;
 
 	data = (struct stat_page *)page_index->page_data;
 
@@ -6686,28 +6755,21 @@ ctl_sap_log_sense_handler(struct ctl_scsiio *ctsio,
 	data->sap.hdr.param_control = SLP_LBIN;
 	data->sap.hdr.param_len = sizeof(struct scsi_log_stat_and_perf) -
 	    sizeof(struct scsi_log_param_header);
-	rn = wn = rb = wb = 0;
-	bintime_clear(&rt);
-	bintime_clear(&wt);
-	for (i = 0; i < CTL_MAX_PORTS; i++) {
-		rn += lun->stats.ports[i].operations[CTL_STATS_READ];
-		wn += lun->stats.ports[i].operations[CTL_STATS_WRITE];
-		rb += lun->stats.ports[i].bytes[CTL_STATS_READ];
-		wb += lun->stats.ports[i].bytes[CTL_STATS_WRITE];
-		bintime_add(&rt, &lun->stats.ports[i].time[CTL_STATS_READ]);
-		bintime_add(&wt, &lun->stats.ports[i].time[CTL_STATS_WRITE]);
-	}
-	scsi_u64to8b(rn, data->sap.read_num);
-	scsi_u64to8b(wn, data->sap.write_num);
-	if (lun->stats.blocksize > 0) {
-		scsi_u64to8b(wb / lun->stats.blocksize,
-		    data->sap.recvieved_lba);
-		scsi_u64to8b(rb / lun->stats.blocksize,
-		    data->sap.transmitted_lba);
-	}
-	scsi_u64to8b((uint64_t)rt.sec * 1000 + rt.frac / (UINT64_MAX / 1000),
+	scsi_u64to8b(lun->stats.operations[CTL_STATS_READ],
+	    data->sap.read_num);
+	scsi_u64to8b(lun->stats.operations[CTL_STATS_WRITE],
+	    data->sap.write_num);
+	if (lun->be_lun->blocksize > 0) {
+		scsi_u64to8b(lun->stats.bytes[CTL_STATS_WRITE] /
+		    lun->be_lun->blocksize, data->sap.recvieved_lba);
+		scsi_u64to8b(lun->stats.bytes[CTL_STATS_READ] /
+		    lun->be_lun->blocksize, data->sap.transmitted_lba);
+	}
+	t = &lun->stats.time[CTL_STATS_READ];
+	scsi_u64to8b((uint64_t)t->sec * 1000 + t->frac / (UINT64_MAX / 1000),
 	    data->sap.read_int);
-	scsi_u64to8b((uint64_t)wt.sec * 1000 + wt.frac / (UINT64_MAX / 1000),
+	t = &lun->stats.time[CTL_STATS_WRITE];
+	scsi_u64to8b((uint64_t)t->sec * 1000 + t->frac / (UINT64_MAX / 1000),
 	    data->sap.write_int);
 	scsi_u64to8b(0, data->sap.weighted_num);
 	scsi_u64to8b(0, data->sap.weighted_int);
@@ -13044,13 +13106,13 @@ static void
 ctl_process_done(union ctl_io *io)
 {
 	struct ctl_softc *softc = CTL_SOFTC(io);
+	struct ctl_port *port = CTL_PORT(io);
 	struct ctl_lun *lun = CTL_LUN(io);
 	void (*fe_done)(union ctl_io *io);
 	union ctl_ha_msg msg;
-	uint32_t targ_port = io->io_hdr.nexus.targ_port;
 
 	CTL_DEBUG_PRINT(("ctl_process_done\n"));
-	fe_done = softc->ctl_ports[targ_port]->fe_done;
+	fe_done = port->fe_done;
 
 #ifdef CTL_TIME_IO
 	if ((time_uptime - io->io_hdr.start_time) > ctl_time_io_secs) {
@@ -13153,11 +13215,13 @@ ctl_process_done(union ctl_io *io)
 	 */
 	if ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_SUCCESS &&
 	    io->io_hdr.io_type == CTL_IO_SCSI) {
-#ifdef CTL_TIME_IO
-		struct bintime cur_bt;
-#endif
 		int type;
+#ifdef CTL_TIME_IO
+		struct bintime bt;
 
+		getbinuptime(&bt);
+		bintime_sub(&bt, &io->io_hdr.start_bt);
+#endif
 		if ((io->io_hdr.flags & CTL_FLAG_DATA_MASK) ==
 		    CTL_FLAG_DATA_IN)
 			type = CTL_STATS_READ;
@@ -13167,18 +13231,38 @@ ctl_process_done(union ctl_io *io)
 		else
 			type = CTL_STATS_NO_IO;
 
-		lun->stats.ports[targ_port].bytes[type] +=
+#ifdef CTL_LEGACY_STATS
+		uint32_t targ_port = port->targ_port;
+		lun->legacy_stats.ports[targ_port].bytes[type] +=
 		    io->scsiio.kern_total_len;
-		lun->stats.ports[targ_port].operations[type]++;
+		lun->legacy_stats.ports[targ_port].operations[type] ++;
+		lun->legacy_stats.ports[targ_port].num_dmas[type] +=
+		    io->io_hdr.num_dmas;
 #ifdef CTL_TIME_IO
-		bintime_add(&lun->stats.ports[targ_port].dma_time[type],
+		bintime_add(&lun->legacy_stats.ports[targ_port].dma_time[type],
 		   &io->io_hdr.dma_bt);
-		getbinuptime(&cur_bt);
-		bintime_sub(&cur_bt, &io->io_hdr.start_bt);
-		bintime_add(&lun->stats.ports[targ_port].time[type], &cur_bt);
+		bintime_add(&lun->legacy_stats.ports[targ_port].time[type],
+		    &bt);
 #endif
-		lun->stats.ports[targ_port].num_dmas[type] +=
-		    io->io_hdr.num_dmas;
+#endif /* CTL_LEGACY_STATS */
+
+		lun->stats.bytes[type] += io->scsiio.kern_total_len;
+		lun->stats.operations[type] ++;
+		lun->stats.dmas[type] += io->io_hdr.num_dmas;
+#ifdef CTL_TIME_IO
+		bintime_add(&lun->stats.dma_time[type], &io->io_hdr.dma_bt);
+		bintime_add(&lun->stats.time[type], &bt);
+#endif
+
+		mtx_lock(&port->port_lock);
+		port->stats.bytes[type] += io->scsiio.kern_total_len;
+		port->stats.operations[type] ++;
+		port->stats.dmas[type] += io->io_hdr.num_dmas;
+#ifdef CTL_TIME_IO
+		bintime_add(&port->stats.dma_time[type], &io->io_hdr.dma_bt);
+		bintime_add(&port->stats.time[type], &bt);
+#endif
+		mtx_unlock(&port->port_lock);
 	}
 
 	/*
diff --git a/sys/cam/ctl/ctl_backend.h b/sys/cam/ctl/ctl_backend.h
index 4177e2d..d8ada83 100644
--- a/sys/cam/ctl/ctl_backend.h
+++ b/sys/cam/ctl/ctl_backend.h
@@ -1,6 +1,6 @@
 /*-
  * Copyright (c) 2003 Silicon Graphics International Corp.
- * Copyright (c) 2014-2015 Alexander Motin <mav@FreeBSD.org>
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -40,54 +40,7 @@
 #ifndef	_CTL_BACKEND_H_
 #define	_CTL_BACKEND_H_
 
-/*
- * XXX KDM move this to another header file?
- */
-#define	CTL_BE_NAME_LEN		32
-
-/*
- * The ID_REQ flag is used to say that the caller has requested a
- * particular LUN ID in the req_lun_id field.  If we cannot allocate that
- * LUN ID, the ctl_add_lun() call will fail.
- *
- * The STOPPED flag tells us that the LUN should default to the powered
- * off state.  It will return 0x04,0x02 until it is powered up.  ("Logical
- * unit not ready, initializing command required.")
- *
- * The NO_MEDIA flag tells us that the LUN has no media inserted.
- *
- * The PRIMARY flag tells us that this LUN is registered as a Primary LUN
- * which is accessible via the Master shelf controller in an HA. This flag
- * being set indicates a Primary LUN. This flag being reset represents a
- * Secondary LUN controlled by the Secondary controller in an HA
- * configuration. Flag is applicable at this time to T_DIRECT types. 
- *
- * The SERIAL_NUM flag tells us that the serial_num field is filled in and
- * valid for use in SCSI INQUIRY VPD page 0x80.
- *
- * The DEVID flag tells us that the device_id field is filled in and
- * valid for use in SCSI INQUIRY VPD page 0x83.
- *
- * The DEV_TYPE flag tells us that the device_type field is filled in.
- *
- * The EJECTED flag tells us that the removable LUN has tray open.
- *
- * The UNMAP flag tells us that this LUN supports UNMAP.
- *
- * The OFFLINE flag tells us that this LUN can not access backing store.
- */
-typedef enum {
-	CTL_LUN_FLAG_ID_REQ		= 0x01,
-	CTL_LUN_FLAG_STOPPED		= 0x02,
-	CTL_LUN_FLAG_NO_MEDIA		= 0x04,
-	CTL_LUN_FLAG_PRIMARY		= 0x08,
-	CTL_LUN_FLAG_SERIAL_NUM		= 0x10,
-	CTL_LUN_FLAG_DEVID		= 0x20,
-	CTL_LUN_FLAG_DEV_TYPE		= 0x40,
-	CTL_LUN_FLAG_UNMAP		= 0x80,
-	CTL_LUN_FLAG_EJECTED		= 0x100,
-	CTL_LUN_FLAG_READONLY		= 0x200
-} ctl_backend_lun_flags;
+#include <cam/ctl/ctl_ioctl.h>
 
 typedef enum {
 	CTL_LUN_SERSEQ_OFF,
diff --git a/sys/cam/ctl/ctl_frontend.c b/sys/cam/ctl/ctl_frontend.c
index 97f4e0b..03ee9da 100644
--- a/sys/cam/ctl/ctl_frontend.c
+++ b/sys/cam/ctl/ctl_frontend.c
@@ -1,5 +1,6 @@
 /*-
  * Copyright (c) 2003 Silicon Graphics International Corp.
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -192,13 +193,14 @@ error:
 		mtx_unlock(&softc->ctl_lock);
 		return (retval);
 	}
+	port->targ_port = port_num;
 	port->ctl_pool_ref = pool;
-
 	if (port->options.stqh_first == NULL)
 		STAILQ_INIT(&port->options);
+	port->stats.item = port_num;
+	mtx_init(&port->port_lock, "CTL port", NULL, MTX_DEF);
 
 	mtx_lock(&softc->ctl_lock);
-	port->targ_port = port_num;
 	STAILQ_INSERT_TAIL(&port->frontend->port_list, port, fe_links);
 	for (tport = NULL, nport = STAILQ_FIRST(&softc->port_list);
 	    nport != NULL && nport->targ_port < port_num;
@@ -218,17 +220,11 @@ int
 ctl_port_deregister(struct ctl_port *port)
 {
 	struct ctl_softc *softc = port->ctl_softc;
-	struct ctl_io_pool *pool;
-	int retval, i;
-
-	retval = 0;
-
-	pool = (struct ctl_io_pool *)port->ctl_pool_ref;
+	struct ctl_io_pool *pool = (struct ctl_io_pool *)port->ctl_pool_ref;
+	int i;
 
-	if (port->targ_port == -1) {
-		retval = 1;
-		goto bailout;
-	}
+	if (port->targ_port == -1)
+		return (1);
 
 	mtx_lock(&softc->ctl_lock);
 	STAILQ_REMOVE(&softc->port_list, port, ctl_port, links);
@@ -251,9 +247,9 @@ ctl_port_deregister(struct ctl_port *port)
 	for (i = 0; i < port->max_initiators; i++)
 		free(port->wwpn_iid[i].name, M_CTL);
 	free(port->wwpn_iid, M_CTL);
+	mtx_destroy(&port->port_lock);
 
-bailout:
-	return (retval);
+	return (0);
 }
 
 void
diff --git a/sys/cam/ctl/ctl_frontend.h b/sys/cam/ctl/ctl_frontend.h
index 3e42ac3..37e6824 100644
--- a/sys/cam/ctl/ctl_frontend.h
+++ b/sys/cam/ctl/ctl_frontend.h
@@ -1,5 +1,6 @@
 /*-
  * Copyright (c) 2003 Silicon Graphics International Corp.
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -39,6 +40,8 @@
 #ifndef	_CTL_FRONTEND_H_
 #define	_CTL_FRONTEND_H_
 
+#include <cam/ctl/ctl_ioctl.h>
+
 typedef enum {
 	CTL_PORT_STATUS_NONE		= 0x00,
 	CTL_PORT_STATUS_ONLINE		= 0x01,
@@ -243,6 +246,8 @@ struct ctl_port {
 	struct ctl_devid *port_devid;		/* passed to CTL */
 	struct ctl_devid *target_devid;		/* passed to CTL */
 	struct ctl_devid *init_devid;		/* passed to CTL */
+	struct ctl_io_stats stats;		/* used by CTL */
+	struct mtx	port_lock;		/* used by CTL */
 	STAILQ_ENTRY(ctl_port) fe_links;	/* used by CTL */
 	STAILQ_ENTRY(ctl_port) links;		/* used by CTL */
 };
diff --git a/sys/cam/ctl/ctl_ioctl.h b/sys/cam/ctl/ctl_ioctl.h
index e8f6b87..ee4ce18 100644
--- a/sys/cam/ctl/ctl_ioctl.h
+++ b/sys/cam/ctl/ctl_ioctl.h
@@ -1,6 +1,7 @@
 /*-
  * Copyright (c) 2003 Silicon Graphics International Corp.
  * Copyright (c) 2011 Spectra Logic Corporation
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -80,6 +81,9 @@
 /* Hopefully this won't conflict with new misc devices that pop up */
 #define	CTL_MINOR	225
 
+/* Legacy statistics accumulated for every port for every LU. */
+#define CTL_LEGACY_STATS	1
+
 typedef enum {
 	CTL_DELAY_TYPE_NONE,
 	CTL_DELAY_TYPE_CONT,
@@ -117,6 +121,18 @@ typedef enum {
 #define	CTL_STATS_NUM_TYPES	3
 
 typedef enum {
+	CTL_SS_OK,
+	CTL_SS_NEED_MORE_SPACE,
+	CTL_SS_ERROR
+} ctl_stats_status;
+
+typedef enum {
+	CTL_STATS_FLAG_NONE		= 0x00,
+	CTL_STATS_FLAG_TIME_VALID	= 0x01
+} ctl_stats_flags;
+
+#ifdef CTL_LEGACY_STATS
+typedef enum {
 	CTL_LUN_STATS_NO_BLOCKSIZE	= 0x01
 } ctl_lun_stats_flags;
 
@@ -137,17 +153,6 @@ struct ctl_lun_io_stats {
 	struct ctl_lun_io_port_stats	ports[CTL_MAX_PORTS];
 };
 
-typedef enum {
-	CTL_SS_OK,
-	CTL_SS_NEED_MORE_SPACE,
-	CTL_SS_ERROR
-} ctl_stats_status;
-
-typedef enum {
-	CTL_STATS_FLAG_NONE		= 0x00,
-	CTL_STATS_FLAG_TIME_VALID	= 0x01
-} ctl_stats_flags;
-
 struct ctl_stats {
 	int			alloc_len;	/* passed to kernel */
 	struct ctl_lun_io_stats	*lun_stats;	/* passed to/from kernel */
@@ -157,6 +162,27 @@ struct ctl_stats {
 	ctl_stats_flags		flags;		/* passed to userland */
 	struct timespec		timestamp;	/* passed to userland */
 };
+#endif /* CTL_LEGACY_STATS */
+
+struct ctl_io_stats {
+	uint32_t			item;
+	uint64_t			bytes[CTL_STATS_NUM_TYPES];
+	uint64_t			operations[CTL_STATS_NUM_TYPES];
+	uint64_t			dmas[CTL_STATS_NUM_TYPES];
+	struct bintime			time[CTL_STATS_NUM_TYPES];
+	struct bintime			dma_time[CTL_STATS_NUM_TYPES];
+};
+
+struct ctl_get_io_stats {
+	struct ctl_io_stats	*stats;		/* passed to/from kernel */
+	size_t			alloc_len;	/* passed to kernel */
+	size_t			fill_len;	/* passed to userland */
+	int			first_item;	/* passed to kernel */
+	int			num_items;	/* passed to userland */
+	ctl_stats_status	status;		/* passed to userland */
+	ctl_stats_flags		flags;		/* passed to userland */
+	struct timespec		timestamp;	/* passed to userland */
+};
 
 /*
  * The types of errors that can be injected:
@@ -342,12 +368,54 @@ typedef enum {
 	CTL_LUNREQ_MODIFY,
 } ctl_lunreq_type;
 
+/*
+ * The ID_REQ flag is used to say that the caller has requested a
+ * particular LUN ID in the req_lun_id field.  If we cannot allocate that
+ * LUN ID, the ctl_add_lun() call will fail.
+ *
+ * The STOPPED flag tells us that the LUN should default to the powered
+ * off state.  It will return 0x04,0x02 until it is powered up.  ("Logical
+ * unit not ready, initializing command required.")
+ *
+ * The NO_MEDIA flag tells us that the LUN has no media inserted.
+ *
+ * The PRIMARY flag tells us that this LUN is registered as a Primary LUN
+ * which is accessible via the Master shelf controller in an HA. This flag
+ * being set indicates a Primary LUN. This flag being reset represents a
+ * Secondary LUN controlled by the Secondary controller in an HA
+ * configuration. Flag is applicable at this time to T_DIRECT types. 
+ *
+ * The SERIAL_NUM flag tells us that the serial_num field is filled in and
+ * valid for use in SCSI INQUIRY VPD page 0x80.
+ *
+ * The DEVID flag tells us that the device_id field is filled in and
+ * valid for use in SCSI INQUIRY VPD page 0x83.
+ *
+ * The DEV_TYPE flag tells us that the device_type field is filled in.
+ *
+ * The EJECTED flag tells us that the removable LUN has tray open.
+ *
+ * The UNMAP flag tells us that this LUN supports UNMAP.
+ *
+ * The OFFLINE flag tells us that this LUN can not access backing store.
+ */
+typedef enum {
+	CTL_LUN_FLAG_ID_REQ		= 0x01,
+	CTL_LUN_FLAG_STOPPED		= 0x02,
+	CTL_LUN_FLAG_NO_MEDIA		= 0x04,
+	CTL_LUN_FLAG_PRIMARY		= 0x08,
+	CTL_LUN_FLAG_SERIAL_NUM		= 0x10,
+	CTL_LUN_FLAG_DEVID		= 0x20,
+	CTL_LUN_FLAG_DEV_TYPE		= 0x40,
+	CTL_LUN_FLAG_UNMAP		= 0x80,
+	CTL_LUN_FLAG_EJECTED		= 0x100,
+	CTL_LUN_FLAG_READONLY		= 0x200
+} ctl_backend_lun_flags;
 
 /*
  * LUN creation parameters:
  *
- * flags:		Various LUN flags, see ctl_backend.h for a
- *			description of the flag values and meanings.
+ * flags:		Various LUN flags, see above.
  *
  * device_type:		The SCSI device type.  e.g. 0 for Direct Access,
  *			3 for Processor, etc.  Only certain backends may
@@ -465,6 +533,7 @@ union ctl_lunreq_data {
  * kern_be_args:	For kernel use only.
  */
 struct ctl_lun_req {
+#define	CTL_BE_NAME_LEN		32
 	char			backend[CTL_BE_NAME_LEN];
 	ctl_lunreq_type		reqtype;
 	union ctl_lunreq_data	reqdata;
@@ -761,6 +830,8 @@ struct ctl_lun_map {
 #define	CTL_PORT_REQ		_IOWR(CTL_MINOR, 0x26, struct ctl_req)
 #define	CTL_PORT_LIST		_IOWR(CTL_MINOR, 0x27, struct ctl_lun_list)
 #define	CTL_LUN_MAP		_IOW(CTL_MINOR, 0x28, struct ctl_lun_map)
+#define	CTL_GET_LUN_STATS	_IOWR(CTL_MINOR, 0x29, struct ctl_get_io_stats)
+#define	CTL_GET_PORT_STATS	_IOWR(CTL_MINOR, 0x2a, struct ctl_get_io_stats)
 
 #endif /* _CTL_IOCTL_H_ */
 
diff --git a/sys/cam/ctl/ctl_private.h b/sys/cam/ctl/ctl_private.h
index 5b71f9a..3f46769 100644
--- a/sys/cam/ctl/ctl_private.h
+++ b/sys/cam/ctl/ctl_private.h
@@ -1,6 +1,6 @@
 /*-
  * Copyright (c) 2003, 2004, 2005, 2008 Silicon Graphics International Corp.
- * Copyright (c) 2014-2015 Alexander Motin <mav@FreeBSD.org>
+ * Copyright (c) 2014-2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -404,7 +404,10 @@ struct ctl_lun {
 	struct callout			ie_callout;	/* INTERVAL TIMER */
 	struct ctl_mode_pages		mode_pages;
 	struct ctl_log_pages		log_pages;
-	struct ctl_lun_io_stats		stats;
+#ifdef CTL_LEGACY_STATS
+	struct ctl_lun_io_stats		legacy_stats;
+#endif /* CTL_LEGACY_STATS */
+	struct ctl_io_stats		stats;
 	uint32_t			res_idx;
 	uint32_t			pr_generation;
 	uint64_t			*pr_keys[CTL_MAX_PORTS];
diff --git a/usr.bin/ctlstat/ctlstat.8 b/usr.bin/ctlstat/ctlstat.8
index 77eb790..038b8ca 100644
--- a/usr.bin/ctlstat/ctlstat.8
+++ b/usr.bin/ctlstat/ctlstat.8
@@ -34,7 +34,7 @@
 .\" $Id: //depot/users/kenm/FreeBSD-test2/usr.bin/ctlstat/ctlstat.8#2 $
 .\" $FreeBSD$
 .\"
-.Dd September 21, 2015
+.Dd January 9, 2017
 .Dt CTLSTAT 8
 .Os
 .Sh NAME
@@ -118,5 +118,6 @@ every 10 seconds.
 .Xr ctld 8 ,
 .Xr iostat 8
 .Sh AUTHORS
-.An Ken Merry Aq ken@FreeBSD.org
-.An Will Andrews Aq will@FreeBSD.org
+.An Ken Merry Aq Mt ken@FreeBSD.org
+.An Will Andrews Aq Mt will@FreeBSD.org
+.An Alexander Motin Aq Mt mav@FreeBSD.org
diff --git a/usr.bin/ctlstat/ctlstat.c b/usr.bin/ctlstat/ctlstat.c
index 3587586..d55e004 100644
--- a/usr.bin/ctlstat/ctlstat.c
+++ b/usr.bin/ctlstat/ctlstat.c
@@ -1,5 +1,6 @@
 /*-
  * Copyright (c) 2004, 2008, 2009 Silicon Graphics International Corp.
+ * Copyright (c) 2017 Alexander Motin <mav@FreeBSD.org>
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -66,17 +67,17 @@ __FBSDID("$FreeBSD$");
 #include <cam/ctl/ctl_ioctl.h>
 
 /*
- * The default amount of space we allocate for LUN storage space.  We
- * dynamically allocate more if needed.
+ * The default amount of space we allocate for stats storage space.
+ * We dynamically allocate more if needed.
  */
-#define	CTL_STAT_NUM_LUNS	30
+#define	CTL_STAT_NUM_ITEMS	256
 
 /*
  * The default number of LUN selection bits we allocate.  This is large
  * because we don't currently increase it if the user specifies a LUN
  * number of 1024 or larger.
  */
-#define	CTL_STAT_LUN_BITS	1024L
+#define	CTL_STAT_BITS		1024L
 
 static const char *ctlstat_opts = "Cc:Ddhjl:n:p:tw:";
 static const char *ctlstat_usage = "Usage:  ctlstat [-CDdjht] [-l lunnum]"
@@ -101,31 +102,32 @@ typedef enum {
 #define	CTLSTAT_FLAG_FIRST_RUN		(1 << 2)
 #define	CTLSTAT_FLAG_TOTALS		(1 << 3)
 #define	CTLSTAT_FLAG_DMA_TIME		(1 << 4)
-#define	CTLSTAT_FLAG_LUN_TIME_VALID	(1 << 5)
-#define	CTLSTAT_FLAG_LUN_MASK		(1 << 6)
-#define	CTLSTAT_FLAG_PORT_MASK		(1 << 7)
+#define	CTLSTAT_FLAG_TIME_VALID		(1 << 5)
+#define	CTLSTAT_FLAG_MASK		(1 << 6)
+#define	CTLSTAT_FLAG_LUNS		(1 << 7)
+#define	CTLSTAT_FLAG_PORTS		(1 << 8)
 #define	F_CPU(ctx) ((ctx)->flags & CTLSTAT_FLAG_CPU)
 #define	F_HDR(ctx) ((ctx)->flags & CTLSTAT_FLAG_HEADER)
 #define	F_FIRST(ctx) ((ctx)->flags & CTLSTAT_FLAG_FIRST_RUN)
 #define	F_TOTALS(ctx) ((ctx)->flags & CTLSTAT_FLAG_TOTALS)
 #define	F_DMA(ctx) ((ctx)->flags & CTLSTAT_FLAG_DMA_TIME)
-#define	F_LUNVAL(ctx) ((ctx)->flags & CTLSTAT_FLAG_LUN_TIME_VALID)
-#define	F_LUNMASK(ctx) ((ctx)->flags & CTLSTAT_FLAG_LUN_MASK)
-#define	F_PORTMASK(ctx) ((ctx)->flags & CTLSTAT_FLAG_PORT_MASK)
+#define	F_TIMEVAL(ctx) ((ctx)->flags & CTLSTAT_FLAG_TIME_VALID)
+#define	F_MASK(ctx) ((ctx)->flags & CTLSTAT_FLAG_MASK)
+#define	F_LUNS(ctx) ((ctx)->flags & CTLSTAT_FLAG_LUNS)
+#define	F_PORTS(ctx) ((ctx)->flags & CTLSTAT_FLAG_PORTS)
 
 struct ctlstat_context {
 	ctlstat_mode_types mode;
 	int flags;
-	struct ctl_lun_io_stats *cur_lun_stats, *prev_lun_stats,
-		*tmp_lun_stats;
-	struct ctl_lun_io_stats cur_total_stats[3], prev_total_stats[3];
+	struct ctl_io_stats *cur_stats, *prev_stats;
+	struct ctl_io_stats cur_total_stats[3], prev_total_stats[3];
 	struct timespec cur_time, prev_time;
 	struct ctl_cpu_stats cur_cpu, prev_cpu;
 	uint64_t cur_total_jiffies, prev_total_jiffies;
 	uint64_t cur_idle, prev_idle;
-	bitstr_t bit_decl(lun_mask, CTL_STAT_LUN_BITS);
-	bitstr_t bit_decl(port_mask, CTL_MAX_PORTS);
-	int num_luns;
+	bitstr_t bit_decl(item_mask, CTL_STAT_BITS);
+	int cur_items, prev_items;
+	int cur_alloc, prev_alloc;
 	int numdevs;
 	int header_interval;
 };
@@ -135,12 +137,11 @@ struct ctlstat_context {
 #endif
 
 static void usage(int error);
-static int getstats(int fd, int *num_luns, struct ctl_lun_io_stats **xlun_stats,
-		    struct timespec *cur_time, int *lun_time_valid);
+static int getstats(int fd, int *alloc_items, int *num_items,
+    struct ctl_io_stats **xstats, struct timespec *cur_time, int *time_valid);
 static int getcpu(struct ctl_cpu_stats *cpu_stats);
-static void compute_stats(struct ctlstat_context *ctx,
-			  struct ctl_lun_io_stats *cur_stats,
-			  struct ctl_lun_io_stats *prev_stats,
+static void compute_stats(struct ctl_io_stats *cur_stats,
+			  struct ctl_io_stats *prev_stats,
 			  long double etime, long double *mbsec,
 			  long double *kb_per_transfer,
 			  long double *transfers_per_second,
@@ -155,64 +156,55 @@ usage(int error)
 }
 
 static int
-getstats(int fd, int *num_luns, struct ctl_lun_io_stats **xlun_stats,
+getstats(int fd, int *alloc_items, int *num_items, struct ctl_io_stats **stats,
 	 struct timespec *cur_time, int *flags)
 {
-	struct ctl_lun_io_stats *lun_stats;
-	struct ctl_stats stats;
-	int more_space_count;
+	struct ctl_get_io_stats get_stats;
+	int more_space_count = 0;
 
-	more_space_count = 0;
-
-	if (*num_luns == 0)
-		*num_luns = CTL_STAT_NUM_LUNS;
-
-	lun_stats = *xlun_stats;
+	if (*alloc_items == 0)
+		*alloc_items = CTL_STAT_NUM_ITEMS;
 retry:
+	if (*stats == NULL)
+		*stats = malloc(sizeof(**stats) * *alloc_items);
 
-	if (lun_stats == NULL) {
-		lun_stats = (struct ctl_lun_io_stats *)malloc(
-			sizeof(*lun_stats) * *num_luns);
-	}
-
-	memset(&stats, 0, sizeof(stats));
-	stats.alloc_len = *num_luns * sizeof(*lun_stats);
-	memset(lun_stats, 0, stats.alloc_len);
-	stats.lun_stats = lun_stats;
+	memset(&get_stats, 0, sizeof(get_stats));
+	get_stats.alloc_len = *alloc_items * sizeof(**stats);
+	memset(*stats, 0, get_stats.alloc_len);
+	get_stats.stats = *stats;
 
-	if (ioctl(fd, CTL_GETSTATS, &stats) == -1)
-		err(1, "error returned from CTL_GETSTATS ioctl");
+	if (ioctl(fd, (*flags & CTLSTAT_FLAG_PORTS) ? CTL_GET_PORT_STATS :
+	    CTL_GET_LUN_STATS, &get_stats) == -1)
+		err(1, "CTL_GET_*_STATS ioctl returned error");
 
-	switch (stats.status) {
+	switch (get_stats.status) {
 	case CTL_SS_OK:
 		break;
 	case CTL_SS_ERROR:
-		err(1, "CTL_SS_ERROR returned from CTL_GETSTATS ioctl");
+		err(1, "CTL_GET_*_STATS ioctl returned CTL_SS_ERROR");
 		break;
 	case CTL_SS_NEED_MORE_SPACE:
-		if (more_space_count > 0) {
-			errx(1, "CTL_GETSTATS returned NEED_MORE_SPACE again");
-		}
-		*num_luns = stats.num_luns;
-		free(lun_stats);
-		lun_stats = NULL;
+		if (more_space_count >= 2)
+			errx(1, "CTL_GET_*_STATS returned NEED_MORE_SPACE again");
+		*alloc_items = get_stats.num_items * 5 / 4;
+		free(*stats);
+		*stats = NULL;
 		more_space_count++;
 		goto retry;
 		break; /* NOTREACHED */
 	default:
-		errx(1, "unknown status %d returned from CTL_GETSTATS ioctl",
-		     stats.status);
+		errx(1, "CTL_GET_*_STATS ioctl returned unknown status %d",
+		     get_stats.status);
 		break;
 	}
 
-	*xlun_stats = lun_stats;
-	*num_luns = stats.num_luns;
-	cur_time->tv_sec = stats.timestamp.tv_sec;
-	cur_time->tv_nsec = stats.timestamp.tv_nsec;
-	if (stats.flags & CTL_STATS_FLAG_TIME_VALID)
-		*flags |= CTLSTAT_FLAG_LUN_TIME_VALID;
+	*num_items = get_stats.fill_len / sizeof(**stats);
+	cur_time->tv_sec = get_stats.timestamp.tv_sec;
+	cur_time->tv_nsec = get_stats.timestamp.tv_nsec;
+	if (get_stats.flags & CTL_STATS_FLAG_TIME_VALID)
+		*flags |= CTLSTAT_FLAG_TIME_VALID;
 	else
-		*flags &= ~CTLSTAT_FLAG_LUN_TIME_VALID;
+		*flags &= ~CTLSTAT_FLAG_TIME_VALID;
 
 	return (0);
 }
@@ -240,14 +232,13 @@ getcpu(struct ctl_cpu_stats *cpu_stats)
 }
 
 static void
-compute_stats(struct ctlstat_context *ctx, struct ctl_lun_io_stats *cur_stats,
-	      struct ctl_lun_io_stats *prev_stats, long double etime,
+compute_stats(struct ctl_io_stats *cur_stats,
+	      struct ctl_io_stats *prev_stats, long double etime,
 	      long double *mbsec, long double *kb_per_transfer,
 	      long double *transfers_per_second, long double *ms_per_transfer,
 	      long double *ms_per_dma, long double *dmas_per_second)
 {
 	uint64_t total_bytes = 0, total_operations = 0, total_dmas = 0;
-	uint32_t port;
 	struct bintime total_time_bt, total_dma_bt;
 	struct timespec total_time_ts, total_dma_ts;
 	int i;
@@ -256,31 +247,18 @@ compute_stats(struct ctlstat_context *ctx, struct ctl_lun_io_stats *cur_stats,
 	bzero(&total_dma_bt, sizeof(total_dma_bt));
 	bzero(&total_time_ts, sizeof(total_time_ts));
 	bzero(&total_dma_ts, sizeof(total_dma_ts));
-	for (port = 0; port < CTL_MAX_PORTS; port++) {
-		if (F_PORTMASK(ctx) &&
-		    bit_test(ctx->port_mask, port) == 0)
-			continue;
-		for (i = 0; i < CTL_STATS_NUM_TYPES; i++) {
-			total_bytes += cur_stats->ports[port].bytes[i];
-			total_operations +=
-			    cur_stats->ports[port].operations[i];
-			total_dmas += cur_stats->ports[port].num_dmas[i];
-			bintime_add(&total_time_bt,
-			    &cur_stats->ports[port].time[i]);
-			bintime_add(&total_dma_bt,
-			    &cur_stats->ports[port].dma_time[i]);
-			if (prev_stats != NULL) {
-				total_bytes -=
-				    prev_stats->ports[port].bytes[i];
-				total_operations -=
-				    prev_stats->ports[port].operations[i];
-				total_dmas -=
-				    prev_stats->ports[port].num_dmas[i];
-				bintime_sub(&total_time_bt,
-				    &prev_stats->ports[port].time[i]);
-				bintime_sub(&total_dma_bt,
-				    &prev_stats->ports[port].dma_time[i]);
-			}
+	for (i = 0; i < CTL_STATS_NUM_TYPES; i++) {
+		total_bytes += cur_stats->bytes[i];
+		total_operations += cur_stats->operations[i];
+		total_dmas += cur_stats->dmas[i];
+		bintime_add(&total_time_bt, &cur_stats->time[i]);
+		bintime_add(&total_dma_bt, &cur_stats->dma_time[i]);
+		if (prev_stats != NULL) {
+			total_bytes -= prev_stats->bytes[i];
+			total_operations -= prev_stats->operations[i];
+			total_dmas -= prev_stats->dmas[i];
+			bintime_sub(&total_time_bt, &prev_stats->time[i]);
+			bintime_sub(&total_dma_bt, &prev_stats->dma_time[i]);
 		}
 	}
 
@@ -340,35 +318,25 @@ compute_stats(struct ctlstat_context *ctx, struct ctl_lun_io_stats *cur_stats,
 static const char *iotypes[] = {"NO IO", "READ", "WRITE"};
 
 static void
-ctlstat_dump(struct ctlstat_context *ctx) {
-	int iotype, lun, port;
-	struct ctl_lun_io_stats *stats = ctx->cur_lun_stats;
+ctlstat_dump(struct ctlstat_context *ctx)
+{
+	int iotype, i;
+	struct ctl_io_stats *stats = ctx->cur_stats;
 
-	for (lun = 0; lun < ctx->num_luns;lun++) {
-		if (F_LUNMASK(ctx) && bit_test(ctx->lun_mask, lun) == 0)
+	for (i = 0; i < ctx->cur_items;i++) {
+		if (F_MASK(ctx) && bit_test(ctx->item_mask, i) == 0)
 			continue;
-		printf("lun %d\n", lun);
-		for (port = 0; port < CTL_MAX_PORTS; port++) {
-			if (F_PORTMASK(ctx) &&
-			    bit_test(ctx->port_mask, port) == 0)
-				continue;
-			printf(" port %d\n",
-			    stats[lun].ports[port].targ_port);
-			for (iotype = 0; iotype < CTL_STATS_NUM_TYPES;
-			    iotype++) {
-				printf("  io type %d (%s)\n", iotype,
-				    iotypes[iotype]);
-				printf("   bytes %ju\n", (uintmax_t)
-				    stats[lun].ports[port].bytes[iotype]);
-				printf("   operations %ju\n", (uintmax_t)
-				    stats[lun].ports[port].operations[iotype]);
-				PRINT_BINTIME("   io time",
-				    stats[lun].ports[port].time[iotype]);
-				printf("   num dmas %ju\n", (uintmax_t)
-				    stats[lun].ports[port].num_dmas[iotype]);
-				PRINT_BINTIME("   dma time",
-				    stats[lun].ports[port].dma_time[iotype]);
-			}
+		printf("%s %d\n", F_PORTS(ctx) ? "port" : "lun", stats[i].item);
+		for (iotype = 0; iotype < CTL_STATS_NUM_TYPES; iotype++) {
+			printf("  io type %d (%s)\n", iotype, iotypes[iotype]);
+			printf("   bytes %ju\n", (uintmax_t)
+			    stats[i].bytes[iotype]);
+			printf("   operations %ju\n", (uintmax_t)
+			    stats[i].operations[iotype]);
+			printf("   dmas %ju\n", (uintmax_t)
+			    stats[i].dmas[iotype]);
+			PRINT_BINTIME("   io time", stats[i].time[iotype]);
+			PRINT_BINTIME("   dma time", stats[i].dma_time[iotype]);
 		}
 	}
 }
@@ -378,63 +346,49 @@ ctlstat_dump(struct ctlstat_context *ctx) {
 	    (uintmax_t)(((bt).frac >> 32) * 1000000 >> 32))
 static void
 ctlstat_json(struct ctlstat_context *ctx) {
-	int iotype, lun, port;
-	struct ctl_lun_io_stats *stats = ctx->cur_lun_stats;
+	int iotype, i;
+	struct ctl_io_stats *stats = ctx->cur_stats;
 
-	printf("{\"luns\":[");
-	for (lun = 0; lun < ctx->num_luns; lun++) {
-		if (F_LUNMASK(ctx) && bit_test(ctx->lun_mask, lun) == 0)
+	printf("{\"%s\":[", F_PORTS(ctx) ? "ports" : "luns");
+	for (i = 0; i < ctx->cur_items; i++) {
+		if (F_MASK(ctx) && bit_test(ctx->item_mask, i) == 0)
 			continue;
-		printf("{\"ports\":[");
-		for (port = 0; port < CTL_MAX_PORTS;port++) {
-			if (F_PORTMASK(ctx) &&
-			    bit_test(ctx->port_mask, port) == 0)
-				continue;
-			printf("{\"num\":%d,\"io\":[",
-			    stats[lun].ports[port].targ_port);
-			for (iotype = 0; iotype < CTL_STATS_NUM_TYPES;
-			    iotype++) {
-				printf("{\"type\":\"%s\",", iotypes[iotype]);
-				printf("\"bytes\":%ju,", (uintmax_t)stats[
-				       lun].ports[port].bytes[iotype]);
-				printf("\"operations\":%ju,", (uintmax_t)stats[
-				       lun].ports[port].operations[iotype]);
-				JSON_BINTIME("io time",
-				    stats[lun].ports[port].time[iotype]);
-				JSON_BINTIME("dma time",
-				    stats[lun].ports[port].dma_time[iotype]);
-				printf("\"num dmas\":%ju}", (uintmax_t)
-				    stats[lun].ports[port].num_dmas[iotype]);
-				if (iotype < (CTL_STATS_NUM_TYPES - 1))
-					printf(","); /* continue io array */
-			}
-			printf("]}"); /* close port */
-			if (port < (CTL_MAX_PORTS - 1))
-				printf(","); /* continue port array */
+		printf("{\"num\":%d,\"io\":[",
+		    stats[i].item);
+		for (iotype = 0; iotype < CTL_STATS_NUM_TYPES; iotype++) {
+			printf("{\"type\":\"%s\",", iotypes[iotype]);
+			printf("\"bytes\":%ju,", (uintmax_t)stats[
+			       i].bytes[iotype]);
+			printf("\"operations\":%ju,", (uintmax_t)stats[
+			       i].operations[iotype]);
+			printf("\"dmas\":%ju}", (uintmax_t)
+			    stats[i].dmas[iotype]);
+			JSON_BINTIME("io time", stats[i].time[iotype]);
+			JSON_BINTIME("dma time", stats[i].dma_time[iotype]);
+			if (iotype < (CTL_STATS_NUM_TYPES - 1))
+				printf(","); /* continue io array */
 		}
-		printf("]}"); /* close lun */
-		if (lun < (ctx->num_luns - 1))
+		printf("]}");
+		if (i < (ctx->cur_items - 1))
 			printf(","); /* continue lun array */
 	}
-	printf("]}"); /* close luns and toplevel */
+	printf("]}");
 }
 
 static void
 ctlstat_standard(struct ctlstat_context *ctx) {
 	long double etime;
 	uint64_t delta_jiffies, delta_idle;
-	uint32_t port;
 	long double cpu_percentage;
-	int i;
-	int j;
+	int i, j;
 
 	cpu_percentage = 0;
 
 	if (F_CPU(ctx) && (getcpu(&ctx->cur_cpu) != 0))
 		errx(1, "error returned from getcpu()");
 
-	etime = ctx->cur_time.tv_sec - ctx->prev_time.tv_sec +                  
-	    (ctx->prev_time.tv_nsec - ctx->cur_time.tv_nsec) * 1e-9; 
+	etime = ctx->cur_time.tv_sec - ctx->prev_time.tv_sec +
+	    (ctx->prev_time.tv_nsec - ctx->cur_time.tv_nsec) * 1e-9;
 
 	if (F_CPU(ctx)) {
 		ctx->prev_total_jiffies = ctx->cur_total_jiffies;
@@ -465,29 +419,28 @@ ctlstat_standard(struct ctlstat_context *ctx) {
 			if (F_TOTALS(ctx)) {
 				fprintf(stdout, "%s     Read       %s"
 					"    Write       %s    Total\n",
-					(F_LUNVAL(ctx) != 0) ? "      " : "",
-					(F_LUNVAL(ctx) != 0) ? "      " : "",
-					(F_LUNVAL(ctx) != 0) ? "      " : "");
+					(F_TIMEVAL(ctx) != 0) ? "      " : "",
+					(F_TIMEVAL(ctx) != 0) ? "      " : "",
+					(F_TIMEVAL(ctx) != 0) ? "      " : "");
 				hdr_devs = 3;
 			} else {
-				for (i = 0; i < min(CTL_STAT_LUN_BITS,
-				     ctx->num_luns); i++) {
-					int lun;
+				for (i = 0; i < min(CTL_STAT_BITS,
+				     ctx->cur_items); i++) {
+					int item;
 
 					/*
 					 * Obviously this won't work with
 					 * LUN numbers greater than a signed
 					 * integer.
 					 */
-					lun = (int)ctx->cur_lun_stats[i
-						].lun_number;
+					item = (int)ctx->cur_stats[i].item;
 
-					if (F_LUNMASK(ctx) &&
-					    bit_test(ctx->lun_mask, lun) == 0)
+					if (F_MASK(ctx) &&
+					    bit_test(ctx->item_mask, item) == 0)
 						continue;
 					fprintf(stdout, "%15.6s%d %s",
-					    "lun", lun,
-					    (F_LUNVAL(ctx) != 0) ? "     " : "");
+					    F_PORTS(ctx) ? "port" : "lun", item,
+					    (F_TIMEVAL(ctx) != 0) ? "     " : "");
 					hdr_devs++;
 				}
 				fprintf(stdout, "\n");
@@ -496,7 +449,7 @@ ctlstat_standard(struct ctlstat_context *ctx) {
 				fprintf(stdout, "    ");
 			for (i = 0; i < hdr_devs; i++)
 				fprintf(stdout, "%s KB/t   %s MB/s",
-					(F_LUNVAL(ctx) != 0) ? "    ms" : "",
+					(F_TIMEVAL(ctx) != 0) ? "    ms" : "",
 					(F_DMA(ctx) == 0) ? "tps" : "dps");
 			fprintf(stdout, "\n");
 			ctx->header_interval = 20;
@@ -519,55 +472,48 @@ ctlstat_standard(struct ctlstat_context *ctx) {
 		memset(&ctx->cur_total_stats, 0, sizeof(ctx->cur_total_stats));
 
 		/* Use macros to make the next loop more readable. */
-#define	ADD_STATS_BYTES(st, p, i, j) \
-	ctx->cur_total_stats[st].ports[p].bytes[j] += \
-	    ctx->cur_lun_stats[i].ports[p].bytes[j]
-#define	ADD_STATS_OPERATIONS(st, p, i, j) \
-	ctx->cur_total_stats[st].ports[p].operations[j] += \
-	    ctx->cur_lun_stats[i].ports[p].operations[j]
-#define	ADD_STATS_NUM_DMAS(st, p, i, j) \
-	ctx->cur_total_stats[st].ports[p].num_dmas[j] += \
-	    ctx->cur_lun_stats[i].ports[p].num_dmas[j]
-#define	ADD_STATS_TIME(st, p, i, j) \
-	bintime_add(&ctx->cur_total_stats[st].ports[p].time[j], \
-	    &ctx->cur_lun_stats[i].ports[p].time[j])
-#define	ADD_STATS_DMA_TIME(st, p, i, j) \
-	bintime_add(&ctx->cur_total_stats[st].ports[p].dma_time[j], \
-	    &ctx->cur_lun_stats[i].ports[p].dma_time[j])
-
-		for (i = 0; i < ctx->num_luns; i++) {
-			if (F_LUNMASK(ctx) && bit_test(ctx->lun_mask,
-			    (int)ctx->cur_lun_stats[i].lun_number) == 0)
+#define	ADD_STATS_BYTES(st, i, j) \
+	ctx->cur_total_stats[st].bytes[j] += \
+	    ctx->cur_stats[i].bytes[j]
+#define	ADD_STATS_OPERATIONS(st, i, j) \
+	ctx->cur_total_stats[st].operations[j] += \
+	    ctx->cur_stats[i].operations[j]
+#define	ADD_STATS_DMAS(st, i, j) \
+	ctx->cur_total_stats[st].dmas[j] += \
+	    ctx->cur_stats[i].dmas[j]
+#define	ADD_STATS_TIME(st, i, j) \
+	bintime_add(&ctx->cur_total_stats[st].time[j], \
+	    &ctx->cur_stats[i].time[j])
+#define	ADD_STATS_DMA_TIME(st, i, j) \
+	bintime_add(&ctx->cur_total_stats[st].dma_time[j], \
+	    &ctx->cur_stats[i].dma_time[j])
+
+		for (i = 0; i < ctx->cur_items; i++) {
+			if (F_MASK(ctx) && bit_test(ctx->item_mask,
+			    (int)ctx->cur_stats[i].item) == 0)
 				continue;
-			for (port = 0; port < CTL_MAX_PORTS; port++) {
-				if (F_PORTMASK(ctx) &&
-				    bit_test(ctx->port_mask, port) == 0)
-					continue;
-				for (j = 0; j < CTL_STATS_NUM_TYPES; j++) {
-					ADD_STATS_BYTES(2, port, i, j);
-					ADD_STATS_OPERATIONS(2, port, i, j);
-					ADD_STATS_NUM_DMAS(2, port, i, j);
-					ADD_STATS_TIME(2, port, i, j);
-					ADD_STATS_DMA_TIME(2, port, i, j);
-				}
-				ADD_STATS_BYTES(0, port, i, CTL_STATS_READ);
-				ADD_STATS_OPERATIONS(0, port, i,
-				    CTL_STATS_READ);
-				ADD_STATS_NUM_DMAS(0, port, i, CTL_STATS_READ);
-				ADD_STATS_TIME(0, port, i, CTL_STATS_READ);
-				ADD_STATS_DMA_TIME(0, port, i, CTL_STATS_READ);
-
-				ADD_STATS_BYTES(1, port, i, CTL_STATS_WRITE);
-				ADD_STATS_OPERATIONS(1, port, i,
-				    CTL_STATS_WRITE);
-				ADD_STATS_NUM_DMAS(1, port, i, CTL_STATS_WRITE);
-				ADD_STATS_TIME(1, port, i, CTL_STATS_WRITE);
-				ADD_STATS_DMA_TIME(1, port, i, CTL_STATS_WRITE);
+			for (j = 0; j < CTL_STATS_NUM_TYPES; j++) {
+				ADD_STATS_BYTES(2, i, j);
+				ADD_STATS_OPERATIONS(2, i, j);
+				ADD_STATS_DMAS(2, i, j);
+				ADD_STATS_TIME(2, i, j);
+				ADD_STATS_DMA_TIME(2, i, j);
 			}
+			ADD_STATS_BYTES(0, i, CTL_STATS_READ);
+			ADD_STATS_OPERATIONS(0, i, CTL_STATS_READ);
+			ADD_STATS_DMAS(0, i, CTL_STATS_READ);
+			ADD_STATS_TIME(0, i, CTL_STATS_READ);
+			ADD_STATS_DMA_TIME(0, i, CTL_STATS_READ);
+
+			ADD_STATS_BYTES(1, i, CTL_STATS_WRITE);
+			ADD_STATS_OPERATIONS(1, i, CTL_STATS_WRITE);
+			ADD_STATS_DMAS(1, i, CTL_STATS_WRITE);
+			ADD_STATS_TIME(1, i, CTL_STATS_WRITE);
+			ADD_STATS_DMA_TIME(1, i, CTL_STATS_WRITE);
 		}
 
 		for (i = 0; i < 3; i++) {
-			compute_stats(ctx, &ctx->cur_total_stats[i],
+			compute_stats(&ctx->cur_total_stats[i],
 				F_FIRST(ctx) ? NULL : &ctx->prev_total_stats[i],
 				etime, &mbsec[i], &kb_per_transfer[i],
 				&transfers_per_sec[i],
@@ -576,7 +522,7 @@ ctlstat_standard(struct ctlstat_context *ctx) {
 			if (F_DMA(ctx) != 0)
 				fprintf(stdout, " %5.1Lf",
 					ms_per_dma[i]);
-			else if (F_LUNVAL(ctx) != 0)
+			else if (F_TIMEVAL(ctx) != 0)
 				fprintf(stdout, " %5.1Lf",
 					ms_per_transfer[i]);
 			fprintf(stdout, " %4.0Lf %5.0Lf %4.0Lf",
@@ -585,25 +531,32 @@ ctlstat_standard(struct ctlstat_context *ctx) {
 				dmas_per_sec[i], mbsec[i]);
 		}
 	} else {
-		for (i = 0; i < min(CTL_STAT_LUN_BITS, ctx->num_luns); i++) {
+		for (i = 0; i < min(CTL_STAT_BITS, ctx->cur_items); i++) {
 			long double mbsec, kb_per_transfer;
 			long double transfers_per_sec;
 			long double ms_per_transfer;
 			long double ms_per_dma;
 			long double dmas_per_sec;
 
-			if (F_LUNMASK(ctx) && bit_test(ctx->lun_mask,
-			    (int)ctx->cur_lun_stats[i].lun_number) == 0)
+			if (F_MASK(ctx) && bit_test(ctx->item_mask,
+			    (int)ctx->cur_stats[i].item) == 0)
 				continue;
-			compute_stats(ctx, &ctx->cur_lun_stats[i],
-			    F_FIRST(ctx) ? NULL : &ctx->prev_lun_stats[i],
+			for (j = 0; j < ctx->prev_items; j++) {
+				if (ctx->prev_stats[j].item ==
+				    ctx->cur_stats[i].item)
+					break;
+			}
+			if (j >= ctx->prev_items)
+				j = -1;
+			compute_stats(&ctx->cur_stats[i],
+			    j >= 0 ? &ctx->prev_stats[j] : NULL,
 			    etime, &mbsec, &kb_per_transfer,
 			    &transfers_per_sec, &ms_per_transfer,
 			    &ms_per_dma, &dmas_per_sec);
 			if (F_DMA(ctx))
 				fprintf(stdout, " %5.1Lf",
 					ms_per_dma);
-			else if (F_LUNVAL(ctx) != 0)
+			else if (F_TIMEVAL(ctx) != 0)
 				fprintf(stdout, " %5.1Lf",
 					ms_per_transfer);
 			fprintf(stdout, " %4.0Lf %5.0Lf %4.0Lf",
@@ -620,6 +573,7 @@ main(int argc, char **argv)
 	int count, waittime;
 	int fd, retval;
 	struct ctlstat_context ctx;
+	struct ctl_io_stats *tmp_stats;
 
 	/* default values */
 	retval = 0;
@@ -658,15 +612,16 @@ main(int argc, char **argv)
 			int cur_lun;
 
 			cur_lun = atoi(optarg);
-			if (cur_lun > CTL_STAT_LUN_BITS)
+			if (cur_lun > CTL_STAT_BITS)
 				errx(1, "Invalid LUN number %d", cur_lun);
 
-			if (!F_LUNMASK(&ctx))
+			if (!F_MASK(&ctx))
 				ctx.numdevs = 1;
 			else
 				ctx.numdevs++;
-			bit_set(ctx.lun_mask, cur_lun);
-			ctx.flags |= CTLSTAT_FLAG_LUN_MASK;
+			bit_set(ctx.item_mask, cur_lun);
+			ctx.flags |= CTLSTAT_FLAG_MASK;
+			ctx.flags |= CTLSTAT_FLAG_LUNS;
 			break;
 		}
 		case 'n':
@@ -676,11 +631,16 @@ main(int argc, char **argv)
 			int cur_port;
 
 			cur_port = atoi(optarg);
-			if (cur_port > CTL_MAX_PORTS)
-				errx(1, "Invalid LUN number %d", cur_port);
+			if (cur_port > CTL_STAT_BITS)
+				errx(1, "Invalid port number %d", cur_port);
 
-			bit_set(ctx.port_mask, cur_port);
-			ctx.flags |= CTLSTAT_FLAG_PORT_MASK;
+			if (!F_MASK(&ctx))
+				ctx.numdevs = 1;
+			else
+				ctx.numdevs++;
+			bit_set(ctx.item_mask, cur_port);
+			ctx.flags |= CTLSTAT_FLAG_MASK;
+			ctx.flags |= CTLSTAT_FLAG_PORTS;
 			break;
 		}
 		case 't':
@@ -697,29 +657,45 @@ main(int argc, char **argv)
 		}
 	}
 
-	if (!F_TOTALS(&ctx) && !F_LUNMASK(&ctx)) {
+	if (F_LUNS(&ctx) && F_PORTS(&ctx))
+		errx(1, "Options -p and -l are exclusive.");
+
+	if (!F_LUNS(&ctx) && !F_PORTS(&ctx)) {
+		if (F_TOTALS(&ctx))
+			ctx.flags |= CTLSTAT_FLAG_PORTS;
+		else
+			ctx.flags |= CTLSTAT_FLAG_LUNS;
+	}
+
+	if (!F_TOTALS(&ctx) && !F_MASK(&ctx)) {
 		/*
 		 * Note that this just selects the first N LUNs to display,
 		 * but at this point we have no knoweledge of which LUN
 		 * numbers actually exist.  So we may select LUNs that
 		 * aren't there.
 		 */
-		bit_nset(ctx.lun_mask, 0, min(ctx.numdevs - 1,
-			 CTL_STAT_LUN_BITS - 1));
-		ctx.flags |= CTLSTAT_FLAG_LUN_MASK;
+		bit_nset(ctx.item_mask, 0, min(ctx.numdevs - 1,
+			 CTL_STAT_BITS - 1));
+		ctx.flags |= CTLSTAT_FLAG_MASK;
 	}
 
 	if ((fd = open(CTL_DEFAULT_DEV, O_RDWR)) == -1)
 		err(1, "cannot open %s", CTL_DEFAULT_DEV);
 
 	for (;count != 0;) {
-		ctx.tmp_lun_stats = ctx.prev_lun_stats;
-		ctx.prev_lun_stats = ctx.cur_lun_stats;
-		ctx.cur_lun_stats = ctx.tmp_lun_stats;
+		tmp_stats = ctx.prev_stats;
+		ctx.prev_stats = ctx.cur_stats;
+		ctx.cur_stats = tmp_stats;
+		c = ctx.prev_alloc;
+		ctx.prev_alloc = ctx.cur_alloc;
+		ctx.cur_alloc = c;
+		c = ctx.prev_items;
+		ctx.prev_items = ctx.cur_items;
+		ctx.cur_items = c;
 		ctx.prev_time = ctx.cur_time;
 		ctx.prev_cpu = ctx.cur_cpu;
-		if (getstats(fd, &ctx.num_luns, &ctx.cur_lun_stats,
-			     &ctx.cur_time, &ctx.flags) != 0)
+		if (getstats(fd, &ctx.cur_alloc, &ctx.cur_items,
+		    &ctx.cur_stats, &ctx.cur_time, &ctx.flags) != 0)
 			errx(1, "error returned from getstats()");
 
 		switch(ctx.mode) {