diff --git a/sys/cam/ata/ata_da.c b/sys/cam/ata/ata_da.c
--- a/sys/cam/ata/ata_da.c
+++ b/sys/cam/ata/ata_da.c
@@ -297,6 +297,8 @@
 	char	announce_buffer[ADA_ANNOUNCE_SZ];
 };
 
+static uma_zone_t ada_ccb_zone;
+
 struct ada_quirk_entry {
 	struct scsi_inquiry_pattern inq_pat;
 	ada_quirks quirks;
@@ -902,6 +904,7 @@
 static int ada_read_ahead = ADA_DEFAULT_READ_AHEAD;
 static int ada_write_cache = ADA_DEFAULT_WRITE_CACHE;
 static int ada_enable_biospeedup = 1;
+static int ada_enable_uma_ccbs = 0;
 
 static SYSCTL_NODE(_kern_cam, OID_AUTO, ada, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
     "CAM Direct Access Disk driver");
@@ -921,6 +924,8 @@
            &ada_write_cache, 0, "Enable disk write cache");
 SYSCTL_INT(_kern_cam_ada, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
 	   &ada_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing");
+SYSCTL_INT(_kern_cam_ada, OID_AUTO, enable_uma_ccbs, CTLFLAG_RWTUN,
+	    &ada_enable_uma_ccbs, 0, "Use UMA for CCBs");
 
 /*
  * ADA_ORDEREDTAG_INTERVAL determines how often, relative
@@ -1178,6 +1183,10 @@
 {
 	cam_status status;
 
+	ada_ccb_zone = uma_zcreate("ada_ccb",
+	    sizeof(struct ccb_ataio), NULL, NULL, NULL, NULL,
+	    UMA_ALIGN_PTR, 0);
+
 	/*
 	 * Install a global async callback.  This callback will
 	 * receive async callbacks like "new device found".
@@ -1855,6 +1864,15 @@
 	    "kern.cam.ada.%d.write_cache", periph->unit_number);
 	TUNABLE_INT_FETCH(announce_buf, &softc->write_cache);
 
+	/*
+	 * Let XPT know we can use UMA-allocated CCBs.
+	 */
+	if (ada_enable_uma_ccbs) {
+		KASSERT(ada_ccb_zone != NULL,
+		    ("%s: NULL ada_ccb_zone", __func__));
+		periph->ccb_zone = ada_ccb_zone;
+	}
+
 	/*
 	 * Set support flags based on the Identify data and quirks.
 	 */
diff --git a/sys/cam/ata/ata_xpt.c b/sys/cam/ata/ata_xpt.c
--- a/sys/cam/ata/ata_xpt.c
+++ b/sys/cam/ata/ata_xpt.c
@@ -1795,6 +1795,13 @@
 ata_action(union ccb *start_ccb)
 {
 
+	if (start_ccb->ccb_h.func_code != XPT_ATA_IO) {
+		KASSERT((start_ccb->ccb_h.alloc_flags & CAM_CCB_FROM_UMA) == 0,
+		    ("%s: ccb %p, func_code %#x should not be allocated "
+		    "from UMA zone\n",
+		    __func__, start_ccb, start_ccb->ccb_h.func_code));
+	}
+
 	switch (start_ccb->ccb_h.func_code) {
 	case XPT_SET_TRAN_SETTINGS:
 	{
diff --git a/sys/cam/cam_ccb.h b/sys/cam/cam_ccb.h
--- a/sys/cam/cam_ccb.h
+++ b/sys/cam/cam_ccb.h
@@ -58,6 +58,12 @@
 /* Struct definitions for CAM control blocks */
 
 /* Common CCB header */
+
+/* CCB memory allocation flags */
+typedef enum {
+	CAM_CCB_FROM_UMA	= 0x00000001,/* CCB from a periph UMA zone */
+} ccb_alloc_flags;
+
 /* CAM CCB flags */
 typedef enum {
 	CAM_CDB_POINTER		= 0x00000001,/* The CDB field is a pointer    */
@@ -341,7 +347,13 @@
 	camq_entry	xpt_links;	/* For chaining in the XPT layer */
 	camq_entry	sim_links;	/* For chaining in the SIM layer */
 	camq_entry	periph_links;	/* For chaining in the type driver */
-	u_int32_t	retry_count;
+#if BYTE_ORDER == LITTLE_ENDIAN
+	u_int16_t       retry_count;
+	u_int16_t       alloc_flags;	/* ccb_alloc_flags */
+#else
+	u_int16_t       alloc_flags;	/* ccb_alloc_flags */
+	u_int16_t       retry_count;
+#endif
 	void		(*cbfcnp)(struct cam_periph *, union ccb *);
 					/* Callback on completion function */
 	xpt_opcode	func_code;	/* XPT function code */
diff --git a/sys/cam/cam_periph.h b/sys/cam/cam_periph.h
--- a/sys/cam/cam_periph.h
+++ b/sys/cam/cam_periph.h
@@ -42,6 +42,8 @@
 #include <sys/sysctl.h>
 #include <sys/taskqueue.h>
 
+#include <vm/uma.h>
+
 #include <cam/cam_xpt.h>
 
 struct devstat;
@@ -147,6 +149,7 @@
 	ac_callback_t		*deferred_callback; 
 	ac_code			 deferred_ac;
 	struct task		 periph_run_task;
+	uma_zone_t		 ccb_zone;
 };
 
 #define CAM_PERIPH_MAXMAPS	2
diff --git a/sys/cam/cam_xpt.c b/sys/cam/cam_xpt.c
--- a/sys/cam/cam_xpt.c
+++ b/sys/cam/cam_xpt.c
@@ -4693,7 +4693,17 @@
 void
 xpt_free_ccb(union ccb *free_ccb)
 {
-	free(free_ccb, M_CAMCCB);
+	struct cam_periph *periph;
+
+	if (free_ccb->ccb_h.alloc_flags & CAM_CCB_FROM_UMA) {
+		/*
+		 * Looks like a CCB allocated from a periph UMA zone.
+		 */
+		periph = free_ccb->ccb_h.path->periph;
+		uma_zfree(periph->ccb_zone, free_ccb);
+	} else {
+		free(free_ccb, M_CAMCCB);
+	}
 }
 
 /* Private XPT functions */
@@ -4707,10 +4717,18 @@
 xpt_get_ccb_nowait(struct cam_periph *periph)
 {
 	union ccb *new_ccb;
+	int alloc_flags;
 
-	new_ccb = malloc(sizeof(*new_ccb), M_CAMCCB, M_ZERO|M_NOWAIT);
+	if (periph->ccb_zone != NULL) {
+		alloc_flags = CAM_CCB_FROM_UMA;
+		new_ccb = uma_zalloc(periph->ccb_zone, M_ZERO|M_NOWAIT);
+	} else {
+		alloc_flags = 0;
+		new_ccb = malloc(sizeof(*new_ccb), M_CAMCCB, M_ZERO|M_NOWAIT);
+	}
 	if (new_ccb == NULL)
 		return (NULL);
+	new_ccb->ccb_h.alloc_flags = alloc_flags;
 	periph->periph_allocated++;
 	cam_ccbq_take_opening(&periph->path->device->ccbq);
 	return (new_ccb);
@@ -4720,9 +4738,17 @@
 xpt_get_ccb(struct cam_periph *periph)
 {
 	union ccb *new_ccb;
+	int alloc_flags;
 
 	cam_periph_unlock(periph);
-	new_ccb = malloc(sizeof(*new_ccb), M_CAMCCB, M_ZERO|M_WAITOK);
+	if (periph->ccb_zone != NULL) {
+		alloc_flags = CAM_CCB_FROM_UMA;
+		new_ccb = uma_zalloc(periph->ccb_zone, M_ZERO|M_WAITOK);
+	} else {
+		alloc_flags = 0;
+		new_ccb = malloc(sizeof(*new_ccb), M_CAMCCB, M_ZERO|M_WAITOK);
+	}
+	new_ccb->ccb_h.alloc_flags = alloc_flags;
 	cam_periph_lock(periph);
 	periph->periph_allocated++;
 	cam_ccbq_take_opening(&periph->path->device->ccbq);
diff --git a/sys/cam/scsi/scsi_da.c b/sys/cam/scsi/scsi_da.c
--- a/sys/cam/scsi/scsi_da.c
+++ b/sys/cam/scsi/scsi_da.c
@@ -403,6 +403,8 @@
 		softc->delete_available &= ~(1 << delete_method);	\
 	}
 
+static uma_zone_t da_ccb_zone;
+
 struct da_quirk_entry {
 	struct scsi_inquiry_pattern inq_pat;
 	da_quirks quirks;
@@ -1557,6 +1559,7 @@
 static int da_send_ordered = DA_DEFAULT_SEND_ORDERED;
 static int da_disable_wp_detection = 0;
 static int da_enable_biospeedup = 1;
+static int da_enable_uma_ccbs = 0;
 
 static SYSCTL_NODE(_kern_cam, OID_AUTO, da, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
     "CAM Direct Access Disk driver");
@@ -1573,6 +1576,8 @@
 	   "Disable detection of write-protected disks");
 SYSCTL_INT(_kern_cam_da, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
 	    &da_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing");
+SYSCTL_INT(_kern_cam_da, OID_AUTO, enable_uma_ccbs, CTLFLAG_RWTUN,
+	    &da_enable_uma_ccbs, 0, "Use UMA for CCBs");
 
 SYSCTL_PROC(_kern_cam_da, OID_AUTO, default_softtimeout,
     CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
@@ -2011,6 +2016,10 @@
 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
 		    printf("dainit: shutdown event registration failed!\n");
 	}
+
+	da_ccb_zone = uma_zcreate("da_ccb",
+	    sizeof(struct ccb_scsiio), NULL, NULL, NULL, NULL,
+	    UMA_ALIGN_PTR, 0);
 }
 
 /*
@@ -2848,6 +2857,15 @@
 
 	TASK_INIT(&softc->sysctl_task, 0, dasysctlinit, periph);
 
+	/*
+	 * Let XPT know we can use UMA-allocated CCBs.
+	 */
+	if (da_enable_uma_ccbs) {
+		KASSERT(da_ccb_zone != NULL,
+		    ("%s: NULL da_ccb_zone", __func__));
+		periph->ccb_zone = da_ccb_zone;
+	}
+
 	/*
 	 * Take an exclusive section lock on the periph while dastart is called
 	 * to finish the probe.  The lock will be dropped in dadone at the end
diff --git a/sys/cam/scsi/scsi_xpt.c b/sys/cam/scsi/scsi_xpt.c
--- a/sys/cam/scsi/scsi_xpt.c
+++ b/sys/cam/scsi/scsi_xpt.c
@@ -2625,6 +2625,13 @@
 scsi_action(union ccb *start_ccb)
 {
 
+	if (start_ccb->ccb_h.func_code != XPT_SCSI_IO) {
+		KASSERT((start_ccb->ccb_h.alloc_flags & CAM_CCB_FROM_UMA) == 0,
+		    ("%s: ccb %p, func_code %#x should not be allocated "
+		    "from UMA zone\n",
+		    __func__, start_ccb, start_ccb->ccb_h.func_code));
+	}
+
 	switch (start_ccb->ccb_h.func_code) {
 	case XPT_SET_TRAN_SETTINGS:
 	{