X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=engines%2Fsg.c;h=33e38c2bd2592d616f0559c991ef36ae5c851594;hp=84c2afe6c354158f6edbbf98fb12f0e0de09d55c;hb=b4b9665e89e2f6a25ca8e2b99207d95d4d7b5154;hpb=da751ca9665bcdeca56d2eec5b629a0953c07662

diff --git a/engines/sg.c b/engines/sg.c
index 84c2afe6..33e38c2b 100644
--- a/engines/sg.c
+++ b/engines/sg.c
@@ -3,24 +3,142 @@
  *
  * IO engine that uses the Linux SG v3 interface to talk to SCSI devices
  *
+ * This ioengine can operate in two modes:
+ *	sync	with block devices (/dev/sdX) or
+ *		with character devices (/dev/sgY) with direct=1 or sync=1
+ *	async	with character devices with direct=0 and sync=0
+ *
+ * What value does queue() return for the different cases?
+ *				queue() return value
+ * In sync mode:
+ *  /dev/sdX		RWT	FIO_Q_COMPLETED
+ *  /dev/sgY		RWT	FIO_Q_COMPLETED
+ *   with direct=1 or sync=1
+ *
+ * In async mode:
+ *  /dev/sgY		RWT	FIO_Q_QUEUED
+ *   direct=0 and sync=0
+ *
+ * Because FIO_SYNCIO is set for this ioengine td_io_queue() will fill in
+ * issue_time *before* each IO is sent to queue()
+ *
+ * Where are the IO counting functions called for the different cases?
+ *
+ * In sync mode:
+ *  /dev/sdX (commit==NULL)
+ *   RWT
+ *    io_u_mark_depth()			called in td_io_queue()
+ *    io_u_mark_submit/complete()	called in td_io_queue()
+ *    issue_time			set in td_io_queue()
+ *
+ *  /dev/sgY with direct=1 or sync=1 (commit does nothing)
+ *   RWT
+ *    io_u_mark_depth()			called in td_io_queue()
+ *    io_u_mark_submit/complete()	called in queue()
+ *    issue_time			set in td_io_queue()
+ *  
+ * In async mode:
+ *  /dev/sgY with direct=0 and sync=0
+ *   RW: read and write operations are submitted in queue()
+ *    io_u_mark_depth()			called in td_io_commit()
+ *    io_u_mark_submit()		called in queue()
+ *    issue_time			set in td_io_queue()
+ *   T: trim operations are queued in queue() and submitted in commit()
+ *    io_u_mark_depth()			called in td_io_commit()
+ *    io_u_mark_submit()		called in commit()
+ *    issue_time			set in commit()
+ *
  */
 #include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>
-#include <assert.h>
-#include <sys/poll.h>
+#include <poll.h>
 
 #include "../fio.h"
-#include "../os.h"
+#include "../optgroup.h"
 
 #ifdef FIO_HAVE_SGIO
 
+enum {
+	FIO_SG_WRITE		= 1,
+	FIO_SG_WRITE_VERIFY	= 2,
+	FIO_SG_WRITE_SAME	= 3
+};
+
+struct sg_options {
+	void *pad;
+	unsigned int readfua;
+	unsigned int writefua;
+	unsigned int write_mode;
+};
+
+static struct fio_option options[] = {
+	{
+		.name	= "readfua",
+		.lname	= "sg engine read fua flag support",
+		.type	= FIO_OPT_BOOL,
+		.off1	= offsetof(struct sg_options, readfua),
+		.help	= "Set FUA flag (force unit access) for all Read operations",
+		.def	= "0",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_SG,
+	},
+	{
+		.name	= "writefua",
+		.lname	= "sg engine write fua flag support",
+		.type	= FIO_OPT_BOOL,
+		.off1	= offsetof(struct sg_options, writefua),
+		.help	= "Set FUA flag (force unit access) for all Write operations",
+		.def	= "0",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_SG,
+	},
+	{
+		.name	= "sg_write_mode",
+		.lname	= "specify sg write mode",
+		.type	= FIO_OPT_STR,
+		.off1	= offsetof(struct sg_options, write_mode),
+		.help	= "Specify SCSI WRITE mode",
+		.def	= "write",
+		.posval = {
+			  { .ival = "write",
+			    .oval = FIO_SG_WRITE,
+			    .help = "Issue standard SCSI WRITE commands",
+			  },
+			  { .ival = "verify",
+			    .oval = FIO_SG_WRITE_VERIFY,
+			    .help = "Issue SCSI WRITE AND VERIFY commands",
+			  },
+			  { .ival = "same",
+			    .oval = FIO_SG_WRITE_SAME,
+			    .help = "Issue SCSI WRITE SAME commands",
+			  },
+		},
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_SG,
+	},
+	{
+		.name	= NULL,
+	},
+};
+
+#define MAX_10B_LBA  0xFFFFFFFFULL
+#define SCSI_TIMEOUT_MS 30000   // 30 second timeout; currently no method to override
+#define MAX_SB 64               // sense block maximum return size
+
 struct sgio_cmd {
-	unsigned char cdb[10];
+	unsigned char cdb[16];      // enhanced from 10 to support 16 byte commands
+	unsigned char sb[MAX_SB];   // add sense block to commands
 	int nr;
 };
 
+struct sgio_trim {
+	char *unmap_param;
+	unsigned int unmap_range_count;
+	struct io_u **trim_io_us;
+};
+
 struct sgio_data {
 	struct sgio_cmd *cmds;
 	struct io_u **events;
@@ -29,8 +147,16 @@ struct sgio_data {
 	void *sgbuf;
 	unsigned int bs;
 	int type_checked;
+	struct sgio_trim **trim_queues;
+	int current_queue;
+	unsigned int *trim_queue_map;
 };
 
+static inline bool sgio_unbuffered(struct thread_data *td)
+{
+	return (td->o.odirect || td->o.sync_io);
+}
+
 static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
 			  struct io_u *io_u, int fs)
 {
@@ -42,8 +168,11 @@ static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
 	hdr->interface_id = 'S';
 	hdr->cmdp = sc->cdb;
 	hdr->cmd_len = sizeof(sc->cdb);
+	hdr->sbp = sc->sb;
+	hdr->mx_sb_len = sizeof(sc->sb);
 	hdr->pack_id = io_u->index;
 	hdr->usr_ptr = io_u;
+	hdr->timeout = SCSI_TIMEOUT_MS;
 
 	if (fs) {
 		hdr->dxferp = io_u->xfer_buf;
@@ -62,18 +191,45 @@ static int pollin_events(struct pollfd *pfds, int fds)
 	return 0;
 }
 
-static int fio_sgio_getevents(struct thread_data *td, int min, int max,
-			      struct timespec fio_unused *t)
+static int sg_fd_read(int fd, void *data, size_t size)
 {
-	/*
-	 * normally hard coding &td->files[0] is a bug that needs to be fixed,
-	 * but it's ok here as all files should point to the same device.
-	 */
-	struct fio_file *f = &td->files[0];
-	struct sgio_data *sd = td->io_ops->data;
-	int left = max, ret, r = 0;
+	int err = 0;
+
+	while (size) {
+		ssize_t ret;
+
+		ret = read(fd, data, size);
+		if (ret < 0) {
+			if (errno == EAGAIN || errno == EINTR)
+				continue;
+			err = errno;
+			break;
+		} else if (!ret)
+			break;
+		else {
+			data += ret;
+			size -= ret;
+		}
+	}
+
+	if (err)
+		return err;
+	if (size)
+		return EAGAIN;
+
+	return 0;
+}
+
+static int fio_sgio_getevents(struct thread_data *td, unsigned int min,
+			      unsigned int max,
+			      const struct timespec fio_unused *t)
+{
+	struct sgio_data *sd = td->io_ops_data;
+	int left = max, eventNum, ret, r = 0, trims = 0;
 	void *buf = sd->sgbuf;
-	unsigned int i, events;
+	unsigned int i, j, events;
+	struct fio_file *f;
+	struct io_u *io_u;
 
 	/*
 	 * Fill in the file descriptors
@@ -82,22 +238,35 @@ static int fio_sgio_getevents(struct thread_data *td, int min, int max,
 		/*
 		 * don't block for min events == 0
 		 */
-		if (!min) {
-			sd->fd_flags[i] = fcntl(f->fd, F_GETFL);
-			fcntl(f->fd, F_SETFL, sd->fd_flags[i] | O_NONBLOCK);
-		}
+		if (!min)
+			sd->fd_flags[i] = fio_set_fd_nonblocking(f->fd, "sg");
+		else
+			sd->fd_flags[i] = -1;
+
 		sd->pfds[i].fd = f->fd;
 		sd->pfds[i].events = POLLIN;
 	}
 
-	while (left) {
-		void *p;
+	/*
+	** There are two counters here:
+	**  - number of SCSI commands completed
+	**  - number of io_us completed
+	**
+	** These are the same with reads and writes, but
+	** could differ with trim/unmap commands because
+	** a single unmap can include multiple io_us
+	*/
+
+	while (left > 0) {
+		char *p;
+
+		dprint(FD_IO, "sgio_getevents: sd %p: min=%d, max=%d, left=%d\n", sd, min, max, left);
 
 		do {
 			if (!min)
 				break;
 
-			ret = poll(sd->pfds, td->nr_files, -1);
+			ret = poll(sd->pfds, td->o.nr_files, -1);
 			if (ret < 0) {
 				if (!r)
 					r = -errno;
@@ -106,7 +275,7 @@ static int fio_sgio_getevents(struct thread_data *td, int min, int max,
 			} else if (!ret)
 				continue;
 
-			if (pollin_events(sd->pfds, td->nr_files))
+			if (pollin_events(sd->pfds, td->o.nr_files))
 				break;
 		} while (1);
 
@@ -117,20 +286,27 @@ re_read:
 		p = buf;
 		events = 0;
 		for_each_file(td, f, i) {
-			ret = read(f->fd, p, left * sizeof(struct sg_io_hdr));
-			if (ret < 0) {
-				if (errno == EAGAIN)
-					continue;
-				r = -errno;
-				td_verror(td, errno, "read");
-				break;
-			} else if (ret) {
-				p += ret;
-				events += ret / sizeof(struct sg_io_hdr);
+			for (eventNum = 0; eventNum < left; eventNum++) {
+				ret = sg_fd_read(f->fd, p, sizeof(struct sg_io_hdr));
+				dprint(FD_IO, "sgio_getevents: sg_fd_read ret: %d\n", ret);
+				if (ret) {
+					r = -ret;
+					td_verror(td, r, "sg_read");
+					break;
+				}
+				io_u = ((struct sg_io_hdr *)p)->usr_ptr;
+				if (io_u->ddir == DDIR_TRIM) {
+					events += sd->trim_queues[io_u->index]->unmap_range_count;
+					eventNum += sd->trim_queues[io_u->index]->unmap_range_count - 1;
+				} else
+					events++;
+
+				p += sizeof(struct sg_io_hdr);
+				dprint(FD_IO, "sgio_getevents: events: %d, eventNum: %d, left: %d\n", events, eventNum, left);
 			}
 		}
 
-		if (r < 0)
+		if (r < 0 && !events)
 			break;
 		if (!events) {
 			usleep(1000);
@@ -142,23 +318,54 @@ re_read:
 
 		for (i = 0; i < events; i++) {
 			struct sg_io_hdr *hdr = (struct sg_io_hdr *) buf + i;
+			sd->events[i + trims] = hdr->usr_ptr;
+			io_u = (struct io_u *)(hdr->usr_ptr);
+
+			if (hdr->info & SG_INFO_CHECK) {
+				/* record if an io error occurred, ignore resid */
+				memcpy(&io_u->hdr, hdr, sizeof(struct sg_io_hdr));
+				sd->events[i + trims]->error = EIO;
+			}
 
-			sd->events[i] = hdr->usr_ptr;
+			if (io_u->ddir == DDIR_TRIM) {
+				struct sgio_trim *st = sd->trim_queues[io_u->index];
+				assert(st->trim_io_us[0] == io_u);
+				dprint(FD_IO, "sgio_getevents: reaping %d io_us from trim queue %d\n", st->unmap_range_count, io_u->index);
+				dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", io_u->index, i+trims);
+				for (j = 1; j < st->unmap_range_count; j++) {
+					++trims;
+					sd->events[i + trims] = st->trim_io_us[j];
+					dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", st->trim_io_us[j]->index, i+trims);
+					if (hdr->info & SG_INFO_CHECK) {
+						/* record if an io error occurred, ignore resid */
+						memcpy(&st->trim_io_us[j]->hdr, hdr, sizeof(struct sg_io_hdr));
+						sd->events[i + trims]->error = EIO;
+					}
+				}
+				events -= st->unmap_range_count - 1;
+				st->unmap_range_count = 0;
+			}
 		}
 	}
 
 	if (!min) {
-		for_each_file(td, f, i)
-			fcntl(f->fd, F_SETFL, sd->fd_flags[i]);
+		for_each_file(td, f, i) {
+			if (sd->fd_flags[i] == -1)
+				continue;
+
+			if (fcntl(f->fd, F_SETFL, sd->fd_flags[i]) < 0)
+				log_err("fio: sg failed to restore fcntl flags: %s\n", strerror(errno));
+		}
 	}
 
 	return r;
 }
 
-static int fio_sgio_ioctl_doio(struct thread_data *td,
-			       struct fio_file *f, struct io_u *io_u)
+static enum fio_q_status fio_sgio_ioctl_doio(struct thread_data *td,
+					     struct fio_file *f,
+					     struct io_u *io_u)
 {
-	struct sgio_data *sd = td->io_ops->data;
+	struct sgio_data *sd = td->io_ops_data;
 	struct sg_io_hdr *hdr = &io_u->hdr;
 	int ret;
 
@@ -166,94 +373,258 @@ static int fio_sgio_ioctl_doio(struct thread_data *td,
 
 	ret = ioctl(f->fd, SG_IO, hdr);
 	if (ret < 0)
-		return -errno;
+		return ret;
+
+	/* record if an io error occurred */
+	if (hdr->info & SG_INFO_CHECK)
+		io_u->error = EIO;
 
 	return FIO_Q_COMPLETED;
 }
 
-static int fio_sgio_rw_doio(struct fio_file *f, struct io_u *io_u, int sync)
+static enum fio_q_status fio_sgio_rw_doio(struct fio_file *f,
+					  struct io_u *io_u, int do_sync)
 {
 	struct sg_io_hdr *hdr = &io_u->hdr;
 	int ret;
 
 	ret = write(f->fd, hdr, sizeof(*hdr));
 	if (ret < 0)
-		return errno;
+		return ret;
 
-	if (sync) {
+	if (do_sync) {
 		ret = read(f->fd, hdr, sizeof(*hdr));
 		if (ret < 0)
-			return -errno;
+			return ret;
+
+		/* record if an io error occurred */
+		if (hdr->info & SG_INFO_CHECK)
+			io_u->error = EIO;
+
 		return FIO_Q_COMPLETED;
 	}
 
 	return FIO_Q_QUEUED;
 }
 
-static int fio_sgio_doio(struct thread_data *td, struct io_u *io_u, int sync)
+static enum fio_q_status fio_sgio_doio(struct thread_data *td,
+				       struct io_u *io_u, int do_sync)
 {
 	struct fio_file *f = io_u->file;
+	enum fio_q_status ret;
+
+	if (f->filetype == FIO_TYPE_BLOCK) {
+		ret = fio_sgio_ioctl_doio(td, f, io_u);
+		td_verror(td, io_u->error, __func__);
+	} else {
+		ret = fio_sgio_rw_doio(f, io_u, do_sync);
+		if (do_sync)
+			td_verror(td, io_u->error, __func__);
+	}
+
+	return ret;
+}
 
-	if (f->filetype == FIO_TYPE_BD)
-		return fio_sgio_ioctl_doio(td, f, io_u);
+static void fio_sgio_rw_lba(struct sg_io_hdr *hdr, unsigned long long lba,
+			    unsigned long long nr_blocks)
+{
+	if (lba < MAX_10B_LBA) {
+		hdr->cmdp[2] = (unsigned char) ((lba >> 24) & 0xff);
+		hdr->cmdp[3] = (unsigned char) ((lba >> 16) & 0xff);
+		hdr->cmdp[4] = (unsigned char) ((lba >>  8) & 0xff);
+		hdr->cmdp[5] = (unsigned char) (lba & 0xff);
+		hdr->cmdp[7] = (unsigned char) ((nr_blocks >> 8) & 0xff);
+		hdr->cmdp[8] = (unsigned char) (nr_blocks & 0xff);
+	} else {
+		hdr->cmdp[2] = (unsigned char) ((lba >> 56) & 0xff);
+		hdr->cmdp[3] = (unsigned char) ((lba >> 48) & 0xff);
+		hdr->cmdp[4] = (unsigned char) ((lba >> 40) & 0xff);
+		hdr->cmdp[5] = (unsigned char) ((lba >> 32) & 0xff);
+		hdr->cmdp[6] = (unsigned char) ((lba >> 24) & 0xff);
+		hdr->cmdp[7] = (unsigned char) ((lba >> 16) & 0xff);
+		hdr->cmdp[8] = (unsigned char) ((lba >>  8) & 0xff);
+		hdr->cmdp[9] = (unsigned char) (lba & 0xff);
+		hdr->cmdp[10] = (unsigned char) ((nr_blocks >> 32) & 0xff);
+		hdr->cmdp[11] = (unsigned char) ((nr_blocks >> 16) & 0xff);
+		hdr->cmdp[12] = (unsigned char) ((nr_blocks >> 8) & 0xff);
+		hdr->cmdp[13] = (unsigned char) (nr_blocks & 0xff);
+	}
 
-	return fio_sgio_rw_doio(f, io_u, sync);
+	return;
 }
 
 static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
 {
 	struct sg_io_hdr *hdr = &io_u->hdr;
-	struct sgio_data *sd = td->io_ops->data;
-	int nr_blocks, lba;
+	struct sg_options *o = td->eo;
+	struct sgio_data *sd = td->io_ops_data;
+	unsigned long long nr_blocks, lba;
+	int offset;
 
 	if (io_u->xfer_buflen & (sd->bs - 1)) {
 		log_err("read/write not sector aligned\n");
 		return EINVAL;
 	}
 
+	nr_blocks = io_u->xfer_buflen / sd->bs;
+	lba = io_u->offset / sd->bs;
+
 	if (io_u->ddir == DDIR_READ) {
 		sgio_hdr_init(sd, hdr, io_u, 1);
 
 		hdr->dxfer_direction = SG_DXFER_FROM_DEV;
-		hdr->cmdp[0] = 0x28;
+		if (lba < MAX_10B_LBA)
+			hdr->cmdp[0] = 0x28; // read(10)
+		else
+			hdr->cmdp[0] = 0x88; // read(16)
+
+		if (o->readfua)
+			hdr->cmdp[1] |= 0x08;
+
+		fio_sgio_rw_lba(hdr, lba, nr_blocks);
+
 	} else if (io_u->ddir == DDIR_WRITE) {
 		sgio_hdr_init(sd, hdr, io_u, 1);
 
 		hdr->dxfer_direction = SG_DXFER_TO_DEV;
-		hdr->cmdp[0] = 0x2a;
-	} else {
-		sgio_hdr_init(sd, hdr, io_u, 0);
+		switch(o->write_mode) {
+		case FIO_SG_WRITE:
+			if (lba < MAX_10B_LBA)
+				hdr->cmdp[0] = 0x2a; // write(10)
+			else
+				hdr->cmdp[0] = 0x8a; // write(16)
+			if (o->writefua)
+				hdr->cmdp[1] |= 0x08;
+			break;
+		case FIO_SG_WRITE_VERIFY:
+			if (lba < MAX_10B_LBA)
+				hdr->cmdp[0] = 0x2e; // write and verify(10)
+			else
+				hdr->cmdp[0] = 0x8e; // write and verify(16)
+			break;
+			// BYTCHK is disabled by virtue of the memset in sgio_hdr_init
+		case FIO_SG_WRITE_SAME:
+			hdr->dxfer_len = sd->bs;
+			if (lba < MAX_10B_LBA)
+				hdr->cmdp[0] = 0x41; // write same(10)
+			else
+				hdr->cmdp[0] = 0x93; // write same(16)
+			break;
+		};
 
-		hdr->dxfer_direction = SG_DXFER_NONE;
-		hdr->cmdp[0] = 0x35;
-	}
+		fio_sgio_rw_lba(hdr, lba, nr_blocks);
 
-	if (hdr->dxfer_direction != SG_DXFER_NONE) {
-		nr_blocks = io_u->xfer_buflen / sd->bs;
-		lba = io_u->offset / sd->bs;
-		hdr->cmdp[2] = (unsigned char) ((lba >> 24) & 0xff);
-		hdr->cmdp[3] = (unsigned char) ((lba >> 16) & 0xff);
-		hdr->cmdp[4] = (unsigned char) ((lba >>  8) & 0xff);
-		hdr->cmdp[5] = (unsigned char) (lba & 0xff);
-		hdr->cmdp[7] = (unsigned char) ((nr_blocks >> 8) & 0xff);
-		hdr->cmdp[8] = (unsigned char) (nr_blocks & 0xff);
-	}
+	} else if (io_u->ddir == DDIR_TRIM) {
+		struct sgio_trim *st;
+
+		if (sd->current_queue == -1) {
+			sgio_hdr_init(sd, hdr, io_u, 0);
+
+			hdr->cmd_len = 10;
+			hdr->dxfer_direction = SG_DXFER_TO_DEV;
+			hdr->cmdp[0] = 0x42; // unmap
+			sd->current_queue = io_u->index;
+			st = sd->trim_queues[sd->current_queue];
+			hdr->dxferp = st->unmap_param;
+			assert(sd->trim_queues[io_u->index]->unmap_range_count == 0);
+			dprint(FD_IO, "sg: creating new queue based on io_u %d\n", io_u->index);
+		}
+		else
+			st = sd->trim_queues[sd->current_queue];
+
+		dprint(FD_IO, "sg: adding io_u %d to trim queue %d\n", io_u->index, sd->current_queue);
+		st->trim_io_us[st->unmap_range_count] = io_u;
+		sd->trim_queue_map[io_u->index] = sd->current_queue;
+
+		offset = 8 + 16 * st->unmap_range_count;
+		st->unmap_param[offset] = (unsigned char) ((lba >> 56) & 0xff);
+		st->unmap_param[offset+1] = (unsigned char) ((lba >> 48) & 0xff);
+		st->unmap_param[offset+2] = (unsigned char) ((lba >> 40) & 0xff);
+		st->unmap_param[offset+3] = (unsigned char) ((lba >> 32) & 0xff);
+		st->unmap_param[offset+4] = (unsigned char) ((lba >> 24) & 0xff);
+		st->unmap_param[offset+5] = (unsigned char) ((lba >> 16) & 0xff);
+		st->unmap_param[offset+6] = (unsigned char) ((lba >>  8) & 0xff);
+		st->unmap_param[offset+7] = (unsigned char) (lba & 0xff);
+		st->unmap_param[offset+8] = (unsigned char) ((nr_blocks >> 32) & 0xff);
+		st->unmap_param[offset+9] = (unsigned char) ((nr_blocks >> 16) & 0xff);
+		st->unmap_param[offset+10] = (unsigned char) ((nr_blocks >> 8) & 0xff);
+		st->unmap_param[offset+11] = (unsigned char) (nr_blocks & 0xff);
+
+		st->unmap_range_count++;
+
+	} else if (ddir_sync(io_u->ddir)) {
+		sgio_hdr_init(sd, hdr, io_u, 0);
+		hdr->dxfer_direction = SG_DXFER_NONE;
+		if (lba < MAX_10B_LBA)
+			hdr->cmdp[0] = 0x35; // synccache(10)
+		else
+			hdr->cmdp[0] = 0x91; // synccache(16)
+	} else
+		assert(0);
 
 	return 0;
 }
 
-static int fio_sgio_queue(struct thread_data *td, struct io_u *io_u)
+static void fio_sgio_unmap_setup(struct sg_io_hdr *hdr, struct sgio_trim *st)
+{
+	hdr->dxfer_len = st->unmap_range_count * 16 + 8;
+	hdr->cmdp[7] = (unsigned char) (((st->unmap_range_count * 16 + 8) >> 8) & 0xff);
+	hdr->cmdp[8] = (unsigned char) ((st->unmap_range_count * 16 + 8) & 0xff);
+
+	st->unmap_param[0] = (unsigned char) (((16 * st->unmap_range_count + 6) >> 8) & 0xff);
+	st->unmap_param[1] = (unsigned char)  ((16 * st->unmap_range_count + 6) & 0xff);
+	st->unmap_param[2] = (unsigned char) (((16 * st->unmap_range_count) >> 8) & 0xff);
+	st->unmap_param[3] = (unsigned char)  ((16 * st->unmap_range_count) & 0xff);
+
+	return;
+}
+
+static enum fio_q_status fio_sgio_queue(struct thread_data *td,
+					struct io_u *io_u)
 {
 	struct sg_io_hdr *hdr = &io_u->hdr;
-	int ret;
+	struct sgio_data *sd = td->io_ops_data;
+	int ret, do_sync = 0;
+
+	fio_ro_check(td, io_u);
 
-	ret = fio_sgio_doio(td, io_u, io_u->ddir == DDIR_SYNC);
+	if (sgio_unbuffered(td) || ddir_sync(io_u->ddir))
+		do_sync = 1;
+
+	if (io_u->ddir == DDIR_TRIM) {
+		if (do_sync || io_u->file->filetype == FIO_TYPE_BLOCK) {
+			struct sgio_trim *st = sd->trim_queues[sd->current_queue];
+
+			/* finish cdb setup for unmap because we are
+			** doing unmap commands synchronously */
+			assert(st->unmap_range_count == 1);
+			assert(io_u == st->trim_io_us[0]);
+			hdr = &io_u->hdr;
+
+			fio_sgio_unmap_setup(hdr, st);
+
+			st->unmap_range_count = 0;
+			sd->current_queue = -1;
+		} else
+			/* queue up trim ranges and submit in commit() */
+			return FIO_Q_QUEUED;
+	}
+
+	ret = fio_sgio_doio(td, io_u, do_sync);
 
 	if (ret < 0)
 		io_u->error = errno;
 	else if (hdr->status) {
 		io_u->resid = hdr->resid;
 		io_u->error = EIO;
+	} else if (td->io_ops->commit != NULL) {
+		if (do_sync && !ddir_sync(io_u->ddir)) {
+			io_u_mark_submit(td, 1);
+			io_u_mark_complete(td, 1);
+		} else if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
+			io_u_mark_submit(td, 1);
+			io_u_queued(td, io_u);
+		}
 	}
 
 	if (io_u->error) {
@@ -264,48 +635,163 @@ static int fio_sgio_queue(struct thread_data *td, struct io_u *io_u)
 	return ret;
 }
 
+static int fio_sgio_commit(struct thread_data *td)
+{
+	struct sgio_data *sd = td->io_ops_data;
+	struct sgio_trim *st;
+	struct io_u *io_u;
+	struct sg_io_hdr *hdr;
+	struct timespec now;
+	unsigned int i;
+	int ret;
+
+	if (sd->current_queue == -1)
+		return 0;
+
+	st = sd->trim_queues[sd->current_queue];
+	io_u = st->trim_io_us[0];
+	hdr = &io_u->hdr;
+
+	fio_sgio_unmap_setup(hdr, st);
+
+	sd->current_queue = -1;
+
+	ret = fio_sgio_rw_doio(io_u->file, io_u, 0);
+
+	if (ret < 0)
+		for (i = 0; i < st->unmap_range_count; i++)
+			st->trim_io_us[i]->error = errno;
+	else if (hdr->status)
+		for (i = 0; i < st->unmap_range_count; i++) {
+			st->trim_io_us[i]->resid = hdr->resid;
+			st->trim_io_us[i]->error = EIO;
+		}
+	else {
+		if (fio_fill_issue_time(td)) {
+			fio_gettime(&now, NULL);
+			for (i = 0; i < st->unmap_range_count; i++) {
+				struct io_u *io_u = st->trim_io_us[i];
+
+				memcpy(&io_u->issue_time, &now, sizeof(now));
+				io_u_queued(td, io_u);
+			}
+		}
+		io_u_mark_submit(td, st->unmap_range_count);
+	}
+
+	if (io_u->error) {
+		td_verror(td, io_u->error, "xfer");
+		return 0;
+	}
+
+	if (ret == FIO_Q_QUEUED)
+		return 0;
+	else
+		return ret;
+}
+
 static struct io_u *fio_sgio_event(struct thread_data *td, int event)
 {
-	struct sgio_data *sd = td->io_ops->data;
+	struct sgio_data *sd = td->io_ops_data;
 
 	return sd->events[event];
 }
 
-static int fio_sgio_get_bs(struct thread_data *td, unsigned int *bs)
+static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
+				  unsigned long long *max_lba)
 {
-	struct sgio_data *sd = td->io_ops->data;
-	struct io_u *io_u;
-	struct sg_io_hdr *hdr;
-	unsigned char buf[8];
+	/*
+	 * need to do read capacity operation w/o benefit of sd or
+	 * io_u structures, which are not initialized until later.
+	 */
+	struct sg_io_hdr hdr;
+	unsigned char cmd[16];
+	unsigned char sb[64];
+	unsigned char buf[32];  // read capacity return
 	int ret;
+	int fd = -1;
 
-	io_u = __get_io_u(td);
-	io_u->file = &td->files[0];
-	assert(io_u);
+	struct fio_file *f = td->files[0];
 
-	hdr = &io_u->hdr;
-	sgio_hdr_init(sd, hdr, io_u, 0);
-	memset(buf, 0, sizeof(buf));
+	/* open file independent of rest of application */
+	fd = open(f->file_name, O_RDONLY);
+	if (fd < 0)
+		return -errno;
 
-	hdr->cmdp[0] = 0x25;
-	hdr->dxfer_direction = SG_DXFER_FROM_DEV;
-	hdr->dxferp = buf;
-	hdr->dxfer_len = sizeof(buf);
+	memset(&hdr, 0, sizeof(hdr));
+	memset(cmd, 0, sizeof(cmd));
+	memset(sb, 0, sizeof(sb));
+	memset(buf, 0, sizeof(buf));
 
-	ret = fio_sgio_doio(td, io_u, 1);
-	if (ret) {
-		put_io_u(td, io_u);
+	/* First let's try a 10 byte read capacity. */
+	hdr.interface_id = 'S';
+	hdr.cmdp = cmd;
+	hdr.cmd_len = 10;
+	hdr.sbp = sb;
+	hdr.mx_sb_len = sizeof(sb);
+	hdr.timeout = SCSI_TIMEOUT_MS;
+	hdr.cmdp[0] = 0x25;  // Read Capacity(10)
+	hdr.dxfer_direction = SG_DXFER_FROM_DEV;
+	hdr.dxferp = buf;
+	hdr.dxfer_len = sizeof(buf);
+
+	ret = ioctl(fd, SG_IO, &hdr);
+	if (ret < 0) {
+		close(fd);
 		return ret;
 	}
 
-	*bs = (buf[4] << 24) | (buf[5] << 16) | (buf[6] << 8) | buf[7];
-	put_io_u(td, io_u);
+	*bs	 = ((unsigned long) buf[4] << 24) | ((unsigned long) buf[5] << 16) |
+		   ((unsigned long) buf[6] << 8) | (unsigned long) buf[7];
+	*max_lba = ((unsigned long) buf[0] << 24) | ((unsigned long) buf[1] << 16) |
+		   ((unsigned long) buf[2] << 8) | (unsigned long) buf[3];
+
+	/*
+	 * If max lba masked by MAX_10B_LBA equals MAX_10B_LBA,
+	 * then need to retry with 16 byte Read Capacity command.
+	 */
+	if (*max_lba == MAX_10B_LBA) {
+		hdr.cmd_len = 16;
+		hdr.cmdp[0] = 0x9e; // service action
+		hdr.cmdp[1] = 0x10; // Read Capacity(16)
+		hdr.cmdp[10] = (unsigned char) ((sizeof(buf) >> 24) & 0xff);
+		hdr.cmdp[11] = (unsigned char) ((sizeof(buf) >> 16) & 0xff);
+		hdr.cmdp[12] = (unsigned char) ((sizeof(buf) >> 8) & 0xff);
+		hdr.cmdp[13] = (unsigned char) (sizeof(buf) & 0xff);
+
+		hdr.dxfer_direction = SG_DXFER_FROM_DEV;
+		hdr.dxferp = buf;
+		hdr.dxfer_len = sizeof(buf);
+
+		ret = ioctl(fd, SG_IO, &hdr);
+		if (ret < 0) {
+			close(fd);
+			return ret;
+		}
+
+		/* record if an io error occurred */
+		if (hdr.info & SG_INFO_CHECK)
+			td_verror(td, EIO, "fio_sgio_read_capacity");
+
+		*bs = (buf[8] << 24) | (buf[9] << 16) | (buf[10] << 8) | buf[11];
+		*max_lba = ((unsigned long long)buf[0] << 56) |
+				((unsigned long long)buf[1] << 48) |
+				((unsigned long long)buf[2] << 40) |
+				((unsigned long long)buf[3] << 32) |
+				((unsigned long long)buf[4] << 24) |
+				((unsigned long long)buf[5] << 16) |
+				((unsigned long long)buf[6] << 8) |
+				(unsigned long long)buf[7];
+	}
+
+	close(fd);
 	return 0;
 }
 
 static void fio_sgio_cleanup(struct thread_data *td)
 {
-	struct sgio_data *sd = td->io_ops->data;
+	struct sgio_data *sd = td->io_ops_data;
+	int i;
 
 	if (sd) {
 		free(sd->events);
@@ -313,44 +799,60 @@ static void fio_sgio_cleanup(struct thread_data *td)
 		free(sd->fd_flags);
 		free(sd->pfds);
 		free(sd->sgbuf);
-		free(sd);
+		free(sd->trim_queue_map);
+
+		for (i = 0; i < td->o.iodepth; i++) {
+			free(sd->trim_queues[i]->unmap_param);
+			free(sd->trim_queues[i]->trim_io_us);
+			free(sd->trim_queues[i]);
+		}
 
-		td->io_ops->data = NULL;
+		free(sd->trim_queues);
+		free(sd);
 	}
 }
 
 static int fio_sgio_init(struct thread_data *td)
 {
 	struct sgio_data *sd;
+	struct sgio_trim *st;
+	int i;
+
+	sd = calloc(1, sizeof(*sd));
+	sd->cmds = calloc(td->o.iodepth, sizeof(struct sgio_cmd));
+	sd->sgbuf = calloc(td->o.iodepth, sizeof(struct sg_io_hdr));
+	sd->events = calloc(td->o.iodepth, sizeof(struct io_u *));
+	sd->pfds = calloc(td->o.nr_files, sizeof(struct pollfd));
+	sd->fd_flags = calloc(td->o.nr_files, sizeof(int));
+	sd->type_checked = 0;
+
+	sd->trim_queues = calloc(td->o.iodepth, sizeof(struct sgio_trim *));
+	sd->current_queue = -1;
+	sd->trim_queue_map = calloc(td->o.iodepth, sizeof(int));
+	for (i = 0; i < td->o.iodepth; i++) {
+		sd->trim_queues[i] = calloc(1, sizeof(struct sgio_trim));
+		st = sd->trim_queues[i];
+		st->unmap_param = calloc(td->o.iodepth + 1, sizeof(char[16]));
+		st->unmap_range_count = 0;
+		st->trim_io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
+	}
 
-	sd = malloc(sizeof(*sd));
-	memset(sd, 0, sizeof(*sd));
-	sd->cmds = malloc(td->iodepth * sizeof(struct sgio_cmd));
-	memset(sd->cmds, 0, td->iodepth * sizeof(struct sgio_cmd));
-	sd->events = malloc(td->iodepth * sizeof(struct io_u *));
-	memset(sd->events, 0, td->iodepth * sizeof(struct io_u *));
-	sd->pfds = malloc(sizeof(struct pollfd) * td->nr_files);
-	memset(sd->pfds, 0, sizeof(struct pollfd) * td->nr_files);
-	sd->fd_flags = malloc(sizeof(int) * td->nr_files);
-	memset(sd->fd_flags, 0, sizeof(int) * td->nr_files);
-	sd->sgbuf = malloc(sizeof(struct sg_io_hdr) * td->iodepth);
-	memset(sd->sgbuf, 0, sizeof(struct sg_io_hdr) * td->iodepth);
-
-	td->io_ops->data = sd;
+	td->io_ops_data = sd;
 
 	/*
 	 * we want to do it, regardless of whether odirect is set or not
 	 */
-	td->override_sync = 1;
+	td->o.override_sync = 1;
 	return 0;
 }
 
 static int fio_sgio_type_check(struct thread_data *td, struct fio_file *f)
 {
-	struct sgio_data *sd = td->io_ops->data;
-	unsigned int bs;
+	struct sgio_data *sd = td->io_ops_data;
+	unsigned int bs = 0;
+	unsigned long long max_lba = 0;
 
-	if (f->filetype == FIO_TYPE_BD) {
+	if (f->filetype == FIO_TYPE_BLOCK) {
 		if (ioctl(f->fd, BLKSSZGET, &bs) < 0) {
 			td_verror(td, errno, "ioctl");
 			return 1;
@@ -363,53 +865,320 @@ static int fio_sgio_type_check(struct thread_data *td, struct fio_file *f)
 			return 1;
 		}
 
-		ret = fio_sgio_get_bs(td, &bs);
-		if (ret)
+		ret = fio_sgio_read_capacity(td, &bs, &max_lba);
+		if (ret) {
+			td_verror(td, td->error, "fio_sgio_read_capacity");
+			log_err("ioengine sg unable to read capacity successfully\n");
 			return 1;
+		}
 	} else {
-		log_err("ioengine sgio only works on block devices\n");
+		td_verror(td, EINVAL, "wrong file type");
+		log_err("ioengine sg only works on block or character devices\n");
 		return 1;
 	}
 
 	sd->bs = bs;
+	// Determine size of commands needed based on max_lba
+	if (max_lba >= MAX_10B_LBA) {
+		dprint(FD_IO, "sgio_type_check: using 16 byte read/write "
+			"commands for lba above 0x%016llx/0x%016llx\n",
+			MAX_10B_LBA, max_lba);
+	}
 
-	if (f->filetype == FIO_TYPE_BD) {
+	if (f->filetype == FIO_TYPE_BLOCK) {
 		td->io_ops->getevents = NULL;
 		td->io_ops->event = NULL;
+		td->io_ops->commit = NULL;
+		/*
+		** Setting these functions to null may cause problems
+		** with filename=/dev/sda:/dev/sg0 since we are only
+		** considering a single file
+		*/
 	}
+	sd->type_checked = 1;
 
 	return 0;
 }
 
 static int fio_sgio_open(struct thread_data *td, struct fio_file *f)
 {
-	struct sgio_data *sd = td->io_ops->data;
+	struct sgio_data *sd = td->io_ops_data;
 	int ret;
 
 	ret = generic_open_file(td, f);
 	if (ret)
 		return ret;
 
-	if (!sd->type_checked && fio_sgio_type_check(td, f)) {
-		generic_close_file(td, f);
+	if (sd && !sd->type_checked && fio_sgio_type_check(td, f)) {
+		ret = generic_close_file(td, f);
 		return 1;
 	}
 
 	return 0;
 }
 
+/*
+ * Build an error string with details about the driver, host or scsi
+ * error contained in the sg header Caller will use as necessary.
+ */
+static char *fio_sgio_errdetails(struct io_u *io_u)
+{
+	struct sg_io_hdr *hdr = &io_u->hdr;
+#define MAXERRDETAIL 1024
+#define MAXMSGCHUNK  128
+	char *msg, msgchunk[MAXMSGCHUNK];
+	int i;
+
+	msg = calloc(1, MAXERRDETAIL);
+	strcpy(msg, "");
+
+	/*
+	 * can't seem to find sg_err.h, so I'll just echo the define values
+	 * so others can search on internet to find clearer clues of meaning.
+	 */
+	if (hdr->info & SG_INFO_CHECK) {
+		if (hdr->host_status) {
+			snprintf(msgchunk, MAXMSGCHUNK, "SG Host Status: 0x%02x; ", hdr->host_status);
+			strlcat(msg, msgchunk, MAXERRDETAIL);
+			switch (hdr->host_status) {
+			case 0x01:
+				strlcat(msg, "SG_ERR_DID_NO_CONNECT", MAXERRDETAIL);
+				break;
+			case 0x02:
+				strlcat(msg, "SG_ERR_DID_BUS_BUSY", MAXERRDETAIL);
+				break;
+			case 0x03:
+				strlcat(msg, "SG_ERR_DID_TIME_OUT", MAXERRDETAIL);
+				break;
+			case 0x04:
+				strlcat(msg, "SG_ERR_DID_BAD_TARGET", MAXERRDETAIL);
+				break;
+			case 0x05:
+				strlcat(msg, "SG_ERR_DID_ABORT", MAXERRDETAIL);
+				break;
+			case 0x06:
+				strlcat(msg, "SG_ERR_DID_PARITY", MAXERRDETAIL);
+				break;
+			case 0x07:
+				strlcat(msg, "SG_ERR_DID_ERROR (internal error)", MAXERRDETAIL);
+				break;
+			case 0x08:
+				strlcat(msg, "SG_ERR_DID_RESET", MAXERRDETAIL);
+				break;
+			case 0x09:
+				strlcat(msg, "SG_ERR_DID_BAD_INTR (unexpected)", MAXERRDETAIL);
+				break;
+			case 0x0a:
+				strlcat(msg, "SG_ERR_DID_PASSTHROUGH", MAXERRDETAIL);
+				break;
+			case 0x0b:
+				strlcat(msg, "SG_ERR_DID_SOFT_ERROR (driver retry?)", MAXERRDETAIL);
+				break;
+			case 0x0c:
+				strlcat(msg, "SG_ERR_DID_IMM_RETRY", MAXERRDETAIL);
+				break;
+			case 0x0d:
+				strlcat(msg, "SG_ERR_DID_REQUEUE", MAXERRDETAIL);
+				break;
+			case 0x0e:
+				strlcat(msg, "SG_ERR_DID_TRANSPORT_DISRUPTED", MAXERRDETAIL);
+				break;
+			case 0x0f:
+				strlcat(msg, "SG_ERR_DID_TRANSPORT_FAILFAST", MAXERRDETAIL);
+				break;
+			case 0x10:
+				strlcat(msg, "SG_ERR_DID_TARGET_FAILURE", MAXERRDETAIL);
+				break;
+			case 0x11:
+				strlcat(msg, "SG_ERR_DID_NEXUS_FAILURE", MAXERRDETAIL);
+				break;
+			case 0x12:
+				strlcat(msg, "SG_ERR_DID_ALLOC_FAILURE", MAXERRDETAIL);
+				break;
+			case 0x13:
+				strlcat(msg, "SG_ERR_DID_MEDIUM_ERROR", MAXERRDETAIL);
+				break;
+			default:
+				strlcat(msg, "Unknown", MAXERRDETAIL);
+				break;
+			}
+			strlcat(msg, ". ", MAXERRDETAIL);
+		}
+		if (hdr->driver_status) {
+			snprintf(msgchunk, MAXMSGCHUNK, "SG Driver Status: 0x%02x; ", hdr->driver_status);
+			strlcat(msg, msgchunk, MAXERRDETAIL);
+			switch (hdr->driver_status & 0x0F) {
+			case 0x01:
+				strlcat(msg, "SG_ERR_DRIVER_BUSY", MAXERRDETAIL);
+				break;
+			case 0x02:
+				strlcat(msg, "SG_ERR_DRIVER_SOFT", MAXERRDETAIL);
+				break;
+			case 0x03:
+				strlcat(msg, "SG_ERR_DRIVER_MEDIA", MAXERRDETAIL);
+				break;
+			case 0x04:
+				strlcat(msg, "SG_ERR_DRIVER_ERROR", MAXERRDETAIL);
+				break;
+			case 0x05:
+				strlcat(msg, "SG_ERR_DRIVER_INVALID", MAXERRDETAIL);
+				break;
+			case 0x06:
+				strlcat(msg, "SG_ERR_DRIVER_TIMEOUT", MAXERRDETAIL);
+				break;
+			case 0x07:
+				strlcat(msg, "SG_ERR_DRIVER_HARD", MAXERRDETAIL);
+				break;
+			case 0x08:
+				strlcat(msg, "SG_ERR_DRIVER_SENSE", MAXERRDETAIL);
+				break;
+			default:
+				strlcat(msg, "Unknown", MAXERRDETAIL);
+				break;
+			}
+			strlcat(msg, "; ", MAXERRDETAIL);
+			switch (hdr->driver_status & 0xF0) {
+			case 0x10:
+				strlcat(msg, "SG_ERR_SUGGEST_RETRY", MAXERRDETAIL);
+				break;
+			case 0x20:
+				strlcat(msg, "SG_ERR_SUGGEST_ABORT", MAXERRDETAIL);
+				break;
+			case 0x30:
+				strlcat(msg, "SG_ERR_SUGGEST_REMAP", MAXERRDETAIL);
+				break;
+			case 0x40:
+				strlcat(msg, "SG_ERR_SUGGEST_DIE", MAXERRDETAIL);
+				break;
+			case 0x80:
+				strlcat(msg, "SG_ERR_SUGGEST_SENSE", MAXERRDETAIL);
+				break;
+			}
+			strlcat(msg, ". ", MAXERRDETAIL);
+		}
+		if (hdr->status) {
+			snprintf(msgchunk, MAXMSGCHUNK, "SG SCSI Status: 0x%02x; ", hdr->status);
+			strlcat(msg, msgchunk, MAXERRDETAIL);
+			// SCSI 3 status codes
+			switch (hdr->status) {
+			case 0x02:
+				strlcat(msg, "CHECK_CONDITION", MAXERRDETAIL);
+				break;
+			case 0x04:
+				strlcat(msg, "CONDITION_MET", MAXERRDETAIL);
+				break;
+			case 0x08:
+				strlcat(msg, "BUSY", MAXERRDETAIL);
+				break;
+			case 0x10:
+				strlcat(msg, "INTERMEDIATE", MAXERRDETAIL);
+				break;
+			case 0x14:
+				strlcat(msg, "INTERMEDIATE_CONDITION_MET", MAXERRDETAIL);
+				break;
+			case 0x18:
+				strlcat(msg, "RESERVATION_CONFLICT", MAXERRDETAIL);
+				break;
+			case 0x22:
+				strlcat(msg, "COMMAND_TERMINATED", MAXERRDETAIL);
+				break;
+			case 0x28:
+				strlcat(msg, "TASK_SET_FULL", MAXERRDETAIL);
+				break;
+			case 0x30:
+				strlcat(msg, "ACA_ACTIVE", MAXERRDETAIL);
+				break;
+			case 0x40:
+				strlcat(msg, "TASK_ABORTED", MAXERRDETAIL);
+				break;
+			default:
+				strlcat(msg, "Unknown", MAXERRDETAIL);
+				break;
+			}
+			strlcat(msg, ". ", MAXERRDETAIL);
+		}
+		if (hdr->sb_len_wr) {
+			snprintf(msgchunk, MAXMSGCHUNK, "Sense Data (%d bytes):", hdr->sb_len_wr);
+			strlcat(msg, msgchunk, MAXERRDETAIL);
+			for (i = 0; i < hdr->sb_len_wr; i++) {
+				snprintf(msgchunk, MAXMSGCHUNK, " %02x", hdr->sbp[i]);
+				strlcat(msg, msgchunk, MAXERRDETAIL);
+			}
+			strlcat(msg, ". ", MAXERRDETAIL);
+		}
+		if (hdr->resid != 0) {
+			snprintf(msgchunk, MAXMSGCHUNK, "SG Driver: %d bytes out of %d not transferred. ", hdr->resid, hdr->dxfer_len);
+			strlcat(msg, msgchunk, MAXERRDETAIL);
+		}
+	}
+
+	if (!(hdr->info & SG_INFO_CHECK) && !strlen(msg))
+		strncpy(msg, "SG Driver did not report a Host, Driver or Device check",
+			MAXERRDETAIL - 1);
+
+	return msg;
+}
+
+/*
+ * get max file size from read capacity.
+ */
+static int fio_sgio_get_file_size(struct thread_data *td, struct fio_file *f)
+{
+	/*
+	 * get_file_size is being called even before sgio_init is
+	 * called, so none of the sg_io structures are
+	 * initialized in the thread_data yet.  So we need to do the
+	 * ReadCapacity without any of those helpers.  One of the effects
+	 * is that ReadCapacity may get called 4 times on each open:
+	 * readcap(10) followed by readcap(16) if needed - just to get
+	 * the file size after the init occurs - it will be called
+	 * again when "type_check" is called during structure
+	 * initialization I'm not sure how to prevent this little
+	 * inefficiency.
+	 */
+	unsigned int bs = 0;
+	unsigned long long max_lba = 0;
+	int ret;
+
+	if (fio_file_size_known(f))
+		return 0;
+
+	if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) {
+		td_verror(td, EINVAL, "wrong file type");
+		log_err("ioengine sg only works on block or character devices\n");
+		return 1;
+	}
+
+	ret = fio_sgio_read_capacity(td, &bs, &max_lba);
+	if (ret ) {
+		td_verror(td, td->error, "fio_sgio_read_capacity");
+		log_err("ioengine sg unable to successfully execute read capacity to get block size and maximum lba\n");
+		return 1;
+	}
+
+	f->real_file_size = (max_lba + 1) * bs;
+	fio_file_set_size_known(f);
+	return 0;
+}
+
+
 static struct ioengine_ops ioengine = {
 	.name		= "sg",
 	.version	= FIO_IOOPS_VERSION,
 	.init		= fio_sgio_init,
 	.prep		= fio_sgio_prep,
 	.queue		= fio_sgio_queue,
+	.commit		= fio_sgio_commit,
 	.getevents	= fio_sgio_getevents,
+	.errdetails	= fio_sgio_errdetails,
 	.event		= fio_sgio_event,
 	.cleanup	= fio_sgio_cleanup,
 	.open_file	= fio_sgio_open,
 	.close_file	= generic_close_file,
+	.get_file_size	= fio_sgio_get_file_size,
 	.flags		= FIO_SYNCIO | FIO_RAWIO,
+	.options	= options,
+	.option_struct_size	= sizeof(struct sg_options)
 };
 
 #else /* FIO_HAVE_SGIO */
@@ -421,12 +1190,12 @@ static struct ioengine_ops ioengine = {
  */
 static int fio_sgio_init(struct thread_data fio_unused *td)
 {
-	fprintf(stderr, "fio: sgio not available\n");
+	log_err("fio: ioengine sg not available\n");
 	return 1;
 }
 
 static struct ioengine_ops ioengine = {
-	.name		= "sgio",
+	.name		= "sg",
 	.version	= FIO_IOOPS_VERSION,
 	.init		= fio_sgio_init,
 };