[RFC] sg: add statistics similar to st

Commit Message

Douglas Gilbert June 3, 2021, 1:12 a.m. UTC

Using the existing statistics gathering framework from the st driver,
collect statistics for access via sysfs. The sysstat package already
has a utility called tapestat for presenting st statistics. Its
author is keen to use the existing tapestat code for showing sg
statistics (rather than write a new utility).

In keeping with the sg driver being SCSI command agnostic, the "read"
statistics are compiled for requests that have "data-in" user data
while write statistics are compiled for requests that have "data-out"
user data.

A new module/driver load time parameter called "no_stats" has been
added. It is boolean, the default is to collect statistics.

Signed-off-by: Douglas Gilbert <[e-mail]>
---

This patch is based on my patchset whose cover is titled: "[PATCH
v19 00/45] sg: add v4 interface" sent to the linux-scsi list on
20210523. That patchset is against Martin Petersen's
5.14/scsi-queue branch.

 drivers/scsi/sg.c | 259 ++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 227 insertions(+), 32 deletions(-)

Comments

Damien Le Moal June 3, 2021, 1:39 a.m. UTC | #1

On 2021/06/03 10:12, Douglas Gilbert wrote:
> Using the existing statistics gathering framework from the st driver,
> collect statistics for access via sysfs. The sysstat package already
> has a utility called tapestat for presenting st statistics. Its
> author is keen to use the existing tapestat code for showing sg
> statistics (rather than write a new utility).
> 
> In keeping with the sg driver being SCSI command agnostic, the "read"
> statistics are compiled for requests that have "data-in" user data
> while write statistics are compiled for requests that have "data-out"
> user data.
> 
> A new module/driver load time parameter called "no_stats" has been
> added. It is boolean, the default is to collect statistics.

This definitely will be useful :)

> 
> Signed-off-by: Douglas Gilbert <[e-mail]>
> ---
> 
> This patch is based on my patchset whose cover is titled: "[PATCH
> v19 00/45] sg: add v4 interface" sent to the linux-scsi list on
> 20210523. That patchset is against Martin Petersen's
> 5.14/scsi-queue branch.
> 
>  drivers/scsi/sg.c | 259 ++++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 227 insertions(+), 32 deletions(-)
> 
> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> index d47efed3a3ca..eedeeb1872f7 100644
> --- a/drivers/scsi/sg.c
> +++ b/drivers/scsi/sg.c
> @@ -143,6 +143,7 @@ int sg_big_buff = SG_DEF_RESERVED_SIZE;
>   */
>  static int def_reserved_size = -1;	/* picks up init parameter */
>  static int sg_allow_dio = SG_ALLOW_DIO_DEF;	/* ignored by code */
> +static bool sg_no_stats;

Most of the tests on this parameter are in the form:

if (!sg_no_stats)

So maybe reverse this and define

static bool sg_stats; (default true)

so that the tests becomes simpler:

if (sg_stats) ...

>  
>  static int scatter_elem_sz = SG_SCATTER_SZ;
>  
> @@ -213,7 +214,7 @@ struct sg_request {	/* active SCSI command or inactive request */
>  	atomic_t rq_st;		/* request state, holds a enum sg_rq_state */
>  	u8 cmd_opcode;		/* first byte of SCSI cdb */
>  	blk_qc_t cookie;	/* ids 1 or more queues for blk_poll() */
> -	u64 start_ns;		/* starting point of command duration calc */
> +	ktime_t start_dur;	/* start time if before completion */
>  	unsigned long frq_bm[1];	/* see SG_FRQ_* defines above */
>  	u8 *sense_bp;		/* mempool alloc-ed sense buffer, as needed */
>  	struct sg_fd *parentfp;	/* pointer to owning fd, even when on fl */
> @@ -258,6 +259,7 @@ struct sg_device { /* holds the state of each scsi generic device */
>  	unsigned long fdev_bm[1];	/* see SG_FDEV_* defines above */
>  	struct gendisk *disk;
>  	struct cdev *cdev;
> +	struct sg_dev_stats *statsp;	/* NULL when no_stats=true */
>  	struct xarray sfp_arr;
>  	struct kref d_ref;
>  };
> @@ -275,6 +277,19 @@ struct sg_comm_wr_t {  /* arguments to sg_common_write() */
>  	const u8 __user *u_cmdp;
>  };
>  
> +struct sg_dev_stats {	/* copied from drivers/scsi/st.h scsi_tape_stats */
> +	atomic64_t read_byte_cnt;	/* data-in bytes */
> +	atomic64_t write_byte_cnt;	/* data-out bytes */
> +	atomic64_t in_flight;		/* Number of I/Os in flight */
> +	atomic64_t read_cnt;		/* Count of data-in requests */
> +	atomic64_t write_cnt;		/* Count of data-out requests */
> +	atomic64_t other_cnt;		/* Count of non-data requests */
> +	atomic64_t resid_cnt;		/* Count of resid_len > 0 */
> +	atomic64_t tot_read_time;	/* time spent completing data-in_s */
> +	atomic64_t tot_write_time;	/* time spent completing data-out_s */
> +	atomic64_t tot_io_time;		/* ktime spent doing any I/O */
> +};
> +
>  /* tasklet or soft irq callback */
>  static void sg_rq_end_io(struct request *rq, blk_status_t status);
>  /* Declarations of other static functions used before they are defined */
> @@ -306,6 +321,8 @@ static struct sg_request *sg_mk_srp_sgat(struct sg_fd *sfp, bool first,
>  static int sg_sfp_blk_poll(struct sg_fd *sfp, int loop_count);
>  static int sg_srp_q_blk_poll(struct sg_request *srp, struct request_queue *q,
>  			     int loop_count);
> +static u32 sg_get_dur(struct sg_request *srp, const enum sg_rq_state *sr_stp,
> +		      bool *is_durp);
>  #if IS_ENABLED(CONFIG_SCSI_LOGGING) && IS_ENABLED(SG_DEBUG)
>  static const char *sg_rq_st_str(enum sg_rq_state rq_st, bool long_str);
>  #endif
> @@ -1020,7 +1037,12 @@ sg_execute_cmd(struct sg_fd *sfp, struct sg_request *srp)
>  	is_v4h = test_bit(SG_FRQ_IS_V4I, srp->frq_bm);
>  	sync = test_bit(SG_FRQ_SYNC_INVOC, srp->frq_bm);
>  	SG_LOG(3, sfp, "%s: is_v4h=%d\n", __func__, (int)is_v4h);
> -	srp->start_ns = ktime_get_boottime_ns();
> +	if (sg_no_stats) {
> +		WRITE_ONCE(srp->start_dur, 0);
> +	} else {
> +		atomic64_inc(&sdp->statsp->in_flight);
> +		WRITE_ONCE(srp->start_dur, ktime_get_boottime());
> +	}
>  	srp->duration = 0;
>  
>  	if (!is_v4h && srp->s_hdr3.interface_id == '\0')
> @@ -1194,11 +1216,45 @@ sg_copy_sense(struct sg_request *srp, bool v4_active)
>  	return sb_len_ret;
>  }
>  
> +static void
> +sg_do_stats(struct sg_fd *sfp, struct sg_request *srp, bool v4_active)
> +{
> +	int dir = v4_active ? srp->s_hdr4.dir : srp->s_hdr3.dxfer_direction;
> +	ktime_t kt = READ_ONCE(srp->start_dur);
> +	struct sg_dev_stats *statsp = sfp->parentdp->statsp;
> +
> +	if (!statsp)
> +		return;
> +	if (dir == SG_DXFER_TO_DEV) {		/* data-out, write-like */
> +		atomic64_add(ktime_to_ns(kt), &statsp->tot_write_time);
> +		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
> +		atomic64_inc(&statsp->write_cnt);
> +		atomic64_add(srp->sgat_h.dlen, &statsp->write_byte_cnt);
> +	} else if (dir == SG_DXFER_FROM_DEV) {	/* data-in, read-like */
> +		atomic64_add(ktime_to_ns(kt), &statsp->tot_read_time);
> +		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
> +		atomic64_inc(&statsp->read_cnt);
> +		atomic64_add(srp->sgat_h.dlen, &statsp->read_byte_cnt);
> +		if (srp->in_resid > 0)
> +			atomic64_inc(&statsp->resid_cnt);
> +	} else {	/* no data transfer (e.g. TEST UNIT READY) */
> +		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
> +		atomic64_inc(&statsp->other_cnt);
> +	}
> +	atomic64_dec(&statsp->in_flight);
> +}

Instead of all these atomic calls, I wonder if it would be less overhead to
protect statsp with a spinlock ?

> +
>  static int
>  sg_rec_state_v3v4(struct sg_fd *sfp, struct sg_request *srp, bool v4_active)
>  {
>  	u32 rq_res = srp->rq_result;
>  
> +	if (!sg_no_stats) {
> +		const enum sg_rq_state sr_st = SG_RS_BUSY;
> +
> +		sg_do_stats(sfp, srp, v4_active);
> +		srp->duration = sg_get_dur(srp, &sr_st, NULL);
> +	}
>  	if (unlikely(srp->rq_result & 0xff)) {
>  		int sb_len_wr = sg_copy_sense(srp, v4_active);
>  
> @@ -1625,49 +1681,43 @@ sg_calc_sgat_param(struct sg_device *sdp)
>  	sdp->max_sgat_sz = sz;
>  }
>  
> -static u32
> -sg_calc_rq_dur(const struct sg_request *srp)
> -{
> -	ktime_t ts0 = ns_to_ktime(srp->start_ns);
> -	ktime_t now_ts;
> -	s64 diff;
> -
> -	if (ts0 == 0)
> -		return 0;
> -	if (unlikely(ts0 == S64_MAX))	/* _prior_ to issuing req */
> -		return 999999999;	/* eye catching */
> -	now_ts = ktime_get_boottime();
> -	if (unlikely(ts0 > now_ts))
> -		return 999999998;
> -	/* unlikely req duration will exceed 2**32 milliseconds */
> -	diff = ktime_ms_delta(now_ts, ts0);
> -	return (diff > (s64)U32_MAX) ? 3999999999U : (u32)diff;
> -}
> -
> -/* Return of U32_MAX means srp is inactive state */
>  static u32
>  sg_get_dur(struct sg_request *srp, const enum sg_rq_state *sr_stp,
>  	   bool *is_durp)
>  {
>  	bool is_dur = false;
> -	u32 res = U32_MAX;
> +	s64 dur_ns;
> +	ktime_t start_dur = READ_ONCE(srp->start_dur);
>  
> +	if (ktime_to_ns(start_dur) <= 0) {
> +		is_dur = true;
> +		dur_ns = 0;
> +		goto fini;
> +	}
>  	switch (sr_stp ? *sr_stp : atomic_read(&srp->rq_st)) {
> -	case SG_RS_INFLIGHT:
>  	case SG_RS_BUSY:
> -		res = sg_calc_rq_dur(srp);
> +		if (test_bit(SG_FRQ_ISSUED, srp->frq_bm)) {
> +			dur_ns = ktime_to_ns(start_dur);
> +			is_dur = true;
> +			break;
> +		}
> +		dur_ns = 1;
> +		break;
> +	case SG_RS_INFLIGHT:
> +		dur_ns = ktime_sub(ktime_get_boottime(), start_dur);
>  		break;
>  	case SG_RS_AWAIT_RCV:
>  	case SG_RS_INACTIVE:
> -		res = srp->duration;
> +		dur_ns = ktime_to_ns(start_dur);
>  		is_dur = true;	/* completion has occurred, timing finished */
>  		break;
> -	default:
> -		break;
>  	}
> +fini:
>  	if (is_durp)
>  		*is_durp = is_dur;
> -	return res;
> +	if (dur_ns < 2)
> +		return 999999999;	/* timekeeping problem */
> +	return ktime_to_ms(ns_to_ktime(dur_ns));
>  }
>  
>  static void
> @@ -1678,8 +1728,6 @@ sg_fill_request_element(struct sg_fd *sfp, struct sg_request *srp,
>  
>  	xa_lock_irqsave(&sfp->srp_arr, iflags);
>  	rip->duration = sg_get_dur(srp, NULL, NULL);
> -	if (rip->duration == U32_MAX)
> -		rip->duration = 0;
>  	rip->orphan = test_bit(SG_FRQ_IS_ORPHAN, srp->frq_bm);
>  	rip->sg_io_owned = test_bit(SG_FRQ_SYNC_INVOC, srp->frq_bm);
>  	rip->problem = !!(srp->rq_result & SG_ML_RESULT_MSK);
> @@ -2502,6 +2550,7 @@ sg_rq_end_io(struct request *rqq, blk_status_t status)
>  	int a_resid, slen;
>  	u32 rq_result;
>  	unsigned long iflags;
> +	ktime_t start_tm;
>  	struct sg_request *srp = rqq->end_io_data;
>  	struct scsi_request *scsi_rp = scsi_req(rqq);
>  	struct sg_device *sdp;
> @@ -2528,7 +2577,9 @@ sg_rq_end_io(struct request *rqq, blk_status_t status)
>  
>  	SG_LOG(6, sfp, "%s: pack_id=%d, res=0x%x\n", __func__, srp->pack_id,
>  	       rq_result);
> -	srp->duration = sg_calc_rq_dur(srp);
> +	start_tm = READ_ONCE(srp->start_dur);
> +	if (start_tm > 0)
> +		WRITE_ONCE(srp->start_dur, ktime_sub(ktime_get_boottime(), start_tm));
>  	if (unlikely((rq_result & SG_ML_RESULT_MSK) && slen > 0 &&
>  		     test_bit(SG_FDEV_LOG_SENSE, sdp->fdev_bm))) {
>  		u32 scsi_stat = rq_result & 0xff;
> @@ -2683,6 +2734,8 @@ sg_add_device_helper(struct gendisk *disk, struct scsi_device *scsidp)
>  	return sdp;
>  }
>  
> +static const struct attribute_group *sg_dev_groups[];
> +
>  static int
>  sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
>  {
> @@ -2714,6 +2767,9 @@ sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
>  		error = PTR_ERR(sdp);
>  		goto out;
>  	}
> +	if (!sg_no_stats)
> +		sdp->statsp = kzalloc(sizeof(*sdp->statsp), GFP_KERNEL);
> +		/* don't worry if NULL, probably a lot of devices */
>  
>  	error = cdev_add(cdev, MKDEV(SCSI_GENERIC_MAJOR, sdp->index), 1);
>  	if (error)
> @@ -2723,6 +2779,7 @@ sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
>  	if (sg_sysfs_valid) {
>  		struct device *sg_class_member;
>  
> +		sg_sysfs_class->dev_groups = sg_dev_groups;
>  		sg_class_member = device_create(sg_sysfs_class, cl_dev->parent,
>  						MKDEV(SCSI_GENERIC_MAJOR,
>  						      sdp->index),
> @@ -2749,6 +2806,7 @@ sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
>  	return 0;
>  
>  cdev_add_err:
> +	kfree(sdp->statsp);
>  	write_lock_irqsave(&sg_index_lock, iflags);
>  	idr_remove(&sg_index_idr, sdp->index);
>  	write_unlock_irqrestore(&sg_index_lock, iflags);
> @@ -2777,6 +2835,7 @@ sg_device_destroy(struct kref *kref)
>  	 */
>  
>  	xa_destroy(&sdp->sfp_arr);
> +	kfree(sdp->statsp);
>  	write_lock_irqsave(&sg_index_lock, flags);
>  	idr_remove(&sg_index_idr, sdp->index);
>  	write_unlock_irqrestore(&sg_index_lock, flags);
> @@ -3919,6 +3978,140 @@ sg_rq_st_str(enum sg_rq_state rq_st, bool long_str)
>  }
>  #endif
>  
> +static ssize_t read_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->read_cnt));

If you switch to use a spinlock for protecting struct sg_dev_stats, these can
probably use READ_ONCE() without taking the spinlock, I think.

> +}
> +static DEVICE_ATTR_RO(read_cnt);
> +
> +static ssize_t read_byte_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->read_byte_cnt));
> +}
> +static DEVICE_ATTR_RO(read_byte_cnt);
> +
> +static ssize_t read_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_read_time));
> +}
> +static DEVICE_ATTR_RO(read_ns);
> +
> +static ssize_t write_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->write_cnt));
> +}
> +static DEVICE_ATTR_RO(write_cnt);
> +
> +static ssize_t write_byte_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->write_byte_cnt));
> +}
> +static DEVICE_ATTR_RO(write_byte_cnt);
> +
> +static ssize_t write_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_write_time));
> +}
> +static DEVICE_ATTR_RO(write_ns);
> +
> +static ssize_t in_flight_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->in_flight));
> +}
> +static DEVICE_ATTR_RO(in_flight);
> +
> +static ssize_t io_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_io_time));
> +}
> +static DEVICE_ATTR_RO(io_ns);
> +
> +static ssize_t other_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->other_cnt));
> +}
> +static DEVICE_ATTR_RO(other_cnt);
> +
> +static ssize_t resid_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
> +{
> +	struct sg_device *sdp = dev_get_drvdata(dev);
> +	struct sg_dev_stats *sp = sdp->statsp;
> +
> +	if (!sdp || !sp)
> +		return -EINVAL;
> +	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->resid_cnt));
> +}
> +static DEVICE_ATTR_RO(resid_cnt);
> +
> +static struct attribute *sg_stats_attrs[] = {
> +	&dev_attr_read_cnt.attr,
> +	&dev_attr_read_byte_cnt.attr,
> +	&dev_attr_read_ns.attr,
> +	&dev_attr_write_cnt.attr,
> +	&dev_attr_write_byte_cnt.attr,
> +	&dev_attr_write_ns.attr,
> +	&dev_attr_in_flight.attr,
> +	&dev_attr_io_ns.attr,
> +	&dev_attr_other_cnt.attr,
> +	&dev_attr_resid_cnt.attr,
> +	NULL,
> +};
> +
> +static struct attribute_group sg_stats_group = {
> +	.name = "stats",
> +	.attrs = sg_stats_attrs,
> +};
> +
> +static const struct attribute_group *sg_dev_groups[] = {
> +	&sg_stats_group,
> +	NULL,
> +};
> +
>  #if IS_ENABLED(SG_PROC_OR_DEBUG_FS)
>  
>  #define SG_SNAPSHOT_DEV_MAX 4
> @@ -4623,6 +4816,7 @@ static void sg_dfs_exit(void) {}
>  module_param_named(scatter_elem_sz, scatter_elem_sz, int, 0644);
>  module_param_named(def_reserved_size, def_reserved_size, int, 0644);
>  module_param_named(allow_dio, sg_allow_dio, int, 0644);
> +module_param_named(no_stats, sg_no_stats, bool, 0644);
>  
>  MODULE_AUTHOR("Douglas Gilbert");
>  MODULE_DESCRIPTION("SCSI generic (sg) driver");
> @@ -4633,5 +4827,6 @@ MODULE_ALIAS_CHARDEV_MAJOR(SCSI_GENERIC_MAJOR);
>  MODULE_PARM_DESC(scatter_elem_sz, "scatter gather element size (default: max(SG_SCATTER_SZ, PAGE_SIZE))");
>  MODULE_PARM_DESC(def_reserved_size, "size of buffer reserved for each fd");
>  MODULE_PARM_DESC(allow_dio, "allow direct I/O (default: 0 (disallow)); now ignored");
> +MODULE_PARM_DESC(no_stats, "don't collect per device statistics (default: 0)");
>  module_init(init_sg);
>  module_exit(exit_sg);
>

Patch

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index d47efed3a3ca..eedeeb1872f7 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -143,6 +143,7 @@  int sg_big_buff = SG_DEF_RESERVED_SIZE;
  */
 static int def_reserved_size = -1;	/* picks up init parameter */
 static int sg_allow_dio = SG_ALLOW_DIO_DEF;	/* ignored by code */
+static bool sg_no_stats;
 
 static int scatter_elem_sz = SG_SCATTER_SZ;
 
@@ -213,7 +214,7 @@  struct sg_request {	/* active SCSI command or inactive request */
 	atomic_t rq_st;		/* request state, holds a enum sg_rq_state */
 	u8 cmd_opcode;		/* first byte of SCSI cdb */
 	blk_qc_t cookie;	/* ids 1 or more queues for blk_poll() */
-	u64 start_ns;		/* starting point of command duration calc */
+	ktime_t start_dur;	/* start time if before completion */
 	unsigned long frq_bm[1];	/* see SG_FRQ_* defines above */
 	u8 *sense_bp;		/* mempool alloc-ed sense buffer, as needed */
 	struct sg_fd *parentfp;	/* pointer to owning fd, even when on fl */
@@ -258,6 +259,7 @@  struct sg_device { /* holds the state of each scsi generic device */
 	unsigned long fdev_bm[1];	/* see SG_FDEV_* defines above */
 	struct gendisk *disk;
 	struct cdev *cdev;
+	struct sg_dev_stats *statsp;	/* NULL when no_stats=true */
 	struct xarray sfp_arr;
 	struct kref d_ref;
 };
@@ -275,6 +277,19 @@  struct sg_comm_wr_t {  /* arguments to sg_common_write() */
 	const u8 __user *u_cmdp;
 };
 
+struct sg_dev_stats {	/* copied from drivers/scsi/st.h scsi_tape_stats */
+	atomic64_t read_byte_cnt;	/* data-in bytes */
+	atomic64_t write_byte_cnt;	/* data-out bytes */
+	atomic64_t in_flight;		/* Number of I/Os in flight */
+	atomic64_t read_cnt;		/* Count of data-in requests */
+	atomic64_t write_cnt;		/* Count of data-out requests */
+	atomic64_t other_cnt;		/* Count of non-data requests */
+	atomic64_t resid_cnt;		/* Count of resid_len > 0 */
+	atomic64_t tot_read_time;	/* time spent completing data-in_s */
+	atomic64_t tot_write_time;	/* time spent completing data-out_s */
+	atomic64_t tot_io_time;		/* ktime spent doing any I/O */
+};
+
 /* tasklet or soft irq callback */
 static void sg_rq_end_io(struct request *rq, blk_status_t status);
 /* Declarations of other static functions used before they are defined */
@@ -306,6 +321,8 @@  static struct sg_request *sg_mk_srp_sgat(struct sg_fd *sfp, bool first,
 static int sg_sfp_blk_poll(struct sg_fd *sfp, int loop_count);
 static int sg_srp_q_blk_poll(struct sg_request *srp, struct request_queue *q,
 			     int loop_count);
+static u32 sg_get_dur(struct sg_request *srp, const enum sg_rq_state *sr_stp,
+		      bool *is_durp);
 #if IS_ENABLED(CONFIG_SCSI_LOGGING) && IS_ENABLED(SG_DEBUG)
 static const char *sg_rq_st_str(enum sg_rq_state rq_st, bool long_str);
 #endif
@@ -1020,7 +1037,12 @@  sg_execute_cmd(struct sg_fd *sfp, struct sg_request *srp)
 	is_v4h = test_bit(SG_FRQ_IS_V4I, srp->frq_bm);
 	sync = test_bit(SG_FRQ_SYNC_INVOC, srp->frq_bm);
 	SG_LOG(3, sfp, "%s: is_v4h=%d\n", __func__, (int)is_v4h);
-	srp->start_ns = ktime_get_boottime_ns();
+	if (sg_no_stats) {
+		WRITE_ONCE(srp->start_dur, 0);
+	} else {
+		atomic64_inc(&sdp->statsp->in_flight);
+		WRITE_ONCE(srp->start_dur, ktime_get_boottime());
+	}
 	srp->duration = 0;
 
 	if (!is_v4h && srp->s_hdr3.interface_id == '\0')
@@ -1194,11 +1216,45 @@  sg_copy_sense(struct sg_request *srp, bool v4_active)
 	return sb_len_ret;
 }
 
+static void
+sg_do_stats(struct sg_fd *sfp, struct sg_request *srp, bool v4_active)
+{
+	int dir = v4_active ? srp->s_hdr4.dir : srp->s_hdr3.dxfer_direction;
+	ktime_t kt = READ_ONCE(srp->start_dur);
+	struct sg_dev_stats *statsp = sfp->parentdp->statsp;
+
+	if (!statsp)
+		return;
+	if (dir == SG_DXFER_TO_DEV) {		/* data-out, write-like */
+		atomic64_add(ktime_to_ns(kt), &statsp->tot_write_time);
+		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
+		atomic64_inc(&statsp->write_cnt);
+		atomic64_add(srp->sgat_h.dlen, &statsp->write_byte_cnt);
+	} else if (dir == SG_DXFER_FROM_DEV) {	/* data-in, read-like */
+		atomic64_add(ktime_to_ns(kt), &statsp->tot_read_time);
+		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
+		atomic64_inc(&statsp->read_cnt);
+		atomic64_add(srp->sgat_h.dlen, &statsp->read_byte_cnt);
+		if (srp->in_resid > 0)
+			atomic64_inc(&statsp->resid_cnt);
+	} else {	/* no data transfer (e.g. TEST UNIT READY) */
+		atomic64_add(ktime_to_ns(kt), &statsp->tot_io_time);
+		atomic64_inc(&statsp->other_cnt);
+	}
+	atomic64_dec(&statsp->in_flight);
+}
+
 static int
 sg_rec_state_v3v4(struct sg_fd *sfp, struct sg_request *srp, bool v4_active)
 {
 	u32 rq_res = srp->rq_result;
 
+	if (!sg_no_stats) {
+		const enum sg_rq_state sr_st = SG_RS_BUSY;
+
+		sg_do_stats(sfp, srp, v4_active);
+		srp->duration = sg_get_dur(srp, &sr_st, NULL);
+	}
 	if (unlikely(srp->rq_result & 0xff)) {
 		int sb_len_wr = sg_copy_sense(srp, v4_active);
 
@@ -1625,49 +1681,43 @@  sg_calc_sgat_param(struct sg_device *sdp)
 	sdp->max_sgat_sz = sz;
 }
 
-static u32
-sg_calc_rq_dur(const struct sg_request *srp)
-{
-	ktime_t ts0 = ns_to_ktime(srp->start_ns);
-	ktime_t now_ts;
-	s64 diff;
-
-	if (ts0 == 0)
-		return 0;
-	if (unlikely(ts0 == S64_MAX))	/* _prior_ to issuing req */
-		return 999999999;	/* eye catching */
-	now_ts = ktime_get_boottime();
-	if (unlikely(ts0 > now_ts))
-		return 999999998;
-	/* unlikely req duration will exceed 2**32 milliseconds */
-	diff = ktime_ms_delta(now_ts, ts0);
-	return (diff > (s64)U32_MAX) ? 3999999999U : (u32)diff;
-}
-
-/* Return of U32_MAX means srp is inactive state */
 static u32
 sg_get_dur(struct sg_request *srp, const enum sg_rq_state *sr_stp,
 	   bool *is_durp)
 {
 	bool is_dur = false;
-	u32 res = U32_MAX;
+	s64 dur_ns;
+	ktime_t start_dur = READ_ONCE(srp->start_dur);
 
+	if (ktime_to_ns(start_dur) <= 0) {
+		is_dur = true;
+		dur_ns = 0;
+		goto fini;
+	}
 	switch (sr_stp ? *sr_stp : atomic_read(&srp->rq_st)) {
-	case SG_RS_INFLIGHT:
 	case SG_RS_BUSY:
-		res = sg_calc_rq_dur(srp);
+		if (test_bit(SG_FRQ_ISSUED, srp->frq_bm)) {
+			dur_ns = ktime_to_ns(start_dur);
+			is_dur = true;
+			break;
+		}
+		dur_ns = 1;
+		break;
+	case SG_RS_INFLIGHT:
+		dur_ns = ktime_sub(ktime_get_boottime(), start_dur);
 		break;
 	case SG_RS_AWAIT_RCV:
 	case SG_RS_INACTIVE:
-		res = srp->duration;
+		dur_ns = ktime_to_ns(start_dur);
 		is_dur = true;	/* completion has occurred, timing finished */
 		break;
-	default:
-		break;
 	}
+fini:
 	if (is_durp)
 		*is_durp = is_dur;
-	return res;
+	if (dur_ns < 2)
+		return 999999999;	/* timekeeping problem */
+	return ktime_to_ms(ns_to_ktime(dur_ns));
 }
 
 static void
@@ -1678,8 +1728,6 @@  sg_fill_request_element(struct sg_fd *sfp, struct sg_request *srp,
 
 	xa_lock_irqsave(&sfp->srp_arr, iflags);
 	rip->duration = sg_get_dur(srp, NULL, NULL);
-	if (rip->duration == U32_MAX)
-		rip->duration = 0;
 	rip->orphan = test_bit(SG_FRQ_IS_ORPHAN, srp->frq_bm);
 	rip->sg_io_owned = test_bit(SG_FRQ_SYNC_INVOC, srp->frq_bm);
 	rip->problem = !!(srp->rq_result & SG_ML_RESULT_MSK);
@@ -2502,6 +2550,7 @@  sg_rq_end_io(struct request *rqq, blk_status_t status)
 	int a_resid, slen;
 	u32 rq_result;
 	unsigned long iflags;
+	ktime_t start_tm;
 	struct sg_request *srp = rqq->end_io_data;
 	struct scsi_request *scsi_rp = scsi_req(rqq);
 	struct sg_device *sdp;
@@ -2528,7 +2577,9 @@  sg_rq_end_io(struct request *rqq, blk_status_t status)
 
 	SG_LOG(6, sfp, "%s: pack_id=%d, res=0x%x\n", __func__, srp->pack_id,
 	       rq_result);
-	srp->duration = sg_calc_rq_dur(srp);
+	start_tm = READ_ONCE(srp->start_dur);
+	if (start_tm > 0)
+		WRITE_ONCE(srp->start_dur, ktime_sub(ktime_get_boottime(), start_tm));
 	if (unlikely((rq_result & SG_ML_RESULT_MSK) && slen > 0 &&
 		     test_bit(SG_FDEV_LOG_SENSE, sdp->fdev_bm))) {
 		u32 scsi_stat = rq_result & 0xff;
@@ -2683,6 +2734,8 @@  sg_add_device_helper(struct gendisk *disk, struct scsi_device *scsidp)
 	return sdp;
 }
 
+static const struct attribute_group *sg_dev_groups[];
+
 static int
 sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
 {
@@ -2714,6 +2767,9 @@  sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
 		error = PTR_ERR(sdp);
 		goto out;
 	}
+	if (!sg_no_stats)
+		sdp->statsp = kzalloc(sizeof(*sdp->statsp), GFP_KERNEL);
+		/* don't worry if NULL, probably a lot of devices */
 
 	error = cdev_add(cdev, MKDEV(SCSI_GENERIC_MAJOR, sdp->index), 1);
 	if (error)
@@ -2723,6 +2779,7 @@  sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
 	if (sg_sysfs_valid) {
 		struct device *sg_class_member;
 
+		sg_sysfs_class->dev_groups = sg_dev_groups;
 		sg_class_member = device_create(sg_sysfs_class, cl_dev->parent,
 						MKDEV(SCSI_GENERIC_MAJOR,
 						      sdp->index),
@@ -2749,6 +2806,7 @@  sg_add_device(struct device *cl_dev, struct class_interface *cl_intf)
 	return 0;
 
 cdev_add_err:
+	kfree(sdp->statsp);
 	write_lock_irqsave(&sg_index_lock, iflags);
 	idr_remove(&sg_index_idr, sdp->index);
 	write_unlock_irqrestore(&sg_index_lock, iflags);
@@ -2777,6 +2835,7 @@  sg_device_destroy(struct kref *kref)
 	 */
 
 	xa_destroy(&sdp->sfp_arr);
+	kfree(sdp->statsp);
 	write_lock_irqsave(&sg_index_lock, flags);
 	idr_remove(&sg_index_idr, sdp->index);
 	write_unlock_irqrestore(&sg_index_lock, flags);
@@ -3919,6 +3978,140 @@  sg_rq_st_str(enum sg_rq_state rq_st, bool long_str)
 }
 #endif
 
+static ssize_t read_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->read_cnt));
+}
+static DEVICE_ATTR_RO(read_cnt);
+
+static ssize_t read_byte_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->read_byte_cnt));
+}
+static DEVICE_ATTR_RO(read_byte_cnt);
+
+static ssize_t read_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_read_time));
+}
+static DEVICE_ATTR_RO(read_ns);
+
+static ssize_t write_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->write_cnt));
+}
+static DEVICE_ATTR_RO(write_cnt);
+
+static ssize_t write_byte_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->write_byte_cnt));
+}
+static DEVICE_ATTR_RO(write_byte_cnt);
+
+static ssize_t write_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_write_time));
+}
+static DEVICE_ATTR_RO(write_ns);
+
+static ssize_t in_flight_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->in_flight));
+}
+static DEVICE_ATTR_RO(in_flight);
+
+static ssize_t io_ns_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->tot_io_time));
+}
+static DEVICE_ATTR_RO(io_ns);
+
+static ssize_t other_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->other_cnt));
+}
+static DEVICE_ATTR_RO(other_cnt);
+
+static ssize_t resid_cnt_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+	struct sg_device *sdp = dev_get_drvdata(dev);
+	struct sg_dev_stats *sp = sdp->statsp;
+
+	if (!sdp || !sp)
+		return -EINVAL;
+	return scnprintf(buf, PAGE_SIZE, "%lld\n", (long long)atomic64_read(&sp->resid_cnt));
+}
+static DEVICE_ATTR_RO(resid_cnt);
+
+static struct attribute *sg_stats_attrs[] = {
+	&dev_attr_read_cnt.attr,
+	&dev_attr_read_byte_cnt.attr,
+	&dev_attr_read_ns.attr,
+	&dev_attr_write_cnt.attr,
+	&dev_attr_write_byte_cnt.attr,
+	&dev_attr_write_ns.attr,
+	&dev_attr_in_flight.attr,
+	&dev_attr_io_ns.attr,
+	&dev_attr_other_cnt.attr,
+	&dev_attr_resid_cnt.attr,
+	NULL,
+};
+
+static struct attribute_group sg_stats_group = {
+	.name = "stats",
+	.attrs = sg_stats_attrs,
+};
+
+static const struct attribute_group *sg_dev_groups[] = {
+	&sg_stats_group,
+	NULL,
+};
+
 #if IS_ENABLED(SG_PROC_OR_DEBUG_FS)
 
 #define SG_SNAPSHOT_DEV_MAX 4
@@ -4623,6 +4816,7 @@  static void sg_dfs_exit(void) {}
 module_param_named(scatter_elem_sz, scatter_elem_sz, int, 0644);
 module_param_named(def_reserved_size, def_reserved_size, int, 0644);
 module_param_named(allow_dio, sg_allow_dio, int, 0644);
+module_param_named(no_stats, sg_no_stats, bool, 0644);
 
 MODULE_AUTHOR("Douglas Gilbert");
 MODULE_DESCRIPTION("SCSI generic (sg) driver");
@@ -4633,5 +4827,6 @@  MODULE_ALIAS_CHARDEV_MAJOR(SCSI_GENERIC_MAJOR);
 MODULE_PARM_DESC(scatter_elem_sz, "scatter gather element size (default: max(SG_SCATTER_SZ, PAGE_SIZE))");
 MODULE_PARM_DESC(def_reserved_size, "size of buffer reserved for each fd");
 MODULE_PARM_DESC(allow_dio, "allow direct I/O (default: 0 (disallow)); now ignored");
+MODULE_PARM_DESC(no_stats, "don't collect per device statistics (default: 0)");
 module_init(init_sg);
 module_exit(exit_sg);