s390/dasd: Derive adaptive ESE fulltrack heuristic from ft_bias

Turn the middle of the ft_bias range (1..99) into an adaptive heuristic
that switches between fulltrack write (ft1) and plain write ft0 depending
on how sparse the device still is.

A sparse device benefits from fulltrack writes (it avoids the format/retry
cycle); once enough tracks are formatted the per-write overhead of
ft1 outweighs that. An state machine  measures the NRF rate in short ft0
probe windows and flips back to ft1 when it is high
(FT1_ACTIVE -> PROBING -> FT0_STABLE, with a backing-off reprobe interval).

The four parameters are derived from ft_bias by linear interpolation,
anchored so ft_bias == 50 derives the following values:

ese_heu_start_interval - 2000 - IOs in ft1, before first ft0-Probe starts
ese_heu_probe_window - 100 - IOs in probe window
ese_heu_nrf_high - 10 ‰ (= 1 %) - TRACK_FORMAT rate that leads to ft1
ese_heu_max_interval - 500000 - Backoff-Cap: max. IOs between two probes

Higher is more eager to use ft1, and 0/100 skips the heuristic.
The NRF counter is bumped in dasd_eckd_ese_format() for both the classic
NRF sense and the HPF INV_TRACK_FORMAT equivalent.
The state machine resets to ft1 on check_characteristics, full format,
and release-space.

A read-only ese_heuristic_state sysfs attribute exposes the current mode.

Reviewed-by: Jan Höppner <hoeppner@linux.ibm.com>
Signed-off-by: Stefan Haberland <sth@linux.ibm.com>
Link: https://patch.msgid.link/20260805111612.1285190-15-sth@linux.ibm.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

3 files changed, 223 insertions, 6 deletions

diff --git a/drivers/s390/block/dasd_devmap.c b/drivers/s390/block/dasd_devmap.c
index 035c022255b6..f6aab94b7be6 100644
--- a/drivers/s390/block/dasd_devmap.c
+++ b/drivers/s390/block/dasd_devmap.c
@@ -1659,8 +1659,14 @@ static ssize_t full_track_bias_store(struct device *dev,
 	if (IS_ERR(device))
 		return -ENODEV;
 
+	/*
+	 * ft_bias is the tuning target; fulltrack is a best-effort mode hint
+	 * that the per-IO heuristic also updates locklessly. A racing writer can
+	 * at most leave a transient mismatch that self-corrects on the next IO,
+	 * never corruption, so the update is left unlocked.
+	 */
 	device->ft_bias = val;
-	device->fulltrack = val ? 1 : 0;
+	dasd_ft_bias_apply(device);
 
 	dasd_put_device(device);
 	return count;
@@ -1668,6 +1674,44 @@ static ssize_t full_track_bias_store(struct device *dev,
 
 static DEVICE_ATTR_RW(full_track_bias);
 
+static const char * const dasd_ese_heu_state_names[] = {
+	[DASD_ESE_HEU_FT1_ACTIVE] = "fulltrack active",
+	[DASD_ESE_HEU_PROBING]    = "probing",
+	[DASD_ESE_HEU_FT0_STABLE] = "fulltrack inactive",
+};
+
+/* read-only: current full-track mode / adaptive FSM state, for observability */
+static ssize_t
+ese_heuristic_state_show(struct device *dev, struct device_attribute *attr,
+			 char *buf)
+{
+	struct dasd_device *device;
+	unsigned int state;
+	int len;
+
+	device = dasd_device_from_cdev(to_ccwdev(dev));
+	if (IS_ERR(device))
+		return -ENODEV;
+	if (device->ft_bias == 0) {
+		len = sysfs_emit(buf, "fulltrack deactivated\n");
+	} else if (device->ft_bias >= DASD_FT_BIAS_MAX) {
+		len = sysfs_emit(buf, "fulltrack permanent active\n");
+	} else if (!dasd_ese_adaptive(device)) {
+		/* adaptive range but not ESE: the heuristic does not run */
+		len = sysfs_emit(buf, "fulltrack deactivated\n");
+	} else {
+		state = device->ese_probe_state;
+		if (state < ARRAY_SIZE(dasd_ese_heu_state_names))
+			len = sysfs_emit(buf, "%s\n", dasd_ese_heu_state_names[state]);
+		else
+			len = sysfs_emit(buf, "unknown\n");
+	}
+	dasd_put_device(device);
+	return len;
+}
+
+static DEVICE_ATTR_RO(ese_heuristic_state);
+
 static ssize_t
 dasd_retries_show(struct device *dev, struct device_attribute *attr, char *buf)
 {
@@ -2464,6 +2508,7 @@ static struct attribute * dasd_attrs[] = {
 	&dev_attr_failfast.attr,
 	&dev_attr_expires.attr,
 	&dev_attr_full_track_bias.attr,
+	&dev_attr_ese_heuristic_state.attr,
 	&dev_attr_retries.attr,
 	&dev_attr_timeout.attr,
 	&dev_attr_reservation_policy.attr,
diff --git a/drivers/s390/block/dasd_eckd.c b/drivers/s390/block/dasd_eckd.c
index c34c3afb55d0..41cb44242204 100644
--- a/drivers/s390/block/dasd_eckd.c
+++ b/drivers/s390/block/dasd_eckd.c
@@ -2160,11 +2160,6 @@ dasd_eckd_check_characteristics(struct dasd_device *device)
 	device->path_interval = DASD_ECKD_PATH_INTERVAL;
 	device->aq_timeouts = DASD_RETRIES_MAX;
 
-	/* default ESE fulltrack write aggressiveness from the module parameter */
-	device->ft_bias = min_t(unsigned int, full_track_bias, DASD_FT_BIAS_MAX);
-	/* only the "always" endpoint forces fulltrack unconditionally here */
-	device->fulltrack = (device->ft_bias >= DASD_FT_BIAS_MAX) ? 1 : 0;
-
 	if (private->conf.gneq) {
 		value = 1;
 		for (i = 0; i < private->conf.gneq->timeout.value; i++)
@@ -2219,6 +2214,13 @@ dasd_eckd_check_characteristics(struct dasd_device *device)
 	/* Read Volume Information */
 	dasd_eckd_read_vol_info(device);
 
+	/*
+	 * is_ese() now reflects the hardware ESE state, so derive the default
+	 * fulltrack write bias from the module parameter.
+	 */
+	device->ft_bias = min_t(unsigned int, full_track_bias, DASD_FT_BIAS_MAX);
+	dasd_ft_bias_apply(device);
+
 	/* Read Extent Pool Information */
 	dasd_eckd_read_ext_pool_info(device);
 
@@ -3171,6 +3173,13 @@ out:
 static int dasd_eckd_format_device(struct dasd_device *base,
 				   struct format_data_t *fdata, int enable_pav)
 {
+	/*
+	 * A full format (start_unit == 0) returns the device to a fully sparse
+	 * state, so restart the heuristic from ft1 without an offline cycle.
+	 */
+	if (fdata->start_unit == 0)
+		dasd_ft_bias_apply(base);
+
 	return dasd_eckd_format_process_data(base, fdata, enable_pav, 0, NULL,
 					     0, NULL);
 }
@@ -3230,6 +3239,69 @@ static void clear_format_track(struct dasd_format_entry *format,
 	spin_unlock_irqrestore(&block->format_lock, flags);
 }
 
+/*
+ * Adaptive ft_bias heuristic, called once per IO from dasd_eckd_build_cp().
+ * Probes the device formatting state by briefly switching to ft0 and measuring
+ * the NRF rate; parameters are derived from ft_bias.
+ */
+static void dasd_ese_heuristic_tick(struct dasd_device *basedev)
+{
+	int ios, nrf, rate;
+
+	if (atomic_inc_return(&basedev->ese_io_cnt) < (int)basedev->ese_probe_interval)
+		return;
+
+	/*
+	 * One wins the race to evaluate, the rest see ios == 0 after the
+	 * xchg and return early, preventing redundant state transitions.
+	 */
+	ios = atomic_xchg(&basedev->ese_io_cnt, 0);
+	if (ios <= 0)
+		return;
+
+	switch (basedev->ese_probe_state) {
+	case DASD_ESE_HEU_FT1_ACTIVE:
+		/* Start ft0 probe window, reset NRF counter for clean measurement */
+		basedev->fulltrack          = 0;
+		basedev->ese_probe_state    = DASD_ESE_HEU_PROBING;
+		basedev->ese_probe_interval = basedev->ese_heu_probe_window;
+		atomic_set(&basedev->ese_nrf_window, 0);
+		break;
+
+	case DASD_ESE_HEU_PROBING:
+	case DASD_ESE_HEU_FT0_STABLE:
+		nrf  = atomic_xchg(&basedev->ese_nrf_window, 0);
+		rate = (int)((u64)nrf * 1000 / ios);
+		if (rate > (int)basedev->ese_heu_nrf_high) {
+			/* NRF rate high: device still sparse, ft1 is better */
+			basedev->fulltrack          = 1;
+			basedev->ese_probe_state    = DASD_ESE_HEU_FT1_ACTIVE;
+			basedev->ese_probe_interval = basedev->ese_heu_start_interval;
+		} else if (basedev->ese_probe_state == DASD_ESE_HEU_PROBING) {
+			/*
+			 * NRF rate low: device mostly formatted, ft0 is faster.
+			 * Re-probe frequently at first, then back off below.
+			 */
+			basedev->fulltrack          = 0;
+			basedev->ese_probe_state    = DASD_ESE_HEU_FT0_STABLE;
+			basedev->ese_probe_interval = basedev->ese_heu_probe_window;
+		} else {
+			/*
+			 * Still stable in ft0: re-assert plain-write mode so a
+			 * fulltrack value left behind by a racing sysfs write
+			 * self-corrects, and back off the re-probe interval
+			 * (double it, capped at max_interval) so a long-lived
+			 * formatted device is not probed more often than needed.
+			 */
+			basedev->fulltrack          = 0;
+			basedev->ese_probe_interval =
+				min(basedev->ese_probe_interval * 2,
+				    basedev->ese_heu_max_interval);
+		}
+		break;
+	}
+}
+
 static void dasd_eckd_ese_format(struct dasd_device *startdev, struct dasd_ccw_req *cqr,
 				 struct irb *irb)
 {
@@ -3254,6 +3326,8 @@ static void dasd_eckd_ese_format(struct dasd_device *startdev, struct dasd_ccw_r
 	block = cqr->block;
 	base = block->base;
 	private = base->private;
+	if (dasd_ese_adaptive(base))
+		atomic_inc(&base->ese_nrf_window);
 	blksize = block->bp_block;
 	recs_per_trk = recs_per_track(&private->rdc_data, 0, blksize);
 
@@ -4017,6 +4091,14 @@ static int dasd_eckd_release_space_full(struct dasd_device *device)
 
 	rc = dasd_sleep_on_interruptible(cqr);
 
+	if (!rc) {
+		/*
+		 * Releasing all space (RAS) wipes every track and the device is fully
+		 * sparse again, so restart the heuristic from ft1.
+		 */
+		dasd_ft_bias_apply(device);
+	}
+
 	dasd_sfree_request(cqr, cqr->memdev);
 
 	return rc;
@@ -5185,6 +5267,11 @@ static struct dasd_ccw_req *dasd_eckd_build_cp(struct dasd_device *startdev,
 	struct dasd_ccw_req *cqr;
 
 	basedev = block->base;
+	if (dasd_ese_adaptive(basedev))
+		dasd_ese_heuristic_tick(basedev);
+	else
+		/* re-assert the endpoint mode: a stale heuristic write cannot stick */
+		basedev->fulltrack = (basedev->ft_bias >= DASD_FT_BIAS_MAX) ? 1 : 0;
 	private = basedev->private;
 
 	/* Calculate number of blocks/records per track. */
diff --git a/drivers/s390/block/dasd_int.h b/drivers/s390/block/dasd_int.h
index b8a3190c9c93..59cce4e7dbc1 100644
--- a/drivers/s390/block/dasd_int.h
+++ b/drivers/s390/block/dasd_int.h
@@ -633,6 +633,15 @@ struct dasd_device {
 	/* ESE fulltrack write control (see full_track_bias sysfs attribute) */
 	unsigned int ft_bias;	/* aggressiveness 0..100: 0=off, 100=always */
 	unsigned int fulltrack;	/* internal: use WRITE_FULL_TRACK for aligned writes */
+	/* adaptive heuristic (active for ft_bias 1..99), derived from ft_bias */
+	unsigned int ese_probe_state;        /* heuristic FSM state */
+	unsigned int ese_probe_interval;     /* IOs between evaluations */
+	atomic_t ese_io_cnt;                 /* IO counter for current window */
+	atomic_t ese_nrf_window;             /* NRF/INV_TRACK_FORMAT events in window */
+	unsigned int ese_heu_start_interval; /* IOs before first probe */
+	unsigned int ese_heu_probe_window;   /* IOs in probe window */
+	unsigned int ese_heu_max_interval;   /* max IOs between probes (backoff cap) */
+	unsigned int ese_heu_nrf_high;       /* NRF per-mille threshold → activate ft1 */
 };
 
 struct dasd_block {
@@ -704,6 +713,25 @@ struct dasd_queue {
 #define DASD_FT_BIAS_MAX	100
 #define DASD_FT_BIAS_DEFAULT	50
 
+/* ESE fulltrack heuristic FSM states (adaptive range, ft_bias 1..99) */
+#define DASD_ESE_HEU_FT1_ACTIVE   0   /* fulltrack write active */
+#define DASD_ESE_HEU_PROBING      1   /* ft0 probe window, measuring NRF rate */
+#define DASD_ESE_HEU_FT0_STABLE   2   /* device formatted, ft0 active */
+
+/*
+ * Heuristic parameters are derived from ft_bias by linear interpolation,
+ * anchored so that ft_bias == 50 reproduces the previously shipped defaults
+ * and ft_bias == 100 is the most aggressive end of the range.
+ * probe_window is constant.
+ */
+#define DASD_ESE_HEU_PROBE_WINDOW	   100
+#define DASD_ESE_HEU_NRF_HIGH_A50	    10	/* NRF per-mille threshold */
+#define DASD_ESE_HEU_NRF_HIGH_A100	     1
+#define DASD_ESE_HEU_START_A50		  2000	/* IOs before first probe */
+#define DASD_ESE_HEU_START_A100		   500
+#define DASD_ESE_HEU_MAX_A50		500000	/* backoff cap */
+#define DASD_ESE_HEU_MAX_A100		 20000
+
 /* per device flags */
 #define DASD_FLAG_OFFLINE	3	/* device is in offline processing */
 #define DASD_FLAG_EER_SNSS	4	/* A SNSS is required */
@@ -866,6 +894,63 @@ static inline bool dasd_req_conflict(struct dasd_ccw_req *cqr1,
 		 cqr2->end_trk < cqr1->format->start_trk);
 }
 
+/*
+ * true when device is ese device and ft_bias selects the adaptive
+ * heuristic (neither hard endpoint)
+ */
+static inline bool dasd_ese_adaptive(struct dasd_device *device)
+{
+	return device->discipline &&
+		device->discipline->is_ese &&
+		device->discipline->is_ese(device) &&
+		device->ft_bias > 0 &&
+		device->ft_bias < DASD_FT_BIAS_MAX;
+}
+
+/*
+ * Linear interpolation of a heuristic parameter between its value at aggr==50
+ * (v50) and its value at aggr==100 (v100).
+ */
+static inline unsigned int dasd_ese_lerp(unsigned int v50, unsigned int v100,
+					 unsigned int aggr)
+{
+	return (unsigned int)((int)v50 +
+		((int)v100 - (int)v50) * ((int)aggr - 50) / 50);
+}
+
+/*
+ * Apply the ft_bias knob. For the hard endpoints just pin the mode; for the
+ * adaptive range derive the heuristic parameters from ft_bias and (re)start
+ * the FSM in ft1 so a freshly sparse device avoids the NRF penalty right away.
+ */
+static inline void dasd_ft_bias_apply(struct dasd_device *device)
+{
+	unsigned int a = device->ft_bias;
+
+	if (!dasd_ese_adaptive(device)) {
+		device->fulltrack = (a >= DASD_FT_BIAS_MAX) ? 1 : 0;
+		device->ese_probe_state = DASD_ESE_HEU_FT1_ACTIVE;
+		return;
+	}
+
+	device->ese_heu_nrf_high =
+		dasd_ese_lerp(DASD_ESE_HEU_NRF_HIGH_A50,
+			      DASD_ESE_HEU_NRF_HIGH_A100, a);
+	device->ese_heu_start_interval =
+		dasd_ese_lerp(DASD_ESE_HEU_START_A50,
+			      DASD_ESE_HEU_START_A100, a);
+	device->ese_heu_max_interval =
+		dasd_ese_lerp(DASD_ESE_HEU_MAX_A50,
+			      DASD_ESE_HEU_MAX_A100, a);
+	device->ese_heu_probe_window = DASD_ESE_HEU_PROBE_WINDOW;
+
+	device->ese_probe_state    = DASD_ESE_HEU_FT1_ACTIVE;
+	device->ese_probe_interval = device->ese_heu_start_interval;
+	device->fulltrack          = 1;
+	atomic_set(&device->ese_io_cnt, 0);
+	atomic_set(&device->ese_nrf_window, 0);
+}
+
 /* externals in dasd.c */
 #define DASD_PROFILE_OFF	 0
 #define DASD_PROFILE_ON 	 1