[PATCH 0/2] ethdev: fix SFF-8472 external Rx power calibration

[PATCH 0/2] ethdev: fix SFF-8472 external Rx power calibration

[Roman Khromenok]

The SFF-8472 decoder does not apply the external Rx power calibration
as defined by the specification: the fourth order polynomial is reduced
to RX_PWR(0) + x * (RX_PWR(1) + RX_PWR(2) + RX_PWR(3)) and RX_PWR(4)
is never used. The code came from ethtool sfpdiag.c.

This was pointed out by Stephen during the review of the module EEPROM
decoding series; this series is based on it (dpdk-next-net).

Patch 1 fixes the formula and is intended for stable.
It also clamps the result to the 16-bit range, as the polynomial
easily overflows it with unexpected coefficients and converting
an out of range double to an integer is undefined behavior.
Patch 2 adds a unit test with all coefficients set.

Note: the calibration formula 1 (Tx bias, Tx power, temperature,
voltage) has the same out of range conversion issue; it is left
for a separate patch.

Depends-on: series-39436 ("ethdev: add API to decode module EEPROM")

Roman Khromenok (2):
  ethdev: fix SFF-8472 external Rx power calibration
  test: check SFF-8472 external Rx power calibration

 app/test/test_ethdev_module_eeprom.c | 37 ++++++++++++++++++++++++++++
 lib/ethdev/sff_8472.c                | 27 +++++++++++++-------
 2 files changed, 55 insertions(+), 9 deletions(-)

-- 
2.47.3

[PATCH 1/2] ethdev: fix SFF-8472 external Rx power calibration

[Roman Khromenok]

With external calibration, SFF-8472 defines the received power
as a fourth order polynomial of the raw reading x:

  RX_PWR(4) * x^4 + RX_PWR(3) * x^3 + RX_PWR(2) * x^2 +
  RX_PWR(1) * x + RX_PWR(0)

The decoder computed RX_PWR(0) + x * (RX_PWR(1) + RX_PWR(2) + RX_PWR(3))
and ignored RX_PWR(4), so the reported power was wrong for modules
using the higher order coefficients.

Evaluate the polynomial in double precision and clamp the result
to the 16-bit range of the stored value, as converting an out of range
floating point value to an integer is undefined behavior.

Fixes: 0caf7f376b08 ("ethdev: support SFF-8472 module telemetry")
Cc: stable@dpdk.org

Signed-off-by: Roman Khromenok <roma55592@yandex.ru>
---
 lib/ethdev/sff_8472.c | 27 ++++++++++++++++++---------
 1 file changed, 18 insertions(+), 9 deletions(-)

diff --git a/lib/ethdev/sff_8472.c b/lib/ethdev/sff_8472.c
index cdb6ef1f7e..3c73a829a8 100644
--- a/lib/ethdev/sff_8472.c
+++ b/lib/ethdev/sff_8472.c
@@ -189,7 +189,8 @@ static float befloattoh(const uint8_t *source)
 static void sff_8472_calibration(const uint8_t *data, struct sff_diags *sd)
 {
 	unsigned long i;
-	uint16_t rx_reading;
+	double rx_reading;
+	double rx_power;
 
 	/* Calibration should occur for all values (threshold and current) */
 	for (i = 0; i < RTE_DIM(sd->bias_cur); ++i) {
@@ -207,16 +208,24 @@ static void sff_8472_calibration(const uint8_t *data, struct sff_diags *sd)
 		sd->sfp_temp[i]    += A2_OFFSET_TO_OFF(SFF_A2_CAL_T_OFF);
 
 		/*
-		 * Apply calibration formula 2 (Rx Power only)
+		 * Apply calibration formula 2 (Rx Power only):
+		 * RX_PWR(4) * x^4 + RX_PWR(3) * x^3 + RX_PWR(2) * x^2 +
+		 * RX_PWR(1) * x + RX_PWR(0)
 		 */
 		rx_reading = sd->rx_power[i];
-		sd->rx_power[i]    = A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR0);
-		sd->rx_power[i]    += rx_reading *
-			A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR1);
-		sd->rx_power[i]    += rx_reading *
-			A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR2);
-		sd->rx_power[i]    += rx_reading *
-			A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR3);
+		rx_power = A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR4);
+		rx_power = rx_power * rx_reading + A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR3);
+		rx_power = rx_power * rx_reading + A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR2);
+		rx_power = rx_power * rx_reading + A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR1);
+		rx_power = rx_power * rx_reading + A2_OFFSET_TO_RXPWRx(SFF_A2_CAL_RXPWR0);
+
+		/* the result is stored in 0.1 uW units, out of range is not representable */
+		if (!(rx_power > 0))
+			sd->rx_power[i] = 0;
+		else if (rx_power >= UINT16_MAX)
+			sd->rx_power[i] = UINT16_MAX;
+		else
+			sd->rx_power[i] = rx_power;
 	}
 }
 
-- 
2.47.3

[PATCH 2/2] test: check SFF-8472 external Rx power calibration

[Roman Khromenok]

Add a test case with all the SFF-8472 Rx power calibration
coefficients set, so that the whole fourth order polynomial
is checked, including RX_PWR(4).

Signed-off-by: Roman Khromenok <roma55592@yandex.ru>
---
 app/test/test_ethdev_module_eeprom.c | 37 ++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/app/test/test_ethdev_module_eeprom.c b/app/test/test_ethdev_module_eeprom.c
index d76bdc0060..b07a18aa06 100644
--- a/app/test/test_ethdev_module_eeprom.c
+++ b/app/test/test_ethdev_module_eeprom.c
@@ -269,6 +269,42 @@ test_module_eeprom_sfp_8472_ext_cal_unaligned(void)
 	return TEST_SUCCESS;
 }
 
+static void
+put_be_float(uint8_t *data, unsigned int offset, uint32_t bits)
+{
+	data[offset] = bits >> 24;
+	data[offset + 1] = (bits >> 16) & 0xff;
+	data[offset + 2] = (bits >> 8) & 0xff;
+	data[offset + 3] = bits & 0xff;
+}
+
+static int
+test_module_eeprom_sfp_8472_rx_power_polynomial(void)
+{
+	uint8_t data[RTE_ETH_MODULE_SFF_8472_LEN];
+	uint8_t *a2 = data + RTE_ETH_MODULE_SFF_8079_LEN;
+
+	fill_sfp(data);
+	/* diagnostics implemented, externally calibrated, average RX power */
+	data[92] = 0x58;
+	a2[76] = 1;	/* TX bias slope */
+	a2[80] = 1;	/* TX power slope */
+	a2[84] = 1;	/* temperature slope */
+	a2[88] = 1;	/* voltage slope */
+	put_u16(a2, 104, 10);	/* raw RX power x = 10 */
+	/* 0.125 x^4 + 0.25 x^3 + 0.5 x^2 + x = 1250 + 250 + 50 + 10 = 1560 */
+	put_be_float(a2, 56, 0x3e000000);	/* RX_PWR(4) = 0.125 */
+	put_be_float(a2, 60, 0x3e800000);	/* RX_PWR(3) = 0.25 */
+	put_be_float(a2, 64, 0x3f000000);	/* RX_PWR(2) = 0.5 */
+	put_be_float(a2, 68, 0x3f800000);	/* RX_PWR(1) = 1.0 */
+	put_be_float(a2, 72, 0);		/* RX_PWR(0) = 0 */
+
+	TEST_ASSERT_SUCCESS(parse(RTE_ETH_MODULE_SFF_8472, data, sizeof(data)),
+		"Failed to parse externally calibrated SFF-8472 data");
+	CHECK_FIELD("Receiver signal average optical power", "0.1560 mW / -8.07 dBm");
+	return TEST_SUCCESS;
+}
+
 static int
 test_module_eeprom_invalid(void)
 {
@@ -318,6 +354,7 @@ static struct unit_test_suite module_eeprom_testsuite = {
 		TEST_CASE(test_module_eeprom_sfp_8472),
 		TEST_CASE(test_module_eeprom_sfp_8472_short),
 		TEST_CASE(test_module_eeprom_sfp_8472_ext_cal_unaligned),
+		TEST_CASE(test_module_eeprom_sfp_8472_rx_power_polynomial),
 		TEST_CASE(test_module_eeprom_qsfp_8636),
 		TEST_CASE(test_module_eeprom_qsfp_8636_thresholds),
 		TEST_CASE(test_module_eeprom_invalid),
-- 
2.47.3

Re: [PATCH 0/2] ethdev: fix SFF-8472 external Rx power calibration

[Stephen Hemminger]

On Tue, 29 Sep 2026 19:22:08 +0200
Roman Khromenok <roma55592@yandex.ru> wrote:

> The SFF-8472 decoder does not apply the external Rx power calibration
> as defined by the specification: the fourth order polynomial is reduced
> to RX_PWR(0) + x * (RX_PWR(1) + RX_PWR(2) + RX_PWR(3)) and RX_PWR(4)
> is never used. The code came from ethtool sfpdiag.c.
> 
> This was pointed out by Stephen during the review of the module EEPROM
> decoding series; this series is based on it (dpdk-next-net).
> 
> Patch 1 fixes the formula and is intended for stable.
> It also clamps the result to the 16-bit range, as the polynomial
> easily overflows it with unexpected coefficients and converting
> an out of range double to an integer is undefined behavior.
> Patch 2 adds a unit test with all coefficients set.
> 
> Note: the calibration formula 1 (Tx bias, Tx power, temperature,
> voltage) has the same out of range conversion issue; it is left
> for a separate patch.
> 
> Depends-on: series-39436 ("ethdev: add API to decode module EEPROM")
> 
> Roman Khromenok (2):
>   ethdev: fix SFF-8472 external Rx power calibration
>   test: check SFF-8472 external Rx power calibration
> 
>  app/test/test_ethdev_module_eeprom.c | 37 ++++++++++++++++++++++++++++
>  lib/ethdev/sff_8472.c                | 27 +++++++++++++-------
>  2 files changed, 55 insertions(+), 9 deletions(-)
> 

Looks ok as is, applied to next-net.

Detailed AI review had a bunch of feedback about range checking.

Review: [PATCH 0/2] ethdev: fix SFF-8472 external Rx power calibration

Patch 1 applies to main on its own and builds with -Dwerror=true, so
it backports as is. Patch 2 needs test_ethdev_module_eeprom.c from
the module EEPROM decoding API series.

The test input decodes to "0.1560 mW / -8.07 dBm" with the fix and
"0.0017 mW / -27.70 dBm" without it, so the test catches the bug.


Patch 1/2 ethdev: fix SFF-8472 external Rx power calibration

Info
- The final conversion truncates:

      sd->rx_power[i] = rx_power;

  Coefficients that are not exact in binary land just below the
  integer. RX_PWR(1) = 0.7 is stored as 0.69999999, so x = 1000
  decodes to 699 (0.0699 mW) instead of 700. rx_power is known
  positive in that branch, so round instead:

      sd->rx_power[i] = rx_power + 0.5;

  The UINT16_MAX check still holds, anything below 65535 rounds to
  at most 65535.

- Pre-existing, not introduced by this patch: formula 1 just above
  has the same out of range float to integer conversion the commit
  message describes.

      sd->bias_cur[i]    *= A2_OFFSET_TO_SLP(SFF_A2_CAL_TXI_SLP);

  The slope is a double up to 255.996, so a large reading leaves
  the uint16_t range. Same for tx_power and sfp_voltage, and for
  sfp_temp (int16_t) in both directions. Worth a follow up patch
  with the same clamping.


Patch 2/2 test: check SFF-8472 external Rx power calibration

Info
- The test covers the polynomial but neither clamp branch. Two more
  cases would cover them:

    x = 10, RX_PWR(1) = -1.0     -> "0.0000 mW / -inf dBm"
    x = 65535, RX_PWR(4) = 1.0   -> "6.5535 mW / 8.16 dBm"