i2c-designware: lcnt/2 SDA hold default is a no-op on rpi-7.x (fs_lcnt still 0), regressing SSD1306 OLED on Pi 5

Start with drivers/i2c/busses/i2c-designware-common.c and trace the order of i2c_dw_probe_master() and i2c_dw_set_sda_hold() in i2c_dw_probe(). Check when fs_lcnt is computed and compare the configured SDA hold value with the reported IC_SDA_HOLD register. Done means the default uses the computed timing and the Pi 5 OLED works; the issue reports a linked pull request with the proposed change.

Disclosure: this report was written by an AI assistant (GitHub Copilot CLI,
Claude Sonnet 5.5) working with a human engineer who owns the hardware and
reviewed the measurements. The kernel experiments were run on the engineer's
Pi 5.

Summary

The Pi-specific default "SDA hold = SCL low count / 2" (#6091, b74c3e122726 on
rpi-6.18.y) no longer takes effect on rpi-7.0.y .. rpi-7.3.y. In 7.x the line sits in
i2c_dw_probe() before i2c_dw_probe_master() has computed fs_lcnt, so it
evaluates to 0. The driver then keeps the hardware reset value (1 cycle, ~5 ns)
and the Pi 5 returns to the behaviour #6091 and #6056 fixed.

Impact

Adafruit PiOLED (128x32 SSD1306, 0x3c) on a Pi 5, no DT i2c-sda-hold-time-ns:

• stock 6.18.60+rpt-rpi-2712: works
• custom 7.3.1 (rpi-7.3.y): SDA stuck at low, OSError: [Errno 121] Remote I/O error
• same Pi, same userspace, same DT, same cmdline

Same symptom as #6056 (AQM0802 LCD: lost arbitration / SDA stuck at low).
Another Pi 5 on 7.3.0-rc3 has the same IC_SDA_HOLD and its OLED happens to work,
so it is marginal and device dependent, not universal.

Root cause

rpi-6.18.y / 6.19.y: the default is set inside i2c_dw_set_timings_master()
right after dev->fs_lcnt = lcnt;.

rpi-7.0.y .. 7.3.y (first bad: bb68f1f3423b on rpi-7.0.y; later rebases
d2587dfeac41, 6b7c1f3d0028, 60199372c10d): the line is in i2c_dw_probe():

if (!dev->sda_hold_time)
    dev->sda_hold_time = dev->fs_lcnt / 2;   /* fs_lcnt == 0 here */
ret = i2c_dw_set_sda_hold(dev);
...
ret = i2c_dw_probe_master(dev);              /* computes fs_lcnt */

This follows mainline 38fa29b01a6a ("Combine the init functions"), which moved
i2c_dw_set_sda_hold() ahead of the timing calculation. Mainline has no such
default, so nothing upstream notices.

Evidence (printk on 7.3.1 + instrumented module)

probe: fs_lcnt=0, sda_hold_time=0 before default
default: sda_hold_time = fs_lcnt/2 = 0
set_sda_hold: sda_hold_time was 0 -> kept HW reset value 0x1
after set_timings_master: fs_lcnt=1186

Register (/dev/mem read of RP1 i2c1 at 0x1f00074000, clk_sys 200 MHz, 100 kHz):

kernel	IC_SDA_HOLD	OLED
6.18.60+rpt-rpi-2712 stock	0x10251	works
7.3.1 unpatched	0x10001	fails
7.3.1 with patch below	0x10251	works

0x251 = 593 = 1186 / 2 cycles, about 3 us.

What was checked

• Ruled out userspace: same Blinka 9.2.0 / ssd1306 libs and daemon on a second, working Pi 5 (7.3.0-rc3), same 100 kHz bus, same GPIO2/3 pinmux, identical DT I2C properties.
• Ruled out hardware: stock 6.18.60+rpt-rpi-2712 on the same Pi and OLED works; the bad kernels fail on the same board.
• Booted the 7.3.0-rc3 kernel from the other Pi on the failing Pi: also fails, IC_SDA_HOLD=0x10001. The working Pi has the same value, so its OLED just tolerates it.
• Compared kernel config and i2c_designware sources between the two Pis: identical (only an unrelated led-modes DT line differs).
• Checked the lcnt / 2 line on current heads: rpi-6.18.y (a4985478de14) and rpi-6.19.y (597a84a013b2) have it in i2c_dw_set_timings_master(); rpi-7.0.y (42d9bb9081f1), 7.1.y (49246ab8cb8a), 7.2.y (dbb0de2b7b04) and 7.3.y (5f369cb6094a) have it in i2c_dw_probe().
• Searched raspberrypi/linux issues and PRs for this symptom on 7.x: found #6091 (the original patch), #6056, #6043, #5784 and #7104 (6.x, different cause) but no report of this regression.
• Not done: a real build bisect (the rpi branches are rebased), tests on other boards, slave mode.

Proposed fix

Compute the default after the timings are known, then run i2c_dw_set_sda_hold()
(which also adds the RX hold bit). Keeps all existing logic; only reorders:

--- a/drivers/i2c/busses/i2c-designware-common.c
+++ b/drivers/i2c/busses/i2c-designware-common.c
@@ -949,18 +949,18 @@ int i2c_dw_probe(struct dw_i2c_dev *dev)
 	if (ret)
 		return ret;
 
-	if (!dev->sda_hold_time)
-		dev->sda_hold_time = dev->fs_lcnt / 2;
-
-	ret = i2c_dw_set_sda_hold(dev);
+	ret = i2c_dw_set_fifo_size(dev);
 	if (ret)
 		return ret;
 
-	ret = i2c_dw_set_fifo_size(dev);
+	ret = i2c_dw_probe_master(dev);
 	if (ret)
 		return ret;
 
-	ret = i2c_dw_probe_master(dev);
+	if (!dev->sda_hold_time)
+		dev->sda_hold_time = dev->fs_lcnt / 2;
+
+	ret = i2c_dw_set_sda_hold(dev);
 	if (ret)
 		return ret;

Tested on Pi 5 with 7.3.1: IC_SDA_HOLD = 0x10251, OLED connects and refreshes,
no restarts. Not tested on other boards or with slave mode.

Caveats: "first bad" comes from reading the branch history (rebased, non-linear),
not a build bisect.

PR with the same change: see the linked pull request from jflecool2:i2c-designware-sda-hold-fix.