[PATCH] clk: sunxi-ng: nkm: avoid 32-bit overflow in rate calculations

[Romain TISSERAND]

The N-K-M rate code computes parent * n * k / m in an unsigned long,
which wraps on 32-bit ARM as soon as parent * n * k passes 4.29 GHz.

On the A33, pll-mipi (parent pll-video at 297 MHz) with N = 14, K = 2,
M = 14 runs at 594 MHz, but 297 MHz * 14 * 2 wraps around to
4021032704 and the code computes 287.2 MHz. ccu_nkm_find_best() then
picks these factors for a 288 MHz request, and clk_get_rate() reports
287.2 MHz while the PLL runs at twice that. On a PocketGo S30 (A33,
ILI9488 panel on one MIPI-DSI lane) the lane ended up at about
594 Mbit/s instead of 288 and the picture broke up.

Do the multiplications in 64 bits, saturating at ULONG_MAX, in
ccu_nkm_recalc_rate(), in both best-rate searches, and for the parent
rate request of the parent-adjusting search, where rate * m overflows
the same way.

Without the saturation, a candidate above 4.29 GHz in the search would
be truncated again on return and could look like a valid low rate.

64-bit kernels are not affected. On 32-bit ones, every N-K-M clock can
hit it (A31, A23, A33, R40, H3, V3s and F1C100s use them), whenever the
parent rate times N * K goes above 2^32.

Fixes: df6561e60244 ("clk: sunxi-ng: Add N-K-M Factor clock")
Fixes: 3492e4f6db5f ("clk: sunxi-ng: nkm: consider alternative parent rates when determining rate")
Signed-off-by: Romain TISSERAND <romain.tisserand@xxxxxxxxx>
---
Notes:
- Testing: built with W=1 for arm (multi_v7_defconfig with the A31, A23,
  A33, R40 and H3 CCUs) and arm64 (defconfig), SUNXI_CCU=y and =m, no
  warnings; on arm the 64-bit divisions go through __do_div64 (no libgcc
  helper). Tested on a PocketGo S30 (A33) running 6.18.54 with this
  change, without the ULONG_MAX saturation (same results for rates that
  fit, which is the only case the hardware hits): before, pll-mipi was
  programmed with N = 14, K = 2, M = 14 (594 MHz measured as a 2x lane
  rate) while clk_get_rate() returned 287216621; after, N = 15, K = 1,
  M = 16, 278437500 Hz in both the register and clk_get_rate(), and the
  panel picture is right. This version, with the saturation, then on
  the same board on 7.3.1: pll-mipi again programmed with N = 15, K = 1,
  M = 16 (278.4 MHz from the 297 MHz pll-video0) and the picture is
  right. Not tested on the other SoCs.
- Tool disclosure: this patch was prepared with Claude Opus 5.5
  (Anthropic) as a coding assistant, in an interactive session where I
  asked it to bring up the PocketGo S30 display. It found the mismatch
  between the pll-mipi register and the rate the clock framework
  reported, traced it to this overflow and drafted the fix and the
  message. I reviewed and understand the change and tested it as
  described above.

 drivers/clk/sunxi-ng/ccu_nkm.c | 30 ++++++++++++++++++++++++++----
 1 file changed, 26 insertions(+), 4 deletions(-)

diff --git a/drivers/clk/sunxi-ng/ccu_nkm.c b/drivers/clk/sunxi-ng/ccu_nkm.c
index 401fbb7524..e5a8f22d50 100644
--- a/drivers/clk/sunxi-ng/ccu_nkm.c
+++ b/drivers/clk/sunxi-ng/ccu_nkm.c
@@ -6,6 +6,7 @@
 
 #include <linux/clk-provider.h>
 #include <linux/io.h>
+#include <linux/math64.h>
 
 #include "ccu_gate.h"
 #include "ccu_nkm.h"
@@ -30,6 +31,26 @@ static bool ccu_nkm_is_valid_rate(struct ccu_common *common, unsigned long paren
 	return true;
 }
 
+static unsigned long ccu_nkm_calc_rate(unsigned long parent,
+				       unsigned long n, unsigned long k,
+				       unsigned long m)
+{
+	u64 rate = (u64)parent * n * k;
+
+	do_div(rate, m);
+	return min_t(u64, rate, ULONG_MAX);
+}
+
+static unsigned long ccu_nkm_calc_parent_rate(unsigned long rate,
+					      unsigned long n, unsigned long k,
+					      unsigned long m)
+{
+	u64 parent = (u64)rate * m;
+
+	do_div(parent, n * k);
+	return min_t(u64, parent, ULONG_MAX);
+}
+
 static unsigned long ccu_nkm_find_best_with_parent_adj(struct ccu_common *common,
 						       struct clk_hw *parent_hw,
 						       unsigned long *parent, unsigned long rate,
@@ -44,12 +65,13 @@ static unsigned long ccu_nkm_find_best_with_parent_adj(struct ccu_common *common
 			for (_m = nkm->min_m; _m <= nkm->max_m; _m++) {
 				unsigned long tmp_rate, tmp_parent;
 
-				tmp_parent = clk_hw_round_rate(parent_hw, rate * _m / (_n * _k));
+				tmp_parent = ccu_nkm_calc_parent_rate(rate, _n, _k, _m);
+				tmp_parent = clk_hw_round_rate(parent_hw, tmp_parent);
 
 				if (!ccu_nkm_is_valid_rate(common, tmp_parent, _n, _m))
 					continue;
 
-				tmp_rate = tmp_parent * _n * _k / _m;
+				tmp_rate = ccu_nkm_calc_rate(tmp_parent, _n, _k, _m);
 
 				if (ccu_is_better_rate(common, rate, tmp_rate, best_rate) ||
 				    (tmp_parent == *parent && tmp_rate == best_rate)) {
@@ -87,7 +109,7 @@ static unsigned long ccu_nkm_find_best(unsigned long parent, unsigned long rate,
 
 				unsigned long tmp_rate;
 
-				tmp_rate = parent * _n * _k / _m;
+				tmp_rate = ccu_nkm_calc_rate(parent, _n, _k, _m);
 
 				if (ccu_is_better_rate(common, rate, tmp_rate, best_rate)) {
 					best_rate = tmp_rate;
@@ -154,7 +176,7 @@ static unsigned long ccu_nkm_recalc_rate(struct clk_hw *hw,
 	if (!m)
 		m++;
 
-	rate = parent_rate * n  * k / m;
+	rate = ccu_nkm_calc_rate(parent_rate, n, k, m);
 
 	if (nkm->common.features & CCU_FEATURE_FIXED_POSTDIV)
 		rate /= nkm->fixed_post_div;

base-commit: 5243bd55dfe63a5f806622878f6fbdd78172fac9
-- 
2.43.0