mtd: spinand: Enhance the logic when picking a variant

Currently the best variant picked in the first one in the list provided
in the manufacturer driver. This worked well while all operations where
performed at the same speed, but with the introduction of DTR transfers
this no longer works correctly.

Let's continue iterating over all the alternatives, even if we find a
match, keeping a reference over the theoretically fastest
operation. Only at the end we can tell which variant is the best.

This logic happening only once at boot.

The patch is based on linux commit
666c299be696 (mtd: spinand: Enhance the logic when picking a variant)
created by Miquel Raynal <miquel.raynal@bootlin.com>

The code was a bit restricted in the functionality since not all
required functionality is supported in the u-boot.

Signed-off-by: Mikhail Kshevetskiy <mikhail.kshevetskiy@iopsys.eu>
Reviewed-by: Frieder Schrempf <frieder.schrempf@kontron.de>
Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com>

drivers/mtd/nand/spi/core.c

@@ -1266,10 +1266,13 @@ spinand_select_op_variant(struct spinand_device *spinand,
 			  const struct spinand_op_variants *variants)
 {
 	struct nand_device *nand = spinand_to_nand(spinand);
+	const struct spi_mem_op *best_variant = NULL;
+	u64 best_op_duration_ns = ULLONG_MAX;
 	unsigned int i;
 
 	for (i = 0; i < variants->nops; i++) {
 		struct spi_mem_op op = variants->ops[i];
+		u64 op_duration_ns = 0;
 		unsigned int nbytes;
 		int ret;
 
@@ -1286,13 +1289,17 @@ spinand_select_op_variant(struct spinand_device *spinand,
 				break;
 
 			nbytes -= op.data.nbytes;
+
+			op_duration_ns += spi_mem_calc_op_duration(&op);
 		}
 
-		if (!nbytes)
-			return &variants->ops[i];
+		if (!nbytes && op_duration_ns < best_op_duration_ns) {
+			best_op_duration_ns = op_duration_ns;
+			best_variant = &variants->ops[i];
+		}
 	}
 
-	return NULL;
+	return best_variant;
 }
 
 static int spinand_setup_slave(struct spinand_device *spinand,

drivers/spi/spi-mem.c

@@ -499,6 +499,38 @@ int spi_mem_adjust_op_size(struct spi_slave *slave, struct spi_mem_op *op)
 }
 EXPORT_SYMBOL_GPL(spi_mem_adjust_op_size);
 
+/**
+ * spi_mem_calc_op_duration() - Derives the theoretical length (in cpu cycles)
+ *				of an operation. This helps finding the best
+ *				variant among a list of possible choices.
+ * @op: the operation to benchmark
+ *
+ * Some chips have per-op frequency limitations, PCBs usually have their own
+ * limitations as well, and controllers can support dual, quad or even octal
+ * modes, sometimes in DTR. All these combinations make it impossible to
+ * statically list the best combination for all situations. If we want something
+ * accurate, all these combinations should be rated (eg. with a time estimate)
+ * and the best pick should be taken based on these calculations.
+ *
+ * Returns a estimate for the time this op would take.
+ */
+u64 spi_mem_calc_op_duration(struct spi_mem_op *op)
+{
+	u64 ncycles = 0;
+
+	ncycles += ((op->cmd.nbytes * 8) / op->cmd.buswidth) / (op->cmd.dtr ? 2 : 1);
+	ncycles += ((op->addr.nbytes * 8) / op->addr.buswidth) / (op->addr.dtr ? 2 : 1);
+
+	/* Dummy bytes are optional for some SPI flash memory operations */
+	if (op->dummy.nbytes)
+		ncycles += ((op->dummy.nbytes * 8) / op->dummy.buswidth) / (op->dummy.dtr ? 2 : 1);
+
+	ncycles += ((op->data.nbytes * 8) / op->data.buswidth) / (op->data.dtr ? 2 : 1);
+
+	return ncycles;
+}
+EXPORT_SYMBOL_GPL(spi_mem_calc_op_duration);
+
 static ssize_t spi_mem_no_dirmap_read(struct spi_mem_dirmap_desc *desc,
 				      u64 offs, size_t len, void *buf)
 {

include/spi-mem.h

@@ -392,6 +392,7 @@ bool spi_mem_default_supports_op(struct spi_mem *mem,
 #endif /* __UBOOT__ */
 
 int spi_mem_adjust_op_size(struct spi_slave *slave, struct spi_mem_op *op);
+u64 spi_mem_calc_op_duration(struct spi_mem_op *op);
 
 bool spi_mem_supports_op(struct spi_slave *slave, const struct spi_mem_op *op);
 bool spi_mem_dtr_supports_op(struct spi_slave *slave,