spi: spi-mem: Estimate the time taken by operations

In the SPI-NAND layer, we currently make list of operation variants from
the fastest one to the slowest and there is a bit of logic in the core
to go over them and pick the first one that is supported by the
controller, ie. the fastest one among the supported ops.

This kind of logic only works if all operations run at the same
frequency, but as soon as we introduce per operation max frequencies it
is not longer as obvious which operation will be faster, especially
since it also depends on the PCB/controller frequency limitation.

One way to make this choice more clever is to go over all the
variants and for each of them derive an indicator which will help derive
the theoretical best. In this case, we derive a theoretical duration for
the entire operation and we take the smallest one.

Add a helper that parses the spi-mem operation and returns this value.

Signed-off-by: Miquel Raynal <miquel.raynal@bootlin.com>
Link: https://patch.msgid.link/20250110-winbond-6-11-rc1-quad-support-v3-20-7ab4bd56cf6e@bootlin.com
Signed-off-by: Mark Brown <broonie@kernel.org>

drivers/spi/spi-mem.c

@@ -562,6 +562,36 @@ void spi_mem_adjust_op_freq(struct spi_mem *mem, struct spi_mem_op *op)
 }
 EXPORT_SYMBOL_GPL(spi_mem_adjust_op_freq);
 
+/**
+ * spi_mem_calc_op_duration() - Derives the theoretical length (in ns) 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 ns estimate for the time this op would take.
+ */
+u64 spi_mem_calc_op_duration(struct spi_mem_op *op)
+{
+	u64 ncycles = 0;
+	u32 ns_per_cycles;
+
+	ns_per_cycles = 1000000000 / op->max_freq;
+	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);
+	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 * ns_per_cycles;
+}
+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/linux/spi/spi-mem.h

@@ -424,6 +424,7 @@ bool spi_mem_default_supports_op(struct spi_mem *mem,
 
 int spi_mem_adjust_op_size(struct spi_mem *mem, struct spi_mem_op *op);
 void spi_mem_adjust_op_freq(struct spi_mem *mem, struct spi_mem_op *op);
+u64 spi_mem_calc_op_duration(struct spi_mem_op *op);
 
 bool spi_mem_supports_op(struct spi_mem *mem,
 			 const struct spi_mem_op *op);