This commit fixes the following issues:

1. The 9th bit of buf was believed to be the LSB of divisor's
exponent, but the hardware interprets it as MSB (9th bit) of the
mantissa. The exponent is actually one bit shorter and applies
to base 4, not 2 as previously believed.

2. Loop iterations doubled the exponent instead of incrementing.

3. The exponent wasn't checked for overflow.

4. The function returned requested rate instead of actual rate.

Due to issue #2, the old code deviated from the wrong formula
described in #1 and actually yielded correct rates when divisor
was lower than 4096 by using exponents of 0, 2 or 4 base-2,
interpreted as 0, 1, 2 base-4 with the 9th mantissa bit clear.
However, at 93.75 kbaud or less the rate turned out too slow
due to #2 or too fast due to #2 and #3.

I tested this patch by sending and validating 0x00,0x01,..,0xff
to an FTDI dongle at 234, 987, 2401, 9601, 31415, 115199, 250k,
500k, 750k, 1M, 1.5M, 3M+1 baud. All rates passed.

I also used pv to check speed at some rates unsupported by FTDI:
45 (the lowest possible), 2M, 4M, 5M and 6M-1. Looked sane.

Signed-off-by: Michal Pecio <michal.pe...@gmail.com>
---
 drivers/usb/serial/pl2303.c | 32 +++++++++++++++++++++++---------
 1 file changed, 23 insertions(+), 9 deletions(-)

diff --git a/drivers/usb/serial/pl2303.c b/drivers/usb/serial/pl2303.c
index f5257af..3f9a808 100644
--- a/drivers/usb/serial/pl2303.c
+++ b/drivers/usb/serial/pl2303.c
@@ -362,22 +362,36 @@ static speed_t pl2303_encode_baud_rate_direct(unsigned 
char buf[4],
 static speed_t pl2303_encode_baud_rate_divisor(unsigned char buf[4],
                                                                speed_t baud)
 {
-       unsigned int tmp;
+       u32 baseline, mantissa, exponent;
 
        /*
         * Apparently the formula is:
-        * baudrate = 12M * 32 / (2^buf[1]) / buf[0]
+        *   baudrate = 12M * 32 / (mantissa * 4^exponent)
+        * where
+        *   mantissa = buf[8:0]
+        *   exponent = buf[11:9]
         */
-       tmp = 12000000 * 32 / baud;
+       baseline = 12000000 * 32;
+       mantissa = baseline / baud;
+       exponent = 0;
+       while (mantissa >= 512) {
+               if (exponent < 7) {
+                       mantissa >>= 2; /* divide by 4 */
+                       exponent++;
+               } else {
+                       /* Exponent is maxed. Trim mantissa and leave. */
+                       mantissa = 511;
+                       break;
+               }
+       }
+
        buf[3] = 0x80;
        buf[2] = 0;
-       buf[1] = (tmp >= 256);
-       while (tmp >= 256) {
-               tmp >>= 2;
-               buf[1] <<= 1;
-       }
-       buf[0] = tmp;
+       buf[1] = exponent << 1 | mantissa >> 8;
+       buf[0] = mantissa & 0xff;
 
+       /* Calculate and return the exact baud rate. */
+       baud = (baseline / mantissa) >> (exponent << 1);
        return baud;
 }
 
-- 
2.4.5

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Reply via email to