Perhaps I can add something useful here:

This is from memory, waayyy back, so forgive me if I miss anything. At one time I worked in a design group that designed a 16550 standard part (not me directly). The 16550 problem was with the first version National 16550(A?) device circa 1990. The transmit FIFO had a bug where sometimes the last character in the TX buffer would not transmit. The part would indicate buffer empty to the host who promptly refilled the buffer, overwriting the remaining previous last byte. The fix was to add a patch to Widows placing the part in 16450 compatibility mode (turning off the FIFO’s). Later device revisions fixed this and as far as I recall it was limited to National. This showed up in late ‘386 and early ‘486 machines, probably in Win3.x .

On XOFF, I worked on a V.42bis modem using an 8052 controller. 100 bytes is about right (measured on a storage scope) for the old ‘486 days, could more now. The 16550 FIFO is only 16 byes, so the latency is from the OS, there is a buffer in the driver “sitting above” the actual hardware. We used an external RAM with about a 2K Rx buffer so we could store an entire fame.

The Xon/Xoff characters were/are user selectable in the data steam in dial-up modems in case of problems. They can’t “show up” as data.

The 16550 (and I believe the 16450) can also do hardware (out of band) flow control, see the /CTS and /RTS pins in the data sheet.

I hope this helps.

Stephan Buck