14. Interrupt Problem Details

While the section Troubleshooting lists problems by symptom, this section explains what will happen if interrupts are set incorrectly. This section helps you understand what caused the symptom, what other symptoms might be due to the same problem, and what to do about it.

14.1 Types of interrupt problems

The "setserial" program will show you how serial driver thinks the interrupts are set. If the serial driver (and setserial) has it right then everything regarding interrupts should be OK. Of course a /dev/ttyS must exist for the device and Plug-and-Play (or jumpers) must have set an address and IRQ in the hardware. Linux will not knowingly permit an interrupt conflict and you will get a "Device or resource busy" error message if you attempt to do something that would create a conflict.

Since the kernel tries to avoid interrupt conflicts and gives you the "resource busy" message if you try to create a conflict, how can interrupt conflicts happen? Easy. "setserial" may have it wrong and erroneously predicts no conflict when there will actually be a real conflict based on what is set in the hardware. When this happens there will be no "... busy" message but a conflict will physically happen. Performance is likely to be extremely slow. Both devices will send identical interrupt signals on the same wire and the CPU will erroneously think that the interrupts only come from one device. This will be explained in detail in the following sections.

Linux doesn't complain when you assign two devices the same IRQ provided that neither device is in use. As each device starts up (initializes), it asks Linux for permission to use its hardware interrupt. Linux keeps track of which interrupt is assigned to whom, and if your interrupt is already in use, you'll see this "... busy" error message. Thus if two devices use the same IRQ and you start up only one of the devices, everything is OK. But when you next try to start the second device (without quitting the first device) you get "... busy" error message.

14.2 Symptoms of Mis-set or Conflicting Interrupts

The symptoms depend on whether or not you have a modern serial port with FIFO buffers or an obsolete serial port without FIFO buffers. It's important to understand the symptoms for the obsolete ones also since sometimes modern ports seem to behave that way.

For the obsolete serial ports, only one character gets thru every several seconds. This is so slow that it seems almost like nothing is working (especially if the character that gets thru is invisible (such a space or newline). For the modern ports with FIFO buffers you will likely see bursts of up to 16 characters every several seconds.

If you have a modem on the port and dial a number, it seemingly may not connect since the CONNECT message may not make it thru. But after a long wait it may finally connect and you may see part of a login message (or the like). The response from your side of the connection may be so delayed that the other side gives up and disconnects you, resulting in a NO CARRIER message.

If you use minicom, a common test to see if things are working is to type the simplest "AT" command and see if the modem responds. Typing just at<enter> should normally (if interrupts are OK) result in an immediate "OK" response from the modem. With bad interrupts you type at<enter> and may see nothing. But then after 10 seconds or so you see the cursor drop down one line. What is going on is that the FIFO is behaving like it can only hold one byte. The "at" you typed caused it to overrun and both letters were lost. But the final <enter> eventually got thru and you "see" this invisible character by noticing that the cursor jumped down one line. If you were to type a single letter and then wait about 10 seconds, you should see it echo back to the screen. This is fine if your typing speed is less that one word per minute :-)

14.3 Mis-set Interrupts

If you don't understand what an interrupt does see Interrupts. If a serial port has one IRQ set in the hardware but a different one set in the device driver, the device driver will not catch any interrupts sent by the serial port. Since the serial port uses interrupts to call its driver to service the port (fetching bytes from its 16-byte receive buffer or putting another 16-bytes in its transmit buffer) one might expect that the serial port would not work at all.

But it still may work anyway --sort of. Why? Well, besides the interrupt method of servicing the port there's a slow polling method that doesn't need interrupts. The way it works is that every so often the device driver checks the serial port to see if it needs anything such as if it has some bytes that need fetching from its receive buffer. If interrupts don't work, the serial driver falls back to this polling method. But this polling method was not intended to be used a substitute for interrupts. It's so slow that it's not practical to use and may cause buffer overruns. Its purpose may have been to get things going again if just one interrupt is lost or fails to do the right thing. It's also useful in showing you that interrupts have failed. Don't confuse this slow polling method with the fast polling method that operates on ports that have their IRQs set to 0.

For the 16-byte transmit buffer, 16 bytes will be transmitted and then it will wait until the next polling takes place (several seconds later) before the next 16 bytes are sent out. Thus transmission is very slow and in small chunks. Receiving is slow too since bytes that are received by the receive buffer are likely to remain there for several seconds until it is polled.

This explains why it takes so long before you see what you typed. When you type say AT to a modem, the AT goes out the serial port to the modem. The modem then echos the AT back thru the serial port to the screen. Thus the AT characters have to pass twice thru the serial port. Normally this happens so fast that AT seems to appear on the screen at the same time you hit the keys on the keyboard. With slow polling delays at the serial port, you don't see what you typed until many seconds later.

What about overruns of the 16-byte receive buffer? This will happen with an external modem since the modem just sends to the serial port at high speed which is likely to overrun the 16-byte buffer. But for an internal modem, the serial port is on the same card and it's likely to check that this receive buffer has room for more bytes before putting received bytes into it. In this case there will be no overrun of this receive buffer, but text will just appear on your screen in 16-byte chunks spaced at intervals of several seconds.

Even with an external modem you might not get overruns. If just a few characters (under 16) are sent you don't get overruns since the buffer likely has room for them. But attempts to send a larger number of bytes from your modem to your screen may result in overruns. However, more than 16 (with no gaps) can get thru without overruns if the timing is right. For example, suppose a burst of 32 bytes is sent into the port from the external cable. The polling might just happen after the first 16 bytes came in so it would pick up these 16 bytes OK. Then there would be space for the next 16 bytes so that entire 32 bytes gets thru OK. While this scenario is not very likely, similar cases where 17 to 31 bytes make thru are more likely. But it's even more likely that only an occasional 16-byte chunk will get thru with possible loss of data.

If you have an obsolete serial port with only a 1-byte buffer (or it's been incorrectly set to work like a 1-byte buffer) then the situation will be much worse than described above and only one character will occasionally make it thru the port. Every character received causes an overrun (and is lost) except for the last character received. This character is likely to be just a line-feed since this is often the last character to be transmitted in a burst of characters sent to your screen. Thus you may type AT<return> to the modem but never see AT on the screen. All you see several seconds later is that the cursor drops down one line (a line feed). This has happened to me with a 16-byte FIFO buffer that was behaving like a 1-byte buffer.

When a communication program starts up, it expects interrupts to be working. It's not geared to using this slow polling-like mode of operation. Thus all sorts of mistakes may be made such as setting up the serial port and/or modem incorrectly. It may fail to realize when a connection has been made. If a script is being used for login, it may fail (caused by timeout) due to the polling delays.

14.4 Interrupt Conflicts

When two devices have the same IRQ number it's called sharing interrupts. Under some conditions this sharing works out OK. Starting with kernel version 2.2, ISA serial ports may, if the hardware is designed for this, share interrupts with other serial ports. Devices on the PCI bus may share the same IRQ interrupt with other devices on the PCI bus (provided the software supports this). In other cases where there is potential for conflict, there should be no problem if no two devices with the same IRQ are ever "in use" at the same time. More precisely, "in use" really means "open" (in programmer jargon). In cases other than the exceptions mentioned above (unless special software and hardware permit sharing), sharing is not allowed and conflicts arise if sharing is attempted.

Even if two processes with conflicting IRQs run at the same time, one of the devices will likely have its interrupts caught by its device driver and may work OK. The other device will not have its interrupts caught by the correct driver and will likely behave just like a process with mis-set interrupts. See Mis-set Interrupts for more details.

14.5 Resolving Interrupt Problems

If you are getting a very slow response as described above, then one test is to change the IRQ to 0 (uses fast polling instead of interrupts) and see if the problem goes away. Note that the polling due to IRQ=0 is orders of magnitude faster than the slow "polling" due to bad interrupts. If IRQ=0 seems to fix the problem, then there was likely something wrong with the interrupts. Using IRQ=0 is very resource intensive and is only a temporary fix. You should try to find the cause of the interrupt problem and not permanently use IRQ=0.

Check /proc/interrupts to see if the IRQ is currently in use by another process. If it's in use by another serial port you could try "top" (type f and then enable the TTY display) or "ps -e" to find out which serial ports are in use. If you suspect that setserial has a wrong IRQ then see What is the current IO address and IRQ of my Serial Port ?