4.2.9 Reduce Poll Frequency

UEFI drivers operate in a polled mode and do not use interrupts. For example, UEFI drivers that implement blocking I/O services can simply poll the device until the request is complete. UEFI drivers that implement non-blocking I/O can create a periodic timer event to poll a device at periodic intervals.

A common mistake in UEFI drivers is polling too often.

Remember that polling, versus interrupts, is a pull model, not a push model. The tradeoff in a polling system is how fast the device is polled (which can degrade system performance) versus how responsive the driver is to that request. For example, in a polling system, the driver should not send a request to a device and wait until that device responds before moving on to another task. In general, the polling interval should be set to the largest possible period for the UEFI driver to complete its I/O services in a reasonable period of time. The overall performance of a UEFI-enabled platform degrades if too many UEFI drivers create high-frequency periodic timer events.

Note: It is recommended that the period of a periodic timer event be at least 10 ms. In general, the period should be as large as possible based upon a specific device's timing requirements. Most drivers can use events with timer periods in the range of 100 ms to several seconds.

When initially writing the driver, an estimate can be made for the initial polling frequency. However, the polling frequency may have to be adjusted based on an analysis of the driver's performance on an actual machine.

TIP: As part of the development process, make sure time is reserved for performance analysis to find out how much time is taken up polling each device.

4.2.9.1 Distinguishing a polling issue versus another type of bug

The symptoms of a polling issue versus some other type of bug can look nearly identical. The key to identifying a polling issue is: Don't assume anything. Begin simply by performing an analysis to get data-the time taken by each task can be measured. If a task is taking longer than expected, the code associated with that task can then be examined more closely.