Just over a year ago I implemented an optimization to the SPI core code in Linux that avoids some needless context switches to a worker thread in the main data path that most clients use. This was really nice, it was simple to do but saved a bunch of work for most drivers using SPI and made things noticeably faster. The code got merged in v4.0 and that was that, I kept on kicking a few more ideas for optimizations in this area around but that was that until the past month.
What happened then was that for whatever reason people started picking up v4.0 and using it in production more. On some systems people started seeing problems when there was heavy SPI flash usage, often during things like distribution installation. In some cases the lockup detector fired, but the most entertaining error was that on Marvell Orion systems (which are single core) when the flash was being heavily used the SATA controller started having trouble handling interrupts. These problems all bisected down to the key commit in that series, 0461a4149836c79 (spi: Pump transfers inside calling context for spi_sync()).
The problem is that there are a number of widely deployed SPI controllers out there that don’t support DMA and instead require the CPU to explicitly read and write everything sent to and from registers in the controller. To make matters worse these accesses to the controller will usually take many CPU cycles to complete, each one stalling the CPU while they happen. This is fine for short transfers or if the CPU has nothing else to do but on a busy multitasking system it’s an issue. Before the optimization the switches between the worker thread interacting with the hardware and the thread initiating the SPI operations provided breaks in this activity which allowed other things to switch in. Unfortunately when we optimize those away then if there’s a lot of work for the controller being done from one thread then that thread can run for a long time without pause. The fix for affected drivers if there are no less CPU intensive ways of driving the hardware is to add some explicit sleeps into the driver itself, either at the end of the transfer_one() or perhaps in an unprepare_message() function.
In a way I was quite pleased to see this, it was a clear demonstration that the optimization was having the intended effect though obviously users of affected systems will not find that so comforting. It’s not the first time that making things faster or fixing a bug has revealed an underlying problem, I’m sure it won’t be the last.