Hi all, The I2C/SMBus chip detection in sensors-detect is quite dumb, each possible device is tested individually, resulting in the same registers being read over and over again. We could rethink the probing logic completely to workaround that, but that's a huge change and it doesn't go without risks. So I propose a different approach: caching the register values. We can do it almost transparently in i2c_smbus_read_byte_data and i2c_smbus_read_word_data, clearing the caches when we switch to the next address. I implemented that, patch is attached. One thing that needed special care is that some devices do not answer consistently to some reads, and we use it as a detection method. So I added a NO_CACHE flag to force actual reads from the chip, for detecting these devices properly. The LM75, LM77 and MAX6657/8/9 are affected, and I could test that my LM75 is still detected properly after the change. I left the extra instrumentation in this version of the patch if you want to play with it. To disable caching, just comment out the intermediate return statements in i2c_smbus_read_byte_data and i2c_smbus_read_word_data. On my systems, I found that the benefit differs greatly depending on the address and chip, from almost none to over 50% of cache hits, effectively dividing the probing time by almost 2. I kept the byte and word caches separated for now, but we might improve the speed even more by sharing them. I wonder if there are drawbacks in doing so. I'll look into that later. The attached patch applies on top of current SVN. Cumulated with the other cleanups I made since the beginning of the week, this should give us a much much faster detection script. Comments anyone? -- Jean Delvare -------------- next part -------------- An embedded and charset-unspecified text was scrubbed... Name: sensors-detect-register-cache.patch Url: http://lists.lm-sensors.org/pipermail/lm-sensors/attachments/20070125/a99c466e/attachment.pl
