On 6/3/20 7:04 PM, Jens Axboe wrote: > On 6/3/20 6:59 PM, Andres Freund wrote: >> Hi, >> >> I was trying to benchmark the benefits of this for the io_uring using >> postgres I am working on. The initial results where quite promising >> (reducing cpu usage significantly, to lower than non-uring sync io). But >> unfortunately trying another workload triggered both panics and before >> that seemingly returned wrong data. >> >> I first saw that problem with b360d424ce02, which was >> linux-block/async-buffered.6 at the time. After hitting the issue, I >> updated to the current linux-block/async-buffered.6, but the problem >> persists. >> >> The workload that triggers the bug within a few seconds is postgres >> doing a parallel sequential scan of a large table (and aggregating the >> data, but that shouldn't matter). In the triggering case that boils down >> to 9 processes sequentially reading a number of 1GB files (we chunk >> tables internally into smaller files). Each process will read a 512kB >> chunk of the file on its own, and then claim the next 512kB from a >> shared memory location. Most of the IO will be READV requests, reading >> 16 * 8kB into postgres' buffer pool (which may or may not be neighboring >> 8kB pages). > > I'll try and reproduce this, any chance you have a test case that can > be run so I don't have to write one from scratch? The more detailed > instructions the better. Can you try with async-buffered.7? I've rebased it on a new mechanism, and doing something like what you describe above I haven't been able to trigger anything bad. I'd try your test case specifically, so do let know if it's something I can run. -- Jens Axboe
