Hi,
On 2023-07-24 09:48:58 -0600, Jens Axboe wrote:
On 7/24/23 9:35?AM, Phil Elwell wrote:
Hi Andres,
With this commit applied to the 6.1 and later kernels (others not tested) the iowait time ("wa" field in top) in an ARM64 build running on a 4 core CPU (a Raspberry Pi 4 B) increases to 25%, as if one core is permanently blocked on I/O. The change can be observed after installing mariadb-server (no configuration or use is required). After reverting just this commit, "wa" drops to zero again.
There are a few other threads on this...
I can believe that this change hasn't negatively affected performance, but the result is misleading. I also think it's pushing the boundaries of what a back-port to stable should do.
FWIW, I think this partially just mpstat reporting something quite bogus. It makes no sense to say that a cpu is 100% busy waiting for IO, when the one process is doing IO is just waiting.
+static bool current_pending_io(void) +{
- struct io_uring_task *tctx = current->io_uring;
- if (!tctx)
return false;- return percpu_counter_read_positive(&tctx->inflight);
+}
/* when returns >0, the caller should retry */ static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, struct io_wait_queue *iowq) {
- int token, ret;
- int io_wait, ret;
if (unlikely(READ_ONCE(ctx->check_cq))) return 1; @@ -2511,17 +2520,19 @@ static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, return 0; /*
* Use io_schedule_prepare/finish, so cpufreq can take into account* that the task is waiting for IO - turns out to be important for low* QD IO.
* Mark us as being in io_wait if we have pending requests, so cpufreq* can take into account that the task is waiting for IO - turns out* to be important for low QD IO.
- token = io_schedule_prepare();
- io_wait = current->in_iowait;
I don't know the kernel "rules" around this, but ->in_iowait is only modified in kernel/sched, so it seemed a tad "unfriendly" to scribble on it here...
Building a kernel to test with the patch applied, will reboot into it once the call I am on has finished. Unfortunately the performance difference didn't reproduce nicely in VM...
Greetings,
Andres Freund