On 3/31/2020 11:43 AM, Bijan Mottahedeh wrote:
Does io_uring though have to deal with BLK_QC_T_NONE at all? Or are
you
saying that it should never receive that result?
That's one of the things I'm not clear about.
BLK_QC_T_* are block cookies, they are only valid in the block layer.
Only the poll handler called should have to deal with them, inside
their f_op->iopoll() handler. It's simply passed from the queue to
the poll side.
So no, io_uring shouldn't have to deal with them at all.
The problem, as I see it, is if the block layer returns BLK_QC_T_NONE
and the IO was actually queued and requires polling to be found. We'd
end up with IO timeouts for handling those requests, and that's not a
good thing...
I see requests in io_do_iopoll() on poll_list with req->res ==
-EAGAIN, I think because the completion happened after an issued
request was added to poll_list in io_iopoll_req_issued().
How should we deal with such a request, reissue unconditionally or
something else?
I mimicked the done processing code in io_iopoll_complete() for -EAGAIN
as a test. I can now get further and don't see polling threads hang; in
fact, I eventually see I/O timeouts as you noted.
It seems that there might be two separate issues here. Makes sense?
Thanks.
--bijan
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 62bd410..a3e3a4e 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1738,11 +1738,24 @@ static void io_iopoll_complete(struct
io_ring_ctx *ctx,
io_free_req_many(ctx, &rb);
}
+static void io_iopoll_queue(struct list_head *again)
+{
+ struct io_kiocb *req;
+
+ while (!list_empty(again)) {
+ req = list_first_entry(again, struct io_kiocb, list);
+ list_del(&req->list);
+ refcount_inc(&req->refs);
+ io_queue_async_work(req);
+ }
+}
+
static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
long min)
{
struct io_kiocb *req, *tmp;
LIST_HEAD(done);
+ LIST_HEAD(again);
bool spin;
int ret;
@@ -1757,9 +1770,9 @@ static int io_do_iopoll(struct io_ring_ctx *ctx,
unsigned
struct kiocb *kiocb = &req->rw.kiocb;
/*
- * Move completed entries to our local list. If we find a
- * request that requires polling, break out and complete
- * the done list first, if we have entries there.
+ * Move completed and retryable entries to our local lists.
+ * If we find a request that requires polling, break out
+ * and complete those lists first, if we have entries there.
*/
if (req->flags & REQ_F_IOPOLL_COMPLETED) {
list_move_tail(&req->list, &done);
@@ -1768,6 +1781,13 @@ static int io_do_iopoll(struct io_ring_ctx *ctx,
unsigned
if (!list_empty(&done))
break;
+ if (req->result == -EAGAIN) {
+ list_move_tail(&req->list, &again);
+ continue;
+ }
+ if (!list_empty(&again))
+ break;
+
ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
if (ret < 0)
break;
@@ -1780,6 +1800,9 @@ static int io_do_iopoll(struct io_ring_ctx *ctx,
unsigned
if (!list_empty(&done))
io_iopoll_complete(ctx, nr_events, &done);
+ if (!list_empty(&again))
+ io_iopoll_queue(&again);
+
return ret;
}