[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Qemu-devel] [PULL 01/17] linux-aio: consume events in userspace instead
From: |
Stefan Hajnoczi |
Subject: |
[Qemu-devel] [PULL 01/17] linux-aio: consume events in userspace instead of calling io_getevents |
Date: |
Mon, 12 Sep 2016 15:08:38 +0100 |
From: Roman Pen <address@hidden>
AIO context in userspace is represented as a simple ring buffer, which
can be consumed directly without entering the kernel, which obviously
can bring some performance gain. QEMU does not use timeout value for
waiting for events completions, so we can consume all events from
userspace.
Signed-off-by: Roman Pen <address@hidden>
Message-id: address@hidden
Cc: Stefan Hajnoczi <address@hidden>
Cc: Paolo Bonzini <address@hidden>
Cc: address@hidden
Signed-off-by: Stefan Hajnoczi <address@hidden>
---
block/linux-aio.c | 129 ++++++++++++++++++++++++++++++++++++++++++------------
1 file changed, 101 insertions(+), 28 deletions(-)
diff --git a/block/linux-aio.c b/block/linux-aio.c
index e906abe..62ee1ea 100644
--- a/block/linux-aio.c
+++ b/block/linux-aio.c
@@ -59,7 +59,6 @@ struct LinuxAioState {
/* I/O completion processing */
QEMUBH *completion_bh;
- struct io_event events[MAX_EVENTS];
int event_idx;
int event_max;
};
@@ -102,6 +101,85 @@ static void qemu_laio_process_completion(struct
qemu_laiocb *laiocb)
}
}
+/**
+ * aio_ring buffer which is shared between userspace and kernel.
+ *
+ * This copied from linux/fs/aio.c, common header does not exist
+ * but AIO exists for ages so we assume ABI is stable.
+ */
+struct aio_ring {
+ unsigned id; /* kernel internal index number */
+ unsigned nr; /* number of io_events */
+ unsigned head; /* Written to by userland or by kernel. */
+ unsigned tail;
+
+ unsigned magic;
+ unsigned compat_features;
+ unsigned incompat_features;
+ unsigned header_length; /* size of aio_ring */
+
+ struct io_event io_events[0];
+};
+
+/**
+ * io_getevents_peek:
+ * @ctx: AIO context
+ * @events: pointer on events array, output value
+
+ * Returns the number of completed events and sets a pointer
+ * on events array. This function does not update the internal
+ * ring buffer, only reads head and tail. When @events has been
+ * processed io_getevents_commit() must be called.
+ */
+static inline unsigned int io_getevents_peek(io_context_t ctx,
+ struct io_event **events)
+{
+ struct aio_ring *ring = (struct aio_ring *)ctx;
+ unsigned int head = ring->head, tail = ring->tail;
+ unsigned int nr;
+
+ nr = tail >= head ? tail - head : ring->nr - head;
+ *events = ring->io_events + head;
+ /* To avoid speculative loads of s->events[i] before observing tail.
+ Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
+ smp_rmb();
+
+ return nr;
+}
+
+/**
+ * io_getevents_commit:
+ * @ctx: AIO context
+ * @nr: the number of events on which head should be advanced
+ *
+ * Advances head of a ring buffer.
+ */
+static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
+{
+ struct aio_ring *ring = (struct aio_ring *)ctx;
+
+ if (nr) {
+ ring->head = (ring->head + nr) % ring->nr;
+ }
+}
+
+/**
+ * io_getevents_advance_and_peek:
+ * @ctx: AIO context
+ * @events: pointer on events array, output value
+ * @nr: the number of events on which head should be advanced
+ *
+ * Advances head of a ring buffer and returns number of elements left.
+ */
+static inline unsigned int
+io_getevents_advance_and_peek(io_context_t ctx,
+ struct io_event **events,
+ unsigned int nr)
+{
+ io_getevents_commit(ctx, nr);
+ return io_getevents_peek(ctx, events);
+}
+
/* The completion BH fetches completed I/O requests and invokes their
* callbacks.
*
@@ -116,43 +194,38 @@ static void qemu_laio_process_completion(struct
qemu_laiocb *laiocb)
static void qemu_laio_completion_bh(void *opaque)
{
LinuxAioState *s = opaque;
-
- /* Fetch more completion events when empty */
- if (s->event_idx == s->event_max) {
- do {
- struct timespec ts = { 0 };
- s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
- s->events, &ts);
- } while (s->event_max == -EINTR);
-
- s->event_idx = 0;
- if (s->event_max <= 0) {
- s->event_max = 0;
- return; /* no more events */
- }
- s->io_q.in_flight -= s->event_max;
- }
+ struct io_event *events;
/* Reschedule so nested event loops see currently pending completions */
qemu_bh_schedule(s->completion_bh);
- /* Process completion events */
- while (s->event_idx < s->event_max) {
- struct iocb *iocb = s->events[s->event_idx].obj;
- struct qemu_laiocb *laiocb =
+ while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
+ s->event_idx))) {
+ for (s->event_idx = 0; s->event_idx < s->event_max; ) {
+ struct iocb *iocb = events[s->event_idx].obj;
+ struct qemu_laiocb *laiocb =
container_of(iocb, struct qemu_laiocb, iocb);
- laiocb->ret = io_event_ret(&s->events[s->event_idx]);
- s->event_idx++;
+ laiocb->ret = io_event_ret(&events[s->event_idx]);
- qemu_laio_process_completion(laiocb);
- }
-
- if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
- ioq_submit(s);
+ /* Change counters one-by-one because we can be nested. */
+ s->io_q.in_flight--;
+ s->event_idx++;
+ qemu_laio_process_completion(laiocb);
+ }
}
qemu_bh_cancel(s->completion_bh);
+
+ /* If we are nested we have to notify the level above that we are done
+ * by setting event_max to zero, upper level will then jump out of it's
+ * own `for` loop. If we are the last all counters droped to zero. */
+ s->event_max = 0;
+ s->event_idx = 0;
+
+ if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
+ ioq_submit(s);
+ }
}
static void qemu_laio_completion_cb(EventNotifier *e)
--
2.7.4
- [Qemu-devel] [PULL 00/17] Block patches, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 01/17] linux-aio: consume events in userspace instead of calling io_getevents,
Stefan Hajnoczi <=
- [Qemu-devel] [PULL 02/17] linux-aio: split processing events function, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 04/17] virtio-blk: rename virtio_device_info to virtio_blk_info, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 03/17] linux-aio: process completions from ioq_submit(), Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 05/17] block: unblock backup operations in backing file, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 06/17] Backup: clear all bitmap when doing block checkpoint, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 07/17] Backup: export interfaces for extra serialization, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 09/17] docs: block replication's description, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 08/17] block: Link backup into block core, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 10/17] mirror: auto complete active commit, Stefan Hajnoczi, 2016/09/12
- [Qemu-devel] [PULL 11/17] configure: support replication, Stefan Hajnoczi, 2016/09/12