qemu-devel
[Top][All Lists]
Advanced
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH] ceph/rbd block driver for qemu-kvm (v4)

From: Anthony Liguori
Subject: Re: [Qemu-devel] [PATCH] ceph/rbd block driver for qemu-kvm (v4)
Date: Fri, 08 Oct 2010 09:06:18 -0500
User-agent: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.9.1.12) Gecko/20100915 Lightning/1.0b1 Thunderbird/3.0.8 
On 10/07/2010 05:45 PM, Sage Weil wrote:

On Thu, 7 Oct 2010, Anthony Liguori wrote:


On 10/07/2010 04:49 PM, Yehuda Sadeh Weinraub wrote:

On Thu, Oct 7, 2010 at 2:04 PM, Anthony Liguori<address@hidden>
wrote:


On 10/07/2010 03:47 PM, Yehuda Sadeh Weinraub wrote:


How is that possible?  Are the callbacks delivered in the context of a
different thread?  If so, don't you need locking?



Not sure I'm completely following you. The callbacks are delivered in
the context of a different thread, but won't run concurrently.


Concurrently to what?  How do you prevent them from running concurrently
with qemu?


There are two types of callbacks. The first is for rados aio
completions, and the second one is the one added later for the fd glue
layer.


This is a bad architecture for something like qemu.  You could create a
pipe and use the pipe to signal to qemu.  Same principle as eventfd.
Ideally, you would do this in the library itself.

I'm sorry, I'm having a hard time understanding what it is you're
objecting to, or what you would prefer, as there are two different things
we're talking about here (callbacks and fd glue/pipes).  (Please bear with
me as I am not a qemu expert!)

The first is the aio completion.  You said a few messages back:


It looks like you just use the eventfd to signal aio completion
callbacks.  A better way to do this would be to schedule a bottom half.

This is what we're doing.  The librados makes a callback to rbd.c's
rbd_finish_aiocb(), which updates some internal rbd accounting and then
calls qemu_bh_schedule().  Is that part right?
No. You're calling qemu_bh_schedule() in a separate thread in parallel to other operations. 

That's absolutely not safe.


The second part is an fd (currently created via eventfd(), but I don't
think it matters where it comes from) that was later added because
qemu_aio_flush() wouldn't trigger when our aio's completed (and scheduled
the bottom halves).  This was proposed by Simone Gotti, who had problems
with live migration:

        http://www.mail-archive.com/address@hidden/msg35516.html

Apparently calling the bottom half isn't sufficient to wake up a blocked
qemu_aio_flush()?  His solution was to create an eventfd() fd, write a
word to it in the aio completion callback (before we schedule the bh), and
add the necessary callbacks to make qemu_aio_flush() behave.

Is the problem simply that we should be using pipe(2) instead of
eventfd(2)?

So far I've heard that we should be scheduling the bottom halves (we are),
and we should be using a pipe to signal qemu (we're using an fd created by
eventfd(2)).
Your fundamental problem is your use of threads. QEMU is single threaded. You cannot call into QEMU code from another thread without introducing locking. Any other solution is going to be intrinsically broken. 

There are two possibilities to fix this:
1) You can change your library interface so that it doesn't generate callbacks via threads. That would be my preference because I think it's a bad interface but it's your library so it's not really my choice :-)
2) You can limit the callbacks to doing nothing other than writing to a file descriptor. You then read the file descriptor somewhere else in the normal QEMU code and you can use the file descriptor to get signals. If you're passing data to callbacks, it's much harder because you're going to have to store that data somewhere and inevitably require locking.
The complexity of (2) is why I think thread-based callbacks is such a bad interface. 

Regards,

Anthony Liguori


Thanks,
sage





Regards,

Anthony Liguori


The first callback, called by librados whenever aio completes, runs in
the context of a single librados thread:

+static void rbd_finish_aiocb(rados_completion_t c, RADOSCB *rcb)
+{
+    RBDAIOCB *acb = rcb->acb;
rcb is per a single aio. Was created  before and will be destroyed
here, whereas acb is shared between a few aios, however, it was
generated before the first aio was created.

+    int64_t r;
+    uint64_t buf = 1;
+    int i;
+
+    acb->aiocnt--;

acb->aiocnt has been set before initiating all the aios, so it's ok to
touch it now. Same goes to all acb fields.

+    r = rados_aio_get_return_value(c);
+    rados_aio_release(c);
+    if (acb->write) {
+        if (r<   0) {
+            acb->ret = r;
+            acb->error = 1;
+        } else if (!acb->error) {
+            acb->ret += rcb->segsize;
+        }
+    } else {
+        if (r == -ENOENT) {
+            memset(rcb->buf, 0, rcb->segsize);
+            if (!acb->error) {
+                acb->ret += rcb->segsize;
+            }
+        } else if (r<   0) {
+            acb->ret = r;
+            acb->error = 1;
+        } else if (r<   rcb->segsize) {
+            memset(rcb->buf + r, 0, rcb->segsize - r);
+            if (!acb->error) {
+                acb->ret += rcb->segsize;
+            }
+        } else if (!acb->error) {
+            acb->ret += r;
+        }
+    }
+    if (write(acb->s->efd,&buf, sizeof(buf))<   0)
This will wake up the io_read()

+        error_report("failed writing to acb->s->efd\n");
+    qemu_free(rcb);
+    i = 0;
+    if (!acb->aiocnt&&   acb->bh) {
+        qemu_bh_schedule(acb->bh);
This is the only qemu related call in here, seems safe to call it.

+    }
+}

The scheduled bh function will be called only after all aios that
relate to this specific aio set are done, so the following seems ok,
as there's no more acb references.
+static void rbd_aio_bh_cb(void *opaque)
+{
+    RBDAIOCB *acb = opaque;
+    uint64_t buf = 1;
+
+    if (!acb->write) {
+        qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
+    }
+    qemu_vfree(acb->bounce);
+    acb->common.cb(acb->common.opaque, (acb->ret>   0 ? 0 : acb->ret));
+    qemu_bh_delete(acb->bh);
+    acb->bh = NULL;
+
+    if (write(acb->s->efd,&buf, sizeof(buf))<   0)
+        error_report("failed writing to acb->s->efd\n");
+    qemu_aio_release(acb);
+}

Now, the second ones are the io_read(), in which we have our glue fd.
We send uint64 per each completed io

+static void rbd_aio_completion_cb(void *opaque)
+{
+    BDRVRBDState *s = opaque;
+
+    uint64_t val;
+    ssize_t ret;
+
+    do {
+        if ((ret = read(s->efd,&val, sizeof(val)))>   0) {
+            s->qemu_aio_count -= val;
There is an issue here with s->qemu_aio_count which needs to be
protected by a mutex. Other than that, it just reads from s->efd.

+       }
+    } while (ret<   0&&   errno == EINTR);
+
+    return;
+}
+
+static int rbd_aio_flush_cb(void *opaque)
+{
+    BDRVRBDState *s = opaque;
+
+    return (s->qemu_aio_count>   0);
Same here as with the previous one, needs a mutex around s->qemu_aio_count.

+}



If you saw lock ups, I bet that's what it was from.



As I explained before, before introducing the fd glue layer, the lack
of fd associated with our block device caused that there was no way
for qemu to check whether all aios were flushed or not, which didn't
work well when doing migration/savevm.

Thanks,
Yehuda


--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to address@hidden
More majordomo info at  http://vger.kernel.org/majordomo-info.html
reply via email to
[Prev in Thread] Current Thread [Next in Thread]