Patchwork [v3,2/2] block: disable I/O throttling on sync api

login
register
mail settings
Submitter Zhiyong Wu
Date March 30, 2012, 8:50 a.m.
Message ID <1333097427-20912-1-git-send-email-zwu.kernel@gmail.com>
Download mbox | patch
Permalink /patch/149589/
State New
Headers show

Comments

Zhiyong Wu - March 30, 2012, 8:50 a.m.
From: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>

Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
---
 block.c |   22 ++++++++++++++++++++++
 1 files changed, 22 insertions(+), 0 deletions(-)
Stefan Hajnoczi - March 30, 2012, 10:29 a.m.
On Fri, Mar 30, 2012 at 9:50 AM,  <zwu.kernel@gmail.com> wrote:
> +     * When all sync I/O drivers are converted to async I/O, it will be restored
> +     * to the original state.

The problem here is that guess_disk_lchs() uses bdrv_read(), not
whether device emulation uses sync or async I/O.  Converting devices
to async I/O will not fix this.

Stefan
Zhiyong Wu - March 31, 2012, 12:07 p.m.
On Fri, Mar 30, 2012 at 6:29 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> On Fri, Mar 30, 2012 at 9:50 AM,  <zwu.kernel@gmail.com> wrote:
>> +     * When all sync I/O drivers are converted to async I/O, it will be restored
>> +     * to the original state.
>
> The problem here is that guess_disk_lchs() uses bdrv_read(), not
> whether device emulation uses sync or async I/O.  Converting devices
> to async I/O will not fix this.
Yeah, but when all sync emulated devices drivers are converted to
async mode, the code change need to be rolled back. This is what i
mean.
>
> Stefan
Zhiyong Wu - March 31, 2012, 12:09 p.m.
On Fri, Mar 30, 2012 at 6:29 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> On Fri, Mar 30, 2012 at 9:50 AM,  <zwu.kernel@gmail.com> wrote:
>> +     * When all sync I/O drivers are converted to async I/O, it will be restored
>> +     * to the original state.
>
> The problem here is that guess_disk_lchs() uses bdrv_read(), not
> whether device emulation uses sync or async I/O.  Converting devices
> to async I/O will not fix this.
By the way, when you did testing, you applied #1 of this patchset? I
think that it make sense for that assert issue to be fixed.

>
> Stefan
Stefan Hajnoczi - April 2, 2012, 7:31 a.m.
On Sat, Mar 31, 2012 at 08:07:25PM +0800, Zhi Yong Wu wrote:
> On Fri, Mar 30, 2012 at 6:29 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> > On Fri, Mar 30, 2012 at 9:50 AM,  <zwu.kernel@gmail.com> wrote:
> >> +     * When all sync I/O drivers are converted to async I/O, it will be restored
> >> +     * to the original state.
> >
> > The problem here is that guess_disk_lchs() uses bdrv_read(), not
> > whether device emulation uses sync or async I/O.  Converting devices
> > to async I/O will not fix this.
> Yeah, but when all sync emulated devices drivers are converted to
> async mode, the code change need to be rolled back. This is what i
> mean.

There are places outside device emulation that call bdrv_read(), like
guest_disk_lchs() and bdrv_commit().  We would need to convert them
before it's safe to remove the code to disable I/O throttling in
bdrv_read()/bdrv_write().

It's not okay to drop the code before all bdrv_read()/bdrv_write()
callers that are in a code path where I/O limits can be enabled have
been fixed (including guest_disk_lchs()) because a request that exceeds
I/O limits would cause QEMU to hang.

Once all devices have been converted it might be possible to temporarily
disable throttling in bdrv_read()/bdrv_write() and never disable it
permanently.  This is because the bdrv_read()/bdrv_write() calls would
only come from QEMU code that is not directly driven by the guest - here
we probably don't need to apply I/O limits anyway.  In other words,
devices use bdrv_aio_*() and I/O limits apply, QEMU internal code might
still use bdrv_read()/bdrv_write() and limits will not apply.

Anyway, I think it's best to drop this part of the comment because it's
more related to the disable code in bdrv_read()/bdrv_write() than to
this patch.

Stefan
Stefan Hajnoczi - April 2, 2012, 7:42 a.m.
On Sat, Mar 31, 2012 at 08:09:46PM +0800, Zhi Yong Wu wrote:
> On Fri, Mar 30, 2012 at 6:29 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> > On Fri, Mar 30, 2012 at 9:50 AM,  <zwu.kernel@gmail.com> wrote:
> >> +     * When all sync I/O drivers are converted to async I/O, it will be restored
> >> +     * to the original state.
> >
> > The problem here is that guess_disk_lchs() uses bdrv_read(), not
> > whether device emulation uses sync or async I/O.  Converting devices
> > to async I/O will not fix this.
> By the way, when you did testing, you applied #1 of this patchset? I
> think that it make sense for that assert issue to be fixed.

Yes, I applied both patches 1 and 2.

Stefan

Patch

diff --git a/block.c b/block.c
index 1fbf4dd..6de6f89 100644
--- a/block.c
+++ b/block.c
@@ -1477,6 +1477,17 @@  static int bdrv_rw_co(BlockDriverState *bs, int64_t sector_num, uint8_t *buf,
 
     qemu_iovec_init_external(&qiov, &iov, 1);
 
+    /**
+     * In sync call context, when the vcpu is blocked, this throttling timer
+     * will not fire; so the I/O throttling function has to be disabled here
+     * if it has been enabled.
+     */
+    if (bs->io_limits_enabled) {
+        fprintf(stderr, "Disabling I/O throttling on '%s' due "
+                        "to synchronous I/O.\n", bdrv_get_device_name(bs));
+        bdrv_io_limits_disable(bs);
+    }
+
     if (qemu_in_coroutine()) {
         /* Fast-path if already in coroutine context */
         bdrv_rw_co_entry(&rwco);
@@ -1983,10 +1994,21 @@  static int guess_disk_lchs(BlockDriverState *bs,
     struct partition *p;
     uint32_t nr_sects;
     uint64_t nb_sectors;
+    bool enabled;
 
     bdrv_get_geometry(bs, &nb_sectors);
 
+    /**
+     * The function will be invoked during startup not only in sync I/O mode,
+     * but also in async I/O mode. So the I/O throttling function has to
+     * be disabled temporarily here, not permanently.
+     * When all sync I/O drivers are converted to async I/O, it will be restored
+     * to the original state.
+     */
+    enabled = bs->io_limits_enabled;
+    bs->io_limits_enabled = false;
     ret = bdrv_read(bs, 0, buf, 1);
+    bs->io_limits_enabled = enabled;
     if (ret < 0)
         return -1;
     /* test msdos magic */