Re: [Xen-devel] [PATCH v2 3/3] xen-disk: use an IOThread per instance

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Fri, 7 Jul 2017, Paul Durrant wrote:
> > -----Original Message-----
> > From: Stefano Stabellini [mailto:sstabellini@xxxxxxxxxx]
> > Sent: 22 June 2017 23:15
> > To: Paul Durrant <Paul.Durrant@xxxxxxxxxx>
> > Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx; qemu-devel@xxxxxxxxxx; qemu-
> > block@xxxxxxxxxx; Stefano Stabellini <sstabellini@xxxxxxxxxx>; Anthony
> > Perard <anthony.perard@xxxxxxxxxx>; Kevin Wolf <kwolf@xxxxxxxxxx>;
> > Max Reitz <mreitz@xxxxxxxxxx>; afaerber@xxxxxxx
> > Subject: Re: [PATCH v2 3/3] xen-disk: use an IOThread per instance
> > 
> > CC'ing Andreas Färber. Could you please give a quick look below at the
> > way the iothread object is instantiate and destroyed? I am no object
> > model expert and would appreaciate a second opinion.
> > 
> 
> I have not seen any response so far.
> 
> > 
> > On Wed, 21 Jun 2017, Paul Durrant wrote:
> > > This patch allocates an IOThread object for each xen_disk instance and
> > > sets the AIO context appropriately on connect. This allows processing
> > > of I/O to proceed in parallel.
> > >
> > > The patch also adds tracepoints into xen_disk to make it possible to
> > > follow the state transtions of an instance in the log.
> > >
> > > Signed-off-by: Paul Durrant <paul.durrant@xxxxxxxxxx>
> > > ---
> > > Cc: Stefano Stabellini <sstabellini@xxxxxxxxxx>
> > > Cc: Anthony Perard <anthony.perard@xxxxxxxxxx>
> > > Cc: Kevin Wolf <kwolf@xxxxxxxxxx>
> > > Cc: Max Reitz <mreitz@xxxxxxxxxx>
> > >
> > > v2:
> > >  - explicitly acquire and release AIO context in qemu_aio_complete() and
> > >    blk_bh()
> > > ---
> > >  hw/block/trace-events |  7 ++++++
> > >  hw/block/xen_disk.c   | 69
> > ++++++++++++++++++++++++++++++++++++++++++++-------
> > >  2 files changed, 67 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/hw/block/trace-events b/hw/block/trace-events
> > > index 65e83dc258..608b24ba66 100644
> > > --- a/hw/block/trace-events
> > > +++ b/hw/block/trace-events
> > > @@ -10,3 +10,10 @@ virtio_blk_submit_multireq(void *mrb, int start, int
> > num_reqs, uint64_t offset,
> > >  # hw/block/hd-geometry.c
> > >  hd_geometry_lchs_guess(void *blk, int cyls, int heads, int secs) "blk %p
> > LCHS %d %d %d"
> > >  hd_geometry_guess(void *blk, uint32_t cyls, uint32_t heads, uint32_t
> > secs, int trans) "blk %p CHS %u %u %u trans %d"
> > > +
> > > +# hw/block/xen_disk.c
> > > +xen_disk_alloc(char *name) "%s"
> > > +xen_disk_init(char *name) "%s"
> > > +xen_disk_connect(char *name) "%s"
> > > +xen_disk_disconnect(char *name) "%s"
> > > +xen_disk_free(char *name) "%s"
> > > diff --git a/hw/block/xen_disk.c b/hw/block/xen_disk.c
> > > index 0e6513708e..8548195195 100644
> > > --- a/hw/block/xen_disk.c
> > > +++ b/hw/block/xen_disk.c
> > > @@ -27,10 +27,13 @@
> > >  #include "hw/xen/xen_backend.h"
> > >  #include "xen_blkif.h"
> > >  #include "sysemu/blockdev.h"
> > > +#include "sysemu/iothread.h"
> > >  #include "sysemu/block-backend.h"
> > >  #include "qapi/error.h"
> > >  #include "qapi/qmp/qdict.h"
> > >  #include "qapi/qmp/qstring.h"
> > > +#include "qom/object_interfaces.h"
> > > +#include "trace.h"
> > >
> > >  /* ------------------------------------------------------------- */
> > >
> > > @@ -128,6 +131,9 @@ struct XenBlkDev {
> > >      DriveInfo           *dinfo;
> > >      BlockBackend        *blk;
> > >      QEMUBH              *bh;
> > > +
> > > +    IOThread            *iothread;
> > > +    AioContext          *ctx;
> > >  };
> > >
> > >  /* ------------------------------------------------------------- */
> > > @@ -599,9 +605,12 @@ static int ioreq_runio_qemu_aio(struct ioreq
> > *ioreq);
> > >  static void qemu_aio_complete(void *opaque, int ret)
> > >  {
> > >      struct ioreq *ioreq = opaque;
> > > +    struct XenBlkDev *blkdev = ioreq->blkdev;
> > > +
> > > +    aio_context_acquire(blkdev->ctx);
> > 
> > I think that Paolo was right that we need a aio_context_acquire here,
> > however the issue is that with the current code:
> > 
> >   blk_handle_requests -> ioreq_runio_qemu_aio -> qemu_aio_complete
> > 
> > leading to aio_context_acquire being called twice on the same lock,
> > which I don't think is allowed?
> 
> It resolves to a qemu_rec_mutex_lock() which I believed is a recursive lock, 
> so I think that's ok.

On Linux it becomes pthread_mutex_lock. The lock is created by
qemu_rec_mutex_init which specifies PTHREAD_MUTEX_RECURSIVE, so yes, it
should be recursive. Good.


> > 
> > I think we need to get rid of the qemu_aio_complete call from
> > ioreq_runio_qemu_aio, but to do that we need to be careful with the
> > accounting of aio_inflight (today it's incremented unconditionally at
> > the beginning of ioreq_runio_qemu_aio, I think we would have to change
> > that to increment it only if presync).
> > 
> 
> If the lock is indeed recursive then I think we can avoid this complication.

OK


> > 
> > >      if (ret != 0) {
> > > -        xen_pv_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
> > > +        xen_pv_printf(&blkdev->xendev, 0, "%s I/O error\n",
> > >                        ioreq->req.operation == BLKIF_OP_READ ? "read" : 
> > > "write");
> > >          ioreq->aio_errors++;
> > >      }
> > > @@ -610,13 +619,13 @@ static void qemu_aio_complete(void *opaque, int
> > ret)
> > >      if (ioreq->presync) {
> > >          ioreq->presync = 0;
> > >          ioreq_runio_qemu_aio(ioreq);
> > > -        return;
> > > +        goto done;
> > >      }
> > >      if (ioreq->aio_inflight > 0) {
> > > -        return;
> > > +        goto done;
> > >      }
> > >
> > > -    if (ioreq->blkdev->feature_grant_copy) {
> > > +    if (blkdev->feature_grant_copy) {
> > >          switch (ioreq->req.operation) {
> > >          case BLKIF_OP_READ:
> > >              /* in case of failure ioreq->aio_errors is increased */
> > > @@ -638,7 +647,7 @@ static void qemu_aio_complete(void *opaque, int
> > ret)
> > >      }
> > >
> > >      ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR :
> > BLKIF_RSP_OKAY;
> > > -    if (!ioreq->blkdev->feature_grant_copy) {
> > > +    if (!blkdev->feature_grant_copy) {
> > >          ioreq_unmap(ioreq);
> > >      }
> > >      ioreq_finish(ioreq);
> > > @@ -650,16 +659,19 @@ static void qemu_aio_complete(void *opaque, int
> > ret)
> > >          }
> > >      case BLKIF_OP_READ:
> > >          if (ioreq->status == BLKIF_RSP_OKAY) {
> > > -            block_acct_done(blk_get_stats(ioreq->blkdev->blk), 
> > > &ioreq->acct);
> > > +            block_acct_done(blk_get_stats(blkdev->blk), &ioreq->acct);
> > >          } else {
> > > -            block_acct_failed(blk_get_stats(ioreq->blkdev->blk), 
> > > &ioreq->acct);
> > > +            block_acct_failed(blk_get_stats(blkdev->blk), &ioreq->acct);
> > >          }
> > >          break;
> > >      case BLKIF_OP_DISCARD:
> > >      default:
> > >          break;
> > >      }
> > > -    qemu_bh_schedule(ioreq->blkdev->bh);
> > > +    qemu_bh_schedule(blkdev->bh);
> > > +
> > > +done:
> > > +    aio_context_release(blkdev->ctx);
> > >  }
> > >
> > >  static bool blk_split_discard(struct ioreq *ioreq, blkif_sector_t
> > sector_number,
> > > @@ -917,17 +929,40 @@ static void blk_handle_requests(struct XenBlkDev
> > *blkdev)
> > >  static void blk_bh(void *opaque)
> > >  {
> > >      struct XenBlkDev *blkdev = opaque;
> > > +
> > > +    aio_context_acquire(blkdev->ctx);
> > >      blk_handle_requests(blkdev);
> > > +    aio_context_release(blkdev->ctx);
> > >  }
> > >
> > >  static void blk_alloc(struct XenDevice *xendev)
> > >  {
> > >      struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev,
> > xendev);
> > > +    Object *obj;
> > > +    char *name;
> > > +    Error *err = NULL;
> > > +
> > > +    trace_xen_disk_alloc(xendev->name);
> > >
> > >      QLIST_INIT(&blkdev->inflight);
> > >      QLIST_INIT(&blkdev->finished);
> > >      QLIST_INIT(&blkdev->freelist);
> > > -    blkdev->bh = qemu_bh_new(blk_bh, blkdev);
> > > +
> > > +    obj = object_new(TYPE_IOTHREAD);
> > > +    name = g_strdup_printf("iothread-%s", xendev->name);
> > > +
> > > +    object_property_add_child(object_get_objects_root(), name, obj,
> > &err);
> > > +    assert(!err);
> > 
> > Would it be enough to call object_ref?
> > 
> 
> You mean to avoid the assert? I guess so but I think any failure here would 
> be indicative of a larger problem.

No, I meant calling object_ref instead of object_property_add_child.


> > 
> > > +    g_free(name);
> > > +
> > > +    user_creatable_complete(obj, &err);
> > 
> > Why do we need to call this?
> > 
> 
> I'm not entirely sure but looking around the object code it seemed to be a 
> necessary part of instantiation. Maybe it is not required for iothread 
> objects, but I could not figure that out from looking at the code and 
> comments in the header suggest it is harmless if it is not required.
>
> > > +    assert(!err);
> > > +
> > > +    blkdev->iothread = (IOThread *)object_dynamic_cast(obj,
> > TYPE_IOTHREAD);
> > > +    blkdev->ctx = iothread_get_aio_context(blkdev->iothread);
> > > +    blkdev->bh = aio_bh_new(blkdev->ctx, blk_bh, blkdev);
> > > +
> > >      if (xen_mode != XEN_EMULATE) {
> > >          batch_maps = 1;
> > >      }
> > > @@ -1288,6 +1327,8 @@ static int blk_connect(struct XenDevice *xendev)
> > >          blkdev->persistent_gnt_count = 0;
> > >      }
> > >
> > > +    blk_set_aio_context(blkdev->blk, blkdev->ctx);
> > > +
> > >      xen_be_bind_evtchn(&blkdev->xendev);
> > >
> > >      xen_pv_printf(&blkdev->xendev, 1, "ok: proto %s, nr-ring-ref %u, "
> > > @@ -1301,13 +1342,20 @@ static void blk_disconnect(struct XenDevice
> > *xendev)
> > >  {
> > >      struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev,
> > xendev);
> > >
> > > +    trace_xen_disk_disconnect(xendev->name);
> > > +
> > > +    aio_context_acquire(blkdev->ctx);
> > > +
> > >      if (blkdev->blk) {
> > > +        blk_set_aio_context(blkdev->blk, qemu_get_aio_context());
> > >          blk_detach_dev(blkdev->blk, blkdev);
> > >          blk_unref(blkdev->blk);
> > >          blkdev->blk = NULL;
> > >      }
> > >      xen_pv_unbind_evtchn(&blkdev->xendev);
> > >
> > > +    aio_context_release(blkdev->ctx);
> > > +
> > >      if (blkdev->sring) {
> > >          xengnttab_unmap(blkdev->xendev.gnttabdev, blkdev->sring,
> > >                          blkdev->nr_ring_ref);
> > > @@ -1358,6 +1408,7 @@ static int blk_free(struct XenDevice *xendev)
> > >      g_free(blkdev->dev);
> > >      g_free(blkdev->devtype);
> > >      qemu_bh_delete(blkdev->bh);
> > > +    object_unparent(OBJECT(blkdev->iothread));
> > 
> > Shouldn't this be object_unref?
> > 
> 
> I don't think so. I think this is required to undo what was done by calling 
> object_property_add_child() on the root object.

Right, so if object_property_add_child is not actually required, then
you might be able to turn object_unparent into object_unref.

Unfortunately I don't know enough about QOM to be able to tell which is
the right way of doing things, but looking at
hw/block/dataplane/virtio-blk.c, it would seem that only object_ref and
object_unref are required?


> Looking at other code such as object_new_with_propv() it looks like the right 
> thing to do is to call object_unref() after calling 
> object_property_add_child() to drop the implicit ref taken by object_new() so 
> I'd need to add the call in blk_alloc().

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel