Today disconnecting xen-blkback is broken in case there are still
I/Os in flight: xen_blkif_disconnect() will bail out early without
releasing all resources in the hope it will be called again when
the last request has terminated. This, however, won't happen as
xen_blkif_free() won't be called on termination of the last running
request: xen_blkif_put() won't decrement the blkif refcnt to 0 as
xen_blkif_disconnect() didn't finish before thus some xen_blkif_put()
calls in xen_blkif_disconnect() didn't happen.
To solve this deadlock xen_blkif_disconnect() and
xen_blkif_alloc_rings() shouldn't use xen_blkif_put() and
xen_blkif_get() but use some other way to do their accounting of
resources.
This at once fixes another error in xen_blkif_disconnect(): when it
returned early with -EBUSY for another ring than 0 it would call
xen_blkif_put() again for already handled rings on a subsequent call.
This will lead to inconsistencies in the refcnt handling.
Cc: [email protected]
Signed-off-by: Juergen Gross <[email protected]>
Tested-by: Steven Haigh <[email protected]>
---
drivers/block/xen-blkback/common.h | 1 +
drivers/block/xen-blkback/xenbus.c | 7 +++++--
2 files changed, 6 insertions(+), 2 deletions(-)
diff --git a/drivers/block/xen-blkback/common.h b/drivers/block/xen-blkback/common.h
index dea61f6ab8cb..953f38802333 100644
--- a/drivers/block/xen-blkback/common.h
+++ b/drivers/block/xen-blkback/common.h
@@ -281,6 +281,7 @@ struct xen_blkif_ring {
wait_queue_head_t wq;
atomic_t inflight;
+ int active;
/* One thread per blkif ring. */
struct task_struct *xenblkd;
unsigned int waiting_reqs;
diff --git a/drivers/block/xen-blkback/xenbus.c b/drivers/block/xen-blkback/xenbus.c
index 8fe61b5dc5a6..411d2ded2456 100644
--- a/drivers/block/xen-blkback/xenbus.c
+++ b/drivers/block/xen-blkback/xenbus.c
@@ -159,7 +159,7 @@ static int xen_blkif_alloc_rings(struct xen_blkif *blkif)
init_waitqueue_head(&ring->shutdown_wq);
ring->blkif = blkif;
ring->st_print = jiffies;
- xen_blkif_get(blkif);
+ ring->active = 1;
}
return 0;
@@ -249,6 +249,9 @@ static int xen_blkif_disconnect(struct xen_blkif *blkif)
struct xen_blkif_ring *ring = &blkif->rings[r];
unsigned int i = 0;
+ if (!ring->active)
+ continue;
+
if (ring->xenblkd) {
kthread_stop(ring->xenblkd);
wake_up(&ring->shutdown_wq);
@@ -296,7 +299,7 @@ static int xen_blkif_disconnect(struct xen_blkif *blkif)
BUG_ON(ring->free_pages_num != 0);
BUG_ON(ring->persistent_gnt_c != 0);
WARN_ON(i != (XEN_BLKIF_REQS_PER_PAGE * blkif->nr_ring_pages));
- xen_blkif_put(blkif);
+ ring->active = 0;
}
blkif->nr_ring_pages = 0;
/*
--
2.12.0
On Fri, Apr 21, 2017 at 10:51:00AM +0200, Juergen Gross wrote:
> Today disconnecting xen-blkback is broken in case there are still
> I/Os in flight: xen_blkif_disconnect() will bail out early without
> releasing all resources in the hope it will be called again when
> the last request has terminated. This, however, won't happen as
> xen_blkif_free() won't be called on termination of the last running
> request: xen_blkif_put() won't decrement the blkif refcnt to 0 as
> xen_blkif_disconnect() didn't finish before thus some xen_blkif_put()
> calls in xen_blkif_disconnect() didn't happen.
>
> To solve this deadlock xen_blkif_disconnect() and
> xen_blkif_alloc_rings() shouldn't use xen_blkif_put() and
> xen_blkif_get() but use some other way to do their accounting of
> resources.
>
> This at once fixes another error in xen_blkif_disconnect(): when it
> returned early with -EBUSY for another ring than 0 it would call
> xen_blkif_put() again for already handled rings on a subsequent call.
> This will lead to inconsistencies in the refcnt handling.
>
> Cc: [email protected]
> Signed-off-by: Juergen Gross <[email protected]>
> Tested-by: Steven Haigh <[email protected]>
Acked-by: Roger Pau Monn? <[email protected]>
TBH, it seems like all this accounting in blkback could be improved quite a
lot. Right now this is just a complete caos.
Thanks, Roger.