When submit async direct-io write operation in function
do_blockdev_direct_IO, 'struct dio' records the info of all bios,
initial value of dio->refcount is set to 1, 'dio->refcount++' is
executed in dio_bio_submit when submit one bio, 'dio->refcount--'
is executed in bio completion handler dio_bio_end_aio.
In do_blockdev_direct_IO, it also calls drop_refcount to do
'dio->refcount--', then judge if dio->refcount is 0, if yes, it
will call dio_complete to complete the dio:
if (drop_refcount(dio) == 0) {
retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
} else
dio_bio_end_aio and drop_refcount will race to judge if dio->refcount
is 0:
1, if dio_bio_end_aio finds dio->refcount is 0, it will queue work if
defer_completion is set, work handler
dio_aio_complete_work->dio_complete will be called:
dio_complete(dio, 0,
DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
if defer_completion not set, it will call:
dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
In above two cases, because DIO_COMPLETE_ASYNC is passed to
dio_complete. So in dio_complete, it will call aio completion handler:
dio->iocb->ki_complete(dio->iocb, ret, 0);
As ki_complete is set to aio_complete for async io, which will fill
an event to ring buffer, then user can use io_getevents to get this
event.
2, if drop_refcount finds dio->refcount is 0, it will call:
dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
As no DIO_COMPLETE_ASYNC is passed to dio_complete. So in dio_complete,
ki_complete(aio_complete) will not be called. Eventually, no one fills
the completion event to ring buffer, so user can't get the completion
event via io_getevents.
Currently, we doesn't meet above issue with existing kernel code,
I think because do_blockdev_direct_IO is called in bio submission path,
it will be quickly completed before all aync bios completion in almost
all cases, so when drop_refcounng is executing, it finds dio->refcount is
not 0 after 'dio->refcount--'. But when the last bio completed,
dio_bio_end_aio will be called, which will find dio->refcount is 0,
then below code will be executed and the async events ring buffer getting
to be filled:
dio_complete(dio, 0, DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE);
or
dio_complete(dio, 0, DIO_COMPLETE_ASYNC);
Make the code logically with this patch and cover above scenario.
Signed-off-by: Xianting Tian <[email protected]>
---
fs/direct-io.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index 1543b5a..552459f 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -1345,7 +1345,9 @@ static inline int drop_refcount(struct dio *dio)
dio_await_completion(dio);
if (drop_refcount(dio) == 0) {
- retval = dio_complete(dio, retval, DIO_COMPLETE_INVALIDATE);
+ retval = dio_complete(dio, retval, dio->is_async ?
+ DIO_COMPLETE_ASYNC | DIO_COMPLETE_INVALIDATE :
+ DIO_COMPLETE_INVALIDATE);
} else
BUG_ON(retval != -EIOCBQUEUED);
--
1.8.3.1