libceph: always signal completion when done

r_safe_completion is currently, and has always been, signaled only if
on-disk ack was requested.  It's there for fsync and syncfs, which wait
for in-flight writes to flush - all data write requests set ONDISK.

However, the pool perm check code introduced in 4.2 sends a write
request with only ACK set.  An unfortunately timed syncfs can then hang
forever: r_safe_completion won't be signaled because only an unsafe
reply was requested.

We could patch ceph_osdc_sync() to skip !ONDISK write requests, but
that is somewhat incomplete and yet another special case.  Instead,
rename this completion to r_done_completion and always signal it when
the OSD client is done with the request, whether unsafe, safe, or
error.  This is a bit cleaner and helps with the cancellation code.

Reported-by: Yan, Zheng <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

3 files changed, 13 insertions, 16 deletions

diff --git a/fs/ceph/file.c b/fs/ceph/file.c
index 12ce2b5..f633165 100644
--- a/fs/ceph/file.c
+++ b/fs/ceph/file.c
@@ -864,7 +864,7 @@ void ceph_sync_write_wait(struct inode *inode)
 
 		dout("sync_write_wait on tid %llu (until %llu)\n",
 		     req->r_tid, last_tid);
-		wait_for_completion(&req->r_safe_completion);
+		wait_for_completion(&req->r_done_completion);
 		ceph_osdc_put_request(req);
 
 		spin_lock(&ci->i_unsafe_lock);
diff --git a/include/linux/ceph/osd_client.h b/include/linux/ceph/osd_client.h
index a8e6634..03a6653 100644
--- a/include/linux/ceph/osd_client.h
+++ b/include/linux/ceph/osd_client.h
@@ -176,7 +176,7 @@ struct ceph_osd_request {
 	struct kref       r_kref;
 	bool              r_mempool;
 	struct completion r_completion;
-	struct completion r_safe_completion;  /* fsync waiter */
+	struct completion r_done_completion;  /* fsync waiter */
 	ceph_osdc_callback_t r_callback;
 	ceph_osdc_unsafe_callback_t r_unsafe_callback;
 	struct list_head  r_unsafe_item;
diff --git a/net/ceph/osd_client.c b/net/ceph/osd_client.c
index 5d812a2..5a8e867 100644
--- a/net/ceph/osd_client.c
+++ b/net/ceph/osd_client.c
@@ -460,7 +460,7 @@ static void request_init(struct ceph_osd_request *req)
 
 	kref_init(&req->r_kref);
 	init_completion(&req->r_completion);
-	init_completion(&req->r_safe_completion);
+	init_completion(&req->r_done_completion);
 	RB_CLEAR_NODE(&req->r_node);
 	RB_CLEAR_NODE(&req->r_mc_node);
 	INIT_LIST_HEAD(&req->r_unsafe_item);
@@ -1772,7 +1772,7 @@ static void complete_request(struct ceph_osd_request *req, int err)
 	req->r_result = err;
 	__finish_request(req);
 	__complete_request(req);
-	complete_all(&req->r_safe_completion);
+	complete_all(&req->r_done_completion);
 	ceph_osdc_put_request(req);
 }
 
@@ -1797,7 +1797,9 @@ static void cancel_request(struct ceph_osd_request *req)
 	dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
 
 	cancel_map_check(req);
-	finish_request(req);
+	__finish_request(req);
+	complete_all(&req->r_done_completion);
+	ceph_osdc_put_request(req);
 }
 
 static void check_pool_dne(struct ceph_osd_request *req)
@@ -2808,12 +2810,12 @@ static bool done_request(const struct ceph_osd_request *req,
  * ->r_unsafe_callback is set?	yes			no
  *
  * first reply is OK (needed	r_cb/r_completion,	r_cb/r_completion,
- * any or needed/got safe)	r_safe_completion	r_safe_completion
+ * any or needed/got safe)	r_done_completion	r_done_completion
  *
  * first reply is unsafe	r_unsafe_cb(true)	(nothing)
  *
  * when we get the safe reply	r_unsafe_cb(false),	r_cb/r_completion,
- *				r_safe_completion	r_safe_completion
+ *				r_done_completion	r_done_completion
  */
 static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg)
 {
@@ -2934,8 +2936,7 @@ static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg)
 			dout("req %p tid %llu cb\n", req, req->r_tid);
 			__complete_request(req);
 		}
-		if (m.flags & CEPH_OSD_FLAG_ONDISK)
-			complete_all(&req->r_safe_completion);
+		complete_all(&req->r_done_completion);
 		ceph_osdc_put_request(req);
 	} else {
 		if (req->r_unsafe_callback) {
@@ -3471,9 +3472,8 @@ int ceph_osdc_start_request(struct ceph_osd_client *osdc,
 EXPORT_SYMBOL(ceph_osdc_start_request);
 
 /*
- * Unregister a registered request.  The request is not completed (i.e.
- * no callbacks or wakeups) - higher layers are supposed to know what
- * they are canceling.
+ * Unregister a registered request.  The request is not completed:
+ * ->r_result isn't set and __complete_request() isn't called.
  */
 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
 {
@@ -3500,9 +3500,6 @@ static int wait_request_timeout(struct ceph_osd_request *req,
 	if (left <= 0) {
 		left = left ?: -ETIMEDOUT;
 		ceph_osdc_cancel_request(req);
-
-		/* kludge - need to to wake ceph_osdc_sync() */
-		complete_all(&req->r_safe_completion);
 	} else {
 		left = req->r_result; /* completed */
 	}
@@ -3549,7 +3546,7 @@ again:
 			up_read(&osdc->lock);
 			dout("%s waiting on req %p tid %llu last_tid %llu\n",
 			     __func__, req, req->r_tid, last_tid);
-			wait_for_completion(&req->r_safe_completion);
+			wait_for_completion(&req->r_done_completion);
 			ceph_osdc_put_request(req);
 			goto again;
 		}