libosd bug fixing

So libosd has decided to sacrifice some code simplicity for the sake of a clean API. One of these things is the possibility for users to call osd_end_request, in any condition at any state. This opens up some problems with calling blk_put_request when out-side of the completion callback but calling __blk_put_request when detecting a from-completion state.

The current hack was working just fine until exofs decided to operate on all devices in parallel and wait for the sum of the requests, before deallocating all osd-requests at once. There are two new possible cases
1. All request in a group are deallocated as part of the last request's async-done, request_queue is locked.
2. All request in a group where executed asynchronously, but de-allocation was delayed to after the async-done, in the context of another thread. Async execution but request_queue is not locked.

The solution I chose was to separate the deallocation of the osd_request which has the information users need, from the deallocation of the internal(2) requests which impose the locking problem. The internal block-requests are freed unconditionally inside the async-done-callback, when we know the queue is always locked. If at osd_end_request time we still have a bock-request, then we know it did not come from within an async-done-callback and we can call the regular blk_put_request.

The internal requests were used for carrying error information after execution. This information is now copied to osd_request members for later analysis by user code.

The external API and behaviour was unchanged, except now it really supports what was previously advertised.

Reported-by Vineet Agarwal <[email protected]>

Signed-off-by Boaz Harrosh <[email protected]>

---
 drivers/scsi/osd/osd_initiator.c |   88 ++++++++++++++++++++------------------
 include/scsi/osd_initiator.h     |    5 +-
 2 files changed, 50 insertions(+), 43 deletions(-)

drivers/scsi/osd/osd_initiator.c +47 -41

index 950202a..2422347 100644
@@ -432,30 +432,23 @@ static void _osd_free_seg(struct osd_request *or __unused,
 	seg->alloc_size = 0;
 }
 
-static void _put_request(struct request *rq , bool is_async)
+static void _put_request(struct request *rq)
 {
-	if (is_async) {
-		WARN_ON(rq->bio);
-		__blk_put_request(rq->q, rq);
-	} else {
-		/*
-		 * If osd_finalize_request() was called but the request was not
-		 * executed through the block layer, then we must release BIOs.
-		 * TODO: Keep error code in or->async_error. Need to audit all
-		 *       code paths.
-		 */
-		if (unlikely(rq->bio))
-			blk_end_request(rq, -ENOMEM, blk_rq_bytes(rq));
-		else
-			blk_put_request(rq);
-	}
+	/*
+	 * If osd_finalize_request() was called but the request was not
+	 * executed through the block layer, then we must release BIOs.
+	 * TODO: Keep error code in or->async_error. Need to audit all
+	 *       code paths.
+	 */
+	if (unlikely(rq->bio))
+		blk_end_request(rq, -ENOMEM, blk_rq_bytes(rq));
+	else
+		blk_put_request(rq);
 }
 
 void osd_end_request(struct osd_request *or)
 {
 	struct request *rq = or->request;
-	/* IMPORTANT: make sure this agrees with osd_execute_request_async */
-	bool is_async = (or->request->end_io_data == or);
 
 	_osd_free_seg(or, &or->set_attr);
 	_osd_free_seg(or, &or->enc_get_attr);

@@ -463,20 +456,34 @@ void osd_end_request(struct osd_request *or)
 
 	if (rq) {
 		if (rq->next_rq) {
-			_put_request(rq->next_rq, is_async);
+			_put_request(rq->next_rq);
 			rq->next_rq = NULL;
 		}
 
-		_put_request(rq, is_async);
+		_put_request(rq);
 	}
 	_osd_request_free(or);
 }
 EXPORT_SYMBOL(osd_end_request);
 
+static void _set_error_resid(struct osd_request *or, struct request *req,
+			     int error)
+{
+	or->async_error = error;
+	or->req_errors = req->errors ? : error;
+	or->sense_len = req->sense_len;
+	if (or->out.req)
+		or->out.residual = or->out.req->resid_len;
+	if (or->in.req)
+		or->in.residual = or->in.req->resid_len;
+}
+
 int osd_execute_request(struct osd_request *or)
 {
-	return or->async_error =
-			blk_execute_rq(or->request->q, NULL, or->request, 0);
+	int error = blk_execute_rq(or->request->q, NULL, or->request, 0);
+
+	_set_error_resid(or, or->request, error);
+	return error;
 }
 EXPORT_SYMBOL(osd_execute_request);
 

@@ -484,15 +491,17 @@ static void osd_request_async_done(struct request *req, int error)
 {
 	struct osd_request *or = req->end_io_data;
 
-	or->async_error = error;
-
-	if (unlikely(error)) {
-		OSD_DEBUG("osd_request_async_done error recieved %d "
-			  "errors 0x%x\n", error, req->errors);
-		if (!req->errors) /* don't miss out on this one */
-			req->errors = error;
+	_set_error_resid(or, req, error);
+	if (req->next_rq) {
+		__blk_put_request(req->q, req->next_rq);
+		req->next_rq = NULL;
 	}
 
+	__blk_put_request(req->q, req);
+	or->request = NULL;
+	or->in.req = NULL;
+	or->out.req = NULL;
+
 	if (or->async_done)
 		or->async_done(or, or->async_private);
 	else

@@ -1489,21 +1498,18 @@ int osd_req_decode_sense_full(struct osd_request *or,
 #endif
 	int ret;
 
-	if (likely(!or->request->errors)) {
-		osi->out_resid = 0;
-		osi->in_resid = 0;
+	if (likely(!or->req_errors))
 		return 0;
-	}
 
 	osi = osi ? : &local_osi;
 	memset(osi, 0, sizeof(*osi));
 
-	ssdb = or->request->sense;
-	sense_len = or->request->sense_len;
+	ssdb = (typeof(ssdb))or->sense;
+	sense_len = or->sense_len;
 	if ((sense_len < (int)sizeof(*ssdb) || !ssdb->sense_key)) {
 		OSD_ERR("Block-layer returned error(0x%x) but "
 			"sense_len(%u) || key(%d) is empty\n",
-			or->request->errors, sense_len, ssdb->sense_key);
+			or->req_errors, sense_len, ssdb->sense_key);
 		goto analyze;
 	}
 

@@ -1525,7 +1531,7 @@ int osd_req_decode_sense_full(struct osd_request *or,
 			"additional_code=0x%x async_error=%d errors=0x%x\n",
 			osi->key, original_sense_len, sense_len,
 			osi->additional_code, or->async_error,
-			or->request->errors);
+			or->req_errors);
 
 	if (original_sense_len < sense_len)
 		sense_len = original_sense_len;

@@ -1695,10 +1701,10 @@ analyze:
 		ret = -EIO;
 	}
 
-	if (or->out.req)
-		osi->out_resid = or->out.req->resid_len ?: or->out.total_bytes;
-	if (or->in.req)
-		osi->in_resid = or->in.req->resid_len ?: or->in.total_bytes;
+	if (!or->out.residual)
+		or->out.residual = or->out.total_bytes;
+	if (!or->in.residual)
+		or->in.residual = or->in.total_bytes;
 
 	return ret;
 }

include/scsi/osd_initiator.h +3 -2

index 39d6d10..a8f3701 100644
@@ -142,6 +142,7 @@ struct osd_request {
 	struct _osd_io_info {
 		struct bio *bio;
 		u64 total_bytes;
+		u64 residual;
 		struct request *req;
 		struct _osd_req_data_segment *last_seg;
 		u8 *pad_buff;

@@ -150,12 +151,14 @@ struct osd_request {
 	gfp_t alloc_flags;
 	unsigned timeout;
 	unsigned retries;
+	unsigned sense_len;
 	u8 sense[OSD_MAX_SENSE_LEN];
 	enum osd_attributes_mode attributes_mode;
 
 	osd_req_done_fn *async_done;
 	void *async_private;
 	int async_error;
+	int req_errors;
 };
 
 static inline bool osd_req_is_ver1(struct osd_request *or)

@@ -297,8 +300,6 @@ enum osd_err_priority {
 };
 
 struct osd_sense_info {
-	u64 out_resid;		/* Zero on success otherwise out residual */
-	u64 in_resid;		/* Zero on success otherwise in residual */
 	enum osd_err_priority osd_err_pri;
 
 	int key;		/* one of enum scsi_sense_keys */

Quoted 29 lines · Boaz Harrosh

So libosd has decided to sacrifice some code simplicity for the sake of a clean API. One of these things is the possibility for users to call osd_end_request, in any condition at any state. This opens up some problems with calling blk_put_request when out-side of the completion callback but calling __blk_put_request when detecting a from-completion state.

The current hack was working just fine until exofs decided to operate on all devices in parallel and wait for the sum of the requests, before deallocating all osd-requests at once. There are two new possible cases
1. All request in a group are deallocated as part of the last request's async-done, request_queue is locked.
2. All request in a group where executed asynchronously, but de-allocation was delayed to after the async-done, in the context of another thread. Async execution but request_queue is not locked.

The solution I chose was to separate the deallocation of the osd_request which has the information users need, from the deallocation of the internal(2) requests which impose the locking problem. The internal block-requests are freed unconditionally inside the async-done-callback, when we know the queue is always locked. If at osd_end_request time we still have a bock-request, then we know it did not come from within an async-done-callback and we can call the regular blk_put_request.

The internal requests were used for carrying error information after execution. This information is now copied to osd_request members for later analysis by user code.

Quoted 5 lines

The external API and behaviour was unchanged, except now it really supports what was previously advertised.

Reported-by Vineet Agarwal <[email protected]>

Signed-off-by Boaz Harrosh <[email protected]>

James Hi

Linus has locked the 2.6.32 Kernel. I absolutely must have this patch submitted in this merge window. The all of exofs patches depend on it. Please submit ASAP

Boaz

James hi.

Users have found an annoying bug in latest osd/exofs code. Actually this deficiency exists for a long time, but only most recent exofs code has exposed it.

Please include the fix to the patches queued for 2.6.33. No need for it to go into stable as current users will not hit it.

[PATCH 1/1] libosd: Fix blk_put_request locking again

Vineet Please test this new patch, the previous fix was crap

Thanks
Boaz

Hi Boaz,

On Tue, Dec 1, 2009 at 9:00 PM, Boaz Harrosh <[email protected] <mailto:[email protected]>> wrote:

James hi.

Users have found an annoying bug in latest osd/exofs code. Actually this deficiency exists for a long time, but only most recent exofs code has exposed it.

Please include the fix to the patches queued for 2.6.33. No need for it to go into stable as current users will not hit it.

[PATCH 1/1] libosd: Fix blk_put_request locking again

Vineet Please test this new patch, the previous fix was crap

I applied the patch and it's working at my end.

Thanks
Boaz

Regards:
Vineet Agarwal