scsi: Do not rely on blk-mq for double completions
authorKeith Busch <keith.busch@intel.com>
Mon, 26 Nov 2018 16:54:29 +0000 (09:54 -0700)
committerJens Axboe <axboe@kernel.dk>
Mon, 26 Nov 2018 17:34:26 +0000 (10:34 -0700)
The scsi timeout error handling had been directly updating the block
layer's request state to prevent a error handling and a natural completion
from completing the same request twice. Fix this layering violation
by having scsi control the fate of its commands with scsi owned flags
rather than use blk-mq's.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
drivers/scsi/scsi_error.c
drivers/scsi/scsi_lib.c
include/scsi/scsi_cmnd.h

index dd338a8cd2757c84698830acf8fd5e6bd942e527..16eef068e9e93bb35a22abff944b1d37ff3a3b66 100644 (file)
@@ -297,19 +297,19 @@ enum blk_eh_timer_return scsi_times_out(struct request *req)
 
        if (rtn == BLK_EH_DONE) {
                /*
-                * For blk-mq, we must set the request state to complete now
-                * before sending the request to the scsi error handler. This
-                * will prevent a use-after-free in the event the LLD manages
-                * to complete the request before the error handler finishes
-                * processing this timed out request.
+                * Set the command to complete first in order to prevent a real
+                * completion from releasing the command while error handling
+                * is using it. If the command was already completed, then the
+                * lower level driver beat the timeout handler, and it is safe
+                * to return without escalating error recovery.
                 *
-                * If the request was already completed, then the LLD beat the
-                * time out handler from transferring the request to the scsi
-                * error handler. In that case we can return immediately as no
-                * further action is required.
+                * If timeout handling lost the race to a real completion, the
+                * block layer may ignore that due to a fake timeout injection,
+                * so return RESET_TIMER to allow error handling another shot
+                * at this command.
                 */
-               if (!blk_mq_mark_complete(req))
-                       return rtn;
+               if (test_and_set_bit(SCMD_STATE_COMPLETE, &scmd->state))
+                       return BLK_EH_RESET_TIMER;
                if (scsi_abort_command(scmd) != SUCCESS) {
                        set_host_byte(scmd, DID_TIME_OUT);
                        scsi_eh_scmd_add(scmd);
index 0df15cb738d2bff71f7685df53cb36eb31ed38bb..0dbf25512778d09f56dd29543d16a765b03d1e96 100644 (file)
@@ -1642,8 +1642,18 @@ static blk_status_t scsi_mq_prep_fn(struct request *req)
 
 static void scsi_mq_done(struct scsi_cmnd *cmd)
 {
+       if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
+               return;
        trace_scsi_dispatch_cmd_done(cmd);
-       blk_mq_complete_request(cmd->request);
+
+       /*
+        * If the block layer didn't complete the request due to a timeout
+        * injection, scsi must clear its internal completed state so that the
+        * timeout handler will see it needs to escalate its own error
+        * recovery.
+        */
+       if (unlikely(!blk_mq_complete_request(cmd->request)))
+               clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
 }
 
 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
@@ -1702,6 +1712,7 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
        if (!scsi_host_queue_ready(q, shost, sdev))
                goto out_dec_target_busy;
 
+       clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
        if (!(req->rq_flags & RQF_DONTPREP)) {
                ret = scsi_mq_prep_fn(req);
                if (ret != BLK_STS_OK)
index d6fd2aba0380d8b6d499e06038c23a448a9224b8..3de905e205ce03f81e90ad3e58d44c2c5769c97e 100644 (file)
@@ -61,6 +61,9 @@ struct scsi_pointer {
 /* flags preserved across unprep / reprep */
 #define SCMD_PRESERVED_FLAGS   (SCMD_UNCHECKED_ISA_DMA | SCMD_INITIALIZED)
 
+/* for scmd->state */
+#define SCMD_STATE_COMPLETE    (1 << 0)
+
 struct scsi_cmnd {
        struct scsi_request req;
        struct scsi_device *device;
@@ -145,6 +148,7 @@ struct scsi_cmnd {
 
        int result;             /* Status code from lower level driver */
        int flags;              /* Command flags */
+       unsigned long state;    /* Command completion state */
 
        unsigned char tag;      /* SCSI-II queued command tag */
 };