block: fix an oops on BLKPREP_KILL
Subsystems (3)
James Bottomley
maintainerPATCH 1/1 block: fix an oops on BLKPREP_KILL
Doing a bit of torture testing, I ran across a BUG in the block subsystem (at blk-core.c:2048): the test for if the request is queued.
It turns out the trigger was a BLKPREP_KILL coming out of the SCSI prep function. Currently for BLKPREP_KILL requests, we send them straight into __blk_end_request_all() with an error, but they've never been dequeued, so they trip the bug. Fix this by starting requests before killing them.
---
block/blk-core.c +1 -0
@@ -1789,6 +1789,7 @@ struct request *blk_peek_request(struct request_queue *q)break ; }else if (ret ==BLKPREP_KILL ) { rq->cmd_flags |=REQ_QUIET ; +blk_start_request (rq);__blk_end_request_all (rq, -EIO ); }else {printk (KERN_ERR "%s: bad return=%d\n" , __func__, ret);
Quoted 18 lines · James Bottomley
Doing a bit of torture testing, I ran across a BUG in the block subsystem (at blk-core.c:2048): the test for if the request is queued.
It turns out the trigger was a
BLKPREP_KILLcoming out of the SCSI prep function. Currently forBLKPREP_KILLrequests, we send them straight into__blk_end_request_all()with an error, but they've never been dequeued, so they trip the bug. Fix this by starting requests before killing them.Signed-off-by James Bottomley <[email protected]>---
block/blk-core.c +0
-0
block/blk-core.c struct request *blk_peek_request(struct request_queue *q) · line 1795
Quoted 7 lines
block/blk-core.c +1
-0@@ -1789,6 +1789,7 @@ struct request *blk_peek_request(struct request_queue *q)break ; }else if (ret ==BLKPREP_KILL ) { rq->cmd_flags |=REQ_QUIET ; +blk_start_request (rq);__blk_end_request_all (rq, -EIO ); }else {printk (KERN_ERR "%s: bad return=%d\n" , __func__, ret);
Given how illogical that now looks, I think it could do with a comment. I'll add that while applying.
Signature · 1 line
Jens Axboe
James Bottomley
maintainerQuoted 14 lines · Jens Axboe
Quoted 13 lines · James Bottomley
Doing a bit of torture testing, I ran across a BUG in the block subsystem (at blk-core.c:2048): the test for if the request is queued.
It turns out the trigger was a
BLKPREP_KILLcoming out of the SCSI prep function. Currently forBLKPREP_KILLrequests, we send them straight into__blk_end_request_all()with an error, but they've never been dequeued, so they trip the bug. Fix this by starting requests before killing them.Signed-off-by James Bottomley <[email protected]>---
Quoted 15 lines
block/blk-core.c struct request *blk_peek_request(struct request_queue *q) · line 1795
Quoted 12 lines
block/blk-core.c +1
-0@@ -1789,6 +1789,7 @@ struct request *blk_peek_request(struct request_queue *q)break ; }else if (ret ==BLKPREP_KILL ) { rq->cmd_flags |=REQ_QUIET ; +blk_start_request (rq);__blk_end_request_all (rq, -EIO ); }else {printk (KERN_ERR "%s: bad return=%d\n" , __func__, ret);Given how illogical that now looks, I think it could do with a comment. I'll add that while applying.
Um, if it looks illogical, then so is the new everything has to be dequeued before completion requirement ... I agree a comment reminding people why it has to work this way doesn't hurt, but if it looks illogical then there might be something wrong with the rules requiring this to happen ...
James
Quoted 31 lines · James Bottomley
Quoted 14 lines · Jens Axboe
Quoted 13 lines · James Bottomley
Doing a bit of torture testing, I ran across a BUG in the block subsystem (at blk-core.c:2048): the test for if the request is queued.
It turns out the trigger was a
BLKPREP_KILLcoming out of the SCSI prep function. Currently forBLKPREP_KILLrequests, we send them straight into__blk_end_request_all()with an error, but they've never been dequeued, so they trip the bug. Fix this by starting requests before killing them.Signed-off-by James Bottomley <[email protected]>---
Quoted 15 lines
block/blk-core.c struct request *blk_peek_request(struct request_queue *q) · line 1795
Quoted 12 lines
block/blk-core.c +1
-0@@ -1789,6 +1789,7 @@ struct request *blk_peek_request(struct request_queue *q)break ; }else if (ret ==BLKPREP_KILL ) { rq->cmd_flags |=REQ_QUIET ; +blk_start_request (rq);__blk_end_request_all (rq, -EIO ); }else {printk (KERN_ERR "%s: bad return=%d\n" , __func__, ret);Given how illogical that now looks, I think it could do with a comment. I'll add that while applying.
Quoted 5 lines
Um, if it looks illogical, then so is the new everything has to be dequeued before completion requirement ... I agree a comment reminding people why it has to work this way doesn't hurt, but if it looks illogical then there might be something wrong with the rules requiring this to happen ...
It's not a bad idea to be a little extra paranoid with such an API change, we can always relax this restriction later.
Signature · 1 line
Jens Axboe