When under memory-pressure it is possible that the mempool which backs
the 'struct request_queue' will make use of up to BLKDEV_MIN_RQ count
emergency buffers - in case it can't get a regular allocation. These
buffers are preallocated and once they are also used, they are
re-supplied with old finished requests from the same request_queue (see
mempool_free()).
The bug is, when re-supplying the emergency pool, the old requests are
not again ran through the callback mempool_t->alloc(), and thus also not
through the callback bsg_init_rq(). Thus we skip initialization, and
while the sense-buffer still should be good, scsi_request->cmd might
have become to be an invalid pointer in the meantime. When the request
is initialized in bsg.c, and the user's CDB is larger than BLK_MAX_CDB,
bsg will replace it with a custom allocated buffer, which is freed when
the user's command is finished, thus it dangles afterwards. When next a
command is sent by the user that has a smaller/similar CDB as
BLK_MAX_CDB, bsg will assume that scsi_request->cmd is backed by
scsi_request->__cmd, will not make a custom allocation, and write into
undefined memory.
Fix this by splitting bsg_init_rq() into two functions:
- bsg_init_job() directly replace bsg_init_rq() and only does the
allocation of the sense-buffer, which is used to back the bsg job's
reply buffer. This pointer should never change during the lifetime of
a scsi_request, so it doesn't need re-initialization.
- bsg_init_rq() is a new function that make use of
'struct request_queue's initialize_rq_fn callback (which was
introduced in v4.12). This is always called before the request is
given out via blk_get_request(). This function does the remaining
initialization that was previously done in bsg_init_rq(), and will
also do it when the request is taken from the emergency-pool of the
backing mempool.
Also rename bsg_exit_rq() into bsg_exit_job(), to make it fit the
name-scheme.
Fixes: 50b4d485528d ("bsg-lib: fix kernel panic resulting from missing
allocation of reply-buffer")
Cc: <sta...@vger.kernel.org> # 4.11+
Signed-off-by: Benjamin Block <bbl...@linux.vnet.ibm.com>
---
Notes:
I did test this on zFCP with FC CT commands send via the ioctl() and
write() system-call. That did work fine. But I would very much
appreciate if anyone could run this against an other HBA or even an
other implementer of bsg-lib, such as now SAS, because I have no access
to such hardware here.
This should make no difference to the normal cases - where each request
is allocated via slab - with- or without this patch; if I didn't miss
anything. Only the order is a bit mixed up - the memset is done after
the sense-allocation, so I have to buffer the sense-pointer for that.
But otherwise there is no difference I am aware of, so it should behave
the same (does for me).
I could not reproduce the memory-pressure case here in the lab.. I
don't see any reason why it should work now, but I am open to
suggestions :)
Beste Grüße / Best regards,
- Benjamin Block
block/bsg-lib.c | 41 ++++++++++++++++++++++++++++-------------
1 file changed, 28 insertions(+), 13 deletions(-)
diff --git a/block/bsg-lib.c b/block/bsg-lib.c
index c82408c7cc3c..634d1557da38 100644
--- a/block/bsg-lib.c
+++ b/block/bsg-lib.c
@@ -203,28 +203,42 @@ static void bsg_request_fn(struct request_queue *q)
spin_lock_irq(q->queue_lock);
}
-static int bsg_init_rq(struct request_queue *q, struct request *req, gfp_t gfp)
+static int bsg_init_job(struct request_queue *q, struct request *req, gfp_t
gfp)
{
struct bsg_job *job = blk_mq_rq_to_pdu(req);
struct scsi_request *sreq = &job->sreq;
- memset(job, 0, sizeof(*job));
+ /* called right after the request is allocated for the request_queue */
- scsi_req_init(sreq);
- sreq->sense_len = SCSI_SENSE_BUFFERSIZE;
- sreq->sense = kzalloc(sreq->sense_len, gfp);
+ sreq->sense = kzalloc(SCSI_SENSE_BUFFERSIZE, gfp);
if (!sreq->sense)
return -ENOMEM;
- job->req = req;
- job->reply = sreq->sense;
- job->reply_len = sreq->sense_len;
- job->dd_data = job + 1;
-
return 0;
}
-static void bsg_exit_rq(struct request_queue *q, struct request *req)
+static void bsg_init_rq(struct request *req)
+{
+ struct bsg_job *job = blk_mq_rq_to_pdu(req);
+ struct scsi_request *sreq = &job->sreq;
+ void *sense = sreq->sense;
+
+ /* called right before the request is given to the request_queue user */
+
+ memset(job, 0, sizeof(*job));
+
+ scsi_req_init(sreq);
+
+ sreq->sense = sense;
+ sreq->sense_len = SCSI_SENSE_BUFFERSIZE;
+
+ job->req = req;
+ job->reply = sense;
+ job->reply_len = sreq->sense_len;
+ job->dd_data = job + 1;
+}
+
+static void bsg_exit_job(struct request_queue *q, struct request *req)
{
struct bsg_job *job = blk_mq_rq_to_pdu(req);
struct scsi_request *sreq = &job->sreq;
@@ -250,8 +264,9 @@ struct request_queue *bsg_setup_queue(struct device *dev,
const char *name,
if (!q)
return ERR_PTR(-ENOMEM);
q->cmd_size = sizeof(struct bsg_job) + dd_job_size;
- q->init_rq_fn = bsg_init_rq;
- q->exit_rq_fn = bsg_exit_rq;
+ q->init_rq_fn = bsg_init_job;
+ q->exit_rq_fn = bsg_exit_job;
+ q->initialize_rq_fn = bsg_init_rq;
q->request_fn = bsg_request_fn;
ret = blk_init_allocated_queue(q);
--
2.14.1
