On 10/13/2016 06:24 PM, Johannes Thumshirn wrote:
On Thu, Oct 13, 2016 at 05:15:25PM +0200, Steffen Maier wrote:
I'm puzzled.

$ git bisect start fc_bsg master

3087864ce3d7282f59021245d8a5f83ef1caef18 is the first bad commit
commit 3087864ce3d7282f59021245d8a5f83ef1caef18
Author: Johannes Thumshirn <jthumsh...@suse.de>
Date:   Wed Oct 12 15:06:28 2016 +0200

    scsi: don't use fc_bsg_job::request and fc_bsg_job::reply directly

    Don't use fc_bsg_job::request and fc_bsg_job::reply directly, but use
    helper variables bsg_request and bsg_reply. This will be helpfull  when
    transitioning to bsg-lib.

    Signed-off-by: Johannes Thumshirn <jthumsh...@suse.de>

:040000 040000 140c4b6829d5cfaec4079716e0795f63f8bc3bd2 
0d9fe225615679550be91fbd9f84c09ab1e280fc M      drivers

From there (on the reverse bisect path) I get the following Oops,
except for the full patch set having another stack trace as in my previous
mail (dying in zfcp code).


[...]


@@ -3937,6 +3944,7 @@ fc_bsg_request_handler(struct request_queue *q, struct 
Scsi_Host *shost,
        struct request *req;
        struct fc_bsg_job *job;
        enum fc_dispatch_result ret;
+       struct fc_bsg_reply *bsg_reply;

        if (!get_device(dev))
                return;
@@ -3973,8 +3981,9 @@ fc_bsg_request_handler(struct request_queue *q, struct 
Scsi_Host *shost,
                /* check if we have the msgcode value at least */
                if (job->request_len < sizeof(uint32_t)) {
                        BUG_ON(job->reply_len < sizeof(uint32_t));
-                       job->reply->reply_payload_rcv_len = 0;
-                       job->reply->result = -ENOMSG;
+                       bsg_reply = job->reply;
+                       bsg_reply->reply_payload_rcv_len = 0;
+                       bsg_reply->result = -ENOMSG;

Compiler optimization re-ordered above two lines and the first pointer derefence is bsg_reply->result [field offset 0] where bsg_reply is NULL. The assignment tries to write to memory at address NULL causing the kernel page fault.

Does your suggested change for [PATCH v3 02/16], shuffling the job->request_len checks, address above kernel page fault?

                        job->reply_len = sizeof(uint32_t);
                        fc_bsg_jobdone(job);
                        spin_lock_irq(q->queue_lock);


Ahm and what exactly can break here? It's just assigning variables. Now
I'm puzzled too.


--
Mit freundlichen Grüßen / Kind regards
Steffen Maier

Linux on z Systems Development

IBM Deutschland Research & Development GmbH
Vorsitzende des Aufsichtsrats: Martina Koederitz
Geschaeftsfuehrung: Dirk Wittkopp
Sitz der Gesellschaft: Boeblingen
Registergericht: Amtsgericht Stuttgart, HRB 243294

Reply via email to