On Wed, 10 Aug 2005, Andrew Morton wrote:
> > What I posted originally; the current SCSI format for a workqueue:
> > scsi_wq_%d hits the bug after the host number rises to 100, which has
> > been seen by some enterprise person with > 100 HBAs.
> >
> > The reason for this name is that the error handler thread is called
> > scsi_eh_%d; so we could rename all our threads to avoid this, but one
> > day someone will come along with a huge enough machine to hit whatever
> > limit we squeeze it down to.
>
> OK, well scsi is using single-threaded workqueues anyway. So we could do:
>
> if (singlethread)
> BUG_ON(strlen(name) > sizeof(task_struct.comm) - 1);
> else
> BUG_ON(strlen(name) > sizeof(task_struct.comm) - 1 - 4);
>
> which gets you 10,000,000 HBAs. Enough?
I suppose so, but the problem is slightly worse:
One does not need 100 HBAs to trigger the BUG_ON:
It is sufficient to have a few HBAs and to insmod/rmmod the driver a few
times.
Since the host_no is choosen with a mere counter increment
in scsi_host_alloc():
shost->host_no = scsi_host_next_hn++; /* XXX(hch): still racy */
Unused `host_no's are not reused and the 100 limit is reached even on
smaller systems.
I have no idea of why someone would do repeated insmod/rmmods, though.
(But someone did).
Simon.
-
: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
