New: SCSI Generic block io queueing can lock up

http://bugzilla.kernel.org/show_bug.cgi?id=15562

           Summary: SCSI Generic block io queueing can lock up
           Product: IO/Storage
           Version: 2.5
    Kernel Version: 2.6.22-2.6.32
          Platform: All
        OS/Version: Linux
              Tree: Mainline
            Status: NEW
          Severity: normal
          Priority: P1
         Component: SCSI
        AssignedTo: [email protected]
        ReportedBy: [email protected]
        Regression: No

When queueing, write() can occassionally return ENOMEM or EBUSY. The SCSI GENERIC HOWTO indicates ENOMEM can be returned for indirect io and that it is extremely rare, however I can typically cause it within an hour even for direct io which shouldn't need to mem copy. The EBUSY return is not even a documented error in these circumstances.

Regardless of which error is received, retrying will never succeed and the fd is wedged at this point. With EBUSY I've noticed several concurrent processes running against different sg block devices to fail simultaneously and never allow a write() to queue a command again.

This happens when there is plenty of swap, only 20% of ram "used ", the rest occupied by buffer cache.

There are no errors logged by the driver. Here is an example of the offending sg_io_hdr, note all values are in hex:

interface_id S
dxfer_direction fffffffd (SG_DXFER_FROM_DEV)

cmd_len         a         (it's a READ 10)
mx_sb_len       fc
iovec_count     0
dxfer_len       200000
dxferp          1c1f400
cmdp            896518
sbp             896528
timeout         20000
flags           1         (SG_FLAG_DIRECT_IO)
pack_id         0
usr_ptr         8964e0

Configure bugmail: http://bugzilla.kernel.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.