Re: Newbie: ENBD on kernel 2.6 - kernel crash!

"Also sprach Roland Paterson-Jones:"
> My gcc is 3.4.3:

Sounds way too risky to use to me, but then if you have compiled the
kernel with it before, so be it! Make sure that you compile at -O2 or
-O1.

> The enbd.ko module seems to be crashing in the call to 
> blk_queue_max_sectors()

Yes that' reasonable. It's in the first loop:

        for (i = 0; i < MAX_NBD; i++) {
                struct enbd_device *lo = &enbd_dev[i];
               /* Pavel says ...
                *  The new linux 2.5 block layer implementation requires
                *  every gendisk to have its very own request_queue
                *  struct.  These structs are big so we dynamically
                *  allocate them.
                */
                struct gendisk *disk = alloc_disk(ENBD_MAXCONN);
                memset (lo, 0, sizeof (*lo));
                if (disk) {
                    lo->disk = disk;
                    spin_lock_init(&lo->lock);
                    disk->queue = blk_init_queue(do_enbd_request, &lo->lock);
                    if (!disk->queue) {
                                // .. back out
                                return -ENOMEM;
                    }
                    enbd_init_queue(lo, disk->queue);
                }
        }


> which is the first thing that enbd_init_queue 
> tries to do (it doesn't reach my debug line immediately after this 
> call). Is there maybe some queue stuff different in 2.6.9+ ?

Quite possibly! The call is:

  static void
  enbd_init_queue (struct enbd_device *lo, struct request_queue *queue)
  {

        // PTB - set up kernel queue struct with default methods
        blk_queue_max_sectors (queue, buf_sectors);     /* max per request */
        ...

Which appears to match the usages I see in 2.6.11:

  nbd:/usr/local/src/linux-2.6.11.6/drivers/block% grep blk_queue_max_sectors *.c
  DAC960.c:       blk_queue_max_sectors(RequestQueue,
  Controller->MaxBlocksPerCommand);
cciss.c:        blk_queue_max_sectors(q, 512);
  floppy.c:       blk_queue_max_sectors(floppy_queue, 64);
  ll_rw_blk.c:    blk_queue_max_sectors(q, MAX_SECTORS);
  ll_rw_blk.c: * blk_queue_max_sectors - set max sectors for a request
  for this queue
  ll_rw_blk.c:void blk_queue_max_sectors(request_queue_t *q, unsigned
  short max_sectors)
  ll_rw_blk.c:EXPORT_SYMBOL(blk_queue_max_sectors);
  pktcdvd.c:              blk_queue_max_sectors(q, pd->settings.size);
  pktcdvd.c:      blk_queue_max_sectors(q, PACKET_MAX_SECTORS);
  ps2esdi.c:      blk_queue_max_sectors(ps2esdi_queue, 128);
  ub.c:   blk_queue_max_sectors(q, UB_MAX_SECTORS);
  viodasd.c:      blk_queue_max_sectors(q, VIODASD_MAXSECTORS);
  xd.c:   blk_queue_max_sectors(xd_queue, xd_maxsectors);
  nbd:/usr/local/src/linux-2.6.11.6/drivers/block% 

and buf_sectors is


   static int buf_sectors = ENBD_MAX_SECTORS;

and the latter constant is defined in enbd.h:

   #define ENBD_MAX_SECTORS 512 /* PTB max number of 512B sectors in a buffer */

so that leaves the queue as the thing to chcke. But we know it was nonzero:

                    disk->queue = blk_init_queue(do_enbd_request, &lo->lock);
                    if (!disk->queue) {
                                // .. back out
                                return -ENOMEM;
                    }
                    enbd_init_queue(lo, disk->queue);

                    ...

                    enbd_init_queue (struct enbd_device *lo, struct request_queue *queue)
                    {
                  
                          // PTB - set up kernel queue struct with default methods
                          blk_queue_max_sectors (queue, buf_sectors);     /* max per request */
 






> it_queue...
> May 17 17:16:58 localhost kernel: ENBD #5153[0]: enbd_init_queue     
> RPJ:      e
> nbd_init_queue 1: lo is 823d0fa0, queue is 69a93d18 ...

yes - queue nonzero.

> May 17 17:16:58 localhost kernel: Unable to handle kernel NULL pointer dereference at virtual address 000003fe


But 0x03fe is what the kernel subroutine gets.

> GF  VLI
> May 17 17:16:58 localhost kernel: EFLAGS: 00210212   (2.6.9-1.667)
> May 17 17:16:58 localhost kernel: EIP is at blk_queue_max_sectors+0x2c/0x3d

About 3/4 through that subroutine in ll_rw_blk.c: But the length is
absurdly long. The subroutine is:

 void blk_queue_max_sectors(request_queue_t *q, unsigned short max_sectors)
 {
        if ((max_sectors << 9) < PAGE_CACHE_SIZE) {
                max_sectors = 1 << (PAGE_CACHE_SHIFT - 9);
                printk("%s: set to minimum %d\n", __FUNCTION__, max_sectors);
        }

        q->max_sectors = q->max_hw_sectors = max_sectors;
 }

and I cannot believe it is anywhere but the last assignment. To me it
looks as though it has got the wrong address for q from your call. It
is possible for a compiler mismatch to cause that (passing data in wrong
registers), but it sounds quite unlikely for a  C call! C++ maybe.

Anyway, try a different compiler. One that matches the one used to
compile the kernel.  Are you SURE you can compile a kernel
with your compiler?



Peter