Thanks Jens,

I can your issue of correcting the position of the buffer.  I've tested 
with this patch and do not hit the BUG().

P.

--- linux-2.5.orig/drivers/ide/ide-io.c 2005-01-04 15:45:17.000000000 -0500
+++ linux-2.5/drivers/ide/ide-io.c      2005-01-04 15:45:27.000000000 -0500

@@ -1205,21 +1205,21 @@
        HWGROUP(drive)->rq = NULL;

        rq->errors = 0;

        if (!rq->bio)
                goto out;

        rq->sector = rq->bio->bi_sector;
        rq->current_nr_sectors = bio_iovec(rq->bio)->bv_len >> 9;
        rq->hard_cur_sectors = rq->current_nr_sectors;
-       rq->buffer = NULL;
+       rq->buffer = bio_data(rq->bio);
 out:
        return ret;
 }

 /**
  *     ide_timer_expiry        -       handle lack of an IDE interrupt
  *     @data: timer callback magic (hwgroup)
  *
  *     An IDE command has timed out before the expected drive return
  *     occurred. At this point we attempt to clean up the current

Jens Axboe wrote:

On Tue, Jan 04 2005, Prarit Bhargava wrote:
 

quoted

Hello,

I have found an IDE bug in the IDE DMA timeout function,
ide-io.c: ide_dma_timeout_retry erroneously sets the first_rq->buffer = 
NULL.

ide_dma_timeout_retry will be called whenever a command is issued, times 
out,
and the drive is waiting for DMA.  The function, ide_dma_timeout_retry,
un-busies the hardware group and attempts to clean up the current request.
As part of this cleanup the current failed first_rq->buffer is set to NULL.

However, as part of this retry process first_rq is retried up to 3 times in
PIO mode (with DMA off).

During the retry, ide-cd.c: cdrom_start_read is called, which in turn
calls, restore_request.  restore_request references first_rq->buffer 
(which is
NULL) in order to calculate hard_cur_sectors, hard_nr_sectors, etc.

ie) All of these values will be bogus because of the first_rq->buffer = 
NULL.

This request will fail and the IDE core will enter error handling.  IDE
core generates a new request, sense_rq, in order to request sense.  Attached
this request is a back pointer to the original first_rq request.

      ie) sense_rq->buffer = first_rq

Eventually ide-cd.c:cdrom_end_request is called on sense_rq, and
then ide-io.c:ide_end_dequeued_request is called on first_rq.  Note
that ide_end_dequeued_request is called with the bogus values from first_rq.

The return value essentially depends on the return value of
ll_rw_blk.c:__end_that_request_first.  The arguements to this function 
include
nr_sectors, which as noted above, is bogus.

This leads to a return of 1 from ll_rw_blk.c:__end_that_request_first
which eventually leads to an erroneous call to BUG() in
ide-cd.c:cdrom_end_request.

I have forced this issue to occur by modifying code to effectively DMA
timeout on CDROM accesses on i686 and ia64 platforms.  I hit the bug
100% of the time.

It appears that the modification should be to rid the ide-io.c code of
the rq->buffer = NULL call.

Patch is based off of latest BK linux-2.5 as of 2005-01-04 09:00.

--- linux-2.5.orig/drivers/ide/ide-io.c 2005-01-04 09:31:45.000000000 -0500
+++ linux-2.5/drivers/ide/ide-io.c      2005-01-04 09:32:23.000000000 -0500

@@ -1205,21 +1205,20 @@
      HWGROUP(drive)->rq = NULL;

      rq->errors = 0;

      if (!rq->bio)
              goto out;

      rq->sector = rq->bio->bi_sector;
      rq->current_nr_sectors = bio_iovec(rq->bio)->bv_len >> 9;
      rq->hard_cur_sectors = rq->current_nr_sectors;
-       rq->buffer = NULL;
   

Probably safer to do

rq->buffer = bio_data(rq->bio);