On 9/12/05, Thomas Kleffel [off-list ref] wrote:

Hi,

Bartlomiej Zolnierkiewicz wrote:

quoted

quoted

[...]
1b) When a disk is released by disk_release(), its queue's reference
count shall be decremented by calling blk_cleanup_queue().

Which conflicts with IDE layer reference counting  & locking scheme
as IDE layer can send (special) requests to the device without any
device driver loaded.

But the IDE layer won't do that anymore after ide_unregister() was
called, while a device driver doesn't know that the device doesn't exist
anymore, so it may still access the queue until it releases the disk.

The point is that you have added blk_cleanup_queue() to the ide-disk
driver release function and it is perfectly fine for IDE layer to send requests
after unloading ide-disk driver.

quoted

quoted

2a) When a physical drive is released by drive_release_dev(), the
corresponding queue is marked dead, so that no further calls to the
physical device's queue-handler are made.

2b) When a request is submitted to a dead queue using
generic_make_request(), the request shall be failed immedaiately with
-ENXIO which causes the caller to recive a "Bus error". This is the same
beaviour as when a USB-Storage device gets pulled while in use.

The fix would be to fail the previous requests during removal of the
device [ somebody already posted a patch to do that but I didn't have
time to give it proper thought ] or alternatively to add the offline state to
the device [ so processing of the requests would be resumed after device
is online again ].

I don't think it'd be wise to resume request processing on the same
device as the CF Card inserted again might not be the same one as the
user pulled out before. Performing old, cached writes on the new card
could destroy innocent data.

Yes, that is why failing old requests is a preferred solution.

 From your responses I read that the correct solution would be to keep
the old pysical device as long as the ide layer still has references to
it and to fail all requests in the meantime.

Is that correct?

Yes, we should just fail all the requests if the device is not present.
[ What do you mean by the old physical device, old 'ide_drive_t *'? ]

Regards,
Bartlomiej

Best regards,

Thomas