My TODO list for SCSI

1. Finish converting to use Scsi_Request. There are a few places (generics) where the Scsi_Cmnd structures are still used, and there is one driver (gdth) still using the Cmnd. Once this is completed, there is some dead code to be removed.
2. Clean up select queue depths interface - make this per-device, not per bus. This solves the issue of add-single-device screwing up other devices.
3. Improve support for shared SCSI busses. Some code went into 2.2.nn to better support reservation conflicts, this needs to be reviewed and brought forward. Someone suggested a "defensive" error recovery mode to help reduce the possibility of reset wars.
4. Clean up the GHOST stuff in the bus scanning code. The current implementation is a bit of a hack. It could be more cleanly solved by changing the device to writable, and then fixing sd.c so that it recognizes TYPE_ROM + writable. Either that or we store the blacklist flags in the device, and tell sd.c to also recognize GHOST devices.
5. Find architecturally satisfactory way to implemenent host blocking. This needs to work well with single_lun as well, as this is an architectural cousin of host blocking.
6. Fix to allow 16-byte commands. Groundwork is essentially laid in the mid-level now. The problem is that some low-level drivers aren't handling this well. Some may have hardcoded assumptions built into the hardware (i.e. in the mailboxes) as to the maximum command length, and for these cases the drivers themselves will need to be modified to reject these commands. If the problems are prevalent enough, we may need to specify a maximum command length in the host template.
7. Add support for smp_safe flag as describe in scsi_eh.html. This will allow low-level drivers to be weaned from io_request_lock. The change to the mid-level is relatively safe/trivial, which will enable the work in low-level drivers.
8. Fix single_lun. A timer should be used so that perhaps 0.5 sec after the last access to a device, control is yielded to the next device with pending commands.
9. Look into and correct the issues related to high-availability.
10. Implement target mode. A number of people have good ideas as to how this should work.
11. Fix overallocation problem of Scsi_Cmnd structures. We probably need one or two SCpnt structures per device, and then a generic pool associated with the host itself which can be assigned to any device. Major problem may well be that we have over-allocated DMA pool. That's actually a lot easier to fix. That's just a thought.
12. Add SMP safety to device add/remove. We are not protecting the linked list insertion/removal with reader-writer locks. This is part of the high-availability work.
13. Clean up architectural issues related to hot-plug/hot-removal of devices.
14. Need to convert all drivers to use new error handling, smp safety.
15. Fix offline condition to make it possible to clear condition and or unmount media.
16. Possibly provide hooks to allow sg driver to issue device reset. Bus reset - ick. I am not entirely wild about this possiblity. Doug has asked for this, I am not wild about it.
17. Try and eliminate DMA pool - keep a small emergency pool on hand, but use normal kernel DMA pool with GFP_ATOMIC. Need to verify that this will not force swapping if the pool is exhausted.
18. Rearrange source tree. Move scsi core up a level, sort of like the way the net tree is set up. Need to divide up low-level drivers in some logical grouping as well. I am not sure what makes the most sense for low-level drivers - some of them drive multiple bus types, I am thinking that chipset manufacturer might be good.
19. Get rid of SR_EXTRA, SD_EXTRA. To do this we need to start by Fix layering violations in ll_rw_blk. Reader-writer locks will work well for this for protecting the datastructures. Once this is done, it will be just a little bit harder than it was for SG, ST.
20. Add support for raw device access to disks. Probably not worthwhile for cdrom. Essentially just character device support. Bypass buffer cache, filesystem, etc.
21. Re-examine the lun scan code. It has been suggested that we do an INQUIRY first, and if this succeds then we scan each and every lun for the associated device to test for an active LUN. Something along those lines.

Things fixed as of 2.3.99

Things fixed as of 2.3.40

1. Fix problem where the scsi error handler thread is counted as active, and thus throws off the load average.
2. Get rid of scsi_init_malloc/free. Replace with kmalloc/__get_free_pages, as appropriate. Trivial replacement, fairly safe.
3. Allow EXTRA stuff to be configurable during kernel config. Trivial change, and a temporary measure until we are able to eliminate them.
4. Fix select_queue_depths - add this to the host template. Tough noogies for people who don't do their host templates the right way.

Things fixed as of 2.3.39

1. Phase out some instances of scsi_do_cmd in favor of scsi_wait_cmd. Get rid of sd_wait_cmd in favor of scsi_wait_cmd (essentially identical).
2. Add generic Scsi_Device that points to the HA itself. Add a single command block. For the moment, don't attach it to any high level drivers - generics would be the only thing which would make sense. Initially this is just a transient device that doesn't appear in the linked lists and is deleted when the user is done with it.

Last updated: 5/15/00.