into mulgrave.(none):/home/jejb/BK/scsi-for-linus-2.5

into mulgrave.(none):/home/jejb/BK/scsi-for-linus-2.5

After way to long of a delay, here's a patch for the queue depths
in the ips driver.  And this patch also shuts up the compiler
warnings.  

Okay, this is the final version of the sd patch to remove sd_dsk_arr.
Instead of keeping a big global array we allocate one structure for
each disk dynamically in sd_attach.  We can use the higher level
private data to access it almost everywhere, for the few other
places we keep a linked list of all disks (these are walked only
in slow-patheß).

This patch also allows to get rid of the sd_init midlayer method and
the CONFIG_SD_EXTRA_DEVS option - sd has no more static limit to
the number of disk except of the number of assigned majors.

Hi James,

Here are the changes that are good outside the other changes :-)

scsi_lib:
	o ->errors is used as the scsi status byte for REQ_BLOCK_PC
	o ->data_len is the residual byte count
	o call __scsi_end_request even for !good_sectors if status is
	  good. This legitimately can happen for REQ_BLOCK_PC commands
	  sent from a user space program, if it gets the command setup
	  wrong (or weird). Right now this will hang that queue.

scsi_merge:
	o set SCpnt->request_bufflen to ->data_len, this is the
	  authoritative io byte count for REQ_BLOCK_PC. Here we deal in
	  bytes and not sectors.

sr + sd:
	o Set transfersize and underlow correctly for REQ_BLOCK_PC

sr_ioctl
	o We want to return -EIO for command failure, not EINVAL. That
 	  is pretty stupid :-)


===== drivers/scsi/scsi_lib.c 1.35 vs edited =====

bring it back in line with sd:
* get rid of typedefs where possible
* tab-align all credits entries
* line-wrap after 80 characters
* use C99-initializers

o  Add sense_len to request, so scsi_ioctl knows how much sense data
   was transferred.

o Add sg_timeout and sg_reserved to queue, we can't have these global...

o And finally kill QUEUE_NR_REQUESTS, it hasn't been used in a while.

Various small bits that make SCSI work well with REQ_BLOCK_PC.

o Use ->errors as the scsi status byte for REQ_BLOCK_PC

o Always call end_io completion, even if 0 sectors, as long as the
  status is good. Otherwise we risk hanging this device if a
  REQ_BLOCK_PC user command didn't specify a transfer size for a command
  that did.

o Remove bouncing checks in scsi_merge for REQ_BLOCK_PC, bio_map_user()
  correctly bounces pages now.

o Decrement req->data_len, it's our residual data count.

o sr/sd: set right transfer and underflow size.

This adds bio_map_user and bio_unmap_user to aid drivers in mapping user
space memory into a bio suitable for block io.

into home.transmeta.com:/home/torvalds/v2.5/linux

into home.transmeta.com:/home/torvalds/v2.5/linux

Sometimes we don't even need a bio->bi_end_io, so make it optional. This
also encourages users to _use_ bio_endio()! I like that, since it means
they don't have to remember to decrement bi_size themselves.

Also clear bi_private in bio_init(), and switch to subsys_initcall().

Make merge_bvec_fn() return number of bytes we can accept at a given
offset, instead of a bool.

o Remove some ide compile warnings, old suspend stuff is not used at
  all, for instance.

o If elv_next_request() returns NULL, remember to clear hwgroup->busy
  if we don't have pending commands. This is important. If a queue
  prep function kills a request, we would before quit with hwgroup busy.
  This essentially froze the hwgroup.

o Don't do own list manipulation in ide_do_drive_cmd(). Use the new and
  great elv_add_request() functions

o Fix race on inspection of request after io completion by bumping the
  request reference count prior to ioscheduler insertion.

o Make ide-floppy understand a REQ_BLOCK_PC eject. More may follow, it's
  ATAPI after all.

o Clear hw before passing to ide_init_hwif_ports(). Fixes oops on
  non-pci controllers.

This adds an end_that_request_chunk() helper, that puts the core
functionality in __end_that_request_first().  This one deals in bytes.
end_that_request_first() does the 512 multiply, and then
end_that_request_chunk() can just use it directly.

This enables ide-cd to do proper dma in sizes that are not sector
aligned.  Some of the most important CD functions (ripping audio dma,
burning audio cds, burning raw cds) _require_ this or we will be in pio.
That stinks.  We simply cannot use pio for these, not on writers much
that are fast!

o Always use list_del_init() on request queuelist, this allows us to
  sanity check the integrity of the request on insertion and removal.
  So we can complain loudly instead of silently corrupting memory.

o Add references to requests. This is cheap, since we dont have to use
  an atomic variable for it (all puts are inside queue lock). We've had
  a bug in IDE for years where we want to inspect request state after
  io completion, but this is not possible to do race free right now.
  REQ_BLOCK_PC and sgio will need this too, for checking io residual
  etc. This is not just a theoretical race, I've seen it happen.

Cleanup the last_merge logic. There are two reasons for clearing
last_merge when we are dealing with integrity, and these are:

o Clear when handing the request to the driver, so we don't merge on
  a started request.

o Clear when a request is taken off the list. This cannot be done from
  the driver (above case would already have been hit), but it can
  happen
  when we merge two requests.

This makes it a lot nicer, it was always peculiar how we cleared in
put_request.

Make it possible for a device to specify the dma alignment restrictions
it has. This will be used by future infrastructure when mapping in user
pages, and allows us to dma to on ATAPI even though user address and
length is not sector size aligned.

Extend q->prep_rq_fn() to return one of three values:

 o BLKPREP_OK: request is good, return it
 o BLKPREP_KILL: request is bad, end it completely
 o BLKPREP_DEFER: request is good, but we can't take it now

We maintain compatability with old prep functions (if any, outside of
ide-cd). This change is needed or SCSI to use prep function for command
init, if sg table allocation fails we can just defer the request.

Add some missing bits, and make it generally a bit more useful outside
of REQ_PC requests.

Request insertion in the current tree is a mess. We have all sorts of
variants of *elv_add_request*, and it's not at all clear who does what
and with what locks (or not). This patch cleans it up to be:

o __elv_add_request(queue, request, at_end, plug)
      Core function, requires queue lock to be held

o elv_add_request(queue, request, at_end, plug)
      Like __elv_add_request(), but grabs queue lock

o __elv_add_request_pos(queue, request, position)
      Insert request at a given location, lock must be held

We must return -EIO if the command fails (the 5/20/00 sense check is
just helping return more sane info), not -EINVAL. Getting -EINVAL return
on an ioctl if a command fails is less than helpful for applications...

o Split blk_queue_bounce() into a slow and fast path. The fast path is
  inlined, only if we actually need to check the bio for possible
  bounces (and bounce) do we enter __blk_queue_bounce() slow path.

o Fix a nasty bug that could cause corruption for file systems not
  using PAGE_CACHE_SIZE blok size! We were not correctly setting the
  'to' bv_offset correctly.

o Add BIO_BOUNCE flag. Later patches will use this for debug checking.

Make deadline_merge() prefetch the next entry on the list.

wake_up_interruptible() and _sync() calls are reversed in pipe_read().

The attached patches only calls _sync if a schedule() call follows.

On Mon, Oct 28, 2002 at 01:50:53AM +0100, Andries Brouwer wrote:
> On Sun, Oct 27, 2002 at 06:05:07PM -0600, James Bottomley wrote:
>
> > If the return type will be ignored by most applications, I don't see
> > what the problem is.
>
> There is no problem. My longish reaction was mostly because you used
> "future proofing", that gave the impression that you did not know
> what this is about.
>
> > (like an obviously wrong truncation)
>
> No, the code I wrote was optimal.
> If you have 16 bits and the value is 70000, I prefer returning
> 65535 over 4464.

Why didn't you write is as patch?  James, this in Andries code
in patch from, it get's rid of the ugly sector_div so it should
be in if not just because of that.

in struct pci_dev.

sector_div has the same slightly strange calling convention do_div has:
it's return value is the modulo of the two operators, the division
result is in the first parameter.  Also optimize one of the expensive
64bit division away (okay, okay - it's not exactly an fast-path :))

Add a pointer to struct scsi_disk instead.  This also obsoletes
sd_dskname().

(suggested by Peter Chubb)

Now that only sd.c includes sd.h it can be safely merged into it.
Also get rid of typedef abuse in sd.c