This patch adds shared library support to the MMU application
loader, binfmt_flat. This is not new, it is a forward port from the
same support in 2.4.x kernels with MMUless support, and has been
running for well over a year now. The code support is conditionally
compiled on CONFIG_BINFMT_FLAT_SHARED. This change also abstracts
a bit more architecture dependent code into the separate flat.h
includes.

Basically relocations within an application also carry a tag to
identify what they refer too (this code or which shared library).
This is patched as before at load/run-time with an appropriate
address.

Unify access_ok for all m68knommu targets. All targets use the
common linker script and have common end symbols. So now we can
just use a simple check.

Remove "%d0" register from clobber list of down_trylock() for
m68knommu. It is not used by the asm code here at all.

Force PAGE_SIZE for the m68knommu architecture to be an unsigned long.
This makes it consistent with all other architectures and cleans up
a load of compiler warnings.

Conditionally copy the ROMfs filesystem on the Motorola M5307C3
target board only if using a ROMfs.

Allow setting boot time parameters at configuration for Motorola
5282 targets.

This improves cold-cache program startup noticeably for me, and
simplifies the read-ahead logic at the same time. The rules for
read-ahead are:

 - if the vma is marked random, we just do the regular one-page case. 
   Obvious.

 - if the vma is marked "linear access", we use the regular readahead
   code. No change in behaviour there (well, we also only consider it a 
   _miss_ if it was marked linear access - the "readahead" and
   "readaround"  things are now totally independent of each other)

 - otherwise, we look at how many hits/misses we've had for this 
   particular file open for mmap, and if we've had noticeably more
   misses than hits, we don't bother with read-around.

In particular, this means that the "real" read-ahead logic literally
only needs to worry about finding sequential accesses, and does not
have to worry about the common executable mmap access patthers that 
have very different behaviour.

Some constant tweaking may be a good idea.

in add_timer_internal() we simply leave the timer pending forever if the
expiry is in more than 0xffffffff jiffies. This means more than 48 days on
eg. ia64 - which is not an unrealistic timeout. IIRC crond is happy to use
extremely large timeouts.

It's better to time out early (if you can call 48 days "early") than to
not time out at all.

This offers a generic do_div64() that actually does the right thing,
unlike some architectures that "optimized" the 64-by-32 divide into
just a 32-bit divide.

Both ppc and sh were already providing an assembly optimized
__div64_32().  I called my function the same, so that their optimized
versions will automatically override mine in lib.a.

I've only tested extensively on m68knommu (uClinux) and made
sure generated code is reasonably short. Should be ok also on
parisc, since it's the same algorithm they were using before.

 - add generic C implementations of the do_div() for 32bit and 64bit
   archs in asm-generic/div64.h;

 - add generic library support function __div64_32() to handle the
   full 64/32 case on 32bit archs;

 - kill multiple copies of generic do_div() in architecture
   specific subdirs. Most copies were either buggy or not doing
   what they were supposed to do;

 - ensure all surviving instances of do_div() have their parameters
   correctly parenthesized to avoid funny side-effects;

Fix arbitration between net open and tty open.

Cleanup missed bits of CUA device removal changes.

Fix arbitration between net open and tty open.

Clean up unused locals resulting from latest tty changes.

Fix arbitration between net open and tty open.

Cleanup unused local resulting from latest tty changes.

From Mikael Petterson:


  Booting kernel 2.5.74 on a PowerMac with CONFIG_BLK_DEV_IDE_PMAC=y
  results in an oops during IDE init, and the box then reboots.

  The patch below updates drivers/ide/ppc/pmac.c to also set up the
  hwif->ide_dma_queued_off and hwif->ide_dma_queued_on function
  pointers, which fixes the oops. Tested on my ancient PM4400.

I no longer have the time/interest in nbd, and Paul agreed to take it
over.

The compat ioctls for device mapper were not being enabled due to an
incorrect config option.

This tweaks the mmap read-ahead behaviour so that the prefaulting
is largely pointless.

 - double the minimum readaround chunksize in page_cache_readaround().

 - when a seek is detected, collapse the window more slowly.

i2o_scsi.c now needs pci.h.

From: ilmari@ilmari.org (Dagfinn Ilmari Mannsaker)

It turns out that net/bluetooth/rfcomm/sock.c (and
net/bluetooth/hci_sock.c) had been left out when net_proto_family gained an
owner field, here's a patch that fixes them both.

From: junkio@cox.net

Sigh.  Is there a gcc option to tell it to not accept this incompatible C99
extension?

From: Arvind Kandhare <arvind.kan@wipro.com>

When switch_uid is called, the reference count of the new user is
incremented twice.  I think the increment in the switch_uid is done because
of the reparent_to_init() function which does not increase the __count for
root user.

But if switch_uid is called from any other function, the reference count is
already incremented by the caller by calling alloc_uid for the new user.
Hence the count is incremented twice.  The user struct will not be deleted
even when there are no processes holding a reference count for it.  This
does not cause any problem currently because nothing is dependent on timely
deletion of the user struct.

From: Davide Libenzi <davidel@xmailserver.org>

- Inline eventpoll_release() so that __fput() does not need to call in
  epoll code if the file itself is not registered inside an epoll fd

- Add <linux/types.h> inclusion due __u32 and __u64 usage

- Fix debug printf that would otherwise panic if enabled with the new
  epoll code

From: Davide Libenzi <davidel@xmailserver.org>

- Remove a couple of impossible debug checks (unsigneds cannot be
  negative!)

- If __alloc_bootmem_core() fails with a goal and unaligned node_boot_start
  it'll loop fovever.

If de_thread() fails in flush_old_exec() then we try to fail the execve().

That is a bad move, because exec_mmap() has already switched the current
process over to the new mm.  The new process is not yet sufficiently set up
to handle the error and the kernel doublefaults and dies.  exec_mmap() is the
point of no return.

Change flush_old_exec() to call de_thread() before running exec_mmap() so the
execing program sees the error.  I added fault injection to both de_thread()
and exec_mmap() - everything now survives OK.

From: Nick Piggin <piggin@cyberone.com.au>

Add some comments to the request allocation code.

- pass gfp_flags to get_io_context(): not all callers are forced to use
  GFP_ATOMIC().

- fix locking in get_io_context(): bump the refcount whilein the exclusive
  region.

- don't go oops in get_io_context() if the kmalloc failed.

- in as_get_io_context(): fail the whole thing if we were unable to
  allocate the AS-specific part.

- as_remove_queued_request() cleanup

From: Nick Piggin <piggin@cyberone.com.au>

The following patch gets batching working how it should be.

After a process is woken up, it is allowed to allocate up to 32 requests
for 20ms.  It does not stop other processes submitting requests if it isn't
submitting though.  This should allow less context switches, and allow
batches of requests from each process to be sent to the io scheduler
instead of 1 request from each process.

tiobench sequential writes are more than tripled, random writes are nearly
doubled over mm1.  In earlier tests I generally saw better CPU efficiency
but it doesn't show here.  There is still debug to be taken out.  Its also
only on UP.

                                Avg     Maximum     Lat%   Lat%   CPU
 Identifier    Rate  (CPU%)  Latency   Latency     >2s    >10s   Eff
 ------------------- ------ --------- ---------- ------- ------ ----
 -2.5.71-mm1   11.13 3.783%    46.10    24668.01   0.84   0.02   294
 +2.5.71-mm1   13.21 4.489%    37.37     5691.66   0.76   0.00   294

 Random Reads
 ------------------- ------ --------- ---------- ------- ------ ----
 -2.5.71-mm1    0.97 0.582%   519.86     6444.66  11.93   0.00   167
 +2.5.71-mm1    1.01 0.604%   484.59     6604.93  10.73   0.00   167

 Sequential Writes
 ------------------- ------ --------- ---------- ------- ------ ----
 -2.5.71-mm1    4.85 4.456%    77.80    99359.39   0.18   0.13   109
 +2.5.71-mm1   14.11 14.19%    10.07    22805.47   0.09   0.04    99

 Random Writes
 ------------------- ------ --------- ---------- ------- ------ ----
 -2.5.71-mm1    0.46 0.371%    14.48     6173.90   0.23   0.00   125
 +2.5.71-mm1    0.86 0.744%    24.08     8753.66   0.31   0.00   115

It decreases context switch rate on IBM's 8-way on ext2 tiobench 64 threads
from ~2500/s to ~140/s on their regression tests.

From: Nick Piggin <piggin@cyberone.com.au>

Generalise the AS-specific per-process IO context so that other IO schedulers
could use it.

From: Nick Piggin <piggin@cyberone.com.au>

This patch fixes the request batching fairness/starvation issue.  Its not
clear what is going on with 2.4, but it seems that its a problem around this
area.

Anyway, previously:

	* request queue fills up
	* process 1 calls get_request, sleeps
	* a couple of requests are freed
	* process 2 calls get_request, proceeds
	* a couple of requests are freed
	* process 2 calls get_request...

Now as unlikely as it seems, it could be a problem.  Its a fairness problem
that process 2 can skip ahead of process 1 anyway.

With the patch:

	* request queue fills up
	* any process calling get_request will sleep
	* once the queue gets below the batch watermark, processes
	  start being worken, and may allocate.


This patch includes Chris Mason's fix to only clear queue_full when all tasks
have been woken.  Previously I think starvation and unfairness could still
occur.

With this change to the blk-fair-batches patch, Chris is showing some much
improved numbers for 2.4 - 170 ms max wait vs 2700ms without blk-fair-batches
for a dbench 90 run.  He didn't indicate how much difference his patch alone
made, but it is an important fix I think.

From: Nick Piggin <piggin@cyberone.com.au>

If there are no requess in flight against the target device and
get_request() fails, nothing will wake us up.  Fix.

From: Nick Piggin <piggin@cyberone.com.au>


This patch implements a hint so that AS can tell the request allocator to
allocate a request even if there are none left (the accounting is quite
flexible and easily handles overallocations).

elv_may_queue semantics have changed from "the elevator does _not_ want
another request allocated" to "the elevator _insists_ that another request is
allocated".  I couldn't see any harm ;)

Now in practice, AS will only allow _1_ request over the limit, because as
soon as the request is sent to AS, it stops anticipating.

From: Nick Piggin <piggin@cyberone.com.au>

Now that we are counting requests (not requests free), this patch changes
the congested & batch watermarks to be more logical.  Also a minor fix to
the sysfs code.

From: Nick Piggin <piggin@cyberone.com.au>

This gets rid of the global queue_nr_requests and usage of BLKDEV_MAX_RQ
(the latter is now only used to set the queues' defaults).

The queue depth becomes per-queue, controlled by a sysfs entry.

Using keventd for running request_fns is risky because keventd itself can
block on disk I/O.  Use the new kblockd kernel threads for the generic
unplugging.

From: Nick Piggin <piggin@cyberone.com.au>

This is the core anticipatory IO scheduler.  There are nearly 100 changesets
in this and five months work.  I really cannot describe it fully here.

Major points:

- It works by recognising that reads are dependent: we don't know where the
  next read will occur, but it's probably close-by the previous one.  So once
  a read has completed we leave the disk idle, anticipating that a request
  for a nearby read will come in.

- There is read batching and write batching logic.

  - when we're servicing a batch of writes we will refuse to seek away
    for a read for some tens of milliseconds.  Then the write stream is
    preempted.

  - when we're servicing a batch of reads (via anticipation) we'll do
    that for some tens of milliseconds, then preempt.

- There are request deadlines, for latency and fairness.
  The oldest outstanding request is examined at regular intervals. If
  this request is older than a specific deadline, it will be the next
  one dispatched. This gives a good fairness heuristic while being simple
  because processes tend to have localised IO.


Just about all of the rest of the complexity involves an array of fixups
which prevent most of teh obvious failure modes with anticipation: trying to
not leave the disk head pointlessly idle.  Some of these algorithms are:

- Process tracking.  If the process whose read we are anticipating submits
  a write, abandon anticipation.

- Process exit tracking.  If the process whose read we are anticipating
  exits, abandon anticipation.

- Process IO history.  We accumulate statistical info on the process's
  recent IO patterns to aid in making decisions about how long to anticipate
  new reads.

  Currently thinktime and seek distance are tracked. Thinktime is the
  time between when a process's last request has completed and when it
  submits another one. Seek distance is simply the number of sectors
  between each read request. If either statistic becomes too high, the
  it isn't anticipated that the process will submit another read.

The above all means that we need a per-process "io context".  This is a fully
refcounted structure.  In this patch it is AS-only.  later we generalise it a
little so other IO schedulers could use the same framework.

- Requests are grouped as synchronous and asynchronous whereas deadline
  scheduler groups requests as reads and writes. This can provide better
  sync write performance, and may give better responsiveness with journalling
  filesystems (although we haven't done that yet).

  We currently detect synchronous writes by nastily setting PF_SYNCWRITE in
  current->flags.  The plan is to remove this later, and to propagate the
  sync hint from writeback_contol.sync_mode into bio->bi_flags thence into
  request->flags.  Once that is done, direct-io needs to set the BIO sync
  hint as well.

- There is also quite a bit of complexity gone into bashing TCQ into
  submission. Timing for a read batch is not started until the first read
  request actually completes. A read batch also does not start until all
  outstanding writes have completed.

AS is the default IO scheduler.  deadline may be chosen by booting with
"elevator=deadline".

There are a few reasons for retaining deadline:

- AS is often slower than deadline in random IO loads with large TCQ
  windows. The usual real world task here is OLTP database loads.

- deadline is presumably more stable.

- deadline is much simpler.



The tunable per-queue entries under /sys/block/*/iosched/ are all in
milliseconds:

* read_expire

  Controls how long until a request becomes "expired".

  It also controls the interval between which expired requests are served,
  so set to 50, a request might take anywhere < 100ms to be serviced _if_ it
  is the next on the expired list.

  Obviously it can't make the disk go faster.  Result is basically the
  timeslice a reader gets in the presence of other IO.  100*((seek time /
  read_expire) + 1) is very roughly the % streaming read efficiency your disk
  should get in the presence of multiple readers.

* read_batch_expire

  Controls how much time a batch of reads is given before pending writes
  are served.  Higher value is more efficient.  Shouldn't really be below
  read_expire.

* write_ versions of the above

* antic_expire

  Controls the maximum amount of time we can anticipate a good read before
  giving up.  Many other factors may cause anticipation to be stopped early,
  or some processes will not be "anticipated" at all.  Should be a bit higher
  for big seek time devices though not a linear correspondance - most
  processes have only a few ms thinktime.

From: Nick Piggin <piggin@cyberone.com.au>

Introduces an elevator_completed_req() callback with which the generic
queueing layer may tell an IO scheduler that a particualr request has
finished.

Introduces the elv_may_queue() predicate with which the IO scheduler may tell
the generic request layer that we may add another request to this queue.

It is used by the CFQ elevator.

keventd is inappropriate for running block request queues because keventd
itself can get blocked on disk I/O.  Via call_usermodehelper()'s vfork and,
presumably, GFP_KERNEL allocations.

So create a new gang of kernel threads whose mandate is for running low-level
disk operations.  It must ever block on disk IO, so any memory allocations
should be GFP_NOIO.

We mainly use it for running unplug operations from interrupt context.

From: Nick Piggin <piggin@cyberone.com.au>

The batch_requests function got lost during the merge of the dynamic request
allocation patch.

We need it for the anticipatory scheduler - when the number of threads
exceeds the number of requests, the anticipated-upon task will undesirably
sleep in get_request_wait().

And apparently some block devices which use small requests need it so they
string a decent number together.

Jens has acked this patch.

From: "Chen, Kenneth W" <kenneth.w.chen@intel.com>

This patch proposes a performance fix for the current IPC semaphore
implementation.

There are two shortcoming in the current implementation:
try_atomic_semop() was called two times to wake up a blocked process,
once from the update_queue() (executed from the process that wakes up
the sleeping process) and once in the retry part of the blocked process
(executed from the block process that gets woken up).

A second issue is that when several sleeping processes that are eligible
for wake up, they woke up in daisy chain formation and each one in turn
to wake up next process in line.  However, every time when a process
wakes up, it start scans the wait queue from the beginning, not from
where it was last scanned.  This causes large number of unnecessary
scanning of the wait queue under a situation of deep wait queue.
Blocked processes come and go, but chances are there are still quite a
few blocked processes sit at the beginning of that queue.

What we are proposing here is to merge the portion of the code in the
bottom part of sys_semtimedop() (code that gets executed when a sleeping
process gets woken up) into update_queue() function.  The benefit is two
folds: (1) is to reduce redundant calls to try_atomic_semop() and (2) to
increase efficiency of finding eligible processes to wake up and higher
concurrency for multiple wake-ups.

We have measured that this patch improves throughput for a large
application significantly on a industry standard benchmark.

This patch is relative to 2.5.72.  Any feedback is very much
appreciated.

Some kernel profile data attached:

  Kernel profile before optimization:
  -----------------------------------------------
                0.05    0.14   40805/529060      sys_semop [133]
                0.55    1.73  488255/529060      ia64_ret_from_syscall
[2]
[52]     2.5    0.59    1.88  529060         sys_semtimedop [52]
                0.05    0.83  477766/817966      schedule_timeout [62]
                0.34    0.46  529064/989340      update_queue [61]
                0.14    0.00 1006740/6473086     try_atomic_semop [75]
                0.06    0.00  529060/989336      ipcperms [149]
  -----------------------------------------------

                0.30    0.40  460276/989340      semctl_main [68]
                0.34    0.46  529064/989340      sys_semtimedop [52]
[61]     1.5    0.64    0.87  989340         update_queue [61]
                0.75    0.00 5466346/6473086     try_atomic_semop [75]
                0.01    0.11  477676/576698      wake_up_process [146]
  -----------------------------------------------
                0.14    0.00 1006740/6473086     sys_semtimedop [52]
                0.75    0.00 5466346/6473086     update_queue [61]
[75]     0.9    0.89    0.00 6473086         try_atomic_semop [75]
  -----------------------------------------------

  Kernel profile with optimization:

  -----------------------------------------------
                0.03    0.05   26139/503178      sys_semop [155]
                0.46    0.92  477039/503178      ia64_ret_from_syscall
[2]
[61]     1.2    0.48    0.97  503178         sys_semtimedop [61]
                0.04    0.79  470724/784394      schedule_timeout [62]
                0.05    0.00  503178/3301773     try_atomic_semop [109]
                0.05    0.00  503178/930934      ipcperms [149]
                0.00    0.03   32454/460210      update_queue [99]
  -----------------------------------------------
                0.00    0.03   32454/460210      sys_semtimedop [61]
                0.06    0.36  427756/460210      semctl_main [75]
[99]     0.4    0.06    0.39  460210         update_queue [99]
                0.30    0.00 2798595/3301773     try_atomic_semop [109]
                0.00    0.09  470630/614097      wake_up_process [146]
  -----------------------------------------------
                0.05    0.00  503178/3301773     sys_semtimedop [61]
                0.30    0.00 2798595/3301773     update_queue [99]
[109]    0.3    0.35    0.00 3301773         try_atomic_semop [109]
  -----------------------------------------------=20

Both number of function calls to try_atomic_semop() and update_queue()
are reduced by 50% as a result of the merge.  Execution time of
sys_semtimedop is reduced because of the reduction in the low level
functions.