The patch arranges for constant 1, 2 and 4-byte copy_*_user() invokations to
be inlined.

It's hard to tell really, but the AIM9 creat_clo, signal_test and dir_rtns_1
numbers went up by 3%-9%, which is to be expected.

set_page_dirty() is racy if the caller has no reference against
page->mapping->host, and if the page is unlocked.  This is because
another CPU could truncate the page off the mapping and then free the
mapping.

Usually, the page _is_ locked, or the caller is a user-space process which
holds a reference on the inode by having an open file.

The exceptional cases are where the page was obtained via
get_user_pages().  The patch changes those to lock the page around the
set_page_dirty() call.

- Fix error-path mem leak in __vfs_follow_link() (From a recent AC->2.4
  patch)

- Make drivers/net/aironet4500_proc.c:driver_lock static.

Here is spin_lock():

#define spin_lock(lock) \
do { \
        preempt_disable(); \
        _raw_spin_lock(lock); \
} while(0)


Here is the scenario:

CPU0:
	spin_lock(some_lock);
	do_very_long_thing();	/* This has cond_resched()s in it */

CPU1:
	spin_lock(some_lock);

Now suppose that the scheduler tries to schedule a task on CPU1.  Nothing
happens, because CPU1 is spinning on the lock with preemption disabled.  CPU0
will happliy hold the lock for a long time because nobody has set
need_resched() against CPU0.

This problem can cause scheduling latencies of many tens of milliseconds on
SMP on kernels which handle UP quite happily.


This patch fixes the problem by changing the spin_lock() and write_lock()
contended slowpath to spin on the lock by hand, while polling for preemption
requests.

I would have done read_lock() too, but we don't seem to have read_trylock()
primitives.

The patch also shrinks the kernel by 30k due to not having separate
out-of-line spinning code for each spin_lock() callsite.

Pagetable teardown can hold page_table_lock for extremely long periods -
hundreds of milliseconds.  This is pretty much the final source of high
scheduling latency in the core kernel.

We fixed it for zap_page_range() by chunking the work up and dropping the
lock occasionally if needed.  But that did not fix exit_mmap() and
unmap_region().

So what this patch does is to create an uber-zapper "unmap_vmas()" which
provides all the vma-walking, page unmapping and low-latency lock-dropping
which zap_page_range(), exit_mmap() and unmap_region() require.  Those three
functions are updated to call unmap_vmas().

It's actually a bit of a cleanup...

touched_by_munmap() returns a reversed list of VMA's.  That makes things
harder in the low-latency-page-zapping patch.

So change touched_by_munmap() to return a VMA list which is in the original
order - ascending virtual addresses.

Oh, and rename it to <hugh>detach_vmas_to_be_unmapped()</hugh>.  It now
returns nothing, because we know that the VMA we passed in is the head of the
to-be-unmapped list.

In the next patch I wish to add to mm.h prototypes of functions which take an
mmu_gather_t* argument.   To do this I must either:

a) include tlb.h in mm.h

   Not good - more nested includes when a simple forward decl is sufficient.

b) Add `typedef struct free_pte_ctx mmu_gather_t;' to mm.h.

   That's silly - it's supposed to be an opaque type.

   or

c) Remove the pesky typedef.

   Bingo.

cond_resched_lock() _used_ to be "if this is the only lock which I am holding
then drop it and schedule if needed".

However with the i_shared_lock->i_shared_sem change, neither of its two
callsites now need those semantics.  So this patch changes it to mean just
"if needed, drop this lock and reschedule".

This allows us to also schedule if CONFIG_PREEMPT=n, which is useful -
zap_page_range() can run for an awfully long time.

The preempt and non-preempt versions of cond_resched_lock() have been
unified.

i_shared_lock is held for a very long time during vmtruncate() and causes
high scheduling latencies when truncating a file which is mmapped.  I've seen
100 milliseconds.

So turn it into a semaphore.  It nests inside mmap_sem.

This change is also needed by the shared pagetable patch, which needs to
unshare pte's on the vmtruncate path - lots of pagetable pages need to
be allocated and they are using __GFP_WAIT.

The patch also makes unmap_vma() static.

This patch from Roland McGrath fixes a threading related ptrace bug:
PTRACE_ATTACH should not stop everybody for each thread attached.

This is needed by some includers of <asm/stat.h>.

Implementing hash_str as hash_mem(..., strlen()) is actually quite slow,
so create a separate hash_str.

Now hash_mem has only one call site, and both are quite small, so we make them
both inline.

blk_plug_device and blk_remove_plug want irqs_disabled,
so we give it to them.

I think this is finally right:
 - we always wake_up when curr_resync decreases.
 - we wait for the other guys curr_resync to be less than ours
 - if we ever wait for someone who might yield, we start checking
   again from the start.

Currently nfsd only worries about read-only exports for
directories and files, which allows device special files
to be chmodded (for example).

This patch extends the test to cover all files, but is careful
to avoid it when an IRIX client is doing an write-permission test
against a pre-existing device special file.

We encode that status in the return value.

Also, don't pass 'proc' parameter to ->accept, as it is
implicit in rqstp.

If one of the callback (e.g. data ready) is called
before socket setup is complete, oops may occur.

With this patch, socket is kept SK_BUSY until all is
ready to avoid this.

Also some code is moved around to make it cleaner.

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