From: Hugh Dickins <hugh@veritas.com>

Pave the way for prio_tree by switching over to its interfaces, but actually
still implement them with the same old lists as before.

Most of the vma_prio_tree interfaces are straightforward.  The interesting one
is vma_prio_tree_next, used to search the tree for all vmas which overlap the
given range: unlike the list_for_each_entry it replaces, it does not find
every vma, just those that match.

But this does leave handling of nonlinear vmas in a very unsatisfactory state:
for now we have to search again over the maximum range to find all the
nonlinear vmas which might contain a page, which of course takes away the
point of the tree.  Fixed in later patch of this batch.

There is no need to initialize vma linkage all over, just do it before
inserting the vma in list or tree.  /proc/pid/statm had an odd test for its
shared count: simplified to an equivalent test on vm_file.

From: Hugh Dickins <hugh@veritas.com>

If file-based vmas are to be kept in a tree, according to the file offsets
they map, then adjusting the vma's start pgoff or its end involves
repositioning in the tree, while holding i_shared_lock (and page_table_lock). 
We used to avoid that if possible, e.g.  when just moving end; but if we're
heading that way, let's now tidy up vma_merge and split_vma, and do all the
locking and adjustment in a new helper vma_adjust.  And please, let's call the
next vma in vma_merge "next" rather than "prev".

Since these patches are diffed over 2.6.6-rc2-mm2, they include the NUMA
mpolicy mods which you'll have to remove to go earlier in the series, sorry
for that nuisance.  I have intentionally changed the one vma_mpol_equal to
mpol_equal, to make the merge cases more alike.

From: Andi Kleen <ak@suse.de>

This fixes a user triggerable crash in mbind() in NUMA API.  It would oops
when running into the end of memory.  Actually not really oops, because a
oops with the mm sem hold for writing always deadlocks.

From: Andi Kleen <ak@suse.de>

Change to core VM to use alloc_page_vma() instead of alloc_page().

Change the swap readahead to follow the policy of the VMA.

From: Andi Kleen <ak@suse.de>

Add NUMA hit/miss statistics to page allocation and display them in sysfs.

This is not 100% required for NUMA API, but without this it is very

The overhead is quite low because all counters are per CPU and only happens
when CONFIG_NUMA is defined.

From: Christoph Hellwig <hch@lst.de>

- don't include mempolicy.h in sched.h and mm.h when a forward delcaration
  is enough.  Andi argued against that in the past, but I'd really hate to add
  another header to two of the includes used in basically every driver when we
  can include it in the six files actually needing it instead (that number is
  for my ppc32 system, maybe other arches need more include in their
  directories)

- make numa api fields in tast_struct conditional on CONFIG_NUMA, this gives
  us a few ugly ifdefs but avoids wasting memory on non-NUMA systems.

From: Andi Kleen <ak@suse.de>

Add support to tmpfs and hugetlbfs to support NUMA API.  Shared memory is a
bit of a special case for NUMA policy.  Normally policy is associated to VMAs
or to processes, but for a shared memory segment you really want to share the
policy.  The core NUMA API has code for that, this patch adds the necessary
changes to tmpfs and hugetlbfs.

First it changes the custom swapping code in tmpfs to follow the policy set
via VMAs.

It is also useful to have a "backing store" of policy that saves the policy
even when nobody has the shared memory segment mapped.  This allows command
line tools to pre configure policy, which is then later used by programs.

Note that hugetlbfs needs more changes - it is also required to switch it to
lazy allocation, otherwise the prefault prevents mbind() from working.

From: Andi Kleen <ak@suse.de>

NUMA API adds a policy to each VMA.  During VMA creattion, merging and
splitting these policies must be handled properly.  This patch adds the calls
to this. 

It is a nop when CONFIG_NUMA is not defined.

From: Andi Kleen <ak@suse.de>

Patch readds the sysfs output of the NUMA API statistics.  All my test
scripts need this and it is very useful to check if the policy actually
works.

This got lost when the huge page numa api changes got dropped.

I decided to not resend the huge pages NUMA API changes for now.  Instead I
will wait for this area to settle when demand paged large pages is merged.

From: Hugh Dickins <hugh@veritas.com>

I think Andi missed the copy_vma I recently added for mremap, and it'll
need something like below....  (Doesn't look like it'll optimize away when
it's not needed - rather bloaty.)

From: Andi Kleen <ak@suse.de>

The following patches add support for configurable NUMA memory policy
for user processes. It is based on the proposal from last kernel summit
with feedback from various people.

This NUMA API doesn't not attempt to implement page migration or anything
else complicated: all it does is to police the allocation when a page
is first allocation or when a page is reallocated after swapping. Currently
only support for shared memory and anonymous memory is there; policy for
file based mappings is not implemented yet (although they get implicitely
policied by the default process policy)

It adds three new system calls: mbind to change the policy of a VMA,
set_mempolicy to change the policy of a process, get_mempolicy to retrieve
memory policy. User tools (numactl, libnuma, test programs, manpages) can be
found in  ftp://ftp.suse.com/pub/people/ak/numa/numactl-0.6.tar.gz

For details on the system calls see the manpages
http://www.firstfloor.org/~andi/mbind.html
http://www.firstfloor.org/~andi/set_mempolicy.html
http://www.firstfloor.org/~andi/get_mempolicy.html
Most user programs should actually not use the system calls directly,
but use the higher level functions in libnuma
(http://www.firstfloor.org/~andi/numa.html) or the command line tools
(http://www.firstfloor.org/~andi/numactl.html

The system calls allow user programs and administors to set various NUMA memory
policies for putting memory on specific nodes. Here is a short description
of the policies copied from the kernel patch:

 * NUMA policy allows the user to give hints in which node(s) memory should
 * be allocated.
 *
 * Support four policies per VMA and per process:
 *
 * The VMA policy has priority over the process policy for a page fault.
 *
 * interleave     Allocate memory interleaved over a set of nodes,
 *                with normal fallback if it fails.
 *                For VMA based allocations this interleaves based on the
 *                offset into the backing object or offset into the mapping
 *                for anonymous memory. For process policy an process counter
 *                is used.
 * bind           Only allocate memory on a specific set of nodes,
 *                no fallback.
 * preferred      Try a specific node first before normal fallback.
 *                As a special case node -1 here means do the allocation
 *                on the local CPU. This is normally identical to default,
 *                but useful to set in a VMA when you have a non default
 *                process policy.
 * default        Allocate on the local node first, or when on a VMA
 *                use the process policy. This is what Linux always did
 *                in a NUMA aware kernel and still does by, ahem, default.
 *
 * The process policy is applied for most non interrupt memory allocations
 * in that process' context. Interrupts ignore the policies and always
 * try to allocate on the local CPU. The VMA policy is only applied for memory
 * allocations for a VMA in the VM.
 *
 * Currently there are a few corner cases in swapping where the policy
 * is not applied, but the majority should be handled. When process policy
 * is used it is not remembered over swap outs/swap ins.
 *
 * Only the highest zone in the zone hierarchy gets policied. Allocations
 * requesting a lower zone just use default policy. This implies that
 * on systems with highmem kernel lowmem allocation don't get policied.
 * Same with GFP_DMA allocations.
 *
 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
 * all users and remembered even when nobody has memory mapped.




This patch:

This is the core NUMA API code. This includes NUMA policy aware
wrappers for get_free_pages and alloc_page_vma(). On non NUMA kernels
these are defined away.

The system calls mbind (see http://www.firstfloor.org/~andi/mbind.html),
get_mempolicy (http://www.firstfloor.org/~andi/get_mempolicy.html) and
set_mempolicy (http://www.firstfloor.org/~andi/set_mempolicy.html) are
implemented here.

Adds a vm_policy field to the VMA and to the process. The process
also has field for interleaving. VMA interleaving uses the offset
into the VMA, but that's not possible for process allocations.

From: Andi Kleen <ak@muc.de>

  > Andi, how come policy_vma() calls ->set_policy under i_shared_sem?

  I think this can be actually dropped now.  In an earlier version I did
  walk the vma shared list to change the policies of other mappings to the
  same shared memory region.  This turned out too complicated with all the
  corner cases, so I eventually gave in and added ->get_policy to the fast
  path.  Also there is still the mmap_sem which prevents races in the same MM.
   

  Patch to remove it attached.  Also adds documentation and removes the
  bogus __alloc_page_vma() prototype noticed by hch.

From: Andi Kleen <ak@suse.de>

  A few incremental fixes for NUMA API.

  - Fix a few comments

  - Add a compat_ function for get_mem_policy I considered changing the
    ABI to avoid this, but that would have made the API too ugly.  I put it
    directly into the file because a mm/compat.c didn't seem worth it just for
    this.

  - Fix the algorithm for VMA interleave.

From: Matthew Dobson <colpatch@us.ibm.com>

  1) Move the extern of alloc_pages_current() into #ifdef CONFIG_NUMA.
    The only references to the function are in NUMA code in mempolicy.c

  2) Remove the definitions of __alloc_page_vma().  They aren't used.

  3) Move forward declaration of struct vm_area_struct to top of file.

From: Andi Kleen <ak@suse.de>

Add NUMA API system calls on IA64 and one bug fix required for it.

From: Andi Kleen <ak@suse.de>

Add NUMA API system calls for i386

From: Andi Kleen <ak@suse.de>

Add NUMA API system calls on x86-64

This includes a bugfix to prevent miscompilation on gcc 3.2 of bitmap.h

From: Hugh Dickins <hugh@veritas.com>

First of batch of six patches which introduce Rajesh Venkatasubramanian's
implementation of a radix priority search tree of vmas, to handle object-based
reverse mapping corner cases well.

rmap 14 i_shared_lock fixes

Start the sequence with a couple of outstanding i_shared_lock fixes.

Since i_shared_sem became i_shared_lock, we've had to shift and then
temporarily remove mremap move's protection of concurrent truncation - if
mremap moves ptes while unmap_mapping_range_list is making its way through the
vmas, there's a danger we'd move a pte from an area yet to be cleaned back
into an area already cleared.

Now site the i_shared_lock with the page_table_lock in move_one_page.  Replace
page_table_present by get_one_pte_map, so we know when it's necessary to
allocate a new page table: in which case have to drop i_shared_lock, trylock
and perhaps reorder locks on the way back.  Yet another fix: must check for
NULL dst before pte_unmap(dst).

And over in rmap.c, try_to_unmap_file's cond_resched amidst its lengthy
nonlinear swapping was now causing might_sleep warnings: moved to a rather
unsatisfactory and less frequent cond_resched_lock on i_shared_lock when we
reach the end of the list; and one before starting on the nonlinears too: the
"cursor" may become out-of-date if we do schedule, but I doubt it's worth
bothering about.

Having a semaphore in there causes modest performance regressions on heavily
mmap-intensive workloads on some hardware.  Specifically, up to 30% in SDET on
NUMAQ and big PPC64.

So switch it back to being a spinlock.  This does mean that unmap_vmas() needs
to be told whether or not it is allowed to schedule away; that's simple to do
via the zap_details structure.

This change means that there will be high scheuling latencies when someone
truncates a large file which is currently mmapped, but nobody does that
anyway.  The scheduling points in unmap_vmas() are mainly for munmap() and
exit(), and they still will work OK for that.

From: Hugh Dickins <hugh@veritas.com>

  Sorry, my premature optimizations (trying to pass down NULL zap_details
  except when needed) have caught you out doubly: unmap_mapping_range_list was
  NULLing the details even though atomic was set; and if it hadn't, then
  zap_pte_range would have missed free_swap_and_cache and pte_clear when pte
  not present.  Moved the optimization into zap_pte_range itself.  Plus
  massive documentation update.

From: Hugh Dickins <hugh@veritas.com>

  Here's a second patch to add to the first: mremap's cows can't come home
  without releasing the i_mmap_lock, better move the whole "Subtle point"
  locking from move_vma into move_page_tables.  And it's possible for the file
  that was behind an anonymous page to be truncated while we drop that lock,
  don't want to abort mremap because of VM_FAULT_SIGBUS.

  (Eek, should we be checking do_swap_page of a vm_file area against the
  truncate_count sequence?  Technically yes, but I doubt we need bother.)


- We cannot hold i_mmap_lock across move_one_page() because
  move_one_page() needs to perform __GFP_WAIT allocations of pagetable pages.

- Move the cond_resched() out so we test it once per page rather than only
  when move_one_page() returns -EAGAIN.

From: Hugh Dickins <hugh@veritas.com>

Delete include/asm*/rmap.h
Delete pte_addr_t typedef from include/asm*/pgtable.h
Delete KM_PTE2 from subset of include/asm*/kmap_types.h
Beware when 4G/4G returns to -mm: i386 may need KM_FILLER for 8K stack.

From: Hugh Dickins <hugh@veritas.com>

Remove the support for pte_chain rmap from page table initialization, just
continue to maintain nr_page_table_pages (but only for user page tables -
it also counted vmalloc page tables before, little need, and I'm unsure if
per-cpu stats are safe early enough on all arches).  mm/memory.c is the
only core file affected.

But ppc and ppc64 have found the old rmap page table initialization useful
to support their ptep_test_and_clear_young: so transfer rmap's
initialization to them (even on kernel page tables?  well, okay).

From: Hugh Dickins <hugh@veritas.com>

A weakness of the anonmm scheme is its difficulty in tracking pages shared
between two or more mms (one being an ancestor of the other), when mremap has
been used to move a range of pages in one of those mms.  mremap move is not
very common anyway, and it's more often used on a page range exclusive to the
mm; but uncommon though it may be, we must not allow unlocked pages to become
unswappable.

This patch follows Linus' suggestion, simply to take a private copy of the
page in such a case: early C-O-W.  My previous implementation was daft with
respect to pages currently on swap: it insisted on swapping them in to copy
them.  No need for that: just take the copy when a page is brought in from
swap, and its intended address is found to clash with what rmap has already
noted.

If do_swap_page has to make this copy in the mremap moved case (simply a call
to do_wp_page), might as well do so also in the case when it's a write access
but the page not exclusive, it's always seemed a little odd that swapin needed
a second fault for that.  A bug even: get_user_pages force imagines that a
single call to handle_mm_fault must break C-O-W.  Another bugfix: swapoff's
unuse_process didn't check is_vm_hugetlb_page.

Andrea's anon_vma has no such problem with mremap moved pages, handling them
with elegant use of vm_pgoff - though at some cost to vma merging.  How
important is it to handle them efficiently?  For now there's a msg
printk(KERN_WARNING "%s: mremap moved %d cows\n", current->comm, cows);

From: Hugh Dickins <hugh@veritas.com>

Hugh's anonmm object-based reverse mapping scheme for anonymous pages.  We
have not yet decided whether to adopt this scheme, or Andrea's more advanced
anon_vma scheme.  anonmm is easier for me to merge quickly, to replace the
pte_chain rmap taken out in the previous patch; a patch to install Andrea's
anon_vma will follow in due course.

Why build up and tear down chains of pte pointers for anonymous pages, when a
page can only appear at one particular address, in a restricted group of mms
that might share it?  (Except: see next patch on mremap.)

Introduce struct anonmm per mm to track anonymous pages, all forks from one
exec sharing the same bundle of linked anonmms.  Anonymous pages originate in
one mm, but may be forked into another mm of the bundle later on.  Callouts
from fork.c to allocate, dup and exit the anonmm structure private to rmap.c.

From: Hugh Dickins <hugh@veritas.com>

  Two concurrent exits (of the last two mms sharing the anonhd).  First
  exit_rmap brings anonhd->count down to 2, gets preempted (at the
  spin_unlock) by second, which brings anonhd->count down to 1, sees it's 1
  and frees the anonhd (without making any change to anonhd->count itself),
  cpu goes on to do something new which reallocates the old anonhd as a new
  struct anonmm (probably not a head, in which case count will start at 1),
  first resumes after the spin_unlock and sees anonhd->count 1, frees "anonhd"
  again, it's used for something else, a later exit_rmap list_del finds list
  corrupt.

From: Hugh Dickins <hugh@veritas.com>

Lots of deletions: the next patch will put in the new anon rmap, which
should look clearer if first we remove all of the old pte-pointer-based
rmap from the core in this patch - which therefore leaves anonymous rmap
totally disabled, anon pages locked in memory until process frees them.

Leave arch files (and page table rmap) untouched for now, clean them up in
a later batch.  A few constructive changes amidst all the deletions:

Choose names (e.g.  page_add_anon_rmap) and args (e.g.  no more pteps) now
so we need not revisit so many files in the next patch.  Inline function
page_dup_rmap for fork's copy_page_range, simply bumps mapcount under lock.
 cond_resched_lock in copy_page_range.  Struct page rearranged: no pte
union, just mapcount moved next to atomic count, so two ints can occupy one
long on 64-bit; i386 struct page now 32 bytes even with PAE.  Never pass
PageReserved to page_remove_rmap, only do_wp_page did so.


From: Hugh Dickins <hugh@veritas.com>

  Move page_add_anon_rmap's BUG_ON(page_mapping(page)) inside the rmap_lock
  (well, might as well just check mapping if !mapcount then): if this page is
  being mapped or unmapped on another cpu at the same time, page_mapping's
  PageAnon(page) and page->mapping are volatile.

  But page_mapping(page) is used more widely: I've a nasty feeling that
  clear_page_anon, page_add_anon_rmap and/or page_mapping need barriers added
  (also in 2.6.6 itself),

Many places do:

	if (kmem_cache_create(...) == NULL)
		panic(...);

We can consolidate all that by passing another flag to kmem_cache_create()
which says "panic if it doesn't work".

From: Hugh Dickins <hugh@veritas.com>

The previous patch let the ptes of file pages be located via page
->mapping->i_mmap and i_mmap_shared lists of vmas; which works well unless
the vma is VM_NONLINEAR - one in which sys_remap_file_pages has been used
to place pages in unexpected places, to avoid an explosion of distinct
unmergable vmas.  Such pages were effectively locked in memory.

page_referenced_file is already skipping nonlinear vmas, they'd just waste
its time, and age unfairly any pages in their proper positions.  Now extend
try_to_unmap_file, to persuade it to swap from nonlinears.

Ignoring the page requested, try to unmap cluster of 32 neighbouring ptes
(in worst case all empty slots) in a nonlinear vma, then move on to the
next vma; stopping when we've unmapped at least as many maps as the
requested page had (vague guide of how hard to try), or have reached the
end.  With large sparse nonlinear vmas, this could take a long time:
inserted a cond_resched while no locks are held, unusual at this level but
I think okay, shrink_list does so.

Use vm_private_data a little like the old mm->swap_address, as a cursor
recording how far we got, so we don't attack the same ptes next time around
(earlier tried inserting an empty marker vma in the list, but that got
messy).  How well this will work on real- life nonlinear vmas remains to be
seen, but should work better than locking them all in memory, or swapping
everything out all the time.

Existing users of vm_private_data have either VM_RESERVED or VM_DONTEXPAND
set, both of which are in the VM_SPECIAL category where we never try to
merge vmas: so removed the vm_private_data test from is_mergeable_vma, so
we can still merge VM_NONLINEARs.  Of course, we could instead add another
field to vm_area_struct.

From: Hugh Dickins <hugh@veritas.com>

Dave McCracken's object-based reverse mapping scheme for file pages: why
build up and tear down chains of pte pointers for file pages, when
page->mapping has i_mmap and i_mmap_shared lists of all the vmas which
might contain that page, and it appears at one deterministic position
within the vma (unless vma is nonlinear - see next patch)?

Has some drawbacks: more work to locate the ptes from page_referenced and
try_to_unmap, especially if the i_mmap lists contain a lot of vmas covering
different ranges; has to down_trylock the i_shared_sem, and hope that
doesn't fail too often.  But attractive in that it uses less lowmem, and
shifts the rmap burden away from the hot paths, to swapout.

Hybrid scheme for the moment: carry on with pte_chains for anonymous pages,
that's unchanged; but file pages keep mapcount in the pte union of struct
page, where anonymous pages keep chain pointer or direct pte address: so
page_mapped(page) works on both.

Hugh massaged it a little: distinct page_add_file_rmap entry point; list
searches check rss so as not to waste time on mms fully swapped out; check
mapcount to terminate once all ptes have been found; and a WARN_ON if
page_referenced should have but couldn't find all the ptes.

Running __mark_inode_dirty() against a swapcache page is illegal and will
oops.

I see a race in set_page_dirty() wherein it can be called with a PageSwapCache
page, but if the page is removed from swapcache after
__set_page_dirty_nobuffers() drops tree_lock(), we have the situation where
PageSwapCache() is false, but local variable `mapping' points at swapcache.

Handle that by checking for non-null mapping->host.  We don't care about the
page state at this point - we're only interested in the inode.



There is a converse case: what if a page is added to swapcache as we are
running set_page_dirty() against it?

In this case the page gets its PG_dirty flag set but it is not tagged as dirty
in the swapper_space radix tree.  The swap writeout code will handle this OK
and test_clear_page_dirty()'s call to
radix_tree_tag_clear(PAGECACHE_TAG_DIRTY) will silently have no effect.  The
only downside is that future radix-tree-based writearound won't notice that
such pages are dirty and swap IO scheduling will be a teensy bit worse.


The patch also fixes the (silly) testing of local variable `mapping' to see if
the page was truncated.  We should test page_mapping() for that.

Revert recent changes to sync_page().  Now that page_mapping() returns
&swapper_space for swapcache pages we don't need to test for PageSwapCache in
sync_page().

Fix up the "may we call writepage" logic for the swapcache changes.

Go back to the 2.6.5 concepts, with rmap additions.  In particular:

- Implement Andrea's flavour of page_mapping().  This function opaquely does
  the right thing for pagecache pages, anon pages and for swapcache pages.

  The critical thing here is that page_mapping() returns &swapper_space for
  swapcache pages without actually requiring the storage at page->mapping. 
  This frees page->mapping for the anonmm/anonvma metadata.

- Andrea and Hugh placed the pagecache index of swapcache pages into
  page->private rather than page->index.  So add new page_index() function
  which hides this.

- Make swapper_space.set_page_dirty() again point at
  __set_page_dirty_buffers().  If we don't do that, a bare set_page_dirty()
  will fall through to __set_page_dirty_buffers(), which is silly.

  This way, __set_page_dirty_buffers() can continue to use page->mapping.
  It should never go near anon or swapcache pages.

- Give swapper_space a ->set_page_dirty address_space_operation method, so
  that set_page_dirty() will not fall through to __set_page_dirty_buffers()
  for swapcache pages.  That function is not set up to handle them.


The main effect of these changes is that swapcache pages are treated more
similarly to pagecache pages.  And we are again tagging swapcache pages as
dirty in their radix tree, which is a requirement if we later wish to
implement swapcache writearound based on tagged radix-tree walks.

Simplify the logic in there a bit.

->tree_lock is supposed to be IRQ-safe.  Hugh worked out that with his
changes, we never actually take it from interrupt context, so spin_lock() is
sufficient.

Apart from kinda freaking me out, the analysis which led to this decision
becomes untrue with later patches.  So make it irq-safe.

into ppc970.osdl.org:/home/torvalds/v2.6/linux

(The difference between two pointers is a "size_t", while
MAX_LEN and the result here are "int"s).

Forwarding table paranoia:
* Solve some potential problems if a device changes address and one or
  more device has the same address.  
* Warn if new device added to a bridge matches a entry that has shown
  up on the network.
* Also don't put static entries in the timer list, they don't time
  out so shouldn't be there.

Replacement 64 bit compatibility code for the new ioctl's.  The new 
ioctl's all pass through clean, but for the old style ioctl's it uses
the mis-feature of the earlier bridge-utils that they check the API version.

So if an old 32bit version of brctl is run on a 64bit platform it will
report
	bridge utilities not compatible with kernel version

Tested on Itanium 1; but should solve issue for sparc, ppc, and x86_64

Add four new ioctl's for the operations that can't be done through sysfs.
The existing bridge ioctl's are multiplexed, and most go through SIOCDEVPRIVATE
so they won't work in a mixed 32/64bit environment.

The new release of bridge-utils will use these if possible, and fall
back to the old interface.

Move the local function timer_residue to br_timer_value so it can
be used by both ioctl and sysfs code.

Change how the read of forwarding table works.  Instead of copying
entries to user one at a time, use an intermediate kernel buffer and
do up to a page at a chunk.

This gets rid of some awkward code dealing with entries getting deleted
during the copy.  And allows same function to be used by later sysfs
hook.

Fix a deadlock where deleting a device call br_del_if with lock held.
br_del_if doesn't want to be called under lock anymore.