linux.git/mm/memory.c, branch v2.6.26-rc7

Reinstate ZERO_PAGE optimization in 'get_user_pages()' and fix XIP

2008-06-20T18:18:25+00:00

KAMEZAWA Hiroyuki and Oleg Nesterov point out that since the commit
557ed1fa2620dc119adb86b34c614e152a629a80 ("remove ZERO_PAGE") removed
the ZERO_PAGE from the VM mappings, any users of get_user_pages() will
generally now populate the VM with real empty pages needlessly.

We used to get the ZERO_PAGE when we did the "handle_mm_fault()", but
since fault handling no longer uses ZERO_PAGE for new anonymous pages,
we now need to handle that special case in follow_page() instead.

In particular, the removal of ZERO_PAGE effectively removed the core
file writing optimization where we would skip writing pages that had not
been populated at all, and increased memory pressure a lot by allocating
all those useless newly zeroed pages.

This reinstates the optimization by making the unmapped PTE case the
same as for a non-existent page table, which already did this correctly.

While at it, this also fixes the XIP case for follow_page(), where the
caller could not differentiate between the case of a page that simply
could not be used (because it had no "struct page" associated with it)
and a page that just wasn't mapped.

We do that by simply returning an error pointer for pages that could not
be turned into a "struct page *".  The error is arbitrarily picked to be
EFAULT, since that was what get_user_pages() already used for the
equivalent IO-mapped page case.

[ Also removed an impossible test for pte_offset_map_lock() failing:
  that's not how that function works ]

Acked-by: Oleg Nesterov 
Acked-by: Nick Piggin 
Cc: KAMEZAWA Hiroyuki 
Cc: Hugh Dickins 
Cc: Andrew Morton 
Cc: Ingo Molnar 
Cc: Roland McGrath 
Signed-off-by: Linus Torvalds

mm: allow pfnmap ->fault()s

2008-05-24T16:56:07+00:00

Take out an assertion to allow ->fault handlers to service PFNMAP regions.
This is required to reimplement .nopfn handlers with .fault handlers and
subsequently remove nopfn.

Signed-off-by: Nick Piggin 
Acked-by: Jes Sorensen 
Cc: Paul Mackerras 
Cc: Benjamin Herrenschmidt 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

fix SMP data race in pagetable setup vs walking

2008-05-14T17:05:18+00:00

There is a possible data race in the page table walking code. After the split
ptlock patches, it actually seems to have been introduced to the core code, but
even before that I think it would have impacted some architectures (powerpc
and sparc64, at least, walk the page tables without taking locks eg. see
find_linux_pte()).

The race is as follows:
The pte page is allocated, zeroed, and its struct page gets its spinlock
initialized. The mm-wide ptl is then taken, and then the pte page is inserted
into the pagetables.

At this point, the spinlock is not guaranteed to have ordered the previous
stores to initialize the pte page with the subsequent store to put it in the
page tables. So another Linux page table walker might be walking down (without
any locks, because we have split-leaf-ptls), and find that new pte we've
inserted. It might try to take the spinlock before the store from the other
CPU initializes it. And subsequently it might read a pte_t out before stores
from the other CPU have cleared the memory.

There are also similar races in higher levels of the page tables. They
obviously don't involve the spinlock, but could see uninitialized memory.

Arch code and hardware pagetable walkers that walk the pagetables without
locks could see similar uninitialized memory problems, regardless of whether
split ptes are enabled or not.

I prefer to put the barriers in core code, because that's where the higher
level logic happens, but the page table accessors are per-arch, and open-coding
them everywhere I don't think is an option. I'll put the read-side barriers
in alpha arch code for now (other architectures perform data-dependent loads
in order).

Signed-off-by: Nick Piggin 
Signed-off-by: Linus Torvalds

x86: fix PAE pmd_bad bootup warning

2008-05-06T20:08:58+00:00

Fix warning from pmd_bad() at bootup on a HIGHMEM64G HIGHPTE x86_32.

That came from 9fc34113f6880b215cbea4e7017fc818700384c2 x86: debug pmd_bad();
but we understand now that the typecasting was wrong for PAE in the previous
version: pagetable pages above 4GB looked bad and stopped Arjan from booting.

And revert that cded932b75ab0a5f9181ee3da34a0a488d1a14fd x86: fix pmd_bad
and pud_bad to support huge pages.  It was the wrong way round: we shouldn't
weaken every pmd_bad and pud_bad check to let huge pages slip through - in
part they check that we _don't_ have a huge page where it's not expected.

Put the x86 pmd_bad() and pud_bad() definitions back to what they have long
been: they can be improved (x86_32 should use PTE_MASK, to stop PAE thinking
junk in the upper word is good; and x86_64 should follow x86_32's stricter
comparison, to stop thinking any subset of required bits is good); but that
should be a later patch.

Fix Hans' good observation that follow_page() will never find pmd_huge()
because that would have already failed the pmd_bad test: test pmd_huge in
between the pmd_none and pmd_bad tests.  Tighten x86's pmd_huge() check?
No, once it's a hugepage entry, it can get quite far from a good pmd: for
example, PROT_NONE leaves it with only ACCESSED of the KERN_PGTABLE bits.

However... though follow_page() contains this and another test for huge
pages, so it's nice to keep it working on them, where does it actually get
called on a huge page?  get_user_pages() checks is_vm_hugetlb_page(vma) to
to call alternative hugetlb processing, as does unmap_vmas() and others.

Signed-off-by: Hugh Dickins 
Earlier-version-tested-by: Ingo Molnar 
Cc: Thomas Gleixner 
Cc: Jeff Chua 
Cc: Hans Rosenfeld 
Cc: Arjan van de Ven 
Signed-off-by: Linus Torvalds

mm: add vm_insert_mixed

2008-04-28T15:58:23+00:00

vm_insert_mixed will insert either a raw pfn or a refcounted struct page into
the page tables, depending on whether vm_normal_page() will return the page or
not.  With the introduction of the new pte bit, this is now a too tricky for
drivers to be doing themselves.

filemap_xip uses this in a subsequent patch.

Signed-off-by: Nick Piggin 
Cc: Jared Hulbert 
Cc: Carsten Otte 
Cc: Martin Schwidefsky 
Cc: Heiko Carstens 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm: introduce pte_special pte bit

2008-04-28T15:58:23+00:00

s390 for one, cannot implement VM_MIXEDMAP with pfn_valid, due to their memory
model (which is more dynamic than most).  Instead, they had proposed to
implement it with an additional path through vm_normal_page(), using a bit in
the pte to determine whether or not the page should be refcounted:

vm_normal_page()
{
	...
        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
                if (vma->vm_flags & VM_MIXEDMAP) {
#ifdef s390
			if (!mixedmap_refcount_pte(pte))
				return NULL;
#else
                        if (!pfn_valid(pfn))
                                return NULL;
#endif
                        goto out;
                }
	...
}

This is fine, however if we are allowed to use a bit in the pte to determine
refcountedness, we can use that to _completely_ replace all the vma based
schemes.  So instead of adding more cases to the already complex vma-based
scheme, we can have a clearly seperate and simple pte-based scheme (and get
slightly better code generation in the process):

vm_normal_page()
{
#ifdef s390
	if (!mixedmap_refcount_pte(pte))
		return NULL;
	return pte_page(pte);
#else
	...
#endif
}

And finally, we may rather make this concept usable by any architecture rather
than making it s390 only, so implement a new type of pte state for this.
Unfortunately the old vma based code must stay, because some architectures may
not be able to spare pte bits.  This makes vm_normal_page a little bit more
ugly than we would like, but the 2 cases are clearly seperate.

So introduce a pte_special pte state, and use it in mm/memory.c.  It is
currently a noop for all architectures, so this doesn't actually result in any
compiled code changes to mm/memory.o.

BTW:
I haven't put vm_normal_page() into arch code as-per an earlier suggestion.
The reason is that, regardless of where vm_normal_page is actually
implemented, the *abstraction* is still exactly the same. Also, while it
depends on whether the architecture has pte_special or not, that is the
only two possible cases, and it really isn't an arch specific function --
the role of the arch code should be to provide primitive functions and
accessors with which to build the core code; pte_special does that. We do
not want architectures to know or care about vm_normal_page itself, and
we definitely don't want them being able to invent something new there
out of sight of mm/ code. If we made vm_normal_page an arch function, then
we have to make vm_insert_mixed (next patch) an arch function too. So I
don't think moving it to arch code fundamentally improves any abstractions,
while it does practically make the code more difficult to follow, for both
mm and arch developers, and easier to misuse.

[akpm@linux-foundation.org: build fix]
Signed-off-by: Nick Piggin 
Acked-by: Carsten Otte 
Cc: Jared Hulbert 
Cc: Martin Schwidefsky 
Cc: Heiko Carstens 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm: introduce VM_MIXEDMAP

2008-04-28T15:58:22+00:00

This series introduces some important infrastructure work.  The overall result
is that:

1. We now support XIP backed filesystems using memory that have no
   struct page allocated to them. And patches 6 and 7 actually implement
   this for s390.

   This is pretty important in a number of cases. As far as I understand,
   in the case of virtualisation (eg. s390), each guest may mount a
   readonly copy of the same filesystem (eg. the distro). Currently,
   guests need to allocate struct pages for this image. So if you have
   100 guests, you already need to allocate more memory for the struct
   pages than the size of the image. I think. (Carsten?)

   For other (eg. embedded) systems, you may have a very large non-
   volatile filesystem. If you have to have struct pages for this, then
   your RAM consumption will go up proportionally to fs size. Even
   though it is just a small proportion, the RAM can be much more costly
   eg in terms of power, so every KB less that Linux uses makes it more
   attractive to a lot of these guys.

2. VM_MIXEDMAP allows us to support mappings where you actually do want
   to refcount _some_ pages in the mapping, but not others, and support
   COW on arbitrary (non-linear) mappings. Jared needs this for his NVRAM
   filesystem in progress. Future iterations of this filesystem will
   most likely want to migrate pages between pagecache and XIP backing,
   which is where the requirement for mixed (some refcounted, some not)
   comes from.

3. pte_special also has a peripheral usage that I need for my lockless
   get_user_pages patch. That was shown to speed up "oltp" on db2 by
   10% on a 2 socket system, which is kind of significant because they
   scrounge for months to try to find 0.1% improvement on these
   workloads. I'm hoping we might finally be faster than AIX on
   pSeries with this :). My reference to lockless get_user_pages is not
   meant to justify this patchset (which doesn't include lockless gup),
   but just to show that pte_special is not some s390 specific thing that
   should be hidden in arch code or xip code: I definitely want to use it
   on at least x86 and powerpc as well.

This patch:

Introduce a new type of mapping, VM_MIXEDMAP.  This is unlike VM_PFNMAP in
that it can support COW mappings of arbitrary ranges including ranges without
struct page *and* ranges with a struct page that we actually want to refcount
(PFNMAP can only support COW in those cases where the un-COW-ed translations
are mapped linearly in the virtual address, and can only support non
refcounted ranges).

VM_MIXEDMAP achieves this by refcounting all pfn_valid pages, and not
refcounting !pfn_valid pages (which is not an option for VM_PFNMAP, because it
needs to avoid refcounting pfn_valid pages eg.  for /dev/mem mappings).

Signed-off-by: Jared Hulbert 
Signed-off-by: Nick Piggin 
Acked-by: Carsten Otte 
Cc: Jared Hulbert 
Cc: Martin Schwidefsky 
Cc: Heiko Carstens 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

mm: remove nopage

2008-04-28T15:58:18+00:00

Nothing in the tree uses nopage any more.  Remove support for it in the
core mm code and documentation (and a few stray references to it in
comments).

Signed-off-by: Nick Piggin 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

memcg: when do_swap's do_wp_page fails

2008-03-05T00:35:14+00:00

Don't uncharge when do_swap_page's call to do_wp_page fails: the page which
was charged for is there in the pagetable, and will be correctly uncharged
when that area is unmapped - it was only its COWing which failed.

And while we're here, remove earlier XXX comment: yes, OR in do_wp_page's
return value (maybe VM_FAULT_WRITE) with do_swap_page's there; but if it
fails, mask out success bits, which might confuse some arches e.g.  sparc.

Signed-off-by: Hugh Dickins 
Cc: David Rientjes 
Acked-by: Balbir Singh 
Acked-by: KAMEZAWA Hiroyuki 
Cc: Hirokazu Takahashi 
Cc: YAMAMOTO Takashi 
Cc: Paul Menage 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds

memcg: page_cache_release not __free_page

2008-03-05T00:35:14+00:00

There's nothing wrong with mem_cgroup_charge failure in do_wp_page and
do_anonymous page using __free_page, but it does look odd when nearby code
uses page_cache_release: use that instead (while turning a blind eye to
ancient inconsistencies of page_cache_release versus put_page).

Signed-off-by: Hugh Dickins 
Cc: David Rientjes 
Acked-by: Balbir Singh 
Acked-by: KAMEZAWA Hiroyuki 
Cc: Hirokazu Takahashi 
Cc: YAMAMOTO Takashi 
Cc: Paul Menage 
Signed-off-by: Andrew Morton 
Signed-off-by: Linus Torvalds