(David:)

If hugetlbfs_file_mmap() returns a failure to do_mmap_pgoff() - for example,
because the given file offset is not hugepage aligned - then do_mmap_pgoff
will go to the unmap_and_free_vma backout path.

But at this stage the vma hasn't been marked as hugepage, and the backout path
will call unmap_region() on it.  That will eventually call down to the
non-hugepage version of unmap_page_range().  On ppc64, at least, that will
cause serious problems if there are any existing hugepage pagetable entries in
the vicinity - for example if there are any other hugepage mappings under the
same PUD.  unmap_page_range() will trigger a bad_pud() on the hugepage pud
entries.  I suspect this will also cause bad problems on ia64, though I don't
have a machine to test it on.

(Hugh:)

prepare_hugepage_range() should check file offset alignment when it checks
virtual address and length, to stop MAP_FIXED with a bad huge offset from
unmapping before it fails further down.  PowerPC should apply the same
prepare_hugepage_range alignment checks as ia64 and all the others do.

Then none of the alignment checks in hugetlbfs_file_mmap are required (nor
is the check for too small a mapping); but even so, move up setting of
VM_HUGETLB and add a comment to warn of what David Gibson discovered - if
hugetlbfs_file_mmap fails before setting it, do_mmap_pgoff's unmap_region
when unwinding from error will go the non-huge way, which may cause bad
behaviour on architectures (powerpc and ia64) which segregate their huge
mappings into a separate region of the address space.
Signed-off-by: NHugh Dickins <hugh@veritas.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Acked-by: NAdam Litke <agl@us.ibm.com>
Acked-by: NDavid Gibson <david@gibson.dropbear.id.au>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Checking source for other get_paca()->field preemption dangers found that
open_high_hpage_areas does a structure copy into its paca while preemption
is enabled: unsafe however gcc accomplishes it.  Just remove that copy:
it's done safely afterwards by on_each_cpu, as in open_low_hpage_areas.
Signed-off-by: NHugh Dickins <hugh@veritas.com>
Acked-by: NDavid Gibson <dwg@au1.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

On Tue, 2006-08-15 at 08:22 -0700, Dave Hansen wrote:
> kernel BUG in cache_free_debugcheck at mm/slab.c:2748!

Alright, this one is only triggered when slab debugging is enabled.  The
slabs are assumed to be aligned on a HUGEPTE_TABLE_SIZE boundary.  The free
path makes use of this assumption and uses the lowest nibble to pass around
an index into an array of kmem_cache pointers.  With slab debugging turned
on, the slab is still aligned, but the "working" object pointer is not.
This would break the assumption above that a full nibble is available for
the PGF_CACHENUM_MASK.

The following patch reduces PGF_CACHENUM_MASK to cover only the two least
significant bits, which is enough to cover the current number of 4 pgtable
cache types.  Then use this constant to mask out the appropriate part of
the huge pte pointer.
Signed-off-by: NAdam Litke <agl@us.ibm.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

At present, ARCH=powerpc kernels can waste considerable space in
pagetables when making large hugepage mappings.  Hugepage PTEs go in
PMD pages, but each PMD page maps 256M and so contains only 16
hugepage PTEs (128 bytes of data), but takes up a 1024 byte
allocation.  With CONFIG_PPC_64K_PAGES enabled (64k base page size),
the situation is worse.  Now hugepage PTEs are at the PTE page level
(also mapping 256M), so we store 16 hugepage PTEs in a 64k allocation.

The PowerPC MMU already means that any 256M region is either all
hugepage, or all normal pages.  Thus, with some care, we can use a
different allocation for the hugepage PTE tables and only allocate the
128 bytes necessary.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Quite a long time back, prepare_hugepage_range() replaced
is_aligned_hugepage_range() as the callback from mm/mmap.c to arch code to
verify if an address range is suitable for a hugepage mapping.
is_aligned_hugepage_range() stuck around, but only to implement
prepare_hugepage_range() on archs which didn't implement their own.

Most archs (everything except ia64 and powerpc) used the same
implementation of is_aligned_hugepage_range().  On powerpc, which
implements its own prepare_hugepage_range(), the custom version was never
used.

In addition, "is_aligned_hugepage_range()" was a bad name, because it
suggests it returns true iff the given range is a good hugepage range,
whereas in fact it returns 0-or-error (so the sense is reversed).

This patch cleans up by abolishing is_aligned_hugepage_range().  Instead
prepare_hugepage_range() is defined directly.  Most archs use the default
version, which simply checks the given region is aligned to the size of a
hugepage.  ia64 and powerpc define custom versions.  The ia64 one simply
checks that the range is in the correct address space region in addition to
being suitably aligned.  The powerpc version (just as previously) checks
for suitable addresses, and if necessary performs low-level MMU frobbing to
set up new areas for use by hugepages.

No libhugetlbfs testsuite regressions on ppc64 (POWER5 LPAR).
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NZhang Yanmin <yanmin.zhang@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: William Lee Irwin III <wli@holomorphy.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Building the arch/powerpc tree currently gives me
two warnings with gcc-4.0:

arch/powerpc/mm/imalloc.c: In function '__im_get_area':
arch/powerpc/mm/imalloc.c:225: warning: 'tmp' may be used uninitialized in this function
arch/powerpc/mm/hugetlbpage.c: In function 'hugetlb_get_unmapped_area':
arch/powerpc/mm/hugetlbpage.c:608: warning: unused variable 'vma'

both fixes are trivial.
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Currently, the powerpc version of hugetlb_get_unmapped_area() entirely
ignores the hint address.  The only way to get a hugepage mapping at a
specified address is with MAP_FIXED, in which case there's no way
(short of parsing /proc/self/maps) for userspace to tell if it will
clobber an existing mapping.  This is inconvenient, so the patch below
makes hugepage mappings use the given hint address if possible.
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

On ppc64, when opening a new hugepage region, we need to make sure any
old normal-page SLBs for the area are flushed on all CPUs.  There was
a bug in this logic - after putting the new hugepage area masks into
the thread structure, we copied it into the paca (read by the SLB miss
handler) only on one CPU, not on all.  This could cause incorrect SLB
entries to be loaded when a multithreaded program was running
simultaneously on several CPUs.  This patch corrects the error,
copying the context information into the PACA on all CPUs using the mm
in question before flushing any existing SLB entries.
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

On most powerpc CPUs, the dcache and icache are not coherent so
between writing and executing a page, the caches must be flushed.
Userspace programs assume pages given to them by the kernel are icache
clean, so we must do this flush between the kernel clearing a page and
it being mapped into userspace for execute.  We were not doing this
for hugepages, this patch corrects the situation.

We use the same lazy mechanism as we use for normal pages, delaying
the flush until userspace actually attempts to execute from the page
in question.

Tested on G5.
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

The 64k pages patch changed the meaning of one argument passed to the
low level hash functions (from "large" it became "psize" or page size
index), but one of the call sites wasn't properly updated, causing
potential random weird problems with huge pages. This fixes it.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Blah.  The patch [0] I recently sent fixing errors with
in_hugepage_area() and prepare_hugepage_range() for powerpc itself has
an off-by-one bug.  Furthermore, the related functions
touches_hugepage_*_range() and within_hugepage_*_range() are also
buggy.  Some of the bugs, like those addressed in [0] originated with
commit 7d24f0b8 where we tweaked the
semantics of where hugepages are allowed.  Other bugs have been there
essentially forever, and are due to the undefined behaviour of '<<'
with shift counts greater than the type width (LOW_ESID_MASK could
return non-zero for high ranges with the right congruences).

The good news is that I now have a testsuite which should pick up
things like this if they creep in again.

[0] "powerpc-fix-for-hugepage-areas-straddling-4gb-boundary"
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Commit 7d24f0b8 fixed bugs in the ppc64 SLB
miss handler with respect to hugepage handling, and in the process tweaked
the semantics of the hugepage address masks in mm_context_t.

Unfortunately, it left out a couple of necessary changes to go with that
change.  First, the in_hugepage_area() macro was not updated to match,
second prepare_hugepage_range() was not updated to correctly handle
hugepages regions which straddled the 4GB point.

The latter appears only to cause process-hangs when attempting to map such
a region, but the former can cause oopses if a get_user_pages() is
triggered at the wrong point.  This patch addresses both bugs.
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Update comments (only) on page_table_lock and mmap_sem in arch/powerpc.
Removed the comment on page_table_lock from hash_huge_page: since it's no
longer taking page_table_lock itself, it's irrelevant whether others are; but
how it is safe (even against huge file truncation?) I can't say.
Signed-off-by: NHugh Dickins <hugh@veritas.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This patch, however, should be applied on top of the 64k-page-size patch to
fix some problems with hugepage (some pre-existing, another introduced by
this patch).

The patch fixes a bug in the SLB miss handler for hugepages on ppc64
introduced by the dynamic hugepage patch (commit id
c594adad) due to a misunderstanding of the
srd instruction's behaviour (mea culpa).  The problem arises when a 64-bit
process maps some hugepages in the low 4GB of the address space (unusual).
In this case, as well as the 256M segment in question being marked for
hugepages, other segments at 32G intervals will be incorrectly marked for
hugepages.

In the process, this patch tweaks the semantics of the hugepage bitmaps to
be more sensible.  Previously, an address below 4G was marked for hugepages
if the appropriate segment bit in the "low areas" bitmask was set *or* if
the low bit in the "high areas" bitmap was set (which would mark all
addresses below 1TB for hugepage).  With this patch, any given address is
governed by a single bitmap.  Addresses below 4GB are marked for hugepage
if and only if their bit is set in the "low areas" bitmap (256M
granularity).  Addresses between 4GB and 1TB are marked for hugepage iff
the low bit in the "high areas" bitmap is set.  Higher addresses are marked
for hugepage iff their bit in the "high areas" bitmap is set (1TB
granularity).

To avoid conflicts, this patch must be applied on top of BenH's pending
patch for 64k base page size [0].  As such, this patch also addresses a
hugepage problem introduced by that patch.  That patch allows hugepages of
1MB in size on hardware which supports it, however, that won't work when
using 4k pages (4 level pagetable), because in that case hugepage PTEs are
stored at the PMD level, and each PMD entry maps 2MB.  This patch simply
disallows hugepages in that case (we can do something cleverer to re-enable
them some other day).

Built, booted, and a handful of hugepage related tests passed on POWER5
LPAR (both ARCH=powerpc and ARCH=ppc64).

[0] http://gate.crashing.org/~benh/ppc64-64k-pages.diffSigned-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Adds a new CONFIG_PPC_64K_PAGES which, when enabled, changes the kernel
base page size to 64K.  The resulting kernel still boots on any
hardware.  On current machines with 4K pages support only, the kernel
will maintain 16 "subpages" for each 64K page transparently.

Note that while real 64K capable HW has been tested, the current patch
will not enable it yet as such hardware is not released yet, and I'm
still verifying with the firmware architects the proper to get the
information from the newer hypervisors.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This moves the remaining files in arch/ppc64/mm to arch/powerpc/mm,
and arranges that we use them when compiling with ARCH=ppc64.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Current kernel has a couple of sneaky bugs in the ppc64 hugetlb code that
cause huge pages to be potentially left stale in the hash table and TLBs
(improperly invalidated), with all the nasty consequences that can have.

One is that we forgot to set the "secondary" bit in the hash PTEs when
hashing a huge page in the secondary bucket (fortunately very rare).

The other one is on non-LPAR machines (like Apple G5s), flush_hash_range()
which is used to flush a batch of PTEs simply did not work for huge pages.
Historically, our huge page code didn't batch, but this was changed without
fixing this routine.  This patch fixes both.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Currently, we set the class bit in kernel SLB entries, and clear it on
user SLB entries.  On POWER5, ERAT entries created in real mode have
the class bit clear.  So to avoid flushing kernel ERAT entries on each
context switch, this patch inverts our usage of the class bit, setting
it on user SLB entries and clearing it on kernel SLB entries.

Booted on POWER5 and G5.
Signed-off-by: NDavid Gibson <dwg@au1.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Paulus, I think this is now a reasonable candidate for the post-2.6.13
queue.

Relax address restrictions for hugepages on ppc64

Presently, 64-bit applications on ppc64 may only use hugepages in the
address region from 1-1.5T.  Furthermore, if hugepages are enabled in
the kernel config, they may only use hugepages and never normal pages
in this area.  This patch relaxes this restriction, allowing any
address to be used with hugepages, but with a 1TB granularity.  That
is if you map a hugepage anywhere in the region 1TB-2TB, that entire
area will be reserved exclusively for hugepages for the remainder of
the process's lifetime.  This works analagously to hugepages in 32-bit
applications, where hugepages can be mapped anywhere, but with 256MB
(mmu segment) granularity.

This patch applies on top of the four level pagetable patch
(http://patchwork.ozlabs.org/linuxppc64/patch?id=1936).
Signed-off-by: NDavid Gibson <dwg@au1.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Implement 4-level pagetables for ppc64

This patch implements full four-level page tables for ppc64, thereby
extending the usable user address range to 44 bits (16T).

The patch uses a full page for the tables at the bottom and top level,
and a quarter page for the intermediate levels.  It uses full 64-bit
pointers at every level, thus also increasing the addressable range of
physical memory.  This patch also tweaks the VSID allocation to allow
matching range for user addresses (this halves the number of available
contexts) and adds some #if and BUILD_BUG sanity checks.
Signed-off-by: NDavid Gibson <dwg@au1.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

This patch removes the use of bitfield types from the ppc64 hash table
manipulation code.
Signed-off-by: NDavid Gibson <dwg@au1.ibm.com>
Acked-by: NStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Ingo recently introduced a great speedup for allocating new mmaps using the
free_area_cache pointer which boosts the specweb SSL benchmark by 4-5% and
causes huge performance increases in thread creation.

The downside of this patch is that it does lead to fragmentation in the
mmap-ed areas (visible via /proc/self/maps), such that some applications
that work fine under 2.4 kernels quickly run out of memory on any 2.6
kernel.

The problem is twofold:

  1) the free_area_cache is used to continue a search for memory where
     the last search ended.  Before the change new areas were always
     searched from the base address on.

     So now new small areas are cluttering holes of all sizes
     throughout the whole mmap-able region whereas before small holes
     tended to close holes near the base leaving holes far from the base
     large and available for larger requests.

  2) the free_area_cache also is set to the location of the last
     munmap-ed area so in scenarios where we allocate e.g.  five regions of
     1K each, then free regions 4 2 3 in this order the next request for 1K
     will be placed in the position of the old region 3, whereas before we
     appended it to the still active region 1, placing it at the location
     of the old region 2.  Before we had 1 free region of 2K, now we only
     get two free regions of 1K -> fragmentation.

The patch addresses thes issues by introducing yet another cache descriptor
cached_hole_size that contains the largest known hole size below the
current free_area_cache.  If a new request comes in the size is compared
against the cached_hole_size and if the request can be filled with a hole
below free_area_cache the search is started from the base instead.

The results look promising: Whereas 2.6.12-rc4 fragments quickly and my
(earlier posted) leakme.c test program terminates after 50000+ iterations
with 96 distinct and fragmented maps in /proc/self/maps it performs nicely
(as expected) with thread creation, Ingo's test_str02 with 20000 threads
requires 0.7s system time.

Taking out Ingo's patch (un-patch available per request) by basically
deleting all mentions of free_area_cache from the kernel and starting the
search for new memory always at the respective bases we observe: leakme
terminates successfully with 11 distinctive hardly fragmented areas in
/proc/self/maps but thread creating is gringdingly slow: 30+s(!) system
time for Ingo's test_str02 with 20000 threads.

Now - drumroll ;-) the appended patch works fine with leakme: it ends with
only 7 distinct areas in /proc/self/maps and also thread creation seems
sufficiently fast with 0.71s for 20000 threads.
Signed-off-by: NWolfgang Wander <wwc@rentec.com>
Credit-to: "Richard Purdie" <rpurdie@rpsys.net>
Signed-off-by: NKen Chen <kenneth.w.chen@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu> (partly)
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

A lot of the code in arch/*/mm/hugetlbpage.c is quite similar.  This patch
attempts to consolidate a lot of the code across the arch's, putting the
combined version in mm/hugetlb.c.  There are a couple of uglyish hacks in
order to covert all the hugepage archs, but the result is a very large
reduction in the total amount of code.  It also means things like hugepage
lazy allocation could be implemented in one place, instead of six.

Tested, at least a little, on ppc64, i386 and x86_64.

Notes:
	- this patch changes the meaning of set_huge_pte() to be more
	  analagous to set_pte()
	- does SH4 need s special huge_ptep_get_and_clear()??
Acked-by: NWilliam Lee Irwin <wli@holomorphy.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This patch converts ppc64 to use the generic pgtable-nopud.h instead of the
"fixup" header.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Once we're strict about clearing away page tables, hugetlb_prefault can assume
there are no page tables left within its range.  Since the other arches
continue if !pte_none here, let i386 do the same.
Signed-off-by: NHugh Dickins <hugh@veritas.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

ia64 and ppc64 had hugetlb_free_pgtables functions which were no longer being
called, and it wasn't obvious what to do about them.

The ppc64 case turns out to be easy: the associated tables are noted elsewhere
and freed later, safe to either skip its hugetlb areas or go through the
motions of freeing nothing.  Since ia64 does need a special case, restore to
ppc64 the special case of skipping them.

The ia64 hugetlb case has been broken since pgd_addr_end went in, though it
probably appeared to work okay if you just had one such area; in fact it's
been broken much longer if you consider a long munmap spanning from another
region into the hugetlb region.

In the ia64 hugetlb region, more virtual address bits are available than in
the other regions, yet the page tables are structured the same way: the page
at the bottom is larger.  Here we need to scale down each addr before passing
it to the standard free_pgd_range.  Was about to write a hugely_scaled_down
macro, but found htlbpage_to_page already exists for just this purpose.  Fixed
off-by-one in ia64 is_hugepage_only_range.

Uninline free_pgd_range to make it available to ia64.  Make sure the
vma-gathering loop in free_pgtables cannot join a hugepage_only_range to any
other (safe to join huges?  probably but don't bother).
Signed-off-by: NHugh Dickins <hugh@veritas.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!