When demoting a process to use 4K HW pages (instead of 64K), which
happens under various circumstances such as doing cache inhibited
mappings on machines that do not support 64K CI pages, the assembly
hash code calls back into the C function flush_hash_page(). This
function prototype was recently changed to accomodate for 1T segments
but the assembly call site was not updated, causing applications that
do demotion to hang. In addition, when updating the per-CPU PACA for
the new sizes, we didn't properly update the slice "map", thus causing
the SLB miss code to re-insert segments for the wrong size.
This fixes both and adds a warning comment next to the C
implementation to try to avoid problems next time someone changes it.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This makes the kernel use 1TB segments for all kernel mappings and for
user addresses of 1TB and above, on machines which support them
(currently POWER5+, POWER6 and PA6T).
We detect that the machine supports 1TB segments by looking at the
ibm,processor-segment-sizes property in the device tree.
We don't currently use 1TB segments for user addresses < 1T, since
that would effectively prevent 32-bit processes from using huge pages
unless we also had a way to revert to using 256MB segments. That
would be possible but would involve extra complications (such as
keeping track of which segment size was used when HPTEs were inserted)
and is not addressed here.
Parts of this patch were originally written by Ben Herrenschmidt.
Signed-off-by: Paul Mackerras <paulus@samba.org>
The code for mapping special 4k pages on kernels using a 64kB base
page size was missing the code for doing the RPN (real page number)
manipulation when inserting the hardware PTE in the secondary hash
bucket. It needs the same code as has already been added to the
code that inserts the HPTE in the primary hash bucket. This adds it.
Spotted by Ben Herrenschmidt.
Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds the ability for a kernel compiled with 4K page size
to have special slices containing 64K pages and hash the right type
of hash PTEs.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Some drivers have resources that they want to be able to map into
userspace that are 4k in size. On a kernel configured with 64k pages
we currently end up mapping the 4k we want plus another 60k of
physical address space, which could contain anything. This can
introduce security problems, for example in the case of an infiniband
adaptor where the other 60k could contain registers that some other
program is using for its communications.
This patch adds a new function, remap_4k_pfn, which drivers can use to
map a single 4k page to userspace regardless of whether the kernel is
using a 4k or a 64k page size. Like remap_pfn_range, it would
typically be called in a driver's mmap function. It only maps a
single 4k page, which on a 64k page kernel appears replicated 16 times
throughout a 64k page. On a 4k page kernel it reduces to a call to
remap_pfn_range.
The way this works on a 64k kernel is that a new bit, _PAGE_4K_PFN,
gets set on the linux PTE. This alters the way that __hash_page_4K
computes the real address to put in the HPTE. The RPN field of the
linux PTE becomes the 4k RPN directly rather than being interpreted as
a 64k RPN. Since the RPN field is 32 bits, this means that physical
addresses being mapped with remap_4k_pfn have to be below 2^44,
i.e. 0x100000000000.
The patch also factors out the code in arch/powerpc/mm/hash_utils_64.c
that deals with demoting a process to use 4k pages into one function
that gets called in the various different places where we need to do
that. There were some discrepancies between exactly what was done in
the various places, such as a call to spu_flush_all_slbs in one case
but not in others.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Some POWER5+ machines can do 64k hardware pages for normal memory but
not for cache-inhibited pages. This patch lets us use 64k hardware
pages for most user processes on such machines (assuming the kernel
has been configured with CONFIG_PPC_64K_PAGES=y). User processes
start out using 64k pages and get switched to 4k pages if they use any
non-cacheable mappings.
With this, we use 64k pages for the vmalloc region and 4k pages for
the imalloc region. If anything creates a non-cacheable mapping in
the vmalloc region, the vmalloc region will get switched to 4k pages.
I don't know of any driver other than the DRM that would do this,
though, and these machines don't have AGP.
When a region gets switched from 64k pages to 4k pages, we do not have
to clear out all the 64k HPTEs from the hash table immediately. We
use the _PAGE_COMBO bit in the Linux PTE to indicate whether the page
was hashed in as a 64k page or a set of 4k pages. If hash_page is
trying to insert a 4k page for a Linux PTE and it sees that it has
already been inserted as a 64k page, it first invalidates the 64k HPTE
before inserting the 4k HPTE. The hash invalidation routines also use
the _PAGE_COMBO bit, to determine whether to look for a 64k HPTE or a
set of 4k HPTEs to remove. With those two changes, we can tolerate a
mix of 4k and 64k HPTEs in the hash table, and they will all get
removed when the address space is torn down.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Our MMU hash management code would not set the "C" bit (changed bit) in
the hardware PTE when updating a RO PTE into a RW PTE. That would cause
the hardware to possibly to a write back to the hash table to set it on
the first store access, which in addition to being a performance issue,
might also hit a bug when running with native hash management (non-HV)
as our code is specifically optimized for the case where no write back
happens.
Thus there is a very small therocial window were a hash PTE can become
corrupted if that HPTE has just been upgraded to read write, a store
access happens on it, and that races with another processor evicting
that same slot. Since eviction (caused by an almost full hash) is
extremely rare, the bug is very unlikely to happen fortunately.
This fixes by allowing the updating of the protection bits in the native
hash handling to also set (but not clear) the "C" bit, and, in order to
also improve performances in the general case, by always setting that
bit on newly inserted hash PTE so that writeback really never happens.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.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: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus 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: Paul Mackerras <paulus@samba.org>
The {BEGIN,END}_FTR_SECTION asm macros used in ppc64 to nop out
sections of code at runtime cannot be nested. However, we do nest
them in hash_low.S. We get away with it there, because there is
nothing between the BEGIN markers for each section. However, that's
confusing to someone reading the code.
This patch removes the nested ifset and ifclr feature sections,
replacing them with a single feature section in the full mask/value
form.
Signed-off-by: David Gibson <dwg@au1.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch removes the use of bitfield types from the ppc64 hash table
manipulation code.
Signed-off-by: David Gibson <dwg@au1.ibm.com>
Acked-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It turns out that our current __hash_page code will do a very hot busy-wait
loop waiting on _PAGE_BUSY to be cleared. It even does ldarx/stdcx in the
loop, which will bounce reservations around like crazy if there's more than
one CPU spinning on the same PTE (or even another PTE in the same
reservation granule). The end result is that each fault takes longer when
there's contention, which in turn increases the chance of another thread
hitting the same fault and also piling up. Not pretty.
There's two options here:
1. Do an out-of-line busy loop a'la spinlocks with just loads (no
reserves)
2. Just bail and refault if needed.
(2) makes sense here: If the PTE is busy, chances are it's in flux anyway
and the other code path making a change might just be ready to hash it.
This fixes a stampede seen on a large-ish system where a multithreaded
HPC app faults in the same text pages on several cpus at the same time.
Signed-off-by: Olof Johansson <olof@lixom.net>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus 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!
Benjamin Herrenschmidt