For some obscure reason, we only set init_bootmem_done after initializing
bootmem when NUMA isn't enabled. We even document this next to the declaration
of that global in system.h which of course I didn't read before I had to
debug why some WIP code wasn't working properly...
This patch changes it so that we always set it after bootmem is initialized
which should have always been the case... go figure !
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The MMU context_lock can be taken from switch_mm() while the
rq->lock is held. The rq->lock can also be taken from interrupts,
thus if we get interrupted in destroy_context() with the context
lock held and that interrupt tries to take the rq->lock, there's
a possible deadlock scenario with another CPU having the rq->lock
and calling switch_mm() which takes our context lock.
The fix is to always ensure interrupts are off when taking our
context lock. The switch_mm() path is already good so this fixes
the destroy_context() path.
While at it, turn the context lock into a new style spinlock.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch fixes a couple of issues that can happen as a result
of steal_context() dropping the context_lock when all possible
PIDs are ineligible for stealing (hopefully an extremely hard to
hit occurence).
This case exposes the possibility of a stale context_mm[] entry
to be seen since destroy_context() doesn't clear it and the free
map isn't re-tested. It also means steal_context() will not notice
a context freed while the lock was help, thus possibly trying to
steal a context when a free one was available.
This fixes it by always returning to the caller from steal_context
when it dropped the lock with a return value that causes the
caller to re-samble the number of free contexts, along with
properly clearing the context_mm[] array for destroyed contexts.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The implementation we just revived has issues, such as using a
Kconfig-defined virtual address area in kernel space that nothing
actually carves out (and thus will overlap whatever is there),
or having some dependencies on being self contained in a single
PTE page which adds unnecessary constraints on the kernel virtual
address space.
This fixes it by using more classic PTE accessors and automatically
locating the area for consistent memory, carving an appropriate hole
in the kernel virtual address space, leaving only the size of that
area as a Kconfig option. It also brings some dma-mask related fixes
from the ARM implementation which was almost identical initially but
grew its own fixes.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Make FIXADDR_TOP a compile time constant and cleanup a
couple of definitions relative to the layout of the kernel
address space on ppc32. We also print out that layout at
boot time for debugging purposes.
This is a pre-requisite for properly fixing non-coherent
DMA allocactions.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The recent rework of the MMU PID handling for non-hash CPUs has a
subtle bug in the !SMP "optimized" variant of the PID stealing
function. It clears the PID in the mm context before it calls
local_flush_tlb_mm(). However, the later will not flush anything
if the PID in the context is clear...
Signed-off-by: Hideo Saito <hsaito.ppc@gmail.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This adds the PowerPC 2.06 tlbie mnemonics and keeps backwards
compatibilty for CPUs before 2.06.
Only useful for bare metal systems.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We're currently choking on mem=4g (and above) due to memory_limit
being specified as an unsigned long. Make memory_limit
phys_addr_t to fix this.
Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Previous gcc versions didn't notice this because one of the preceding
#ifs always evaluated to true.
gcc 4.4.0 produced this error:
arch/powerpc/mm/tlb_nohash_low.S:206:6: error: #elif with no expression
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Acked-by: Josh Boyer <jwboyer@linux.vnet.ibm.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
This reverts commit e996557740. Our HW
guys were able to fix this so it never sees the light of day.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
early_init_mmu_secondary() is called at CPU hotplug time, so it
must be marked as __cpuinit, not __init.
Caused by 757c74d2 ("powerpc/mm: Introduce early_init_mmu() on 64-bit").
Tested-by: Sachin Sant <sachinp@in.ibm.com>
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
arch/powerpc/mm/tlb_nohash.c: In function 'flush_tlb_mm':
arch/powerpc/mm/tlb_nohash.c:128: warning: unused variable 'cpu_mask'
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
During the ISA 2.06 development the opcode for tlbilx changed and some
early implementations used to old opcode. Add support for a MMU_FTR
fixup to deal with this.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
This moves some MMU related init code out of setup_64.c into hash_utils_64.c
and calls it early_init_mmu() and early_init_mmu_secondary(). This will
make it easier to plug in a new MMU type.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch tweaks the way some PTE bit combinations are defined, in such a
way that the 32 and 64-bit variant become almost identical and that will
make it easier to bring in a new common pte-* file for the new variant
of the Book3-E support.
The combination of bits defining access to kernel pages are now clearly
separated from the combination used by userspace and the core VM. The
resulting generated code should remain identical unless I made a mistake.
Note: While at it, I removed a non-sensical statement related to CONFIG_KGDB
in ppc_mmu_32.c which could cause kernel mappings to be user accessible when
that option is enabled. Probably something that bitrot.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Makes code futureproof against the impending change to mm->cpu_vm_mask.
It's also a chance to use the new cpumask_ ops which take a pointer
(the older ones are deprecated, but there's no hurry for arch code).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
While we did add support for _PAGE_SPECIAL on some 32-bit platforms,
we never actually built get_user_pages_fast() on them. This fixes
it which requires a little bit of ifdef'ing around.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This adds the necessary bits and pieces to powerpc implementation of
ioremap to benefit from caller tracking in /proc/vmallocinfo, at least
for ioremap's done after mem init as the older ones aren't tracked.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The e500mc core supports the new tlbilx instructions that do core
local invalidates and also provide us the ability to take down
all TLB entries matching a given PID.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
On 64bit there is a possibility our stack and mmap randomisation will put
the two close enough such that we can't expand our stack to match the ulimit
specified.
To avoid this, start the upper mmap address at 1GB + 128MB below the top of our
address space, so in the worst case we end up with the same ~128MB hole as in
32bit. This works because we randomise the stack over a 1GB range.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
get_random_int() returns the same value within a 1 jiffy interval. This means
that the mmap and stack regions will almost always end up the same distance
apart, making a relative offset based attack possible.
To fix this, shift the randomness we use for the mmap region by 1 bit.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Rearrange mmap.c to better match the x86 version.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch reworks the hot_add_scn_to_nid and its supporting functions
to make them easier to understand. There are no functional changes in
this patch and has been tested on machine with memory represented in the
device tree as memory nodes and in the ibm,dynamic-memory property.
My previous patch that introduced support for hotplug memory add on
systems whose memory was represented by the ibm,dynamic-memory property
of the device tree only left the code more unintelligible. This
will hopefully makes things easier to understand.
Signed-off-by: Nathan Fontenot <nfont@austin.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
At the moment we size the hashtable based on 4kB pages / 2, even on a
64kB kernel. This results in a hashtable that is much larger than it
needs to be.
Grab the real page size and size the hashtable based on that
Note: This only has effect on non hypervisor machines.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Fix the powerpc NUMA reserve bootmem page selection logic.
commit 8f64e1f2d1 (powerpc: Reserve
in bootmem lmb reserved regions that cross NUMA nodes) changed
the logic for how the powerpc LMB reserved regions were converted
to bootmen reserved regions. As the folowing discussion reports,
the new logic was not correct.
mark_reserved_regions_for_nid() goes through each LMB on the
system that specifies a reserved area. It searches for
active regions that intersect with that LMB and are on the
specified node. It attempts to bootmem-reserve only the area
where the active region and the reserved LMB intersect. We
can not reserve things on other nodes as they may not have
bootmem structures allocated, yet.
We base the size of the bootmem reservation on two possible
things. Normally, we just make the reservation start and
stop exactly at the start and end of the LMB.
However, the LMB reservations are not aware of NUMA nodes and
on occasion a single LMB may cross into several adjacent
active regions. Those may even be on different NUMA nodes
and will require separate calls to the bootmem reserve
functions. So, the bootmem reservation must be trimmed to
fit inside the current active region.
That's all fine and dandy, but we trim the reservation
in a page-aligned fashion. That's bad because we start the
reservation at a non-page-aligned address: physbase.
The reservation may only span 2 bytes, but that those bytes
may span two pfns and cause a reserve_size of 2*PAGE_SIZE.
Take the case where you reserve 0x2 bytes at 0x0fff and
where the active region ends at 0x1000. You'll jump into
that if() statment, but node_ar.end_pfn=0x1 and
start_pfn=0x0. You'll end up with a reserve_size=0x1000,
and then call
reserve_bootmem_node(node, physbase=0xfff, size=0x1000);
0x1000 may not be on the same node as 0xfff. Oops.
In almost all the vm code, end_<anything> is not inclusive.
If you have an end_pfn of 0x1234, page 0x1234 is not
included in the range. Using PFN_UP instead of the
(>> >> PAGE_SHIFT) will make this consistent with the other VM
code.
We also need to do math for the reserved size with physbase
instead of start_pfn. node_ar.end_pfn << PAGE_SHIFT is
*precisely* the end of the node. However,
(start_pfn << PAGE_SHIFT) is *NOT* precisely the beginning
of the reserved area. That is, of course, physbase.
If we don't use physbase here, the reserve_size can be
made too large.
From: Dave Hansen <dave@linux.vnet.ibm.com>
Tested-by: Geoff Levand <geoffrey.levand@am.sony.com> Tested on PS3.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/mm/fsl_booke_mmu.c: In function 'adjust_total_lowmem':
arch/powerpc/mm/fsl_booke_mmu.c:221: warning: format '%ld' expects type 'long int', but argument 3 has type 'phys_addr_t'
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
The Power ISA 2.06 added power of two page sizes to the embedded MMU
architecture. Its done it such a way to be code compatiable with the
existing HW. Made the minor code changes to support both power of two
and power of four page sizes. Also added some new MAS bits and macros
that are defined as part of the 2.06 ISA. Renamed some things to use
the 'Book-3e' concept to convey the new MMU that is based on the
Freescale Book-E MMU programming model.
Note, its still invalid to try and use a page size that isn't supported
by cpu.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
The following commit:
commit 64b3d0e812
Author: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Date: Thu Dec 18 19:13:51 2008 +0000
powerpc/mm: Rework usage of _PAGE_COHERENT/NO_CACHE/GUARDED
broke setting of the _PAGE_COHERENT bit in the PPC HW PTE. Since we now
actually set _PAGE_COHERENT in the Linux PTE we shouldn't be clearing it
out before we propogate it to the PPC HW PTE.
Reported-by: Martyn Welch <martyn.welch@gefanuc.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This patch reworks the way we do I and D cache coherency on PowerPC.
The "old" way was split in 3 different parts depending on the processor type:
- Hash with per-page exec support (64-bit and >= POWER4 only) does it
at hashing time, by preventing exec on unclean pages and cleaning pages
on exec faults.
- Everything without per-page exec support (32-bit hash, 8xx, and
64-bit < POWER4) does it for all page going to user space in update_mmu_cache().
- Embedded with per-page exec support does it from do_page_fault() on
exec faults, in a way similar to what the hash code does.
That leads to confusion, and bugs. For example, the method using update_mmu_cache()
is racy on SMP where another processor can see the new PTE and hash it in before
we have cleaned the cache, and then blow trying to execute. This is hard to hit but
I think it has bitten us in the past.
Also, it's inefficient for embedded where we always end up having to do at least
one more page fault.
This reworks the whole thing by moving the cache sync into two main call sites,
though we keep different behaviours depending on the HW capability. The call
sites are set_pte_at() which is now made out of line, and ptep_set_access_flags()
which joins the former in pgtable.c
The base idea for Embedded with per-page exec support, is that we now do the
flush at set_pte_at() time when coming from an exec fault, which allows us
to avoid the double fault problem completely (we can even improve the situation
more by implementing TLB preload in update_mmu_cache() but that's for later).
If for some reason we didn't do it there and we try to execute, we'll hit
the page fault, which will do a minor fault, which will hit ptep_set_access_flags()
to do things like update _PAGE_ACCESSED or _PAGE_DIRTY if needed, we just make
this guys also perform the I/D cache sync for exec faults now. This second path
is the catch all for things that weren't cleaned at set_pte_at() time.
For cpus without per-pag exec support, we always do the sync at set_pte_at(),
thus guaranteeing that when the PTE is visible to other processors, the cache
is clean.
For the 64-bit hash with per-page exec support case, we keep the old mechanism
for now. I'll look into changing it later, once I've reworked a bit how we
use _PAGE_EXEC.
This is also a first step for adding _PAGE_EXEC support for embedded platforms
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Use of_get_cpu_node, which is a superset of numa.c's find_cpu_node in
a less restrictive section (text vs cpuinit).
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
find_min_common_depth() was checking the property length incorrectly.
The value is in bytes not cells, and it is using the second entry.
Signed-off-By: Milton Miller <miltonm@bga.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Fixed v_mapped_by_tlbcam() and p_mapped_by_tlbcam() to use phys_addr_t
instead of unsigned long. In 36-bit physical mode we really need these
functions to deal with phys_addr_t when trying to match a physical
address or when returning one.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
On booke processors, the code that maps low memory only uses up to three
CAM entries, even though there are sixteen and nothing else uses them.
Make this number configurable in the advanced options menu along with max
low memory size. If one wants 1 GB of lowmem, then it's typically
necessary to have four CAM entries.
Signed-off-by: Trent Piepho <tpiepho@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
The code that maps kernel low memory would only use page sizes up to 256
MB. On E500v2 pages up to 4 GB are supported.
However, a page must be aligned to a multiple of the page's size. I.e.
256 MB pages must aligned to a 256 MB boundary. This was enforced by a
requirement that the physical and virtual addresses of the start of lowmem
be aligned to 256 MB. Clearly requiring 1GB or 4GB alignment to allow
pages of that size isn't acceptable.
To solve this, I simply have adjust_total_lowmem() take alignment into
account when it decides what size pages to use. Give it PAGE_OFFSET =
0x7000_0000, PHYSICAL_START = 0x3000_0000, and 2GB of RAM, and it will map
pages like this:
PA 0x3000_0000 VA 0x7000_0000 Size 256 MB
PA 0x4000_0000 VA 0x8000_0000 Size 1 GB
PA 0x8000_0000 VA 0xC000_0000 Size 256 MB
PA 0x9000_0000 VA 0xD000_0000 Size 256 MB
PA 0xA000_0000 VA 0xE000_0000 Size 256 MB
Because the lowmem mapping code now takes alignment into account,
PHYSICAL_ALIGN can be lowered from 256 MB to 64 MB. Even lower might be
possible. The lowmem code will work down to 4 kB but it's possible some of
the boot code will fail before then. Poor alignment will force small pages
to be used, which combined with the limited number of TLB1 pages available,
will result in very little memory getting mapped. So alignments less than
64 MB probably aren't very useful anyway.
Signed-off-by: Trent Piepho <tpiepho@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
The code to map lowmem uses three CAM aka TLB[1] entries to cover it. The
size of each is stored in three globals named __cam0, __cam1, and __cam2.
All the code that uses them is duplicated three times for each of the three
variables.
We have these things called arrays and loops....
Once converted to use an array, it will be easier to make the number of
CAMs configurable.
Signed-off-by: Trent Piepho <tpiepho@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
_PAGE_COHERENT is now always set in _PAGE_RAM resp. PAGE_KERNEL.
Thus it has to be masked out, if the BAT mapping should be non
cacheable or CPU_FTR_NEED_COHERENT is not set.
This will work on normal SMP setups because we force-set
CPU_FTR_NEED_COHERENT as part of CPU_FTR_COMMON on SMP.
Signed-off-by: Gerhard Pircher <gerhard_pircher@gmx.net>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
powerpc: is_hugepage_only_range() must account for both 4kB and 64kB slices
The subpage_prot syscall fails on second and subsequent calls for a given
region, because is_hugepage_only_range() is mis-identifying the 4 kB
slices when the process has a 64 kB page size.
Signed-off-by: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Convert arch/powerpc/ over to long long based u64:
-#ifdef __powerpc64__
-# include <asm-generic/int-l64.h>
-#else
-# include <asm-generic/int-ll64.h>
-#endif
+#include <asm-generic/int-ll64.h>
This will avoid reoccuring spurious warnings in core kernel code that
comes when people test on their own hardware. (i.e. x86 in ~98% of the
cases) This is what x86 uses and it generally helps keep 64-bit code
32-bit clean too.
[Adjusted to not impact user mode (from paulus) - sfr]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The clear_fixmap() routine issues map_page() with flags set to 0.
Currently this causes a BUG_ON() inside the map_page(), as it assumes
that a PTE should be clear before mapping.
This patch makes the map_page() to trigger the BUG_ON() only if the
flags were set.
Signed-off-by: Anton Vorontsov <avorontsov@ru.mvista.com>
Acked-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
