__find_linux_pte() is the only function in hugetlbpage.c
which is compiled in regardless on CONFIG_HUGETLBPAGE

This patch moves it in pgtable.c.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

As per Kconfig.cputype, only CONFIG_PPC_FSL_BOOK3E gets to
select SYS_SUPPORTS_HUGETLBFS so simplify accordingly.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

CONFIG_PPC_64K_PAGES cannot be selected by nohash/64.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This function is not used anymore, drop it.

Fixes: b42279f0 ("powerpc/mm/nohash: MM_SLICE is only used by book3s 64")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This patch defines a subarch specific SLB_ADDR_LIMIT_DEFAULT
to remove the #ifdefs around the setup of mm->context.slb_addr_limit

It also generalises the use of mm_ctx_set_slb_addr_limit() helper.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

get_slice_psize() can be defined regardless of CONFIG_PPC_MM_SLICES
to avoid ifdefs
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

The 8xx only selects CONFIG_PPC_MM_SLICES when CONFIG_HUGETLB_PAGE
is set.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This patch replaces a couple of #ifdef CONFIG_PPC_64K_PAGES
by IS_ENABLED(CONFIG_PPC_64K_PAGES) to improve code maintainability.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

For PPC32 that's a noop, gcc should be smart enough to ignore it.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Now that slice_mask_for_size() is in mmu.h, the mm_ctx_slice_mask_xxx()
are not needed anymore, so drop them. Note that the 8xx ones where
not used anyway.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Move slice_mask_for_size() into subarch mmu.h
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
[mpe: Retain the BUG_ON()s, rather than converting to VM_BUG_ON()]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

slice_mask_for_size() only uses mm->context, so hand directly a
pointer to the context. This will help moving the function in
subarch mmu.h in the next patch by avoiding having to include
the definition of struct mm_struct
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Only nohash/32 and book3s/64 support mm slices.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Commit 67fda38f ("powerpc/mm: Move slb_addr_linit to
early_init_mmu") moved slb_addr_limit init out of setup_arch().

Commit 70110186 ("powerpc/mm: Reduce memory usage for mm_context_t
for radix") brought it back into setup_arch() by error.

This patch reverts that erroneous regress.

Fixes: 70110186 ("powerpc/mm: Reduce memory usage for mm_context_t for radix")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Many files in arch/powerpc/mm are only for nohash. This patch
creates a subdirectory for them.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
[mpe: Shorten new filenames]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Several files in arch/powerpc/mm are only for book3S32. This patch
creates a subdirectory for them.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
[mpe: Shorten new filenames]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Many files in arch/powerpc/mm are only for book3S64. This patch
creates a subdirectory for them.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
[mpe: Update the selftest sym links, shorten new filenames, cleanup some
      whitespace and formatting in the new files.]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This patch make inclusion of mmu_decl.h independant of the location
of the file including it.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

TRANSPARENT_HUGEPAGE is only supported by book3s

VMEMMAP_REGION_ID is never used
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

early_alloc_pgtable() never returns NULL as it panics on failure.

This patch drops the three BUG_ON() which check the non nullity
of early_alloc_pgtable() returned value.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

There is a kernel crash that happens if rt_sigreturn() is called inside
a transactional block.

This crash happens if the kernel hits an in-kernel page fault when
accessing userspace memory, usually through copy_ckvsx_to_user(). A
major page fault calls might_sleep() function, which can cause a task
reschedule. A task reschedule (switch_to()) reclaim and recheckpoint
the TM states, but, in the signal return path, the checkpointed memory
was already reclaimed, thus the exception stack has MSR that points to
MSR[TS]=0.

When the code returns from might_sleep() and a task reschedule
happened, then this task is returned with the memory recheckpointed,
and CPU MSR[TS] = suspended.

This means that there is a side effect at might_sleep() if it is
called with CPU MSR[TS] = 0 and the task has regs->msr[TS] != 0.

This side effect can cause a TM bad thing, since at the exception
entrance, the stack saves MSR[TS]=0, and this is what will be used at
RFID, but, the processor has MSR[TS] = Suspended, and this transition
will be invalid and a TM Bad thing will be raised, causing the
following crash:

  Unexpected TM Bad Thing exception at c00000000000e9ec (msr 0x8000000302a03031) tm_scratch=800000010280b033
  cpu 0xc: Vector: 700 (Program Check) at [c00000003ff1fd70]
      pc: c00000000000e9ec: fast_exception_return+0x100/0x1bc
      lr: c000000000032948: handle_rt_signal64+0xb8/0xaf0
      sp: c0000004263ebc40
     msr: 8000000302a03031
    current = 0xc000000415050300
    paca    = 0xc00000003ffc4080	 irqmask: 0x03	 irq_happened: 0x01
      pid   = 25006, comm = sigfuz
  Linux version 5.0.0-rc1-00001-g3bd6e94b (breno@debian) (gcc version 8.2.0 (Debian 8.2.0-3)) #899 SMP Mon Jan 7 11:30:07 EST 2019
  WARNING: exception is not recoverable, can't continue
  enter ? for help
  [c0000004263ebc40] c000000000032948 handle_rt_signal64+0xb8/0xaf0 (unreliable)
  [c0000004263ebd30] c000000000022780 do_notify_resume+0x2f0/0x430
  [c0000004263ebe20] c00000000000e844 ret_from_except_lite+0x70/0x74
  --- Exception: c00 (System Call) at 00007fffbaac400c
  SP (7fffeca90f40) is in userspace

The solution for this problem is running the sigreturn code with
regs->msr[TS] disabled, thus, avoiding hitting the side effect above.
This does not seem to be a problem since regs->msr will be replaced by
the ucontext value, so, it is being flushed already. In this case, it
is flushed earlier.
Signed-off-by: Breno Leitao <leitao@debian.org>
Acked-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This patch fixes the below crash by making sure we touch the subpage
protection related structures only if we know they are allocated on
the platform. With radix translation we don't allocate hash context at
all and trying to access subpage_prot_table results in:

  Faulting instruction address: 0xc00000000008bdb4
  Oops: Kernel access of bad area, sig: 11 [#1]
  LE PAGE_SIZE=64K MMU=Radix MMU=Hash SMP NR_CPUS=2048 NUMA PowerNV
  ....
  NIP [c00000000008bdb4] sys_subpage_prot+0x74/0x590
  LR [c00000000000b688] system_call+0x5c/0x70
  Call Trace:
  [c00020002c6b7d30] [c00020002c6b7d90] 0xc00020002c6b7d90 (unreliable)
  [c00020002c6b7e20] [c00000000000b688] system_call+0x5c/0x70
  Instruction dump:
  fb61ffd8 fb81ffe0 fba1ffe8 fbc1fff0 fbe1fff8 f821ff11 e92d1178 f9210068
  39200000 e92d0968 ebe90630 e93f03e8 <eb891038> 60000000 3860fffe e9410068

We also move the subpage_prot_table with mmp_sem held to avoid race
between two parallel subpage_prot syscall.

Fixes: 70110186 ("powerpc/mm: Reduce memory usage for mm_context_t for radix")
Reported-by: Sachin Sant <sachinp@linux.ibm.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Tested-by: Sachin Sant <sachinp@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Print more information about MCE error whether it is an hardware or
software error.

Some of the MCE errors can be easily categorized as hardware or
software errors e.g. UEs are due to hardware error, where as error
triggered due to invalid usage of tlbie is a pure software bug. But
not all the MCE errors can be easily categorize into either software
or hardware. There are errors like multihit errors which are usually
result of a software bug, but in some rare cases a hardware failure
can cause a multihit error. In past, we have seen case where after
replacing faulty chip, multihit errors stopped occurring. Same with
parity errors, which are usually due to faulty hardware but there are
chances where multihit can also cause an parity error. Such errors are
difficult to determine what really caused it. Hence this patch
classifies MCE errors into following four categorize:

  1. Hardware error:
  	UE and Link timeout failure errors.
  2. Probable hardware error (some chance of software cause)
  	SLB/ERAT/TLB Parity errors.
  3. Software error
  	Invalid tlbie form.
  4. Probable software error (some chance of hardware cause)
  	SLB/ERAT/TLB Multihit errors.

Sample output:

  MCE: CPU80: machine check (Warning) Guest SLB Multihit DAR: 000001001b6e0320 [Recovered]
  MCE: CPU80: PID: 24765 Comm: qemu-system-ppc Guest NIP: [00007fffa309dc60]
  MCE: CPU80: Probable Software error (some chance of hardware cause)
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Currently all machine check errors are printed as severe errors which
isn't correct. Print soft errors as warning instead of severe errors.
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Also add cpu number while displaying MCE log. This will help cleaner
logs when MCE hits on multiple cpus simultaneously.

Before the changes the MCE output was:

  Severe Machine check interrupt [Recovered]
    NIP [d00000000ba80280]: insert_slb_entry.constprop.0+0x278/0x2c0 [mcetest_slb]
    Initiator: CPU
    Error type: SLB [Multihit]
      Effective address: d00000000ba80280

After this patch series changes the MCE output will be:

  MCE: CPU80: machine check (Warning) Host SLB Multihit [Recovered]
  MCE: CPU80: NIP: [d00000000b550280] insert_slb_entry.constprop.0+0x278/0x2c0 [mcetest_slb]
  MCE: CPU80: Probable software error (some chance of hardware cause)

UE in host application:

  MCE: CPU48: machine check (Severe) Host UE Load/Store DAR: 00007fffc6079a80 paddr: 0000000f8e260000 [Not recovered]
  MCE: CPU48: PID: 4584 Comm: find NIP: [0000000010023368]
  MCE: CPU48: Hardware error

and for MCE in Guest:

  MCE: CPU80: machine check (Warning) Guest SLB Multihit DAR: 000001001b6e0320 [Recovered]
  MCE: CPU80: PID: 24765 Comm: qemu-system-ppc Guest NIP: [00007fffa309dc60]
  MCE: CPU80: Probable software error (some chance of hardware cause)
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

When analysing sources of OS jitter, I noticed that doorbells cannot be
traced.
Signed-off-by: Anton Blanchard <anton@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Fixes gcc '-Wunused-but-set-variable' warning:

  drivers/misc/ocxl/link.c: In function 'xsl_fault_handler':
  drivers/misc/ocxl/link.c:187:17: warning: variable 'tid' set but not used
  drivers/misc/ocxl/link.c:187:6: warning: variable 'lpid' set but not used

They are never used and can be removed.
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Acked-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

In commit 2bf1071a ("powerpc/64s: Remove POWER9 DD1 support") the
function __switch_to remove usage for 'dummy_copy_buffer'. Since it is
not used anywhere else, remove it completely.

This remove the following warning:
  arch/powerpc/kernel/process.c:1156:17: error: 'dummy_copy_buffer' defined but not used
Suggested-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Mathieu Malaterre <malat@debian.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Currently error return from kobject_init_and_add() is not followed by
a call to kobject_put(). This means there is a memory leak.

Add call to kobject_put() in error path of kobject_init_and_add().
Signed-off-by: Tobin C. Harding <tobin@kernel.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Towards the goal of removing cc-ldoption, it seems that --hash-style=
was added to binutils 2.17.50.0.2 in 2006. The minimal required
version of binutils for the kernel according to
Documentation/process/changes.rst is 2.20.
Suggested-by: Masahiro Yamada <yamada.masahiro@socionext.com>
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

When the return value type was changed from int to bool, few places
were left unchanged, this fixes them. We did not hit these failures as
the first one is not happening at all and the second one is little
more likely to happen if the user switches a 33..58bit DMA capable
device between the VFIO and vendor drivers and there are not so many
of these.

Fixes: 2d6ad41b ("powerpc/powernv: use the generic iommu bypass code")
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Merge our topic branch shared with KVM. In particular this includes the
rewrite of the idle code into C.

This is an implementation of commits 53a712ba
("powerpc/powernv/idle: Restore AMR/UAMOR/AMOR after idle") and
a3f3072d ("powerpc/powernv/idle: Restore IAMR after idle") using
the new C-based idle code.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Extract from Nick's patch]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Reimplement Book3S idle code in C, moving POWER7/8/9 implementation
speific HV idle code to the powernv platform code.

Book3S assembly stubs are kept in common code and used only to save
the stack frame and non-volatile GPRs before executing architected
idle instructions, and restoring the stack and reloading GPRs then
returning to C after waking from idle.

The complex logic dealing with threads and subcores, locking, SPRs,
HMIs, timebase resync, etc., is all done in C which makes it more
maintainable.

This is not a strict translation to C code, there are some
significant differences:

- Idle wakeup no longer uses the ->cpu_restore call to reinit SPRs,
  but saves and restores them itself.

- The optimisation where EC=ESL=0 idle modes did not have to save GPRs
  or change MSR is restored, because it's now simple to do. ESL=1
  sleeps that do not lose GPRs can use this optimization too.

- KVM secondary entry and cede is now more of a call/return style
  rather than branchy. nap_state_lost is not required because KVM
  always returns via NVGPR restoring path.

- KVM secondary wakeup from offline sequence is moved entirely into
  the offline wakeup, which avoids a hwsync in the normal idle wakeup
  path.

Performance measured with context switch ping-pong on different
threads or cores, is possibly improved a small amount, 1-3% depending
on stop state and core vs thread test for shallow states. Deep states
it's in the noise compared with other latencies.

KVM improvements:

- Idle sleepers now always return to caller rather than branch out
  to KVM first.

- This allows optimisations like very fast return to caller when no
  state has been lost.

- KVM no longer requires nap_state_lost because it controls NVGPR
  save/restore itself on the way in and out.

- The heavy idle wakeup KVM request check can be moved out of the
  normal host idle code and into the not-performance-critical offline
  code.

- KVM nap code now returns from where it is called, which makes the
  flow a bit easier to follow.
Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Squash the KVM changes in]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

Using a jiffies timer creates a dependency on the tick_do_timer_cpu
incrementing jiffies. If that CPU has locked up and jiffies is not
incrementing, the watchdog heartbeat timer for all CPUs stops and
creates false positives and confusing warnings on local CPUs, and
also causes the SMP detector to stop, so the root cause is never
detected.

Fix this by using hrtimer based timers for the watchdog heartbeat,
like the generic kernel hardlockup detector.

Cc: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Reported-by: Ravikumar Bangoria <ravi.bangoria@in.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Reported-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

When removing memory we need to remove the memory from the node
it was added to instead of looking up the node it should be in
in the device tree.

During testing we have seen scenarios where the affinity for a
LMB changes due to a partition migration or PRRN event. In these
cases the node the LMB exists in may not match the node the device
tree indicates it belongs in. This can lead to a system crash
when trying to DLPAR remove the LMB after a migration or PRRN
event. The current code looks up the node in the device tree to
remove the LMB from, the crash occurs when we try to offline this
node and it does not have any data, i.e. node_data[nid] == NULL.

36:mon> e
cpu 0x36: Vector: 300 (Data Access) at [c0000001828b7810]
    pc: c00000000036d08c: try_offline_node+0x2c/0x1b0
    lr: c0000000003a14ec: remove_memory+0xbc/0x110
    sp: c0000001828b7a90
   msr: 800000000280b033
   dar: 9a28
 dsisr: 40000000
  current = 0xc0000006329c4c80
  paca    = 0xc000000007a55200   softe: 0        irq_happened: 0x01
    pid   = 76926, comm = kworker/u320:3

36:mon> t
[link register   ] c0000000003a14ec remove_memory+0xbc/0x110
[c0000001828b7a90] c00000000006a1cc arch_remove_memory+0x9c/0xd0 (unreliable)
[c0000001828b7ad0] c0000000003a14e0 remove_memory+0xb0/0x110
[c0000001828b7b20] c0000000000c7db4 dlpar_remove_lmb+0x94/0x160
[c0000001828b7b60] c0000000000c8ef8 dlpar_memory+0x7e8/0xd10
[c0000001828b7bf0] c0000000000bf828 handle_dlpar_errorlog+0xf8/0x160
[c0000001828b7c60] c0000000000bf8cc pseries_hp_work_fn+0x3c/0xa0
[c0000001828b7c90] c000000000128cd8 process_one_work+0x298/0x5a0
[c0000001828b7d20] c000000000129068 worker_thread+0x88/0x620
[c0000001828b7dc0] c00000000013223c kthread+0x1ac/0x1c0
[c0000001828b7e30] c00000000000b45c ret_from_kernel_thread+0x5c/0x80

To resolve this we need to track the node a LMB belongs to when
it is added to the system so we can remove it from that node instead
of the node that the device tree indicates it should belong to.
Signed-off-by: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

There are several identical spelling mistakes in warning messages,
fix these.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This patch fix the below section mismatch warnings.

WARNING: vmlinux.o(.text+0x2d1f44): Section mismatch in reference from the function devm_memremap_pages_release() to the function .meminit.text:arch_remove_memory()
WARNING: vmlinux.o(.text+0x2d265c): Section mismatch in reference from the function devm_memremap_pages() to the function .meminit.text:arch_add_memory()
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

The region actually point to linear map. Rename the #define to
clarify thati.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

This helps in debugging. We can look at the dmesg to find out
different kernel mapping details.

On 4K config this shows

 kernel vmalloc start   = 0xc000100000000000
 kernel IO start        = 0xc000200000000000
 kernel vmemmap start   = 0xc000300000000000

On 64K config:

 kernel vmalloc start   = 0xc008000000000000
 kernel IO start        = 0xc00a000000000000
 kernel vmemmap start   = 0xc00c000000000000
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>