openwrt/staging/blogic.git
7 years agopowerpc/64s: Expand core idle state bits
Nicholas Piggin [Wed, 19 Apr 2017 13:05:48 +0000 (23:05 +1000)]
powerpc/64s: Expand core idle state bits

In preparation for adding more bits to the core idle state word, move
the lock bit up, and unlock by flipping the lock bit rather than masking
off all but the thread bits.

Add branch hints for atomic operations while we're here.

Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Fix POWER9 machine check handler from stop state
Nicholas Piggin [Wed, 19 Apr 2017 13:05:47 +0000 (23:05 +1000)]
powerpc/64s: Fix POWER9 machine check handler from stop state

The ISA specifies power save wakeup due to a machine check exception can
cause a machine check interrupt (rather than the usual system reset
interrupt).

The machine check handler copes with this by doing low level machine
check recovery without restoring full state from idle, then queues up a
machine check event for logging, then directly executes the same idle
instruction it woke from. This minimises the work done before recovery
is performed.

The problem is that it requires machine specific instructions and
knowledge of the book3s idle code. Currently it only has code to handle
POWER8 idle, so POWER9 crashes when trying to execute the P8 idle
instructions which don't exist in ISAv3.0B.

cpu 0x0: Vector: e40 (Emulation Assist) at [c0000000008f3810]
    pc: c000000000008380: machine_check_handle_early+0x130/0x2f0
    lr: c00000000053a098: stop_loop+0x68/0xd0
    sp: c0000000008f3a90
   msr: 9000000000081001
  current = 0xc0000000008a1080
  paca    = 0xc00000000ffd0000   softe: 0        irq_happened: 0x01
    pid   = 0, comm = swapper/0

Instead of going to sleep after recovery, do the usual idle wakeup and
state restoration by calling into the normal idle wakeup path. This
reuses the normal idle wakeup paths.

Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Reviewed-by: Mahesh J Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Use alternative feature patching
Nicholas Piggin [Wed, 19 Apr 2017 13:05:46 +0000 (23:05 +1000)]
powerpc/64s: Use alternative feature patching

This reduces the number of nops for POWER8.

Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Stop using bit in HSPRG0 to test winkle
Nicholas Piggin [Wed, 19 Apr 2017 13:05:45 +0000 (23:05 +1000)]
powerpc/64s: Stop using bit in HSPRG0 to test winkle

The POWER8 idle code has a neat trick of programming the power on engine
to restore a low bit into HSPRG0, so idle wakeup code can test and see
if it has been programmed this way and therefore lost all state. Restore
time can be reduced if winkle has not been reached.

However this messes with our r13 PACA pointer, and requires HSPRG0 to be
written to. It also optimizes the slowest and most uncommon case at the
expense of another SPR write in the common nap state wakeup.

Remove this complexity and assume winkle sleeps always require a state
restore. This speedup could be made entirely contained within the winkle
idle code by counting per-core winkles and setting a thread bitmap when
all have gone to winkle.

Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Move remaining system reset idle code into idle_book3s.S
Nicholas Piggin [Wed, 19 Apr 2017 13:05:44 +0000 (23:05 +1000)]
powerpc/64s: Move remaining system reset idle code into idle_book3s.S

No functional change.

Reviewed-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Remove unnecessary relocation branch from idle handler
Nicholas Piggin [Wed, 19 Apr 2017 13:05:43 +0000 (23:05 +1000)]
powerpc/64s: Remove unnecessary relocation branch from idle handler

The system reset idle handler system_reset_idle_common is relocated, so
relocation is not required to branch to kvm_start_guest. The superfluous
relocation does not result in incorrect code, but it does not compile
outside of exception-64s.S (with fixed section definitions).

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Add support for runtime configuration of ASLR limits
Michael Ellerman [Thu, 20 Apr 2017 14:36:20 +0000 (00:36 +1000)]
powerpc/mm: Add support for runtime configuration of ASLR limits

Add powerpc support for mmap_rnd_bits and mmap_rnd_compat_bits, which are two
sysctls that allow a user to configure the number of bits of randomness used for
ASLR.

Because of the way the Kconfig for ARCH_MMAP_RND_BITS is defined, we have to
construct at least the MIN value in Kconfig, vs in a header which would be more
natural. Given that we just go ahead and do it all in Kconfig.

At least according to the code (the documentation makes no mention of it), the
value is defined as the number of bits of randomisation *of the page*, not the
address. This makes some sense, with larger page sizes more of the low bits are
forced to zero, which would reduce the randomisation if we didn't take the
PAGE_SIZE into account. However it does mean the min/max values have to change
depending on the PAGE_SIZE in order to actually limit the amount of address
space consumed by the randomisation.

The result of that is that we have to define the default values based on both
32-bit vs 64-bit, but also the configured PAGE_SIZE. Furthermore now that we
have 128TB address space support on Book3S, we also have to take that into
account.

Finally we can wire up the value in arch_mmap_rnd().

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Bhupesh Sharma <bhsharma@redhat.com>
Tested-by: Bhupesh Sharma <bhsharma@redhat.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
7 years agopowerpc/mm: Wire up ioremap_cache()
Oliver O'Halloran [Tue, 11 Apr 2017 17:42:31 +0000 (03:42 +1000)]
powerpc/mm: Wire up ioremap_cache()

The default implementation of ioremap_cache() is aliased to ioremap().
On powerpc ioremap() creates cache-inhibited mappings by default which
is almost certainly not what you wanted.

Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Use relon prolog for EXC_VIRT_OOL_MASKABLE_HV handlers
Nicholas Piggin [Thu, 13 Apr 2017 09:45:48 +0000 (19:45 +1000)]
powerpc/64s: Use relon prolog for EXC_VIRT_OOL_MASKABLE_HV handlers

Hypervisor Virtualization and Directed Hypervisor Doorbell interrupt handlers
use the macro EXC_VIRT_OOL_MASKABLE_HV for their relocation-on handlers, which
calls MASKABLE_RELON_EXCEPTION_HV_OOL, which uses the *real mode* interrupt
prolog. This means we needlessly rfid from virtual mode to virtual mode.

For POWER8 it only affects doorbell IPIs. Context switch microbenchmark between
threads with snooze disabled (which causes IPI) gets about 3% faster, about 370
cycles. Should be more important on POWER9 with global doorbells and HVI for
host interrupts.

Use the RELON variant instead to reduce overhead.

Fixes: 1707dd1613 ("powerpc: Save CFAR before branching in interrupt entry paths")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Fold some more detail into the change log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/xive: Fix missing check of rc != OPAL_BUSY
Michael Ellerman [Thu, 20 Apr 2017 04:43:19 +0000 (14:43 +1000)]
powerpc/xive: Fix missing check of rc != OPAL_BUSY

Dan Carpenter noticed that the code in __xive_native_disable_queue() has a for
loop with an unconditional break in the middle, which doesn't make a lot of
sense.

What the code's supposed to do is loop as long as OPAL says it's busy, if we get
any other return code, either success or failure, then we should break the loop.

So add the missing check.

Fixes: 243e25112d06 ("powerpc/xive: Native exploitation of the XIVE interrupt controller")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Remove SAO feature from Power9 DD1
Nicholas Piggin [Wed, 19 Apr 2017 02:27:38 +0000 (12:27 +1000)]
powerpc/64s: Remove SAO feature from Power9 DD1

Power9 DD1 does not implement SAO. Although it's not widely used, its presence
or absence is visible to user space via arch_validate_prot() so it's moderately
important that we get the value right.

Fixes: 7dccfbc325bb ("powerpc/book3s: Add a cpu table entry for different POWER9 revs")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Remove ICSWX feature from Power9
Nicholas Piggin [Wed, 19 Apr 2017 02:27:37 +0000 (12:27 +1000)]
powerpc/64s: Remove ICSWX feature from Power9

Power9 does not implement the icswx instruction. This CPU feature is not visible
to userspace and is only used in the CONFIG_PPC_ICSWX code, which is generally
not enabled, and can only be triggered by other code using icswx, which should
not happen on Power9 systems in the first place. So impact should be minimal.

Fixes: c3ab300ea5 ("powerpc: Add POWER9 cputable entry")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Add Power8 mem_access event to sysfs
Madhavan Srinivasan [Tue, 11 Apr 2017 01:51:10 +0000 (07:21 +0530)]
powerpc/perf: Add Power8 mem_access event to sysfs

Patch add "mem_access" event to sysfs. This as-is not a raw event
supported by Power8 pmu. Instead, it is formed based on
raw event encoding specificed in isa207-common.h.

Primary PMU event used here is PM_MRK_INST_CMPL.
This event tracks only the completed marked instructions.

Random sampling mode (MMCRA[SM]) with Random Instruction
Sampling (RIS) is enabled to mark type of instructions.

With Random sampling in RLS mode with PM_MRK_INST_CMPL event,
the LDST /DATA_SRC fields in SIER identifies the memory
hierarchy level (eg: L1, L2 etc) statisfied a data-cache
miss for a marked instruction.

Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Support to export SIERs bit in Power9
Madhavan Srinivasan [Tue, 11 Apr 2017 01:51:09 +0000 (07:21 +0530)]
powerpc/perf: Support to export SIERs bit in Power9

Patch to export SIER bits to userspace via
perf_mem_data_src and perf_sample_data struct.

Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Support to export SIERs bit in Power8
Madhavan Srinivasan [Tue, 11 Apr 2017 01:51:08 +0000 (07:21 +0530)]
powerpc/perf: Support to export SIERs bit in Power8

Patch to export SIER bits to userspace via
perf_mem_data_src and perf_sample_data struct.

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Support to export MMCRA[TEC*] field to userspace
Madhavan Srinivasan [Tue, 11 Apr 2017 01:51:07 +0000 (07:21 +0530)]
powerpc/perf: Support to export MMCRA[TEC*] field to userspace

Threshold feature when used with MMCRA [Threshold Event Counter Event],
MMCRA[Threshold Start event] and MMCRA[Threshold End event] will update
MMCRA[Threashold Event Counter Exponent] and MMCRA[Threshold Event
Counter Multiplier] with the corresponding threshold event count values.
Patch to export MMCRA[TECX/TECM] to userspace in 'weight' field of
struct perf_sample_data.

Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Export memory hierarchy info to user space
Madhavan Srinivasan [Tue, 11 Apr 2017 01:51:06 +0000 (07:21 +0530)]
powerpc/perf: Export memory hierarchy info to user space

The LDST field and DATA_SRC in SIER identifies the memory hierarchy level
(eg: L1, L2 etc), from which a data-cache miss for a marked instruction
was satisfied. Use the 'perf_mem_data_src' object to export this
hierarchy level to user space.

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/perf: Define big-endian version of perf_mem_data_src
Sukadev Bhattiprolu [Tue, 11 Apr 2017 01:51:05 +0000 (07:21 +0530)]
powerpc/perf: Define big-endian version of perf_mem_data_src

perf_mem_data_src is a union that is initialized in the kernel via the ->val
field and accessed by userspace via the mem_xxx bitfields. For this to work
correctly on big endian platforms, we need a big-endian definition for the
bitfields.

Currently on a big endian system, if a user requests PERF_SAMPLE_DATA_SRC (perf
report -d), they will get the default value from perf_sample_data_init(), which
is PERF_MEM_NA. The value for PERF_MEM_NA is constructed using shifts:

  /* TLB access */
  #define PERF_MEM_TLB_NA 0x01 /* not available */
  ...
  #define PERF_MEM_TLB_SHIFT 26

  #define PERF_MEM_S(a, s) \
(((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)

  #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
    PERF_MEM_S(LVL, NA)   |\
    PERF_MEM_S(SNOOP, NA) |\
    PERF_MEM_S(LOCK, NA)  |\
    PERF_MEM_S(TLB, NA))

Which works out as:

  ((0x01 << 0) | (0x01 << 5) | (0x01 << 19) | (0x01 << 24) | (0x01 << 26))

Which means the PERF_MEM_NA value comes out of the kernel as 0x5080021
in CPU endian.

But then in the perf tool, the code uses the bitfields to inspect the value, and
currently the bitfields are defined using little endian ordering.

So eg. in perf_mem__tlb_scnprintf() we see:
  data_src->val = 0x5080021
             op = 0x0
            lvl = 0x0
          snoop = 0x0
           lock = 0x0
           dtlb = 0x0
           rsvd = 0x5080021

Because of the way the perf tool code is written this is still displayed to the
user as "N/A", so there is no bug visible at the UI level.

Currently there are no big endian architectures which export a meaningful
value (ie. other than PERF_MEM_NA), so the extent of the bug on big endian
platforms is that the PERF_MEM_NA value is exported incorrectly as described
above. Subsequent patches will add support on big endian powerpc for populating
the data source value.

This patch does a minimal fix of adding big endian definition of the bitfields
to match the values that are already exported by the kernel on big endian. And
it makes no change on little endian.

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/iommu: Do not call PageTransHuge() on tail pages
Alexey Kardashevskiy [Tue, 11 Apr 2017 07:54:57 +0000 (17:54 +1000)]
powerpc/iommu: Do not call PageTransHuge() on tail pages

The CMA pages migration code does not support compound pages at
the moment so it performs few tests before proceeding to actual page
migration.

One of the tests - PageTransHuge() - has VM_BUG_ON_PAGE(PageTail()) as
it is designed to be called on head pages only. Since we also test for
PageCompound(), and it contains PageTail() and PageHead(), we can
simplify the check by leaving just PageCompound() and therefore avoid
possible VM_BUG_ON_PAGE.

Fixes: 2e5bbb5461f1 ("KVM: PPC: Book3S HV: Migrate pinned pages out of CMA")
Cc: stable@vger.kernel.org # v4.9+
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Balbir Singh <bsingharora@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mmap: Any hint > 128TB searches the full VA space
Aneesh Kumar K.V [Tue, 18 Apr 2017 07:01:27 +0000 (12:31 +0530)]
powerpc/mmap: Any hint > 128TB searches the full VA space

As part of the new large address space support, processes start out life with a
128TB virtual address space. However when calling mmap() a process can pass a
hint address, and if that hint is > 128TB the kernel will use the full 512TB
address space to try and satisfy the mmap() request.

Currently we have a check that the hint is > 128TB and < 512TB (TASK_SIZE),
which was added as an optimisation to avoid updating addr_limit unnecessarily
and also to avoid calling slice_flush_segments() on all CPUs more than
necessary.

However this has the user-visible side effect that an mmap() hint above 512TB
does not search the full address space unless a preceding mmap() used a hint
value > 128TB && < 512TB.

So fix it to treat any hint above 128TB as a hint to search the full address
space, instead of checking the hint against TASK_SIZE, we instead check if the
addr_limit is already == TASK_SIZE.

This also brings the ABI in-line with what is proposed on x86. ie, that a hint
address above 128TB up to and including (2^64)-1 is an indication to search the
full address space.

Fixes: f4ea6dcb08ea2c (powerpc/mm: Enable mappings above 128TB)
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Enable PCI device IDs for future IBM CXL adapters
Matthew R. Ochs [Fri, 24 Mar 2017 16:03:19 +0000 (11:03 -0500)]
cxl: Enable PCI device IDs for future IBM CXL adapters

Add support for future IBM Coherent Accelerator (CXL) devices
with an IDs of 0x0623 and 0x0628.

Signed-off-by: Matthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: Uma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Minor fix for MCE TLB flush for radix
Nicholas Piggin [Sun, 16 Apr 2017 14:21:19 +0000 (00:21 +1000)]
powerpc/64s: Minor fix for MCE TLB flush for radix

The TLB flush for radix first flushes TLB for radix configuration,
then flushes for hash configuration. The second flush is unnecessary
but does not affect correctness.

Fixes: 1a472c9dba6b9 ("powerpc/mm/radix: Add tlbflush routines")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/radix: Use mm->task_size for boundary checking instead of addr_limit
Aneesh Kumar K.V [Thu, 13 Apr 2017 19:18:21 +0000 (00:48 +0530)]
powerpc/mm/radix: Use mm->task_size for boundary checking instead of addr_limit

We don't init addr_limit correctly for 32 bit applications. So default to using
mm->task_size for boundary condition checking. We use addr_limit to only control
free space search. This makes sure that we do the right thing with 32 bit
applications.

We should consolidate the usage of TASK_SIZE/mm->task_size and
mm->context.addr_limit later.

This partially reverts commit fbfef9027c2a7ad (powerpc/mm: Switch some
TASK_SIZE checks to use mm_context addr_limit).

Fixes: fbfef9027c2a ("powerpc/mm: Switch some TASK_SIZE checks to use mm_context addr_limit")
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Revert setting of LPCR[LPES] on POWER9
Nicholas Piggin [Tue, 18 Apr 2017 19:12:16 +0000 (05:12 +1000)]
powerpc/64s: Revert setting of LPCR[LPES] on POWER9

The XIVE enablement patches included a change to set the LPES (Logical
Partitioning Environment Selector) bit (bit # 3) in LPCR (Logical Partitioning
Control Register) on POWER9 hosts. This bit sets external interrupts to guest
delivery mode, which uses SRR0/1. The host's EE interrupt handler is written to
expect HSRR0/1 (for earlier CPUs). This should be fine because XIVE is
configured not to deliver EEs to the host (Hypervisor Virtulization Interrupt is
used instead) so the EE handler should never be executed.

However a bug in interrupt controller code, hardware, or odd configuration of a
simulator could result in the host getting an EE incorrectly. Keeping the EE
delivery mode matching the host EE handler prevents strange crashes due to using
the wrong exception registers.

KVM will configure the LPCR to set LPES prior to running a guest so that EEs are
delivered to the guest using SRR0/1.

Fixes: 08a1e650cc ("powerpc: Fixup LPCR:PECE and HEIC setting on POWER9")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Massage change log to avoid referring to LPES0 which is now renamed LPES]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/pseries: Always enable SMP when building pseries
Michael Ellerman [Wed, 5 Apr 2017 02:44:51 +0000 (12:44 +1000)]
powerpc/pseries: Always enable SMP when building pseries

The pseries platform supports Power4 and later CPUs, all of which are
multithreaded and/or multicore.

In practice no one ever builds a SMP=n kernel for these machines. So as
we did for powernv, have the pseries platform imply SMP=y.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Always enable SMP when building powernv
Michael Ellerman [Wed, 5 Apr 2017 02:44:50 +0000 (12:44 +1000)]
powerpc/powernv: Always enable SMP when building powernv

The powernv platform supports Power7 and later CPUs, all of which are
multithreaded and multicore.

As such we never build a SMP=n kernel for those machines, other than
possibly for debugging or running in a simulator.

In the debugging case we can get a similar effect by booting with
nr_cpus=1, or there's always the option of building a custom kernel with
SMP hacked out.

For running in simulators the code size reduction from building without
SMP is not particularly important, what matters is the number of
instructions executed. A quick test shows that a SMP=y kernel takes ~6%
more instructions to boot to a shell. Booting with nr_cpus=1 recovers
about half that deficit.

On the flip side, keeping the SMP=n kernel building can be a pain at
times. And although we've mostly kept it building in recent years, no
one is regularly testing that the SMP=n kernel actually boots and works
well on these machines.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Allow platforms to force-enable CONFIG_SMP
Michael Ellerman [Wed, 5 Apr 2017 02:44:49 +0000 (12:44 +1000)]
powerpc: Allow platforms to force-enable CONFIG_SMP

Of the 64-bit Book3S platforms, only powermac supports booting on an
actual non-SMP system. The other platforms can be built with SMP
disabled, but it doesn't make a lot of sense given the CPUs they support
are all multicore or multithreaded.

So give platforms the option of forcing SMP=y.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Drop include of linux/io.h from asm/io.h
Michael Ellerman [Thu, 13 Apr 2017 13:14:36 +0000 (23:14 +1000)]
powerpc: Drop include of linux/io.h from asm/io.h

Currently powerpc's asm/io.h includes linux/io.h, and linux/io.h
includes asm/io.h.

This can cause problems because depending on which is included first the
order of definitions between the two files will change.

The include of linux/io.h was added back in 2008 in commit b41e5fffe8b8
("[POWERPC] devres: Add devm_ioremap_prot()"). It's not entirely clear
it was needed then, but devm_ioremap_prot() has since been removed
entirely as unused, in dedd24a12fe6 ("powerpc: Remove unused
devm_ioremap_prot()").

So it seems to be unnecessary and can potentially cause problems, so
remove the include of linux/io.h from asm/io.h

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: POWER9 support for msgsnd/doorbell IPI
Nicholas Piggin [Thu, 13 Apr 2017 10:16:24 +0000 (20:16 +1000)]
powerpc/powernv: POWER9 support for msgsnd/doorbell IPI

POWER9 requires msgsync for receiver-side synchronization, and a DD1
workaround restricts IPIs to core-local.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Drop no longer needed asm feature macro changes]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Avoid a branch for ppc_msgsnd
Nicholas Piggin [Thu, 13 Apr 2017 10:16:23 +0000 (20:16 +1000)]
powerpc/64s: Avoid a branch for ppc_msgsnd

IPIs are a pretty hot path and we already have the ability to do asm feature
patching, so use it.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Change log detail]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Introduce msgsnd/doorbell barrier primitives
Nicholas Piggin [Thu, 13 Apr 2017 10:16:22 +0000 (20:16 +1000)]
powerpc: Introduce msgsnd/doorbell barrier primitives

POWER9 changes requirements and adds new instructions for
synchronization.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Change the doorbell IPI calling convention
Nicholas Piggin [Thu, 13 Apr 2017 10:16:21 +0000 (20:16 +1000)]
powerpc: Change the doorbell IPI calling convention

Change the doorbell callers to know about their msgsnd addressing,
rather than have them set a per-cpu target data tag at boot that gets
sent to the cause_ipi functions. The data is only used for doorbell IPI
functions, no other IPI types, so it makes sense to keep that detail
local to doorbell.

Have the platform code understand doorbell IPIs, rather than the
interrupt controller code understand them. Platform code can look at
capabilities it has available and decide which to use.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Add SCV FSCR bit for ISA v3.0
Nicholas Piggin [Fri, 7 Apr 2017 01:27:44 +0000 (11:27 +1000)]
powerpc/64s: Add SCV FSCR bit for ISA v3.0

Add the bit definition and use it in facility_unavailable_exception() so we can
intelligently report the cause if we take a fault for SCV. This doesn't actually
enable SCV.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Drop whitespace changes to the existing entries, flush out change log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/64s: Add msgp facility unavailable log string
Nicholas Piggin [Fri, 7 Apr 2017 01:27:43 +0000 (11:27 +1000)]
powerpc/64s: Add msgp facility unavailable log string

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Don't open code VMALLOC_INDEX
Aneesh Kumar K.V [Wed, 12 Apr 2017 04:40:22 +0000 (10:10 +0530)]
powerpc/mm/hash: Don't open code VMALLOC_INDEX

We have a #define for it, so use it.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Add psl9 specific code
Christophe Lombard [Wed, 12 Apr 2017 14:34:07 +0000 (16:34 +0200)]
cxl: Add psl9 specific code

The new Coherent Accelerator Interface Architecture, level 2, for the
IBM POWER9 brings new content and features:
- POWER9 Service Layer
- Registers
- Radix mode
- Process element entry
- Dedicated-Shared Process Programming Model
- Translation Fault Handling
- CAPP
- Memory Context ID
    If a valid mm_struct is found the memory context id is used for each
    transaction associated with the process handle. The PSL uses the
    context ID to find the corresponding process element.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
[mpe: Fixup comment formatting, unsplit long strings]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Isolate few psl8 specific calls
Christophe Lombard [Fri, 7 Apr 2017 14:11:58 +0000 (16:11 +0200)]
cxl: Isolate few psl8 specific calls

Point out the specific Coherent Accelerator Interface Architecture,
level 1, registers.
Code and functions specific to PSL8 (CAIA1) must be framed.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
[mpe: Don't split long strings, it makes them hard to grep for]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Rename some psl8 specific functions
Christophe Lombard [Fri, 7 Apr 2017 14:11:57 +0000 (16:11 +0200)]
cxl: Rename some psl8 specific functions

Rename a few functions, changing the '_psl' suffix to '_psl8', to make
clear that the implementation is psl8 specific.
Those functions will have an equivalent implementation for the psl9 in
a later patch.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Update implementation service layer
Christophe Lombard [Fri, 7 Apr 2017 14:11:56 +0000 (16:11 +0200)]
cxl: Update implementation service layer

The service layer API (in cxl.h) lists some low-level functions whose
implementation is different on PSL8, PSL9 and XSL:
- Init implementation for the adapter and the afu.
- Invalidate TLB/SLB.
- Attach process for dedicated/directed models.
- Handle psl interrupts.
- Debug registers for the adapter and the afu.
- Traces.
Each environment implements its own functions, and the common code uses
them through function pointers, defined in cxl_service_layer_ops.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Keep track of mm struct associated with a context
Christophe Lombard [Fri, 7 Apr 2017 14:11:55 +0000 (16:11 +0200)]
cxl: Keep track of mm struct associated with a context

The mm_struct corresponding to the current task is acquired each time
an interrupt is raised. So to simplify the code, we only get the
mm_struct when attaching an AFU context to the process.
The mm_count reference is increased to ensure that the mm_struct can't
be freed. The mm_struct will be released when the context is detached.
A reference on mm_users is not kept to avoid a circular dependency if
the process mmaps its cxl mmio and forget to unmap before exiting.
The field glpid (pid of the group leader associated with the pid), of
the structure cxl_context, is removed because it's no longer useful.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Remove unused values in bare-metal environment.
Christophe Lombard [Fri, 7 Apr 2017 14:11:54 +0000 (16:11 +0200)]
cxl: Remove unused values in bare-metal environment.

The two previously fields pid and tid, located in the structure
cxl_irq_info, are only used in the guest environment. To avoid confusion,
it's not necessary to fill the fields in the bare-metal environment.
Pid_tid is now renamed to 'reserved' to avoid undefined behavior on
bare-metal. The PSL Process and Thread Identification Register
(CXL_PSL_PID_TID_An) is only used when attaching a dedicated process
for PSL8 only. This register goes away in CAIA2.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agocxl: Read vsec perst load image
Christophe Lombard [Fri, 7 Apr 2017 14:11:53 +0000 (16:11 +0200)]
cxl: Read vsec perst load image

This bit is used to cause a flash image load for programmable
CAIA-compliant implementation. If this bit is set to â€˜0’, a power
cycle of the adapter is required to load a programmable CAIA-com-
pliant implementation from flash.
This field will be used by the following patches.

Signed-off-by: Christophe Lombard <clombard@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Fix hash table dump when memory is not contiguous
Rashmica Gupta [Mon, 10 Apr 2017 02:04:04 +0000 (12:04 +1000)]
powerpc/mm: Fix hash table dump when memory is not contiguous

The current behaviour of the hash table dump assumes that memory is contiguous
and iterates from the start of memory to (start + size of memory). When memory
isn't physically contiguous, this doesn't work.

If memory exists at 0-5 GB and 6-10 GB then the current approach will check if
entries exist in the hash table from 0GB to 9GB. This patch changes the
behaviour to iterate over any holes up to the end of memory.

Fixes: 1515ab932156 ("powerpc/mm: Dump hash table")
Signed-off-by: Rashmica Gupta <rashmica.g@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Add physical address to Linux page table dump
Oliver O'Halloran [Fri, 31 Mar 2017 01:37:49 +0000 (12:37 +1100)]
powerpc/mm: Add physical address to Linux page table dump

The current page table dumper scans the Linux page tables and coalesces mappings
with adjacent virtual addresses and similar PTE flags. This behaviour is
somewhat broken when you consider the IOREMAP space where entirely unrelated
mappings will appear to be virtually contiguous. This patch modifies the range
coalescing so that only ranges that are both physically and virtually contiguous
are combined. This patch also adds to the dump output the physical address at
the start of each range.

Fixes: 8eb07b187000 ("powerpc/mm: Dump linux pagetables")
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
[mpe: Print the physicall address with 0x like the other addresses]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Fix missing _PAGE_NON_IDEMPOTENT in pgtable dump
Oliver O'Halloran [Fri, 31 Mar 2017 01:37:48 +0000 (12:37 +1100)]
powerpc/mm: Fix missing _PAGE_NON_IDEMPOTENT in pgtable dump

On Book3s we have two PTE flags used to mark cache-inhibited mappings:
_PAGE_TOLERANT and _PAGE_NON_IDEMPOTENT. Currently the kernel page table dumper
only looks at the generic _PAGE_NO_CACHE which is defined to be _PAGE_TOLERANT.
This patch modifies the dumper so both flags are shown in the dump.

Fixes: 8eb07b187000 ("powerpc/mm: Dump linux pagetables")
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/tracing: Allow tracing of mmap syscalls
Balbir Singh [Wed, 12 Apr 2017 06:35:19 +0000 (16:35 +1000)]
powerpc/tracing: Allow tracing of mmap syscalls

Currently sys_mmap() and sys_mmap2() (32-bit only), are not visible to the
syscall tracing machinery. This means users are not able to see the execution of
mmap() syscalls using the syscall tracer.

Fix that by using SYSCALL_DEFINE6 for sys_mmap() and sys_mmap2() so that the
meta-data associated with these syscalls is visible to the syscall tracer.

A side-effect of this change is that the return type has changed from unsigned
long to long. However this should have no effect, the only code in the kernel
which uses the result of these syscalls is in the syscall return path, which is
written in asm and treats the result as unsigned regardless.

Example output:
  cat-3399  [001] ....   196.542410: sys_mmap(addr: 7fff922a0000, len: 20000, prot: 3, flags: 812, fd: 3, offset: 1b0000)
  cat-3399  [001] ....   196.542443: sys_mmap -> 0x7fff922a0000
  cat-3399  [001] ....   196.542668: sys_munmap(addr: 7fff922c0000, len: 6d2c)
  cat-3399  [001] ....   196.542677: sys_munmap -> 0x0

Signed-off-by: Balbir Singh <bsingharora@gmail.com>
[mpe: Massage change log, add detail on return type change]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Fix swapper_pg_dir size on 64-bit hash w/64K pages
Michael Ellerman [Wed, 12 Apr 2017 04:56:36 +0000 (14:56 +1000)]
powerpc/mm: Fix swapper_pg_dir size on 64-bit hash w/64K pages

Recently in commit f6eedbba7a26 ("powerpc/mm/hash: Increase VA range to 128TB"),
we increased H_PGD_INDEX_SIZE to 15 when we're building with 64K pages. This
makes it larger than RADIX_PGD_INDEX_SIZE (13), which means the logic to
calculate MAX_PGD_INDEX_SIZE in book3s/64/pgtable.h is wrong.

The end result is that the PGD (Page Global Directory, ie top level page table)
of the kernel (aka. swapper_pg_dir), is too small.

This generally doesn't lead to a crash, as we don't use the full range in normal
operation. However if we try to dump the kernel pagetables we can trigger a
crash because we walk off the end of the pgd into other memory and eventually
try to dereference something bogus:

  $ cat /sys/kernel/debug/kernel_pagetables
  Unable to handle kernel paging request for data at address 0xe8fece0000000000
  Faulting instruction address: 0xc000000000072314
  cpu 0xc: Vector: 380 (Data SLB Access) at [c0000000daa13890]
      pc: c000000000072314: ptdump_show+0x164/0x430
      lr: c000000000072550: ptdump_show+0x3a0/0x430
     dar: e802cf0000000000
  seq_read+0xf8/0x560
  full_proxy_read+0x84/0xc0
  __vfs_read+0x6c/0x1d0
  vfs_read+0xbc/0x1b0
  SyS_read+0x6c/0x110
  system_call+0x38/0xfc

The root cause is that MAX_PGD_INDEX_SIZE isn't actually computed to be
the max of H_PGD_INDEX_SIZE or RADIX_PGD_INDEX_SIZE. To fix that move
the calculation into asm-offsets.c where we can do it easily using
max().

Fixes: f6eedbba7a26 ("powerpc/mm/hash: Increase VA range to 128TB")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agoMerge branch 'topic/xive' (early part) into next
Michael Ellerman [Wed, 12 Apr 2017 12:25:02 +0000 (22:25 +1000)]
Merge branch 'topic/xive' (early part) into next

This merges the arch part of the XIVE support, leaving the final commit
with the KVM specific pieces dangling on the branch for Paul to merge
via the kvm-ppc tree.

7 years agopowerpc/powernv: Recover correct PACA on wakeup from a stop on P9 DD1
Gautham R. Shenoy [Wed, 22 Mar 2017 15:04:17 +0000 (20:34 +0530)]
powerpc/powernv: Recover correct PACA on wakeup from a stop on P9 DD1

POWER9 DD1.0 hardware has a bug where the SPRs of a thread waking up
from stop 0,1,2 with ESL=1 can endup being misplaced in the core. Thus
the HSPRG0 of a thread waking up from can contain the paca pointer of
its sibling.

This patch implements a context recovery framework within threads of a
core, by provisioning space in paca_struct for saving every sibling
threads's paca pointers. Basically, we should be able to arrive at the
right paca pointer from any of the thread's existing paca pointer.

At bootup, during powernv idle-init, we save the paca address of every
CPU in each one its siblings paca_struct in the slot corresponding to
this CPU's index in the core.

On wakeup from a stop, the thread will determine its index in the core
from the TIR register and recover its PACA pointer by indexing into
the correct slot in the provisioned space in the current PACA.

Furthermore, ensure that the NVGPRs are restored from the stack on the
way out by setting the NAPSTATELOST in paca.

[Changelog written with inputs from svaidy@linux.vnet.ibm.com]
Signed-off-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Call it a bug]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv/idle: Don't override default/deepest directly in kernel
Gautham R. Shenoy [Wed, 22 Mar 2017 15:04:16 +0000 (20:34 +0530)]
powerpc/powernv/idle: Don't override default/deepest directly in kernel

Currently during idle-init on power9, if we don't find suitable stop
states in the device tree that can be used as the
default_stop/deepest_stop, we set stop0 (ESL=1,EC=1) as the default
stop state psscr to be used by power9_idle and deepest stop state
which is used by CPU-Hotplug.

However, if the platform firmware has not configured or enabled a stop
state, the kernel should not make any assumptions and fallback to a
default choice.

If the kernel uses a stop state that is not configured by the platform
firmware, it may lead to further failures which should be avoided.

In this patch, we modify the init code to ensure that the kernel uses
only the stop states exposed by the firmware through the device
tree. When a suitable default stop state isn't found, we disable
ppc_md.power_save for power9. Similarly, when a suitable
deepest_stop_state is not found in the device tree exported by the
firmware, fall back to the default busy-wait loop in the CPU-Hotplug
code.

[Changelog written with inputs from svaidy@linux.vnet.ibm.com]
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv/smp: Add busy-wait loop as fall back for CPU-Hotplug
Gautham R. Shenoy [Wed, 22 Mar 2017 15:04:15 +0000 (20:34 +0530)]
powerpc/powernv/smp: Add busy-wait loop as fall back for CPU-Hotplug

Currently, the powernv cpu-offline function assumes that platform idle
states such as stop on POWER9, winkle/sleep/nap on POWER8 are always
available. On POWER8, it picks nap as the default state if other deep
idle states like sleep/winkle are not available and enabled in the
platform.

On POWER9, nap is not available and all idle states are managed by
STOP instruction.  The parameters to the idle state are passed through
processor stop status control register (PSSCR).  Hence as such
executing STOP would take parameters from current PSSCR. We do not
want to make any assumptions in kernel on what STOP states and PSSCR
features are configured by the platform.

Ideally platform will configure a good set of stop states that can be
used in the kernel.  We would like to start with a clean slate, if the
platform choose to not configure any state or there is an error in
platform firmware that lead to no stop states being configured or
allowed to be requested.

This patch adds a fallback method for CPU-Hotplug that is similar to
snooze loop at idle where the threads are left to spin at low priority
and hence reduce the cycles consumed.

This is a safe fallback mechanism in the case when no stop state would
be requested if the platform firmware did not configure them most
likely due to an error condition.

Requesting a stop state when the platform has not configured them or
enabled them would lead to further error conditions which could be
difficult to debug.

[Changelog written with inputs from svaidy@linux.vnet.ibm.com]
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Move CPU-Offline idle state invocation from smp.c to idle.c
Gautham R. Shenoy [Wed, 22 Mar 2017 15:04:14 +0000 (20:34 +0530)]
powerpc/powernv: Move CPU-Offline idle state invocation from smp.c to idle.c

Move the piece of code in powernv/smp.c::pnv_smp_cpu_kill_self() which
transitions the CPU to the deepest available platform idle state to a
new function named pnv_cpu_offline() in powernv/idle.c. The rationale
behind this code movement is that the data required to determine the
deepest available platform state resides in powernv/idle.c.

Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/hugetlb: Add ABI defines for supported HugeTLB page sizes
Anshuman Khandual [Fri, 7 Apr 2017 03:55:39 +0000 (09:25 +0530)]
powerpc/hugetlb: Add ABI defines for supported HugeTLB page sizes

Add user space exported API definitions for 512KB, 1MB, 2MB, 8MB, 16MB,
1GB, 16GB non default huge page sizes to be used with mmap() system
call.

Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
[mpe: Reword the comment to emphasise that these are only needed to use
 the non-default huge page size, and updated the change log.]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Remove reduntant initmem information from log
Anshuman Khandual [Fri, 7 Apr 2017 06:53:11 +0000 (12:23 +0530)]
powerpc/mm: Remove reduntant initmem information from log

Generic core VM already prints these information in the log
buffer, hence there is no need for a second print. This just
removes the second print from arch powerpc NUMA init path.

Before the patch:

  $ dmesg | grep "Initmem"

  numa: Initmem setup node 0 [mem 0x00000000-0xffffffff]
  numa: Initmem setup node 1 [mem 0x100000000-0x1ffffffff]
  numa: Initmem setup node 2 [mem 0x200000000-0x2ffffffff]
  numa: Initmem setup node 3 [mem 0x300000000-0x3ffffffff]
  numa: Initmem setup node 4 [mem 0x400000000-0x4ffffffff]
  numa: Initmem setup node 5 [mem 0x500000000-0x5ffffffff]
  numa: Initmem setup node 6 [mem 0x600000000-0x6ffffffff]
  numa: Initmem setup node 7 [mem 0x700000000-0x7ffffffff]
  Initmem setup node 0 [mem 0x0000000000000000-0x00000000ffffffff]
  Initmem setup node 1 [mem 0x0000000100000000-0x00000001ffffffff]
  Initmem setup node 2 [mem 0x0000000200000000-0x00000002ffffffff]
  Initmem setup node 3 [mem 0x0000000300000000-0x00000003ffffffff]
  Initmem setup node 4 [mem 0x0000000400000000-0x00000004ffffffff]
  Initmem setup node 5 [mem 0x0000000500000000-0x00000005ffffffff]
  Initmem setup node 6 [mem 0x0000000600000000-0x00000006ffffffff]
  Initmem setup node 7 [mem 0x0000000700000000-0x00000007ffffffff]

After the patch just the latter set is printed.

Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Make sparsemem the default on 64-bit Book3S
Michael Ellerman [Wed, 5 Apr 2017 06:10:48 +0000 (16:10 +1000)]
powerpc: Make sparsemem the default on 64-bit Book3S

Make sparsemem the default on all 64-bit Book3S platforms. It already is
for pseries and ps3, and we need to enable it for powernv because on
POWER9 memory between chips is discontiguous.

For the other platforms sparsemem should work fine, though it might add
a small amount of overhead. We can always force FLATMEM in the
defconfigs if necessary.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/nohash: Fix use of mmu_has_feature() in setup_initial_memory_limit()
Michael Ellerman [Mon, 3 Apr 2017 02:05:55 +0000 (12:05 +1000)]
powerpc/nohash: Fix use of mmu_has_feature() in setup_initial_memory_limit()

setup_initial_memory_limit() is called from early_init_devtree(), which
runs prior to feature patching. If the kernel is built with CONFIG_JUMP_LABEL=y
and CONFIG_JUMP_LABEL_FEATURE_CHECKS=y then we will potentially get the
wrong value.

If we also have CONFIG_JUMP_LABEL_FEATURE_CHECK_DEBUG=y we get a warning
and backtrace:

  Warning! mmu_has_feature() used prior to jump label init!
  CPU: 0 PID: 0 Comm: swapper Not tainted 4.11.0-rc4-gccN-next-20170331-g6af2434 #1
  Call Trace:
  [c000000000fc3d50] [c000000000a26c30] .dump_stack+0xa8/0xe8 (unreliable)
  [c000000000fc3de0] [c00000000002e6b8] .setup_initial_memory_limit+0xa4/0x104
  [c000000000fc3e60] [c000000000d5c23c] .early_init_devtree+0xd0/0x2f8
  [c000000000fc3f00] [c000000000d5d3b0] .early_setup+0x90/0x11c
  [c000000000fc3f90] [c000000000000520] start_here_multiplatform+0x68/0x80

Fix it by using early_mmu_has_feature().

Fixes: c12e6f24d413 ("powerpc: Add option to use jump label for mmu_has_feature()")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Remove unnecessary includes of asm/debug.h
Michael Ellerman [Fri, 10 Feb 2017 01:12:44 +0000 (12:12 +1100)]
powerpc: Remove unnecessary includes of asm/debug.h

These files don't seem to have any need for asm/debug.h, now that all it
includes are the debugger hooks and breakpoint definitions.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Create asm/debugfs.h and move powerpc_debugfs_root there
Michael Ellerman [Fri, 10 Feb 2017 01:04:56 +0000 (12:04 +1100)]
powerpc: Create asm/debugfs.h and move powerpc_debugfs_root there

powerpc_debugfs_root is the dentry representing the root of the
"powerpc" directory tree in debugfs.

Currently it sits in asm/debug.h, a long with some other things that
have "debug" in the name, but are otherwise unrelated.

Pull it out into a separate header, which also includes linux/debugfs.h,
and convert all the users to include debugfs.h instead of debug.h.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Require MMU_NOTIFIER to fix NPU build
Alistair Popple [Mon, 10 Apr 2017 05:24:35 +0000 (15:24 +1000)]
powerpc/powernv: Require MMU_NOTIFIER to fix NPU build

In the recent commit 1ab66d1fbada ("powerpc/powernv: Introduce address
translation services for Nvlink2") the NPU code gained a dependency on MMU
notifiers.

All our defconfigs have KVM enabled, which selects MMU_NOTIFIER, but if KVM is
not enabled then the build breaks.

Fix it by always selecting MMU_NOTIFIER when we're building powernv.

Fixes: 1ab66d1fbada ("powerpc/powernv: Introduce address translation services for Nvlink2")
Signed-off-by: Alistair Popple <alistair@popple.id.au>
[mpe: Reword change log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/radix: Remove unnecessary ptesync
Aneesh Kumar K.V [Sat, 1 Apr 2017 14:41:48 +0000 (20:11 +0530)]
powerpc/mm/radix: Remove unnecessary ptesync

For a tlbiel with pid, we need to issue tlbiel with set number encoded. We
don't need to do ptesync for each of those. Instead we need one for the entire
tlbiel pid operation.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/radix: Don't do page walk cache flush when doing full mm flush
Aneesh Kumar K.V [Sat, 1 Apr 2017 14:41:47 +0000 (20:11 +0530)]
powerpc/mm/radix: Don't do page walk cache flush when doing full mm flush

For fullmm tlb flush, we do a flush with RIC_FLUSH_ALL which will invalidate all
related caches (radix__tlb_flush()). Hence the pwc flush is not needed.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Acked-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Fixup LPCR:PECE and HEIC setting on POWER9
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:55 +0000 (17:54 +1000)]
powerpc: Fixup LPCR:PECE and HEIC setting on POWER9

We need to set LPES in order for normal external interrupts (0x500)
to be directed to the guest while running in guest state.

We also need HEIC set to prevent them to be sent to the host while
in host state.

With XIVE the host never gets one of these and wouldn't know how to
handle it. All host external interrupts come in via the new
hypervisor virtualization interrupts vector.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Consolidate variants of real-mode MMIOs
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:54 +0000 (17:54 +1000)]
powerpc: Consolidate variants of real-mode MMIOs

We have all sort of variants of MMIO accessors for the real mode
instructions. This creates a clean set of accessors based on
Linux normal naming conventions, replacing all occurrences of
the old ones in the tree.

I have purposefully removed the "out/in" variants in favor of
only including __raw variants. Any code using these is already
pretty much hand tuned to operate in a very specific environment.
I've fixed up the 2 users (only one of them actually needed
a barrier in the first place).

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/kvm: Remove obsolete kvm_vm_ioctl_xics_irq declaration
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:53 +0000 (17:54 +1000)]
powerpc/kvm: Remove obsolete kvm_vm_ioctl_xics_irq declaration

The function doesn't exist anymore

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/kvm: Make kvmppc_xics_create_icp static
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:52 +0000 (17:54 +1000)]
powerpc/kvm: Make kvmppc_xics_create_icp static

It's only used within the same file it's defined

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/kvm: Massage order of #include
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:51 +0000 (17:54 +1000)]
powerpc/kvm: Massage order of #include

We traditionally have linux/ before asm/

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/xive: Native exploitation of the XIVE interrupt controller
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:50 +0000 (17:54 +1000)]
powerpc/xive: Native exploitation of the XIVE interrupt controller

The XIVE interrupt controller is the new interrupt controller
found in POWER9. It supports advanced virtualization capabilities
among other things.

Currently we use a set of firmware calls that simulate the old
"XICS" interrupt controller but this is fairly inefficient.

This adds the framework for using XIVE along with a native
backend which OPAL for configuration. Later, a backend allowing
the use in a KVM or PowerVM guest will also be provided.

This disables some fast path for interrupts in KVM when XIVE is
enabled as these rely on the firmware emulation code which is no
longer available when the XIVE is used natively by Linux.

A latter patch will make KVM also directly exploit the XIVE, thus
recovering the lost performance (and more).

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
[mpe: Fixup pr_xxx("XIVE:"...), don't split pr_xxx() strings,
 tweak Kconfig so XIVE_NATIVE selects XIVE and depends on POWERNV,
 fix build errors when SMP=n, fold in fixes from Ben:
   Don't call cpu_online() on an invalid CPU number
   Fix irq target selection returning out of bounds cpu#
   Extra sanity checks on cpu numbers
 ]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/smp: Remove migrate_irq() custom implementation
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:49 +0000 (17:54 +1000)]
powerpc/smp: Remove migrate_irq() custom implementation

Some powerpc platforms use this to move IRQs away from a CPU being
unplugged. This function has several bugs such as not taking the right
locks or failing to NULL check pointers.

There's a new generic function doing exactly the same thing without all
the bugs, so let's use it instead.

mpe: The obvious place for the select of GENERIC_IRQ_MIGRATION is on
HOTPLUG_CPU, but that doesn't work. On some configs PM_SLEEP_SMP will
select HOTPLUG_CPU even though its dependencies are not met, which means
the select of GENERIC_IRQ_MIGRATION doesn't happen. That leads to the
build breaking. Fix it by moving the select of GENERIC_IRQ_MIGRATION to
SMP.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Add optional smp_ops->prepare_cpu SMP callback
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:48 +0000 (17:54 +1000)]
powerpc: Add optional smp_ops->prepare_cpu SMP callback

Some platforms (will) need to perform allocations before bringing
a new CPU online. Doing it from smp_ops->setup_cpu is the wrong
thing to do:

 - It has no useful failure path (too late)
 - Calling any allocator will enable interrupts prematurely
   causing problems with large decrementer among others

Instead, add a new callback that is called from __cpu_up (so from
the context trying to online the new CPU) at a point where we
can safely allocate and handle failures.

This will be used by XIVE support.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Add more PPC bit conversion macros
Benjamin Herrenschmidt [Wed, 5 Apr 2017 07:54:47 +0000 (17:54 +1000)]
powerpc: Add more PPC bit conversion macros

Add 32 and 8 bit variants

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Add XIVE related definitions to opal-api.h
Benjamin Herrenschmidt [Wed, 5 Apr 2017 23:01:33 +0000 (09:01 +1000)]
powerpc/powernv: Add XIVE related definitions to opal-api.h

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Add OPAL exports attributes to sysfs
Matt Brown [Wed, 29 Mar 2017 23:28:01 +0000 (10:28 +1100)]
powerpc/powernv: Add OPAL exports attributes to sysfs

New versions of OPAL have a device node /ibm,opal/firmware/exports, each
property of which describes a range of memory in OPAL that Linux might
want to export to userspace for debugging.

This patch adds a sysfs file under 'opal/exports' for each property
found there, and makes it read-only by root.

Signed-off-by: Matt Brown <matthew.brown.dev@gmail.com>
[mpe: Drop counting of props, rename to attr, free on sysfs error, c'log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/prom: Increase minimum RMA size to 512MB
Sukadev Bhattiprolu [Mon, 27 Mar 2017 23:43:14 +0000 (19:43 -0400)]
powerpc/prom: Increase minimum RMA size to 512MB

When booting very large systems with a large initrd, we run out of
space early in boot for either RTAS or the flattened device tree (FDT).
Boot fails with messages like:

Could not allocate memory for RTAS
or
No memory for flatten_device_tree (no room)

Increasing the minimum RMA size to 512MB fixes the problem. This
should not have an impact on smaller LPARs (with 256MB memory),
as the firmware will cap the RMA to the memory assigned to the LPAR.

Fix is based on input/discussions with Michael Ellerman. Thanks to
Praveen K. Pandey for testing on a large system.

Reported-by: Praveen K. Pandey <preveen.pandey@in.ibm.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Introduce address translation services for Nvlink2
Alistair Popple [Mon, 3 Apr 2017 09:51:44 +0000 (19:51 +1000)]
powerpc/powernv: Introduce address translation services for Nvlink2

Nvlink2 supports address translation services (ATS) allowing devices
to request address translations from an mmu known as the nest MMU
which is setup to walk the CPU page tables.

To access this functionality certain firmware calls are required to
setup and manage hardware context tables in the nvlink processing unit
(NPU). The NPU also manages forwarding of TLB invalidates (known as
address translation shootdowns/ATSDs) to attached devices.

This patch exports several methods to allow device drivers to register
a process id (PASID/PID) in the hardware tables and to receive
notification of when a device should stop issuing address translation
requests (ATRs). It also adds a fault handler to allow device drivers
to demand fault pages in.

Signed-off-by: Alistair Popple <alistair@popple.id.au>
[mpe: Fix up comment formatting, use flush_tlb_mm()]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/powernv: Add sanity checks to pnv_pci_get_{gpu|npu}_dev
Alistair Popple [Mon, 3 Apr 2017 09:51:43 +0000 (19:51 +1000)]
powerpc/powernv: Add sanity checks to pnv_pci_get_{gpu|npu}_dev

The pnv_pci_get_{gpu|npu}_dev functions are used to find associations
between nvlink PCIe devices and standard PCIe devices. However they
lacked basic sanity checking which results in NULL pointer
dereferencing if they are incorrect called can be harder to spot than
an explicit WARN_ON.

Signed-off-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agodrivers/of/base.c: Add of_property_read_u64_index
Alistair Popple [Mon, 3 Apr 2017 09:51:42 +0000 (19:51 +1000)]
drivers/of/base.c: Add of_property_read_u64_index

There is of_property_read_u32_index but no u64 variant. This patch
adds one similar to the u32 version for u64.

Signed-off-by: Alistair Popple <alistair@popple.id.au>
Acked-by: Rob Herring <robh@kernel.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Remove stale comment about the DART hole
Oliver O'Halloran [Mon, 3 Apr 2017 08:09:06 +0000 (18:09 +1000)]
powerpc/mm: Remove stale comment about the DART hole

The code to fix the problem it describes was removed in commit
c40785ad305b ("powerpc/dart: Use a cachable DART"), and it uses the
stupid comment style. Away it goooooooooooooes!

Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc: Avoid taking a data miss on every userspace instruction miss
Anton Blanchard [Mon, 3 Apr 2017 06:41:02 +0000 (16:41 +1000)]
powerpc: Avoid taking a data miss on every userspace instruction miss

Early on in do_page_fault() we call store_updates_sp(), regardless of
the type of exception. For an instruction miss this doesn't make
sense, because we only use this information to detect if a data miss
is the result of a stack expansion instruction or not.

Worse still, it results in a data miss within every userspace
instruction miss handler, because we try and load the very instruction
we are about to install a pte for!

A simple exec microbenchmark runs 6% faster on POWER8 with this fix:

 #include <stdlib.h>
 #include <stdio.h>
 #include <unistd.h>

int main(int argc, char *argv[])
{
unsigned long left = atol(argv[1]);
char leftstr[16];

if (left-- == 0)
return 0;

sprintf(leftstr, "%ld", left);
execlp(argv[0], argv[0], leftstr, NULL);
perror("exec failed\n");

return 0;
}

Pass the number of iterations on the command line (eg 10000) and time
how long it takes to execute.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/book3s: Print task info if we take a machine check in user mode
Michael Ellerman [Mon, 3 Apr 2017 05:29:34 +0000 (15:29 +1000)]
powerpc/book3s: Print task info if we take a machine check in user mode

For an MCE (Machine Check Exception) that hits while in user mode
MSR(PR=1), print the task info to the console MCE error log. This may
help to identify an application that triggered the MCE.

After this patch the MCE console looks like:

  Severe Machine check interrupt [Recovered]
    NIP: [0000000010039778] PID: 762 Comm: ebizzy
    Initiator: CPU
    Error type: SLB [Multihit]
      Effective address: 0000000010039778

  Severe Machine check interrupt [Not recovered]
    NIP: [0000000010039778] PID: 763 Comm: ebizzy
    Initiator: CPU
    Error type: UE [Page table walk ifetch]
      Effective address: 0000000010039778
  ebizzy[763]: unhandled signal 7 at 0000000010039778 nip 0000000010039778 lr 0000000010001b44 code 30004

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/book3s: Print the kernel function name in machine check
Mahesh Salgaonkar [Tue, 28 Mar 2017 13:45:04 +0000 (19:15 +0530)]
powerpc/book3s: Print the kernel function name in machine check

For D-side errors we print the load/store address that caused the
machine check as 'Effective address'. But the instruction that may have
caused the machine check can also be helpful, so in addition to printing
the NIP, also print the kernel function name as well.

After this patch the MCE console log would look like:

  Severe Machine check interrupt [Recovered]
    NIP [d00000001bc70194]: init_module+0x194/0x2b0 [bork_kernel]
    Initiator: CPU
    Error type: SLB [Parity]
      Effective address: d000000026de0000

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Enable mappings above 128TB
Aneesh Kumar K.V [Thu, 30 Mar 2017 11:05:21 +0000 (16:35 +0530)]
powerpc/mm: Enable mappings above 128TB

Not all user space application is ready to handle wide addresses. It's
known that at least some JIT compilers use higher bits in pointers to
encode their information. It collides with valid pointers with 512TB
addresses and leads to crashes.

To mitigate this, we are not going to allocate virtual address space
above 128TB by default.

But userspace can ask for allocation from full address space by
specifying hint address (with or without MAP_FIXED) above 128TB.

If hint address set above 128TB, but MAP_FIXED is not specified, we try
to look for unmapped area by specified address. If it's already
occupied, we look for unmapped area in *full* address space, rather than
from 128TB window.

This approach helps to easily make application's memory allocator aware
about large address space without manually tracking allocated virtual
address space.

This is going to be a per mmap decision. ie, we can have some mmaps with
larger addresses and other that do not.

A sample memory layout looks like:

  10000000-10010000 r-xp 00000000 fc:00 9057045          /home/max_addr_512TB
  10010000-10020000 r--p 00000000 fc:00 9057045          /home/max_addr_512TB
  10020000-10030000 rw-p 00010000 fc:00 9057045          /home/max_addr_512TB
  10029630000-10029660000 rw-p 00000000 00:00 0          [heap]
  7fff834a0000-7fff834b0000 rw-p 00000000 00:00 0
  7fff834b0000-7fff83670000 r-xp 00000000 fc:00 9177190  /lib/powerpc64le-linux-gnu/libc-2.23.so
  7fff83670000-7fff83680000 r--p 001b0000 fc:00 9177190  /lib/powerpc64le-linux-gnu/libc-2.23.so
  7fff83680000-7fff83690000 rw-p 001c0000 fc:00 9177190  /lib/powerpc64le-linux-gnu/libc-2.23.so
  7fff83690000-7fff836a0000 rw-p 00000000 00:00 0
  7fff836a0000-7fff836c0000 r-xp 00000000 00:00 0        [vdso]
  7fff836c0000-7fff83700000 r-xp 00000000 fc:00 9177193  /lib/powerpc64le-linux-gnu/ld-2.23.so
  7fff83700000-7fff83710000 r--p 00030000 fc:00 9177193  /lib/powerpc64le-linux-gnu/ld-2.23.so
  7fff83710000-7fff83720000 rw-p 00040000 fc:00 9177193  /lib/powerpc64le-linux-gnu/ld-2.23.so
  7fffdccf0000-7fffdcd20000 rw-p 00000000 00:00 0        [stack]
  1000000000000-1000000010000 rw-p 00000000 00:00 0
  1ffff83710000-1ffff83720000 rw-p 00000000 00:00 0

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Switch some TASK_SIZE checks to use mm_context addr_limit
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:37:01 +0000 (09:07 +0530)]
powerpc/mm: Switch some TASK_SIZE checks to use mm_context addr_limit

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/pseries: Skip using reserved virtual address range
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:37:00 +0000 (09:07 +0530)]
powerpc/pseries: Skip using reserved virtual address range

Now that we use all the available virtual address range, we need to make
sure we don't generate VSID such that it overlaps with the reserved vsid
range. Reserved vsid range include the virtual address range used by the
adjunct partition and also the VRMA virtual segment. We find the context
value that can result in generating such a VSID and reserve it early in
boot.

We don't look at the adjunct range, because for now we disable the
adjunct usage in a Linux LPAR via CAS interface.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
[mpe: Rewrite hash__reserve_context_id(), move the rest into pseries]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Store addr_limit in PACA
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:59 +0000 (09:06 +0530)]
powerpc/mm/hash: Store addr_limit in PACA

We optmize the slice page size array copy to paca by copying only the
range based on addr_limit. This will require us to not look at page size
array beyond addr_limit in PACA on slb fault. To enable that copy task
size to paca which will be used during slb fault.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
[mpe: Rename from task_size to addr_limit, consolidate #ifdefs]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Add addr_limit to mm_context and use it to derive max slice index
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:58 +0000 (09:06 +0530)]
powerpc/mm: Add addr_limit to mm_context and use it to derive max slice index

In the followup patch, we will increase the slice array size to handle
512TB range, but will limit the max addr to 128TB. Avoid doing
unnecessary computation and avoid doing slice mask related operation
above address limit.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Increase VA range to 128TB
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:57 +0000 (09:06 +0530)]
powerpc/mm/hash: Increase VA range to 128TB

We update the hash linux page table layout such that we can support
512TB. But we limit the TASK_SIZE to 128TB. We can switch to 128TB by
default without conditional because that is the max virtual address
supported by other architectures. We will later add a mechanism to
on-demand increase the application's effective address range to 512TB.

Having the page table layout changed to accommodate 512TB makes testing
large memory configuration easier with less code changes to kernel

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Convert mask to unsigned long
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:56 +0000 (09:06 +0530)]
powerpc/mm/hash: Convert mask to unsigned long

This doesn't have any functional change. But helps in avoiding mistakes
in case the shift bit changes

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Support 68 bit VA
Aneesh Kumar K.V [Wed, 29 Mar 2017 06:21:53 +0000 (17:21 +1100)]
powerpc/mm/hash: Support 68 bit VA

Inorder to support large effective address range (512TB), we want to
increase the virtual address bits to 68. But we do have platforms like
p4 and p5 that can only do 65 bit VA. We support those platforms by
limiting context bits on them to 16.

The protovsid -> vsid conversion is verified to work with both 65 and 68
bit va values. I also documented the restrictions in a table format as
part of code comments.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Check for non-kernel address in get_kernel_vsid()
Michael Ellerman [Wed, 29 Mar 2017 12:10:34 +0000 (23:10 +1100)]
powerpc/mm/hash: Check for non-kernel address in get_kernel_vsid()

get_kernel_vsid() has a very stern comment saying that it's only valid
for kernel addresses, but there's nothing in the code to enforce that.

Rather than hoping our callers are well behaved, add a check and return
a VSID of 0 (invalid).

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Use context ids 1-4 for the kernel
Aneesh Kumar K.V [Wed, 29 Mar 2017 12:10:22 +0000 (23:10 +1100)]
powerpc/mm/hash: Use context ids 1-4 for the kernel

Currently we use the top 4 context ids (0x7fffc-0x7ffff) for the kernel.
Kernel VSIDs are built using these top context values and effective the
segement ID. In subsequent patches we want to increase the max effective
address to 512TB. We will achieve that by increasing the effective
segment IDs there by increasing virtual address range.

We will be switching to a 68bit virtual address in the following patch.
But platforms like Power4 and Power5 only support a 65 bit virtual
address. We will handle that by limiting the context bits to 16 instead
of 19 on those platforms. That means the max context id will have a
different value on different platforms.

So that we don't have to deal with the kernel context ids changing
between different platforms, move the kernel context ids down to use
context ids 1-4.

We can't use segment 0 of context-id 0, because that maps to VSID 0,
which we want to keep as invalid, so we avoid context-id 0 entirely.
Similarly we can't use the last segment of the maximum context, so we
avoid it too.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
[mpe: Switch from 0-3 to 1-4 so VSID=0 remains invalid]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Split radix vs hash mm context initialisation
Michael Ellerman [Wed, 29 Mar 2017 11:36:56 +0000 (22:36 +1100)]
powerpc/mm: Split radix vs hash mm context initialisation

Complete the split of the radix vs hash mm context initialisation.

This is mostly code movement, with the exception that we now limit the
context allocation to PRTB_ENTRIES - 1 on radix.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Pull hash constants into hash__alloc_context_id()
Michael Ellerman [Wed, 29 Mar 2017 11:10:45 +0000 (22:10 +1100)]
powerpc/mm/hash: Pull hash constants into hash__alloc_context_id()

The min and max context id values used in alloc_context_id() are
currently the right values for use on hash, and happen to also be safe
for use on radix.

But we need to change that in a subsequent patch, so make the min/max
ids parameters and pull the hash values into hsah__alloc_context_id().

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/hash: Abstract context id allocation for KVM
Michael Ellerman [Wed, 29 Mar 2017 11:00:46 +0000 (22:00 +1100)]
powerpc/mm/hash: Abstract context id allocation for KVM

KVM wants to be able to allocate an MMU context id, which it does
currently by calling __init_new_context().

We're about to rework that code, so provide a wrapper for KVM so it
can not worry about the details.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/slice: Update slice mask printing to use bitmap printing.
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:52 +0000 (09:06 +0530)]
powerpc/mm/slice: Update slice mask printing to use bitmap printing.

We now get output like below which is much better.

[    0.935306]  good_mask low_slice: 0-15
[    0.935360]  good_mask high_slice: 0-511

Compared to

[    0.953414]  good_mask:1111111111111111 - 1111111111111.........

I also fixed an error with slice_dbg printing.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/slice: Move slice_mask struct definition to slice.c
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:51 +0000 (09:06 +0530)]
powerpc/mm/slice: Move slice_mask struct definition to slice.c

This structure definition need not be in a header since this is used only by
slice.c file. So move it to slice.c. This also allow us to use SLICE_NUM_HIGH
instead of 64.

I also switch the low_slices type to u64 from u16. This doesn't have an impact
on size of struct due to padding added with u16 type. This helps in using
bitmap printing function for printing slice mask.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Remove checks that TASK_SIZE_USER64 is too small
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:50 +0000 (09:06 +0530)]
powerpc/mm: Remove checks that TASK_SIZE_USER64 is too small

Remove the checks that TASK_SIZE_USER64 is smaller than H_PGTABLE_RANGE
and USER_VSID_RANGE.

In a following patch we will deliberately add support for a TASK_SIZE
smaller than both ranges, so this will no longer be an error condition.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
[mpe: Keep the check in pgtable_64.c that we don't exceed USER_VSID_RANGE]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Move copy_mm_to_paca to paca.c
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:49 +0000 (09:06 +0530)]
powerpc/mm: Move copy_mm_to_paca to paca.c

We also update the function arg to struct mm_struct. Move this so that function
finds the definition of struct mm_struct. No functional change in this patch.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/slice: Update the function prototype
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:48 +0000 (09:06 +0530)]
powerpc/mm/slice: Update the function prototype

This avoid copying the slice_mask struct as function return value

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm/slice: Convert slice_mask high slice to a bitmap
Aneesh Kumar K.V [Wed, 22 Mar 2017 03:36:47 +0000 (09:06 +0530)]
powerpc/mm/slice: Convert slice_mask high slice to a bitmap

In followup patch we want to increase the va range which will result
in us requiring high_slices to have more than 64 bits. To enable this
convert high_slices to bitmap. We keep the number bits same in this patch
and later change that to higher value

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
[mpe: Fold in fix to use bitmap_empty()]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
7 years agopowerpc/mm: Move hash specific pte bits to be top bits of RPN
Aneesh Kumar K.V [Tue, 28 Mar 2017 04:21:12 +0000 (15:21 +1100)]
powerpc/mm: Move hash specific pte bits to be top bits of RPN

We don't support the full 57 bits of physical address and hence can
overload the top bits of RPN as hash specific pte bits.

Add a BUILD_BUG_ON() to enforce the relationship between H_PAGE_F_SECOND
and H_PAGE_F_GIX.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Reviewed-by: Paul Mackerras <paulus@ozlabs.org>
[mpe: Move the BUILD_BUG_ON() into hash_utils_64.c and comment it]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>