platform/kernel/linux-rpi.git
10 years agopowerpc/vphn: NUMA node code expects big-endian
Greg Kurz [Wed, 15 Oct 2014 10:42:58 +0000 (12:42 +0200)]
powerpc/vphn: NUMA node code expects big-endian

The associativity domain numbers are obtained from the hypervisor through
registers and written into memory by the guest: the packed array passed to
vphn_unpack_associativity() is then native-endian, unlike what was assumed
in the following commit:

commit b08a2a12e44eaec5024b2b969f4fcb98169d1ca3
Author: Alistair Popple <alistair@popple.id.au>
Date:   Wed Aug 7 02:01:44 2013 +1000

    powerpc: Make NUMA device node code endian safe

This issue fills the topology with bogus data and makes it unusable. It may
lead to severe performance breakdowns.

We should ideally patch the vphn_unpack_associativity() function to do the
64-bit loads, but this requires some more brain storming.

In the meantime, let's go for a suboptimal and temporary bug fix: this patch
converts each 64-bit value of the packed array to big endian, as expected by
the current parsing code in vphn_unpack_associativity().

Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/msi: Use WARN_ON() in msi bitmap selftests
Michael Ellerman [Fri, 10 Oct 2014 08:04:25 +0000 (19:04 +1100)]
powerpc/msi: Use WARN_ON() in msi bitmap selftests

As demonstrated in the previous commit, the failure message from the msi
bitmap selftests is a bit subtle, it's easy to miss a failure in a busy
boot log.

So drop our check() macro and use WARN_ON() instead. This necessitates
inverting all the conditions as well.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/msi: Fix the msi bitmap alignment tests
Michael Ellerman [Fri, 10 Oct 2014 08:04:24 +0000 (19:04 +1100)]
powerpc/msi: Fix the msi bitmap alignment tests

When we added the alignment tests recently we failed to check they were
actually passing - oops.

They weren't passing, because the bitmap was full. We should also be a
bit more careful when checking the return code, a negative error return
could by divisible by our alignment value.

Fixes: b0345bbc6d09 ("powerpc/msi: Improve IRQ bitmap allocator")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Block CFG upon frozen Shiner adapter
Gavin Shan [Fri, 3 Oct 2014 04:58:32 +0000 (14:58 +1000)]
powerpc/eeh: Block CFG upon frozen Shiner adapter

The Broadcom Shiner 2-ports 10G ethernet adapter has same problem
commit 6f20bda0 ("powerpc/eeh: Block PCI config access upon frozen
PE") fixes. Put it to the black list as well.

   # lspci -s 0004:01:00.0
   0004:01:00.0 Ethernet controller: Broadcom Corporation \
                NetXtreme II BCM57810 10 Gigabit Ethernet (rev 10)
   # lspci -n -s 0004:01:00.0
   0004:01:00.0 0200: 14e4:168e (rev 10)

Reported-by: John Walthour <jwalthour@us.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Don't collect logs on PE with blocked config space
Gavin Shan [Wed, 1 Oct 2014 07:07:54 +0000 (17:07 +1000)]
powerpc/eeh: Don't collect logs on PE with blocked config space

When the PE's config space is marked as blocked, PCI config read
requests always return 0xFF's. It's pointless to collect logs in
this case.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Block PCI config access upon frozen PE
Gavin Shan [Wed, 1 Oct 2014 07:07:53 +0000 (17:07 +1000)]
powerpc/eeh: Block PCI config access upon frozen PE

The problem was found when I tried to inject PCI config error by
PHB3 PAPR error injection registers into Broadcom Austin 4-ports
NIC adapter. The frozen PE was reported successfully and EEH core
started to recover it. However, I run into fenced PHB when dumping
PCI config space as EEH logs. I was told that PCI config requests
should not be progagated to the adapter until PE reset is done
successfully. Otherise, we would run out of PHB internal credits
and trigger PCT (PCIE Completion Timeout), which leads to the
fenced PHB.

The patch introduces another PE flag EEH_PE_CFG_RESTRICTED, which
is set during PE initialization time if the PE includes the specific
PCI devices that need block PCI config access until PE reset is done.
When the PE becomes frozen for the first time, EEH_PE_CFG_BLOCKED is
set if the PE has flag EEH_PE_CFG_RESTRICTED. Then the PCI config
access to the PE will be dropped by platform PCI accessors until
PE reset is done successfully. The mechanism is shared by PowerNV
platform owned PE or userland owned ones. It's not used on pSeries
platform yet.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pseries: Drop config requests in EEH accessors
Gavin Shan [Wed, 1 Oct 2014 07:07:52 +0000 (17:07 +1000)]
powerpc/pseries: Drop config requests in EEH accessors

The pSeires EEH config accessors rely on rtas_{read, write}_config()
and the condition to check if the PE's config space is blocked
should be moved to those 2 functions so that config requests from
kernel, userland, EEH core can be dropped to avoid recursive EEH error
if necessary.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Drop config requests in EEH accessors
Gavin Shan [Wed, 1 Oct 2014 07:07:51 +0000 (17:07 +1000)]
powerpc/powernv: Drop config requests in EEH accessors

It's bad idea to access the PCI config registers of the adapters,
which is experiencing reset. It leads to recursive EEH error without
exception. The patch drops PCI config requests in EEH accessors if
the PE has been marked to accept PCI config requests, for example
during PE reseet time.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Rename flag EEH_PE_RESET to EEH_PE_CFG_BLOCKED
Gavin Shan [Wed, 1 Oct 2014 07:07:50 +0000 (17:07 +1000)]
powerpc/eeh: Rename flag EEH_PE_RESET to EEH_PE_CFG_BLOCKED

The flag EEH_PE_RESET indicates blocking config space of the PE
during reset time. We potentially need block PE's config space
other than reset time. So it's reasonable to replace it with
EEH_PE_CFG_BLOCKED to indicate its usage.

There are no substantial code or logic changes in this patch.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Fix condition for isolated state
Gavin Shan [Wed, 1 Oct 2014 07:07:49 +0000 (17:07 +1000)]
powerpc/eeh: Fix condition for isolated state

Function eeh_pe_state_mark() could possibly have combination of
multiple EEH PE state as its argument. The patch fixes the condition
used to check if EEH_PE_ISOLATED is included.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pseries: Make CPU hotplug path endian safe
Bharata B Rao [Tue, 16 Sep 2014 20:15:45 +0000 (15:15 -0500)]
powerpc/pseries: Make CPU hotplug path endian safe

- ibm,rtas-configure-connector should treat the RTAS data as big endian.
- Treat ibm,ppc-interrupt-server#s as big-endian when setting
  smp_processor_id during hotplug.

Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Acked-by: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pseries: Use dump_stack instead of show_stack
Anton Blanchard [Mon, 13 Oct 2014 08:41:40 +0000 (19:41 +1100)]
powerpc/pseries: Use dump_stack instead of show_stack

We can use the simpler dump_stack() instead of
show_stack(current, __get_SP())

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Rename __get_SP() to current_stack_pointer()
Anton Blanchard [Mon, 13 Oct 2014 08:41:39 +0000 (19:41 +1100)]
powerpc: Rename __get_SP() to current_stack_pointer()

Michael points out that __get_SP() is a pretty horrible
function name. Let's give it a better name.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Reimplement __get_SP() as a function not a define
Anton Blanchard [Mon, 13 Oct 2014 08:41:38 +0000 (19:41 +1100)]
powerpc: Reimplement __get_SP() as a function not a define

Li Zhong points out an issue with our current __get_SP()
implementation. If ftrace function tracing is enabled (ie -pg
profiling using _mcount) we spill a stack frame on 64bit all the
time.

If a function calls __get_SP() and later calls a function that is
tail call optimised, we will pop the stack frame and the value
returned by __get_SP() is no longer valid. An example from Li can
be found in save_stack_trace -> save_context_stack:

c0000000000432c0 <.save_stack_trace>:
c0000000000432c0:       mflr    r0
c0000000000432c4:       std     r0,16(r1)
c0000000000432c8:       stdu    r1,-128(r1) <-- stack frame for _mcount
c0000000000432cc:       std     r3,112(r1)
c0000000000432d0:       bl      <._mcount>
c0000000000432d4:       nop

c0000000000432d8:       mr      r4,r1 <-- __get_SP()

c0000000000432dc:       ld      r5,632(r13)
c0000000000432e0:       ld      r3,112(r1)
c0000000000432e4:       li      r6,1

c0000000000432e8:       addi    r1,r1,128 <-- pop stack frame

c0000000000432ec:       ld      r0,16(r1)
c0000000000432f0:       mtlr    r0
c0000000000432f4:       b       <.save_context_stack> <-- tail call optimized

save_context_stack ends up with a stack pointer below the current
one, and it is likely to be scribbled over.

Fix this by making __get_SP() a function which returns the
callers stack frame. Also replace inline assembly which grabs
the stack pointer in save_stack_trace and show_stack with
__get_SP().

This also fixes an issue with perf_arch_fetch_caller_regs().
It currently unwinds the stack once, which will skip a
valid stack frame on a leaf function. With the __get_SP() fixes
in this patch, we never need to unwind the stack frame to get
to the first interesting frame.

We have to export __get_SP() because perf_arch_fetch_caller_regs()
(which is used in modules) calls it from a header file.

Reported-by: Li Zhong <zhong@linux.vnet.ibm.com>
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/numa: Add ability to disable and debug topology updates
Nishanth Aravamudan [Fri, 10 Oct 2014 16:04:49 +0000 (09:04 -0700)]
powerpc/numa: Add ability to disable and debug topology updates

We have hit a few customer issues with the topology update code (VPHN
and PRRN). It would be nice to be able to debug the notifications coming
from the hypervisor in both cases to the LPAR, as well as to disable
responding to the notifications at boot-time, to narrow down the source
of the problems. Add a basic level of such functionality, similar to the
numa= command-line parameter. We already have a toggle in
/proc/powerpc/topology_updates that allows run-time enabling/disabling,
so the updates can be started at run-time if desired. But the bugs we've
run into have occured during boot or very shortly after coming to login,
and have resulted in a broken NUMA topology.

Signed-off-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/numa: check error return from proc_create
Nishanth Aravamudan [Thu, 9 Oct 2014 23:41:28 +0000 (16:41 -0700)]
powerpc/numa: check error return from proc_create

proc_create can fail, we should check the return value and pass up the
failure.

Suggested-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Fallback to old HMI handling behavior for old firmware
Mahesh Salgaonkar [Fri, 10 Oct 2014 15:58:26 +0000 (21:28 +0530)]
powerpc/powernv: Fallback to old HMI handling behavior for old firmware

Recently we moved HMI handling into Linux kernel instead of taking
HMI directly in OPAL. This new change is dependent on new OPAL call
for HMI recovery which was introduced in newer firmware. While this new
change works fine with latest OPAL firmware, we broke the HMI handling
if we run newer kernel on old OPAL firmware that results in system hang.

This patch fixes this issue by falling back to old HMI behavior on older
OPAL firmware.

This patch introduces a check for opal token OPAL_HANDLE_HMI to see
if we are running on newer firmware or old firmware. On newer firmware
this check would return OPAL_TOKEN_PRESENT, otherwise we are running on
old firmware and fallback to old HMI behavior.

Old firmware: POWER8 System Firmware Release as of today <= SV810_087
Action: Let OPAL handle HMIs

Newer firmware: in development/yet to be released.
Action: Let Linux host handle HMIs.

This patch depends on opal check token patch posted at ppc-devel
https://lists.ozlabs.org/pipermail/linuxppc-dev/2014-August/120224.html

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
[mpe: Minor comment and printk rewording]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/book3s: Don't clear MSR_RI in hmi handler.
Mahesh Salgaonkar [Tue, 7 Oct 2014 10:25:33 +0000 (15:55 +0530)]
powerpc/book3s: Don't clear MSR_RI in hmi handler.

In HMI interrupt handler we don't touch SRR0/SRR1, instead we touch
HSRR0/HSRR1. Hence we don't need to clear MSR_RI bit.

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Fix sys_call_table declaration to enable syscall tracing
Romeo Cane [Thu, 2 Oct 2014 14:41:39 +0000 (15:41 +0100)]
powerpc: Fix sys_call_table declaration to enable syscall tracing

Declaring sys_call_table as a pointer causes the compiler to generate
the wrong lookup code in arch_syscall_addr().

     <arch_syscall_addr>:
        lis     r9,-16384
        rlwinm  r3,r3,2,0,29
  -     lwz     r11,30640(r9)
  -     lwzx    r3,r11,r3
  +     addi    r9,r9,30640
  +     lwzx    r3,r9,r3
        blr

The actual sys_call_table symbol, declared in assembler, is an
array. If we lie about that to the compiler we get the wrong code
generated, as above.

This definition seems only to be used by the syscall tracing code in
kernel/trace/trace_syscalls.c. With this patch I can successfully use
the syscall tracepoints:

  bash-3815  [002] ....   333.239082: sys_write -> 0x2
  bash-3815  [002] ....   333.239087: sys_dup2(oldfd: a, newfd: 1)
  bash-3815  [002] ....   333.239088: sys_dup2 -> 0x1
  bash-3815  [002] ....   333.239092: sys_fcntl(fd: a, cmd: 1, arg: 0)
  bash-3815  [002] ....   333.239093: sys_fcntl -> 0x1
  bash-3815  [002] ....   333.239094: sys_close(fd: a)
  bash-3815  [002] ....   333.239094: sys_close -> 0x0

Signed-off-by: Romeo Cane <romeo.cane.ext@coriant.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Fix afu_read() not doing finish_wait() on signal or non-blocking
Ian Munsie [Thu, 9 Oct 2014 00:17:46 +0000 (11:17 +1100)]
cxl: Fix afu_read() not doing finish_wait() on signal or non-blocking

If afu_read() returned due to a signal or the AFU file descriptor being
opened non-blocking it would not call finish_wait() before returning,
which could lead to a crash later when something else wakes up the wait
queue.

This patch restructures the wait logic to ensure that the cleanup is
done correctly.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Add documentation for userspace APIs
Ian Munsie [Wed, 8 Oct 2014 08:55:05 +0000 (19:55 +1100)]
cxl: Add documentation for userspace APIs

This documentation gives an overview of the hardware architecture, userspace
APIs via /dev/cxl/afuM.N and the syfs files. It also adds a MAINTAINERS file
entry for cxl.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Add driver to Kbuild and Makefiles
Ian Munsie [Wed, 8 Oct 2014 08:55:04 +0000 (19:55 +1100)]
cxl: Add driver to Kbuild and Makefiles

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Add userspace header file
Ian Munsie [Wed, 8 Oct 2014 08:55:03 +0000 (19:55 +1100)]
cxl: Add userspace header file

This adds a header file for use by userspace programs wanting to interact with
the kernel cxl driver.  It defines structs and magic numbers required for
userspace to interact with devices in /dev/cxl/afuM.N.

Further documentation on this interface is added in a subsequent patch in
Documentation/powerpc/cxl.txt.

It also adds this new userspace header file to Kbuild so it's exported when
doing "make headers_installs".

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Driver code for powernv PCIe based cards for userspace access
Ian Munsie [Wed, 8 Oct 2014 08:55:02 +0000 (19:55 +1100)]
cxl: Driver code for powernv PCIe based cards for userspace access

This is the core of the cxl driver.

It adds support for using cxl cards in the powernv environment only (ie POWER8
bare metal). It allows access to cxl accelerators by userspace using the
/dev/cxl/afuM.N char devices.

The kernel driver has no knowledge of the function implemented by the
accelerator. It provides services to userspace via the /dev/cxl/afuM.N
devices. When a program opens this device and runs the start work IOCTL, the
accelerator will have coherent access to that processes memory using the same
virtual addresses. That process may mmap the device to access any MMIO space
the accelerator provides.  Also, reads on the device will allow interrupts to
be received. These services are further documented in a later patch in
Documentation/powerpc/cxl.txt.

Documentation of the cxl hardware architecture and userspace API is provided in
subsequent patches.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Add base builtin support
Ian Munsie [Wed, 8 Oct 2014 08:55:01 +0000 (19:55 +1100)]
cxl: Add base builtin support

This adds the base cxl support that cannot be built as a module. Specifically
it adds the cxl callbacks that are called from the core powerpc mm code which
must always exist irrespective of if the cxl module is loaded or not. This is
similar to how cell works with CONFIG_SPU_BASE.

This adds a cxl_slbia() call (similar to spu_flush_all_slbs()) which checks if
the cxl module is loaded and in use, returning immediately if it is not. If it
is in use it calls into the cxl SLB invalidation code.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/mm: Add hooks for cxl
Ian Munsie [Wed, 8 Oct 2014 08:55:00 +0000 (19:55 +1100)]
powerpc/mm: Add hooks for cxl

This adds hooks into the core powerpc mm code for cxl.

The core powerpc code sometimes uses local tlbie. Unfortunately this won't
work with the current cxl driver as it relies on snooping tlbie broadcasts.

The cxl hardware can have TLB entries invalidated via MMIO but this is not
currently supported by the driver. In future we can make local tlbie smarter so
that it invalidates cxl contexts via MMIO when it needs to but for now we have
this workaround.

This workaround checks for any active cxl contexts and if so, disables local
tlbie.

This also adds a hook for when SLBs are invalidated. This ensures any
corresponding SLBs in cxl are also invalidated at the same time. This is
required for segment demotion.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/opal: Add PHB to cxl mode call
Ian Munsie [Wed, 8 Oct 2014 08:54:59 +0000 (19:54 +1100)]
powerpc/opal: Add PHB to cxl mode call

This adds the OPAL call to change a PHB into cxl mode.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/mm: Add new hash_page_mm()
Ian Munsie [Wed, 8 Oct 2014 08:54:58 +0000 (19:54 +1100)]
powerpc/mm: Add new hash_page_mm()

This adds a new function hash_page_mm() based on the existing hash_page().
This version allows any struct mm to be passed in, rather than assuming
current. This is useful for servicing co-processor faults which are not in the
context of the current running process.

We need to be careful here as the current hash_page() assumes current in a few
places.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powerpc: Add new PCIe functions for allocating cxl interrupts
Ian Munsie [Wed, 8 Oct 2014 08:54:57 +0000 (19:54 +1100)]
powerpc/powerpc: Add new PCIe functions for allocating cxl interrupts

This adds a number of functions for allocating IRQs under powernv PCIe for cxl.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agocxl: Add new header for call backs and structs
Ian Munsie [Wed, 8 Oct 2014 08:54:56 +0000 (19:54 +1100)]
cxl: Add new header for call backs and structs

This new header adds callbacks and structs needed by the rest of the kernel to
hook into the cxl infrastructure.

This adds the cxl_ctx_in_use() function for use in the mm code to see if any
cxl contexts are currently in use. This is used by the tlbie() to determine if
it can do local TLB invalidations or not. This also adds get/put calls for the
cxl driver module to refcount the active cxl contexts.

cxl_ctx_get/put/in_use are static inlined here as they are called in tlbie
which we want to be fast (mpe's suggestion).

Empty functions are provided when CONFIG_CXL_BASE is not enabled.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Split out set MSI IRQ chip code
Ian Munsie [Wed, 8 Oct 2014 08:54:55 +0000 (19:54 +1100)]
powerpc/powernv: Split out set MSI IRQ chip code

Some of the MSI IRQ code in pnv_pci_ioda_msi_setup() is generically useful so
split it out.

This will be used by some of the cxl PCIe code later.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/mm: Export mmu_kernel_ssize and mmu_linear_psize
Ian Munsie [Wed, 8 Oct 2014 08:54:54 +0000 (19:54 +1100)]
powerpc/mm: Export mmu_kernel_ssize and mmu_linear_psize

Export mmu_kernel_ssize and mmu_linear_psize.  These are needed by the cxl
driver which has it's own MMU.  To setup the MMU cxl needs access to these.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/msi: Improve IRQ bitmap allocator
Ian Munsie [Wed, 8 Oct 2014 08:54:53 +0000 (19:54 +1100)]
powerpc/msi: Improve IRQ bitmap allocator

Currently msi_bitmap_alloc_hwirqs() will round up any IRQ allocation requests
to the nearest power of 2. eg. ask for 5 IRQs and you'll get 8. This wastes a
lot of IRQs which can be a scarce resource.

For cxl we may require multiple IRQs for every context that is attached to the
accelerator. There may be 1000s of contexts attached, hence we can easily run
out of IRQs, especially if we are needlessly wasting them.

This changes the msi_bitmap_alloc_hwirqs() to allocate only the required number
of IRQs, hence avoiding this wastage. It keeps the natural alignment
requirement though.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/cell: Make spu_flush_all_slbs() generic
Ian Munsie [Wed, 8 Oct 2014 08:54:52 +0000 (19:54 +1100)]
powerpc/cell: Make spu_flush_all_slbs() generic

This moves spu_flush_all_slbs() into a generic call copro_flush_all_slbs().

This will be useful when we add cxl which also needs a similar SLB flush call.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/cell: Move data segment faulting code out of cell platform
Ian Munsie [Wed, 8 Oct 2014 08:54:51 +0000 (19:54 +1100)]
powerpc/cell: Move data segment faulting code out of cell platform

__spu_trap_data_seg() currently contains code to determine the VSID and ESID
required for a particular EA and mm struct.

This code is generically useful for other co-processors. This moves the code of
the cell platform so it can be used by other powerpc code. It also adds 1TB
segment handling which Cell didn't support.  The new function is called
copro_calculate_slb().

This also moves the internal struct spu_slb to a generic struct copro_slb which
is now used in the Cell and copro code.  We use this new struct instead of
passing around esid and vsid parameters.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/cell: Move spu_handle_mm_fault() out of cell platform
Ian Munsie [Wed, 8 Oct 2014 08:54:50 +0000 (19:54 +1100)]
powerpc/cell: Move spu_handle_mm_fault() out of cell platform

Currently spu_handle_mm_fault() is in the cell platform.

This code is generically useful for other non-cell co-processors on powerpc.

This patch moves this function out of the cell platform into arch/powerpc/mm so
that others may use it.

Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pseries: Use new defines when calling H_SET_MODE
Michael Neuling [Thu, 29 May 2014 07:45:47 +0000 (17:45 +1000)]
powerpc/pseries: Use new defines when calling H_SET_MODE

Now that we define these in the KVM code, use these defines when we call
H_SET_MODE. No functional change.

Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Update contact info in Documentation files
sukadev@linux.vnet.ibm.com [Wed, 1 Oct 2014 06:03:21 +0000 (23:03 -0700)]
powerpc: Update contact info in Documentation files

Cody's email address has changed. Update the contact information for
the 24x7 and GPCI counters to the PowerPC developers mailing list.

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/perf/hv-24x7: Simplify catalog_read()
sukadev@linux.vnet.ibm.com [Wed, 1 Oct 2014 06:03:18 +0000 (23:03 -0700)]
powerpc/perf/hv-24x7: Simplify catalog_read()

catalog_read() implements the read interface for the sysfs file

/sys/bus/event_source/devices/hv_24x7/interface/catalog

It essentially takes a buffer, an offset and count as parameters
to the read() call.  It makes a hypervisor call to read a specific
page from the catalog and copy the required bytes into the given
buffer. Each call to catalog_read() returns at most one 4K page.

Given these requirements, we should be able to simplify the
catalog_read().

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/perf/hv-24x7: use kmem_cache instead of aligned stack allocations
Cody P Schafer [Wed, 1 Oct 2014 06:03:17 +0000 (23:03 -0700)]
powerpc/perf/hv-24x7: use kmem_cache instead of aligned stack allocations

Ian pointed out the use of __aligned(4096) caused rather large stack
consumption in single_24x7_request(), so use the kmem_cache
hv_page_cache (which we've already got set up for other allocations)
insead of allocating locally.

CC: Haren Myneni <hbabu@us.ibm.com>
Reported-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Cody P Schafer <dev@codyps.com>
Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Fix endian bug in LPC bus debugfs accessors
Benjamin Herrenschmidt [Fri, 3 Oct 2014 07:12:25 +0000 (17:12 +1000)]
powerpc/powernv: Fix endian bug in LPC bus debugfs accessors

When reading from the LPC, the OPAL FW calls return the value via pointer
to a uint32_t which is always returned big endian. Our internal inb/outb
implementation byteswaps that fine but our debugfs code is still broken.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: <stable@vger.kernel.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agoMerge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/scottwood/linux.git
Michael Ellerman [Fri, 3 Oct 2014 22:59:06 +0000 (08:59 +1000)]
Merge branch 'next' of git://git./linux/kernel/git/scottwood/linux.git

Freescale updates from Scott (27 commits):

  "Highlights include DMA32 zone support (SATA, USB, etc now works on 64-bit
   FSL kernels), MSI changes, 8xx optimizations and cleanup, t104x board
   support, and PrPMC PCI enumeration."

10 years agopowerpc: Enable CONFIG_CRASH_DUMP=y for ppc64_defconfig
Michael Ellerman [Wed, 24 Sep 2014 05:57:11 +0000 (15:57 +1000)]
powerpc: Enable CONFIG_CRASH_DUMP=y for ppc64_defconfig

It pulls in more code, including causing us to build a relocatable
kernel, which is good for testing.

The resulting kernel is still usable as a non-crash dump kernel.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/kdump: crash_dump.c needs to include io.h
Michael Ellerman [Wed, 24 Sep 2014 05:57:12 +0000 (15:57 +1000)]
powerpc/kdump: crash_dump.c needs to include io.h

For __ioremap().

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Don't build powernv for other platform defconfigs
Michael Ellerman [Wed, 24 Sep 2014 05:57:10 +0000 (15:57 +1000)]
powerpc: Don't build powernv for other platform defconfigs

Because powernv arrived after these other platforms, the defconfigs
didn't have PPC_POWERNV disabled, and being default y it gets turned on.

If we're going to bother having defconfigs for the specific platforms
then they should only build the code required for those platforms.

The grab bag of everything config is ppc64_defconfig.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pci: remove duplicate declaration of pci_bus_find_capability
Wei Yang [Fri, 19 Sep 2014 09:25:22 +0000 (17:25 +0800)]
powerpc/pci: remove duplicate declaration of pci_bus_find_capability

pci_bus_find_capability() is decleared in pci.h, so it is not necessary to do
it again.

This patch removes it.

Signed-off-by: Wei Yang <weiyang@linux.vnet.ibm.com>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/iommu/ddw: Fix endianness
Alexey Kardashevskiy [Thu, 25 Sep 2014 06:39:18 +0000 (16:39 +1000)]
powerpc/iommu/ddw: Fix endianness

rtas_call() accepts and returns values in CPU endianness.
The ddw_query_response and ddw_create_response structs members are
defined and treated as BE but as they are passed to rtas_call() as
(u32 *) and they get byteswapped automatically, the data is CPU-endian.
This fixes ddw_query_response and ddw_create_response definitions and use.

of_read_number() is designed to work with device tree cells - it assumes
the input is big-endian and returns data in CPU-endian. However due
to the ddw_create_response struct fix, create.addr_hi/lo are already
CPU-endian so do not byteswap them.

ddw_avail is a pointer to the "ibm,ddw-applicable" property which contains
3 cells which are big-endian as it is a device tree. rtas_call() accepts
a RTAS token in CPU-endian. This makes use of of_property_read_u32_array
to byte swap and avoid the need for a number of be32_to_cpu calls.

Cc: stable@vger.kernel.org # v3.13+
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
[aik: folded Anton's patch with of_property_read_u32_array]
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Add printk levels to powerpc code
Anton Blanchard [Wed, 17 Sep 2014 04:39:39 +0000 (14:39 +1000)]
powerpc: Add printk levels to powerpc code

Add printk levels to some places in the powerpc port.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Add printk levels to powernv platform code
Anton Blanchard [Wed, 17 Sep 2014 04:39:38 +0000 (14:39 +1000)]
powerpc: Add printk levels to powernv platform code

Add printk levels to powernv platform code, and convert to
pr_err() etc while here.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Remove powerpc specific cmd_line
Anton Blanchard [Wed, 17 Sep 2014 04:39:36 +0000 (14:39 +1000)]
powerpc: Remove powerpc specific cmd_line

There is no need for yet another copy of the command line, just
use boot_command_line like everyone else.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Use pr_fmt in module loader code
Anton Blanchard [Wed, 17 Sep 2014 04:39:35 +0000 (14:39 +1000)]
powerpc: Use pr_fmt in module loader code

Use pr_fmt to give some context to the error messages in the
module code, and convert open coded debug printk to pr_debug.

Use pr_err for error messages.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Fill in si_addr_lsb siginfo field
Anton Blanchard [Wed, 24 Sep 2014 06:59:58 +0000 (16:59 +1000)]
powerpc: Fill in si_addr_lsb siginfo field

Fill in the si_addr_lsb siginfo field so the hwpoison code can
pass to userspace the length of memory that has been corrupted.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Add VM_FAULT_HWPOISON handling to powerpc page fault handler
Anton Blanchard [Wed, 24 Sep 2014 06:59:57 +0000 (16:59 +1000)]
powerpc: Add VM_FAULT_HWPOISON handling to powerpc page fault handler

do_page_fault was missing knowledge of HWPOISON, and we would oops
if userspace tried to access a poisoned page:

kernel BUG at arch/powerpc/mm/fault.c:180!

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Simplify do_sigbus
Anton Blanchard [Wed, 24 Sep 2014 06:59:56 +0000 (16:59 +1000)]
powerpc: Simplify do_sigbus

Exit out early for a kernel fault, avoiding indenting of
most of the function.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Speed up clear_page by unrolling it
Anton Blanchard [Thu, 2 Oct 2014 05:44:21 +0000 (15:44 +1000)]
powerpc: Speed up clear_page by unrolling it

Unroll clear_page 8 times. A simple microbenchmark which
allocates and frees a zeroed page:

for (i = 0; i < iterations; i++) {
unsigned long p = __get_free_page(GFP_KERNEL | __GFP_ZERO);
free_page(p);
}

improves 20% on POWER8.

This assumes cacheline sizes won't grow beyond 512 bytes or
page sizes wont drop below 1kB, which is unlikely, but we could
add a runtime check during early init if it makes people nervous.

Michael found that some versions of gcc produce quite bad code
(all multiplies), so we give gcc a hand by using shifts and adds.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Show hex prefix for PE state sysfs
Gavin Shan [Wed, 1 Oct 2014 04:34:51 +0000 (14:34 +1000)]
powerpc/eeh: Show hex prefix for PE state sysfs

As Michael suggested, the hex prefix for the output of EEH PE
state sysfs entry (/sys/bus/pci/devices/xxx/eeh_pe_state) is
always informative to users.

Suggested-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Override dma_get_required_mask()
Gavin Shan [Tue, 30 Sep 2014 02:39:10 +0000 (12:39 +1000)]
powerpc/powernv: Override dma_get_required_mask()

The dma_get_required_mask() function is used by some drivers to
query the platform about what DMA mask is needed to cover all of
memory. This is a bit of a strange semantic when we have to choose
between IOMMU translation or bypass, but essentially what it means
is "what DMA mask will give best performances".

Currently, our IOMMU backend always returns a 32-bit mask here, we
don't do anything special to it when we have bypass available. This
causes some drivers to choose a 32-bit mask, thus losing the ability
to use the bypass window, thinking this is more efficient. The problem
was reported from the driver of following device:

0004:03:00.0 0107: 1000:0087 (rev 05)
0004:03:00.0 Serial Attached SCSI controller: LSI Logic / Symbios \
             Logic SAS2308 PCI-Express Fusion-MPT SAS-2 (rev 05)

This patch adds an override of that function in order to, instead,
return a 64-bit mask whenever a bypass window is available in order
for drivers to prefer this configuration.

Reported-by: Murali N. Iyer <mniyer@us.ibm.com>
Suggested-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Fetch frozen PE on top level
Gavin Shan [Tue, 30 Sep 2014 02:39:09 +0000 (12:39 +1000)]
powerpc/powernv: Fetch frozen PE on top level

It should have been part of commit 1ad7a72c5 ("powerpc/eeh: Report
frozen parent PE prior to child PE"). There are 2 ways to report
EEH errors: proactively polling because of 0xFF's returned from
PCI config or IO read, or interrupt driven event. We missed to
report and handle parent frozen PE prior to child frozen PE for
the later case on PowerNV platform.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Dump PCI config space for all child devices
Gavin Shan [Tue, 30 Sep 2014 02:39:08 +0000 (12:39 +1000)]
powerpc/eeh: Dump PCI config space for all child devices

The PEs can be organized as nested. Current implementation doesn't
dump PCI config space for subordinate devices of child PEs. However,
the frozen PE could be caused by those subordinate devices of its
child PEs.

The patch dumps PCI config space for all subordinate devices of the
problematic PE.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Emulate EEH recovery for VFIO devices
Gavin Shan [Tue, 30 Sep 2014 02:39:07 +0000 (12:39 +1000)]
powerpc/eeh: Emulate EEH recovery for VFIO devices

When enabling EEH functionality on passed through devices (PE)
with VFIO, the devices in the PE would be removed permanently
from guest side. In that case, the PE remains frozen state.
When returning PE to host, or restarting the guest again, we
had mechanism unfreezing the PE by clearing PESTA/B frozen
bits. However, that's not enough for some adapters, which are
indicated as following "lspci" shows. Those adapters require
hot reset on the parent bus to bring their firmware back to
workable state. Otherwise, those adaptrs won't be operative
and the host (for returning case) or the guest will fail to
load the drivers for those adapters without exception.

0000:01:00.0 Ethernet controller: Emulex Corporation OneConnect \
             10Gb NIC (be3) (rev 02)
0000:01:00.0 0200: 19a2:0710 (rev 02)
0001:03:00.0 Ethernet controller: Emulex Corporation OneConnect \
             NIC (Lancer) (rev 10)
0001:03:00.0 0200: 10df:e220 (rev 10)

The patch adds mechanism to emulate EEH recovery (for hot reset
on parent PCI bus) on 3 gates to fix the issue: open/release one
adapter of the PE, enable EEH functionality on one adapter of the
PE.

Reported-by: Murilo Fossa Vicentini <muvic@br.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Tag reset state for user owned PE
Gavin Shan [Tue, 30 Sep 2014 02:39:06 +0000 (12:39 +1000)]
powerpc/eeh: Tag reset state for user owned PE

PE would be owned by userland, which probably request PE reset
done in host side. During the reset, we should drop the PCI
config accesses to the PE with help of flag EEH_PE_RESET.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Sync OpalPciResetScope with firmware
Gavin Shan [Tue, 30 Sep 2014 02:39:05 +0000 (12:39 +1000)]
powerpc/powernv: Sync OpalPciResetScope with firmware

The names of PCI reset scopes aren't sychronized with firmware.
The patch fixes it.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/pseries: Decrease message level on EEH initialization
Gavin Shan [Tue, 30 Sep 2014 02:39:04 +0000 (12:39 +1000)]
powerpc/pseries: Decrease message level on EEH initialization

As Anton suggested, the patch decreases the message level on EEH
initialization to avoid unnecessary messages if required. Also,
we have unified hint if any of needful RTAS calls is missed, and
then we can check /proc/device-tree to figure out the missed RTAS
calls.

Suggested-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Block PCI config access during reset
Gavin Shan [Tue, 30 Sep 2014 02:39:03 +0000 (12:39 +1000)]
powerpc/eeh: Block PCI config access during reset

Function pcibios_set_pcie_reset_state() can be used to do PCI
reset. PCI config access during the reset usually causes EEH
errors unexpectedly. In order to avoid the EEH error, the patch
blocks PCI config access during reset with the help of flag
EEH_PE_RESET, which is similar to what we did in EEH PE reset
path.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Use eeh_unfreeze_pe()
Gavin Shan [Tue, 30 Sep 2014 02:39:02 +0000 (12:39 +1000)]
powerpc/eeh: Use eeh_unfreeze_pe()

The patch uses eeh_unfreeze_pe() to replace the logic clearing
frozen IO and DMA, in order to simplify the code.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Unfreeze PE on enabling EEH functionality
Gavin Shan [Tue, 30 Sep 2014 02:39:01 +0000 (12:39 +1000)]
powerpc/eeh: Unfreeze PE on enabling EEH functionality

When passing through PE to guest, that's possibly in frozen
state. The driver for the pass-through devices on guest side
can't be loaded successfully as reported. We already had one
gate in eeh_dev_open() to clear PE frozen state accordingly,
but that's not enough because the function is only called at
QEMU startup for once.

The patch adds another gate in eeh_pe_set_option() so that the
PE frozen state can be cleared at QEMU restart time.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Fix improper condition in eeh_pci_enable()
Gavin Shan [Tue, 30 Sep 2014 02:39:00 +0000 (12:39 +1000)]
powerpc/eeh: Fix improper condition in eeh_pci_enable()

The function eeh_pci_enable() is called to apply various requests
to one particular PE: Enabling EEH, Disabling EEH, Enabling IO,
Enabling DMA, Freezing PE. When enabling IO or DMA on one specific
PE, we need check that IO or DMA isn't enabled previously. But
the condition used to do the check isn't completely correct because
one PE would be in DMA frozen state with workable IO path, or vice
versa.

The patch fixes the improper condition.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Clear frozen device state in time
Gavin Shan [Tue, 30 Sep 2014 02:38:59 +0000 (12:38 +1000)]
powerpc/eeh: Clear frozen device state in time

The problem was reported by Carol: In the scenario of passing mlx4
adapter to guest, EEH error could be recovered successfully. When
returning the device back to host, the driver (mlx4_core.ko)
couldn't be loaded successfully because of error number -5 (-EIO)
returned from mlx4_get_ownership(), which hits offlined PCI device.
The root cause is that we missed to put the affected devices into
normal state on clearing PE isolated state right after PE reset.

The patch fixes above issue by putting the affected devices to
normal state when clearing PE isolated state in eeh_pe_state_clear().

Cc: stable@vger.kernel.org
Reported-by: Carol L. Soto <clsoto@us.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Clear PAPR error injection registers
Gavin Shan [Tue, 30 Sep 2014 02:38:58 +0000 (12:38 +1000)]
powerpc/powernv: Clear PAPR error injection registers

The frozen state on one specific PE is probably caused by error
injection, which is done with help of PAPR error injection registers.
According to the hardware spec, those registers should be cleared
automatically after one-shot frozen PE. However, that's not always
true, at least on P7IOC of Firebird-L. So we have to clear them
before doing PE reset to avoid recursive EEH errors at recovery
stage.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Add PCI error injection debugfs entry
Mike Qiu [Tue, 30 Sep 2014 02:38:57 +0000 (12:38 +1000)]
powerpc/powernv: Add PCI error injection debugfs entry

The patch adds debugfs file (/sys/kernel/debug/powerpc/PCIxxxx/
err_injct), which accepts following formated string, to support
error injection. It will be used to support userland utility
"errinjct" in future.

  "pe_no:0:function:address:mask" - 32-bits PCI errors
  "pe_no:1:function:address:mask" - 64-bits PCI errors

Signed-off-by: Mike Qiu <qiudayu@linux.vnet.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Introduce eeh_ops::err_inject
Gavin Shan [Tue, 30 Sep 2014 02:38:56 +0000 (12:38 +1000)]
powerpc/eeh: Introduce eeh_ops::err_inject

The patch introduces eeh_ops::err_inject(), which allows to inject
specified errors to indicated PE for testing purpose. The functionality
isn't support on pSeries platform. On PowerNV, the functionality
relies on OPAL API opal_pci_err_inject().

Signed-off-by: Mike Qiu <qiudayu@linux.vnet.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Sync header with firmware
Gavin Shan [Tue, 30 Sep 2014 02:38:55 +0000 (12:38 +1000)]
powerpc/powernv: Sync header with firmware

The patch synchronizes firmware header file (opal.h) for PCI error
injection.

Signed-off-by: Mike Qiu <qiudayu@linux.vnet.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Clear frozen state on passing device
Gavin Shan [Tue, 30 Sep 2014 02:38:54 +0000 (12:38 +1000)]
powerpc/eeh: Clear frozen state on passing device

When passing through device, its PE might have been put into frozen
state. One obvious example would be: the passed PE is forced to be
offline because of hitting maximal allowed EEH errors in userland.
In that case, the frozen state won't be cleared and then the PE is
returned back to host, which might not have chance detecting and
recovering from it.

The patch adds more check when passing through device and clear the
PE frozen state if necessary.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Reenable PCI devices after reset
Gavin Shan [Tue, 30 Sep 2014 02:38:53 +0000 (12:38 +1000)]
powerpc/eeh: Reenable PCI devices after reset

The PCI devices that have been passed through are enabled before
reset, we need restore to the enabled state after reset. Otherwise,
MMIO access might be issued to disabled devices after reset and
causes exceptional recursive EEH error.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Freeze PE before PE reset
Gavin Shan [Tue, 30 Sep 2014 02:38:52 +0000 (12:38 +1000)]
powerpc/eeh: Freeze PE before PE reset

The patch adds one more option (EEH_OPT_FREEZE_PE) to set_option()
method to proactively freeze PE, which will be issued before resetting
pass-throughed PE to drop MMIO access during reset because it's
always contributing to recursive EEH error.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Add eeh_pe_state sysfs entry
Gavin Shan [Tue, 30 Sep 2014 02:38:51 +0000 (12:38 +1000)]
powerpc/eeh: Add eeh_pe_state sysfs entry

The patch adds sysfs entry "eeh_pe_state". Reading on it returns
the PE's state while writing to it clears the frozen state. It's
used to check or clear the PE frozen state from userland for
debugging purpose.

The patch also replaces printk(KERN_WARNING ...) with pr_warn() in
eeh_sysfs.c

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Drop unused argument in eeh_check_failure()
Gavin Shan [Tue, 30 Sep 2014 02:38:50 +0000 (12:38 +1000)]
powerpc/eeh: Drop unused argument in eeh_check_failure()

eeh_check_failure() is used to check frozen state of the PE which
owns the indicated I/O address. The argument "val" of the function
isn't used. The patch drops it and return the frozen state of the
PE as expected.

Cc: Vishal Mansur <vmansur@linux.vnet.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Enable DCACHE_WORD_ACCESS on ppc64le
Anton Blanchard [Thu, 18 Sep 2014 23:40:21 +0000 (09:40 +1000)]
powerpc: Enable DCACHE_WORD_ACCESS on ppc64le

Enable on DCACHE_WORD_ACCESS on ppc64le. It should work on
ppc64 and ppc32 but we need to do some testing first.

A somewhat reasonable testcase used to show the performance
improvement - a repeated stat of a 33 byte filename that
doesn't exist:

 #include <sys/types.h>
 #include <sys/stat.h>
 #include <unistd.h>

 #define ITERATIONS 10000000

 #define PATH "123456781234567812345678123456781"

 int main(void)
 {
  unsigned long i;
  struct stat buf;

  for (i = 0; i < ITERATIONS; i++)
  stat(PATH, &buf);

  return 0;
 }

runs 27% faster on POWER8.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: ppc64le optimised word at a time
Anton Blanchard [Thu, 18 Sep 2014 23:40:20 +0000 (09:40 +1000)]
powerpc: ppc64le optimised word at a time

Use cmpb which compares each byte in two 64 bit values and
for each matching byte places 0xff in the target and 0x00
otherwise.

A simple hash_name microbenchmark:

http://ozlabs.org/~anton/junkcode/hash_name_bench.c

shows this version to be 10-20% faster than running the x86
version on POWER8, depending on the length.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agoselftests/powerpc: Add test of load_unaligned_zero_pad()
Michael Ellerman [Thu, 25 Sep 2014 06:45:11 +0000 (16:45 +1000)]
selftests/powerpc: Add test of load_unaligned_zero_pad()

It is a rarely exercised case, so we want to have a test to ensure it
works as required.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Implement load_unaligned_zeropad
Anton Blanchard [Thu, 18 Sep 2014 23:40:19 +0000 (09:40 +1000)]
powerpc: Implement load_unaligned_zeropad

Implement a bi-arch and bi-endian version of load_unaligned_zeropad.

Since the fallback case is so rare, a userspace test harness was used
to test this on ppc64le, ppc64 and ppc32:

http://ozlabs.org/~anton/junkcode/test_load_unaligned_zeropad.c

It uses mprotect to force a SEGV across a page boundary, and a SEGV
handler to lookup the exception tables and run the fixup routine.
It also compares the result against a normal load.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: pci-ioda: Use a single function to emit logging messages
Joe Perches [Sun, 21 Sep 2014 17:55:06 +0000 (10:55 -0700)]
powerpc: pci-ioda: Use a single function to emit logging messages

No need for 3 functions when a single one will do.

Modify the function declaring macros to call the single function.

Reduces object code size a little:

$ size arch/powerpc/platforms/powernv/pci-ioda.o*
   text    data     bss     dec     hex filename
  22303    1073    6680   30056    7568 arch/powerpc/platforms/powernv/pci-ioda.o.new
  22840    1121    6776   30737    7811 arch/powerpc/platforms/powernv/pci-ioda.o.old

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: pci-ioda: Remove unnecessary return value from printk
Joe Perches [Sun, 21 Sep 2014 17:55:05 +0000 (10:55 -0700)]
powerpc: pci-ioda: Remove unnecessary return value from printk

The return value is unnecessary and unused, so make the functions
void instead of int.

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/eeh: Fix kernel crash when passing through VF
Wei Yang [Wed, 17 Sep 2014 02:48:26 +0000 (10:48 +0800)]
powerpc/eeh: Fix kernel crash when passing through VF

When doing vfio passthrough a VF, the kernel will crash with following
message:

[  442.656459] Unable to handle kernel paging request for data at address 0x00000060
[  442.656593] Faulting instruction address: 0xc000000000038b88
[  442.656706] Oops: Kernel access of bad area, sig: 11 [#1]
[  442.656798] SMP NR_CPUS=1024 NUMA PowerNV
[  442.656890] Modules linked in: vfio_pci mlx4_core nf_conntrack_netbios_ns nf_conntrack_broadcast ipt_MASQUERADE ip6t_REJECT xt_conntrack bnep bluetooth rfkill ebtable_nat ebtable_broute bridge stp llc ebtable_filter ebtables ip6table_nat nf_conntrack_ipv6 nf_defrag_ipv6 nf_nat_ipv6 ip6table_mangle ip6table_security ip6table_raw ip6table_filter ip6_tables iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 nf_nat_ipv4 nf_nat nf_conntrack iptable_mangle iptable_security iptable_raw tg3 nfsd be2net nfs_acl ses lockd ptp enclosure pps_core kvm_hv kvm_pr shpchp binfmt_misc kvm sunrpc uinput lpfc scsi_transport_fc ipr scsi_tgt [last unloaded: mlx4_core]
[  442.658152] CPU: 40 PID: 14948 Comm: qemu-system-ppc Not tainted 3.10.42yw-pkvm+ #37
[  442.658219] task: c000000f7e2a9a00 ti: c000000f6dc3c000 task.ti: c000000f6dc3c000
[  442.658287] NIP: c000000000038b88 LR: c0000000004435a8 CTR: c000000000455bc0
[  442.658352] REGS: c000000f6dc3f580 TRAP: 0300   Not tainted  (3.10.42yw-pkvm+)
[  442.658419] MSR: 9000000000009032 <SF,HV,EE,ME,IR,DR,RI>  CR: 28004882  XER: 20000000
[  442.658577] CFAR: c00000000000908c DAR: 0000000000000060 DSISR: 40000000 SOFTE: 1
GPR00: c0000000004435a8 c000000f6dc3f800 c0000000012b1c10 c00000000da24000
GPR04: 0000000000000003 0000000000001004 00000000000015b3 000000000000ffff
GPR08: c00000000127f5d8 0000000000000000 000000000000ffff 0000000000000000
GPR12: c000000000068078 c00000000fdd6800 000001003c320c80 000001003c3607f0
GPR16: 0000000000000001 00000000105480c8 000000001055aaa8 000001003c31ab18
GPR20: 000001003c10fb40 000001003c360ae8 000000001063bcf0 000000001063bdb0
GPR24: 000001003c15ed70 0000000010548f40 c000001fe5514c88 c000001fe5514cb0
GPR28: c00000000da24000 0000000000000000 c00000000da24000 0000000000000003
[  442.659471] NIP [c000000000038b88] .pcibios_set_pcie_reset_state+0x28/0x130
[  442.659530] LR [c0000000004435a8] .pci_set_pcie_reset_state+0x28/0x40
[  442.659585] Call Trace:
[  442.659610] [c000000f6dc3f800] [00000000000719e0] 0x719e0 (unreliable)
[  442.659677] [c000000f6dc3f880] [c0000000004435a8] .pci_set_pcie_reset_state+0x28/0x40
[  442.659757] [c000000f6dc3f900] [c000000000455bf8] .reset_fundamental+0x38/0x80
[  442.659835] [c000000f6dc3f980] [c0000000004562a8] .pci_dev_specific_reset+0xa8/0xf0
[  442.659913] [c000000f6dc3fa00] [c0000000004448c4] .__pci_dev_reset+0x44/0x430
[  442.659980] [c000000f6dc3fab0] [c000000000444d5c] .pci_reset_function+0x7c/0xc0
[  442.660059] [c000000f6dc3fb30] [d00000001c141ab8] .vfio_pci_open+0xe8/0x2b0 [vfio_pci]
[  442.660139] [c000000f6dc3fbd0] [c000000000586c30] .vfio_group_fops_unl_ioctl+0x3a0/0x630
[  442.660219] [c000000f6dc3fc90] [c000000000255fbc] .do_vfs_ioctl+0x4ec/0x7c0
[  442.660286] [c000000f6dc3fd80] [c000000000256364] .SyS_ioctl+0xd4/0xf0
[  442.660354] [c000000f6dc3fe30] [c000000000009e54] syscall_exit+0x0/0x98
[  442.660420] Instruction dump:
[  442.660454] 4bfffce9 4bfffee4 7c0802a6 fbc1fff0 fbe1fff8 f8010010 f821ff81 7c7e1b78
[  442.660566] 7c9f2378 60000000 60000000 e93e02c8 <e86900602fa30000 41de00c4 2b9f0002
[  442.660679] ---[ end trace a64ac9546bcf0328 ]---
[  442.660724]

The reason is current VF is not EEH enabled.

This patch introduces a macro to convert eeh_dev to eeh_pe. By doing so, it
will prevent converting with NULL pointer.

Signed-off-by: Wei Yang <weiyang@linux.vnet.ibm.com>
Acked-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
CC: Michael Ellerman <mpe@ellerman.id.au>
V3 -> V4:
   1. move the macro definition from include/linux/pci.h to
      arch/powerpc/include/asm/eeh.h

V2 -> V3:
   1. rebased on 3.17-rc4
   2. introduce a macro
   3. use this macro in several other places

V1 -> V2:
   1. code style and patch subject adjustment

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/mm: Unindent htab_dt_scan_page_sizes()
Michael Ellerman [Thu, 7 Aug 2014 07:26:33 +0000 (17:26 +1000)]
powerpc/mm: Unindent htab_dt_scan_page_sizes()

We can unindent the bulk of htab_dt_scan_page_sizes() by returning early
if the property is not found. That is nice in and of itself, but also
has the advantage of making it clear that we always return success once
we have found the ibm,segment-page-sizes property.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/ppc64: Print CPU/MMU/FW features at boot
Michael Ellerman [Wed, 6 Aug 2014 09:08:22 +0000 (19:08 +1000)]
powerpc/ppc64: Print CPU/MMU/FW features at boot

"Helps debug funky firmware issues".

After:
  Starting Linux PPC64 #108 SMP Wed Aug 6 19:04:51 EST 2014
  -----------------------------------------------------
  ppc64_pft_size    = 0x1a
  phys_mem_size     = 0x200000000
  cpu_features      = 0x17fc7a6c18500249
    possible        = 0x1fffffff18700649
    always          = 0x0000000000000040
  cpu_user_features = 0xdc0065c2 0xee000000
  mmu_features      = 0x5a000001
  firmware_features = 0x00000001405a440b
  htab_hash_mask    = 0x7ffff
  -----------------------------------------------------

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/ppc64: Clean up the boot-time settings display
Michael Ellerman [Wed, 6 Aug 2014 09:08:21 +0000 (19:08 +1000)]
powerpc/ppc64: Clean up the boot-time settings display

At boot we display a bunch of low level settings which can be useful to
know, and can help to spot bugs when things are fundamentally
misconfigured.

At the moment they are very widely spaced, so that we can accommodate
the line:

  ppc64_caches.dcache_line_size = 0xYY

But we only print that line when the cache line size is not 128, ie.
almost never, so it just makes the display look odd usually.

The ppc64_caches prefix is redundant so remove it, which means we can
align things a bit closer for the common case. While we're there
replace the last use of camelCase (physicalMemorySize), and use
phys_mem_size.

Before:
  Starting Linux PPC64 #104 SMP Wed Aug 6 18:41:34 EST 2014
  -----------------------------------------------------
  ppc64_pft_size                = 0x1a
  physicalMemorySize            = 0x200000000
  ppc64_caches.dcache_line_size = 0xf0
  ppc64_caches.icache_line_size = 0xf0
  htab_address                  = 0xdeadbeef
  htab_hash_mask                = 0x7ffff
  physical_start                = 0xf000bar
  -----------------------------------------------------

After:
  Starting Linux PPC64 #103 SMP Wed Aug 6 18:38:04 EST 2014
  -----------------------------------------------------
  ppc64_pft_size    = 0x1a
  phys_mem_size     = 0x200000000
  dcache_line_size  = 0xf0
  icache_line_size  = 0xf0
  htab_address      = 0xdeadbeef
  htab_hash_mask    = 0x7ffff
  physical_start    = 0xf000bar
  -----------------------------------------------------

This patch is final, no bike shedding ;)

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Fix build failure when CONFIG_USB=y
Pranith Kumar [Thu, 21 Aug 2014 13:16:04 +0000 (09:16 -0400)]
powerpc: Fix build failure when CONFIG_USB=y

We are enabling USB unconditionally which results in following build failure

drivers/built-in.o: In function `tb_drom_read':
(.text+0x1b62b70): undefined reference to `usb_speed_string'
make: *** [vmlinux] Error

Enable USB only if USB_SUPPORT is set to avoid such failures

Signed-off-by: Pranith Kumar <bobby.prani@gmail.com>
Acked-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Fix build failure on 44x
Pranith Kumar [Thu, 21 Aug 2014 13:04:31 +0000 (09:04 -0400)]
powerpc: Fix build failure on 44x

Fix the following build failure

drivers/built-in.o: In function `nhi_init':
nhi.c:(.init.text+0x63390): undefined reference to `ehci_init_driver'

by adding a dependency on USB_EHCI_HCD which supplies the ehci_init_driver().

Also we need to depend on USB_OHCI_HCD similarly

Signed-off-by: Pranith Kumar <bobby.prani@gmail.com>
Acked-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: some changes in numa_setup_cpu()
Li Zhong [Wed, 27 Aug 2014 09:34:01 +0000 (17:34 +0800)]
powerpc: some changes in numa_setup_cpu()

this patches changes some error handling logics in numa_setup_cpu(),
when cpu node is not found, so:

if the cpu is possible, but not present, -1 is kept in numa_cpu_lookup_table,
so later, if the cpu is added, we could set correct numa information for it.

if the cpu is present, then we set the first online node to
numa_cpu_lookup_table instead of 0 ( in case 0 might not be an online node? )

Cc: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Cc: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Acked-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Only set numa node information for present cpus at boottime
Li Zhong [Wed, 27 Aug 2014 09:34:00 +0000 (17:34 +0800)]
powerpc: Only set numa node information for present cpus at boottime

As Nish suggested, it makes more sense to init the numa node informatiion
for present cpus at boottime, which could also avoid WARN_ON(1) in
numa_setup_cpu().

With this change, we also need to change the smp_prepare_cpus() to set up
numa information only on present cpus.

For those possible, but not present cpus, their numa information
will be set up after they are started, as the original code did before commit
2fabf084b6ad.

Cc: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Cc: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Acked-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Tested-by: Cyril Bur <cyril.bur@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Fix warning reported by verify_cpu_node_mapping()
Li Zhong [Wed, 27 Aug 2014 09:33:59 +0000 (17:33 +0800)]
powerpc: Fix warning reported by verify_cpu_node_mapping()

With commit 2fabf084b6ad ("powerpc: reorder per-cpu NUMA information's
initialization"), during boottime, cpu_numa_callback() is called
earlier(before their online) for each cpu, and verify_cpu_node_mapping()
uses cpu_to_node() to check whether siblings are in the same node.

It skips the checking for siblings that are not online yet. So the only
check done here is for the bootcpu, which is online at that time. But
the per-cpu numa_node cpu_to_node() uses hasn't been set up yet (which
will be set up in smp_prepare_cpus()).

So I saw something like following reported:
[    0.000000] CPU thread siblings 1/2/3 and 0 don't belong to the same
node!

As we don't actually do the checking during this early stage, so maybe
we could directly call numa_setup_cpu() in do_init_bootmem().

Cc: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Cc: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Li Zhong <zhong@linux.vnet.ibm.com>
Acked-by: Nishanth Aravamudan <nacc@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Implement emulation of string loads and stores
Paul Mackerras [Tue, 2 Sep 2014 04:35:09 +0000 (14:35 +1000)]
powerpc: Implement emulation of string loads and stores

The size field of the op.type word is now the total number of bytes
to be loaded or stored.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Emulate icbi, mcrf and conditional-trap instructions
Paul Mackerras [Tue, 2 Sep 2014 04:35:08 +0000 (14:35 +1000)]
powerpc: Emulate icbi, mcrf and conditional-trap instructions

This extends the instruction emulation done by analyse_instr() and
emulate_step() to handle a few more instructions that are found in
the kernel.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Split out instruction analysis part of emulate_step()
Paul Mackerras [Tue, 2 Sep 2014 04:35:07 +0000 (14:35 +1000)]
powerpc: Split out instruction analysis part of emulate_step()

This splits out the instruction analysis part of emulate_step() into
a separate analyse_instr() function, which decodes the instruction,
but doesn't execute any load or store instructions.  It does execute
integer instructions and branches which can be executed purely by
updating register values in the pt_regs struct.  For other instructions,
it returns the instruction type and other details in a new
instruction_op struct.  emulate_step() then uses that information
to execute loads, stores, cache operations, mfmsr, mtmsr[d], and
(on 64-bit) sc instructions.

The reason for doing this is so that the KVM code can use it instead
of having its own separate instruction emulation code.  Possibly the
alignment interrupt handler could also use this.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Check flat device tree version at boot
Michael Ellerman [Thu, 28 Aug 2014 08:40:47 +0000 (18:40 +1000)]
powerpc: Check flat device tree version at boot

In commit e6a6928c3ea1 "of/fdt: Convert FDT functions to use libfdt",
the kernel stopped supporting old flat device tree formats. The minimum
supported version is now 0x10.

There was a checking function added, early_init_dt_verify(), but it's
not called on powerpc.

The result is, if you boot with an old flat device tree, the kernel will
fail to parse it correctly, think you have no memory etc. and hilarity
ensues.

We can't really fix it, but we can at least catch the fact that the
device tree is in an unsupported format and panic(). We can't call
BUG(), it's too early.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc/powernv: Don't call generic code on offline cpus
Paul Mackerras [Tue, 2 Sep 2014 04:23:16 +0000 (14:23 +1000)]
powerpc/powernv: Don't call generic code on offline cpus

On PowerNV platforms, when a CPU is offline, we put it into nap mode.
It's possible that the CPU wakes up from nap mode while it is still
offline due to a stray IPI.  A misdirected device interrupt could also
potentially cause it to wake up.  In that circumstance, we need to clear
the interrupt so that the CPU can go back to nap mode.

In the past the clearing of the interrupt was accomplished by briefly
enabling interrupts and allowing the normal interrupt handling code
(do_IRQ() etc.) to handle the interrupt.  This has the problem that
this code calls irq_enter() and irq_exit(), which call functions such
as account_system_vtime() which use RCU internally.  Use of RCU is not
permitted on offline CPUs and will trigger errors if RCU checking is
enabled.

To avoid calling into any generic code which might use RCU, we adopt
a different method of clearing interrupts on offline CPUs.  Since we
are on the PowerNV platform, we know that the system interrupt
controller is a XICS being driven directly (i.e. not via hcalls) by
the kernel.  Hence this adds a new icp_native_flush_interrupt()
function to the native-mode XICS driver and arranges to call that
when an offline CPU is woken from nap.  This new function reads the
interrupt from the XICS.  If it is an IPI, it clears the IPI; if it
is a device interrupt, it prints a warning and disables the source.
Then it does the end-of-interrupt processing for the interrupt.

The other thing that briefly enabling interrupts did was to check and
clear the irq_happened flag in this CPU's PACA.  Therefore, after
flushing the interrupt from the XICS, we also clear all bits except
the PACA_IRQ_HARD_DIS (interrupts are hard disabled) bit from the
irq_happened flag.  The PACA_IRQ_HARD_DIS flag is set by power7_nap()
and is left set to indicate that interrupts are hard disabled.  This
means we then have to ignore that flag in power7_nap(), which is
reasonable since it doesn't indicate that any interrupt event needs
servicing.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Use CONFIG_ARCH_HAS_FAST_MULTIPLIER
Anton Blanchard [Tue, 16 Sep 2014 00:49:14 +0000 (10:49 +1000)]
powerpc: Use CONFIG_ARCH_HAS_FAST_MULTIPLIER

I ran some tests to compare hash_64 using shifts and multiplies.
The results:

POWER6: ~2x slower
POWER7: ~2x faster
POWER8: ~2x faster

Now we have a proper config option, select
CONFIG_ARCH_HAS_FAST_MULTIPLIER on POWER7 and POWER8.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopowerpc: Add POWER8 CPU selection
Anton Blanchard [Tue, 16 Sep 2014 00:47:49 +0000 (10:47 +1000)]
powerpc: Add POWER8 CPU selection

This allows the user to build a kernel targeted at POWER8
(ie gcc -mcpu=power8).

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
10 years agopseries: Fix endian issues in cpu hot-removal
Thomas Falcon [Fri, 12 Sep 2014 19:11:42 +0000 (14:11 -0500)]
pseries: Fix endian issues in cpu hot-removal

When removing a cpu, this patch makes sure that values
gotten from or passed to firmware are in the correct
endian format.

Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>