This extends the check that the adapter is in a CAPI capable slot so
that it may be called by external users in the kernel API. This will be
used by the upcoming Mellanox CX4 support, which needs to know ahead of
time if the card can be switched to cxl mode so that it can leave it in
PCI mode if it is not.
This API takes a parameter to check if CAPP DMA mode is supported, which
it currently only allows on P8NVL systems, since that mode currently has
issues accessing memory < 4GB on P8, and we cannot realistically avoid
that.
This API does not currently check if a CAPP unit is available (i.e. not
already assigned to another PHB) on P8. Doing so would be racy since it
is assigned on a first come first serve basis, and so long as CAPP DMA
mode is not supported on P8 we don't need this, since the only
anticipated user of this API requires CAPP DMA mode.
Cc: Philippe Bergheaud <felix@linux.vnet.ibm.com>
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The support for using the Mellanox CX4 in cxl mode will require
additions to the PHB code. In preparation for this, move the existing
cxl code out of pci-ioda.c into a separate pci-cxl.c file to keep things
more organised.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Use for_each_compatible_node() macro instead of open coding it.
Generated by Coccinelle.
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
PROT_SAO is a powerpc-specific flag to mmap(), and we rely on arch
specific logic to allow it to be passed to mmap().
Add a small test to ensure mmap() accepts PROT_SAO. We don't have a good
way to test that it actually causes the mapping to be created with the
right flags, so for now we just touch the mapping so it's faulted in. In
future we might be able to do something better.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The array crash_shutdown_handles[] has size CRASH_HANDLER_MAX, thus when
we loop over the elements of the list we check crash_shutdown_handles[i]
&& i < CRASH_HANDLER_MAX. However this means that when we increment i to
CRASH_HANDLER_MAX we will perform an out of bound array access checking
the first condition before exiting on the second condition.
To avoid the out of bounds access, simply reorder the loop conditions.
Fixes: 1d1451655b ("powerpc: Add array bounds checking to crash_shutdown_handlers")
Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@gmail.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The subsequent test for RTAS along with the LPAR test are sufficient
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The test is unnecessary, the FW_FEATURE_LPAR is sufficient as there
exist no other LPAR type that has RTAS.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
ge_imp3a_pic_init() is called way beyond the unflattening of
the tree, it shouldn't be using of_flat_dt_*
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Some bit of SPU code was using the FDT rather than the expanded
device-tree. Fix it.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The function is called by both 32-bit and 64-bit early setup right
after early_init_devtree(). All it does is run yet another early
DT parser which is precisely what early_init_devtree() is about,
so move it in there.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Anything in early_setup() needs to be justified to be there, in
this case, we need the trampolines before we can take exceptions
and thus before we turn on the MMU.
Also remove a pretty meaningless and misplaced debug message
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
[mpe: Fix comment formatting]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
early_init() is called in-place before kernel relocation and using
whatever MMU setup exists at the point the kernel is entered.
machine_init() is called after relocation and after some initial
mapping of PAGE_OFFSET has been established (typically using BATs
on 6xx/7xx/7xxx processors or some form of bolted TLB on others).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
One should not attempt to switch a PHB into CAPI mode if there is
a switch between the PHB and the adapter. This patch modifies the
cxl driver to ignore CAPI adapters misplaced in switched slots.
Signed-off-by: Philippe Bergheaud <felix@linux.vnet.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The Kconfig/Makefile currently controlling compilation of this code is:
drivers/misc/cxl/Kconfig:config CXL_BASE
drivers/misc/cxl/Kconfig: bool
drivers/misc/cxl/Makefile:obj-$(CONFIG_CXL_BASE) += base.o
...meaning that it currently is not being built as a module by anyone.
Lets convert the one module_init into device_initcall so that
when reading the driver it more clear that it is builtin-only.
Since module_init translates to device_initcall in the non-modular
case, the init ordering remains unchanged with this commit.
We don't replace module.h with init.h since the file is doing
other modular stuff (module_get/put) even though it is built-in.
Cc: Ian Munsie <imunsie@au1.ibm.com>
Cc: Michael Neuling <mikey@neuling.org>
Cc: linuxppc-dev@lists.ozlabs.org
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The PSL Slice Error Register (PSL_SERR_An) reports implementation
dependent AFU errors, in the form of a bitmap. The PSL_SERR_An
register content is printed in the form of hex dump debug message.
This patch decodes the PSL_ERR_An register contents, and prints a
specific error message for each possible error bit. It also dumps
the secondary registers AFU_ERR_An and PSL_DSISR_An, that may
contain extra debug information.
This patch also removes the large WARN message that used to report
the cxl slice error interrupt, and replaces it by a short informative
message, that draws attention to AFU implementation errors.
Signed-off-by: Philippe Bergheaud <felix@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If a kernel context is initialised and does not have any AFU interrupts
allocated it will cause a NULL pointer dereference when the context is
detached since the irq_names list will not have been initialised.
Move the initialisation of the irq_names list into the cxl_context_init
routine so that it will be valid for the entire lifetime of the context
and will not cause a NULL pointer dereference.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
An issue was noted in our debug logs where the XSL would leave the RA
bit asserted after an AFU reset operation, which would effectively
prevent further AFU reset operations from working.
Workaround the issue by clearing the RA bit with an MMIO write if it is
still asserted after any AFU control operation.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The AFU disable operation has a bug where it will not clear the enable
bit and therefore will have no effect. To date this has likely been
masked by fact that we perform an AFU reset before the disable, which
also has the effect of clearing the enable bit, making the following
disable operation effectively a noop on most hardware. This patch
modifies the afu_control function to take a parameter to clear from the
AFU control register so that the disable operation can clear the
appropriate bit.
This bug was uncovered on the Mellanox CX4, which uses an XSL rather
than a PSL. On the XSL the reset operation will not complete while the
AFU is enabled, meaning the enable bit was still set at the start of the
disable and as a result this bug was hit and the disable also timed out.
Because of this difference in behaviour between the PSL and XSL, this
patch now makes the reset dependent on the card using a PSL to avoid
waiting for a timeout on the XSL. It is entirely possible that we may be
able to drop the reset altogether if it turns out we only ever needed it
due to this bug - however I am not willing to drop it without further
regression testing and have added comments to the code explaining the
background.
This also fixes a small issue where the AFU_Cntl register was read
outside of the lock that protects it.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The Scheduled Process Area is allocated dynamically with enough pages to
fit at least as many processes as the AFU descriptor indicated. Since
the calculation is non-trivial, it does this by calculating how many
processes could fit in an allocation of a given order, and increasing
that order until it can fit enough processes or hits the maximum
supported size.
Currently, it will start this search using a SPA of 2 pages instead of
1. This can waste a page of memory if the AFU's maximum number of
supported processes was small enough to fit in one page.
Fix the algorithm to start the search at 1 page.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If the AFU descriptor of an AFU directed AFU indicates that it supports
0 maximum processes, we will accept that value and attempt to use it.
The SPA will still be allocated (with 2 pages due to another minor bug
and room for 958 processes), and when a context is allocated we will
pass the value of 0 to idr_alloc as the maximum. However, idr_alloc will
treat that as meaning no maximum and will allocate a context number and
we return a valid context.
Conceivably, this could lead to a buffer overflow of the SPA if more
than 958 contexts were allocated, however this is mitigated by the fact
that there are no known AFUs in the wild with a bogus AFU descriptor
like this, and that only the root user is allowed to flash an AFU image
to a card.
Add a check when validating the AFU descriptor to reject any with 0
maximum processes.
We do still allow a dedicated process only AFU to indicate that it
supports 0 contexts even though that is forbidden in the architecture,
as in that case we ignore the value and use 1 instead. This is just on
the off-chance that such a dedicated process AFU may exist (not that I
am aware of any), since their developers are less likely to have cared
about this value at all.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
Reviewed-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
eeh_cache.c doesn't build cleanly with -DDEBUG when
CONFIG_PHYS_ADDR_T_64BIT is set, as a couple of pr_debug()s use "%lx" for
resource_size_t parameters.
Use "%pap" instead, as it's the correct format specifier for types deriving
from phys_addr_t.
Signed-off-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
On some environments (prototype machines, some simulators, etc...)
there is no functional interrupt source to signal completion, so
we rely on the fairly slow OPAL heartbeat.
In a number of cases, the calls complete very quickly or even
immediately. We've observed that it helps a lot to wakeup the OPAL
heartbeat thread before waiting for event in those cases, it will
call OPAL immediately to collect completions for anything that
finished fast enough.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Acked-By: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This is so we can use the powernv_flash mtd driver as an block
device.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
A strange behaviour is observed when comparing PCI hotplug in QEMU, between
x86 and pseries. If you consider the following steps:
- start a VM
- add a PCI device via the QEMU monitor before the rtasd has started (for
example starting the VM in paused state, or hotplug during FW or boot
loader)
- resume the VM execution
The x86 kernel detects the PCI device, but the pseries one does not.
This happens because the rtasd kernel worker is currently started under
device_initcall, while PCI probing happens earlier under subsys_initcall.
As a consequence, if we have a pending RTAS event at boot time, a message
is printed and the event is dropped.
This patch moves all the initialization of rtasd to arch_initcall, which is
run before subsys_call: this way, logging_enabled is true when the RTAS
event pops up and it is not lost anymore.
The proc fs bits stay at device_initcall because they cannot be run before
fs_initcall.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The domain/PHB field of PCI addresses has its value obtained from a
global variable, incremented each time a new domain (represented by
struct pci_controller) is added on the system. The domain addition
process happens during boot or due to PHB hotplug add.
As recent kernels are using predictable naming for network interfaces,
the network stack is more tied to PCI naming. This can be a problem in
hotplug scenarios, because PCI addresses will change if devices are
removed and then re-added. This situation seems unusual, but it can
happen if a user wants to replace a NIC without rebooting the machine,
for example.
This patch changes the way PCI domain values are generated: now, we use
device-tree properties to assign fixed PHB numbers to PCI addresses
when available (meaning pSeries and PowerNV cases). We also use a bitmap
to allow dynamic PHB numbering when device-tree properties are not
used. This bitmap keeps track of used PHB numbers and if a PHB is
released (by hotplug operations for example), it allows the reuse of
this PHB number, avoiding PCI address to change in case of device remove
and re-add soon after. No functional changes were introduced.
Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com>
Reviewed-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Ian Munsie <imunsie@au1.ibm.com>
Acked-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
[mpe: Drop unnecessary machine_is(pseries) test]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Despite attempting to fix this in commit fb36e90736 ("powerpc/pci: Fix
SRIOV not building without EEH enabled"), the build is still broken when
PCI_IOV=y and EEH=n (eg. g5_defconfig with PCI_IOV=y):
arch/powerpc/kernel/pci_dn.c: In function ‘remove_dev_pci_data’:
arch/powerpc/kernel/pci_dn.c:230:18: error: unused variable ‘edev’
Incorporate Ben's idea of using __maybe_unused to avoid so many #ifdefs.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
For memory hotplug to work, the MMU code needs to provide the functions
create_section_mapping() and remove_section_mapping() to respectively
map and unmap portions of the linear mapping.
At the moment only hash64 provides these, so we provide weak stubs that
just error out. This fixes the build with configurations such as 64-bit
BookE with CONFIG_MEMORY_HOTPLUG enabled.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Use "Delta" to refer to the difference between measurements, rather than
"Error", so scripts that look for "Error" aren't confused into thinking
there was a failure.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch adds an OPAL console backend to the powerpc boot wrapper so
that decompression failures inside the wrapper can be reported to the
user. This is important since it typically indicates data corruption in
the firmware and other nasty things.
Currently this only works when building a little endian kernel. When
compiling a 64 bit BE kernel the wrapper is always build 32 bit to be
compatible with some 32 bit firmwares. BE support will be added at a
later date. Another limitation of this is that only the "raw" type of
OPAL console is supported, however machines that provide a hvsi console
also provide a raw console so this is not an issue in practice.
Actually-written-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
[mpe: Move #ifdef __powerpc64__ to avoid warnings on 32-bit]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch adds the kernel command line parameter "no_tb_segs" which
forces the kernel to use 256MB rather than 1TB segments. Forcing the use
of 256MB segments makes it considerably easier to test code that depends
on an SLB miss occurring.
Suggested-by: Michael Neuling <mikey@neuling.org>
Suggested-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Power ISAv3 adds a large decrementer (LD) mode which increases the size
of the decrementer register. The size of the enlarged decrementer
register is between 32 and 64 bits with the exact size being dependent
on the implementation. When in LD mode, reads are sign extended to 64
bits and a decrementer exception is raised when the high bit is set (i.e
the value goes below zero). Writes however are truncated to the physical
register width so some care needs to be taken to ensure that the high
bit is not set when reloading the decrementer. This patch adds support
for using the LD inside the host kernel on processors that support it.
When LD mode is supported firmware will supply the ibm,dec-bits property
for CPU nodes to allow the kernel to determine the maximum decrementer
value. Enabling LD mode is a hypervisor privileged operation so the kernel
can only enable it manually when running in hypervisor mode. Guests that
support LD mode can request it using the "ibm,client-architecture-support"
firmware call (not implemented in this patch) or some other platform
specific method. If this property is not supplied then the traditional
decrementer width of 32 bit is assumed and LD mode will not be enabled.
This patch was based on initial work by Jack Miller.
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Balbir Singh <bsingharora@gmail.com>
Acked-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Check the assembler supports -maltivec by wrapping it with
call as-option.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
cmm_mem_going_offline() is (only) called from cmm_memory_cb(), which
sends the return value through notifier_from_errno(). The latter
expects 0 or -errno (notifier_to_errno(notifier_from_errno(x)) is 0
for any x >= 0, so passing a positive value cannot make sense). Hence
negate ENOMEM.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If ppc_rtas() is called with args.nargs == 16 and args.nret == 0,
args.rets is set to point to &args.args[16], which is beyond the end of
the args.args array. This results in a minor read overrun of the array
when we check the first return code (which, per PAPR, is a required
output of all RTAS calls) to see if there's been a hardware error.
Change the nargs/nret check to ensure nargs is <= 15, allowing room for
the status code. Users shouldn't be calling with nret == 0, but there's
no real harm if they do, so we don't stop them.
Signed-off-by: Andrew Donnellan <andrew.donnellan@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Test that an ISA 3.0 compliant machine performing an unaligned copy,
copy_first, paste or paste_last is sent a SIGBUS.
Signed-off-by: Chris Smart <chris@distroguy.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Calling ISA 3.0 instructions copy, copy_first, paste and paste_last
generates an alignment fault when copying or pasting unaligned
data (128 byte). We catch this and send SIGBUS to the userspace
process that caused it.
We do not emulate these because paste may contain additional metadata
when pasting to a co-processor and paste_last is the synchronisation
point for preceding copy/paste sequences.
Thanks to Michael Neuling <mikey@neuling.org> for his help.
Signed-off-by: Chris Smart <chris@distroguy.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
These are useful little loops for smoke testing performance.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently it doesn't appear the resulting binary actually uses any
Altivec or VSX instructions the solution is to explicitly tell GCC to
use vector instructions and use vector types in the code.
Part of this this issue can be GCC version specific:
GCC 4.9.x is happy to use Altivec and VSX instructions if altivec.h is
includedi (and possibly if vector types are used), this also means that
4.9.x will use VSX instructions even if only -maltivec is passed. It is
also possible that Altivec instructions will be used even without
-maltivec or -mabi=altivec.
GCC 5.2.x complains about the lack of -maltivec parameter if altivec.h
is included and will not use VSX unless -mvsx is present on commandline.
GCC 5.3.0 has a regression that means __attribute__((__target__("no-vsx"))
fails to build. A fix is targeted for 5.4.
Furthermore LTO (Link Time Optimisation) doesn't play well with
__attribute__((__target__("no-vsx")), LTO can cause GCC to forget about
the attribute and compile with VSX instructions regardless. Be wary when
enabling -flfo for this test.
Signed-off-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
