Commit 0195659 introduced a NEON accelerated version of the xor_blocks()
function, but it needs the changes in this patch to allow it to be built
as a module rather than statically into the kernel.
This patch creates a separate module xor-neon.ko which exports the NEON
inner xor_blocks() functions depended upon by the regular xor.ko if it
is built with CONFIG_KERNEL_MODE_NEON=y
Reported-by: Josh Boyer <jwboyer@fedoraproject.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Comments from Ard Biesheuvel:
I have included two use cases that I have been using, XOR and RAID-6
checksumming. The former gets a 60% performance boost on the NEON, the
latter over 400%.
ARM: add support for kernel mode NEON
Adds kernel_neon_begin/end (renamed from kernel_vfp_begin/end in the
previous version to de-emphasize the VFP part as VFP code that needs
software assistance is not supported currently.)
Introduces <asm/neon.h> and the Kconfig symbol KERNEL_MODE_NEON. This
has been aligned with Catalin for arm64, so any NEON code that does
not use assembly but intrinsics or the GCC vectorizer (such as my
examples) can potentially be shared between arm and arm64 archs.
ARM: move VFP init to an earlier boot stage
This is needed so the NEON is enabled when the XOR and RAID-6 algo
boot time benchmarks are run.
ARM: be strict about FP exceptions in kernel mode
This adds a check to vfp_support_entry() to flag unsupported uses of
the NEON/VFP in kernel mode. FP exceptions (bounces) are flagged as
a bug, this is because of their potentially intermittent nature.
Exceptions caused by the fact that kernel_neon_begin has not been
called are just routed through the undef handler.
ARM: crypto: add NEON accelerated XOR implementation
This is the xor_blocks() implementation built with -ftree-vectorize,
60% faster than optimized ARM code. It calls in_interrupt() to check
whether the NEON flavor can be used: this should really not be
necessary, but due to xor_blocks'squite generic nature, there is no
telling how exactly people may be using it in the real world.
lib/raid6: add ARM-NEON accelerated syndrome calculation
This is a port of the RAID-6 checksumming code in altivec.uc ported
to use NEON intrinsics. It is about 4x faster than the sequential
code.
The __cpuinit type of throwaway sections might have made sense
some time ago when RAM was more constrained, but now the savings
do not offset the cost and complications. For example, the fix in
commit 5e427ec2d0 ("x86: Fix bit corruption at CPU resume time")
is a good example of the nasty type of bugs that can be created
with improper use of the various __init prefixes.
After a discussion on LKML[1] it was decided that cpuinit should go
the way of devinit and be phased out. Once all the users are gone,
we can then finally remove the macros themselves from linux/init.h.
Note that some harmless section mismatch warnings may result, since
notify_cpu_starting() and cpu_up() are arch independent (kernel/cpu.c)
and are flagged as __cpuinit -- so if we remove the __cpuinit from
the arch specific callers, we will also get section mismatch warnings.
As an intermediate step, we intend to turn the linux/init.h cpuinit
related content into no-ops as early as possible, since that will get
rid of these warnings. In any case, they are temporary and harmless.
This removes all the ARM uses of the __cpuinit macros from C code,
and all __CPUINIT from assembly code. It also had two ".previous"
section statements that were paired off against __CPUINIT
(aka .section ".cpuinit.text") that also get removed here.
[1] https://lkml.org/lkml/2013/5/20/589
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Will Deacon <will.deacon@arm.com>
Cc: linux-arm-kernel@lists.infradead.org
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Add a source file xor-neon.c (which is really just the reference
C implementation passed through the GCC vectorizer) and hook it
up to the XOR framework.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Nicolas Pitre <nico@linaro.org>
Commit 70264367a2 ("ARM: 7653/2: do not scale loops_per_jiffy when
using a constant delay clock") fixed a problem with our timer-based
delay loop, where loops_per_jiffy is scaled by cpufreq yet used directly
by the timer delay ops.
This patch fixes the problem in a more elegant way by keeping a private
ticks_per_jiffy field in the delay ops, independent of loops_per_jiffy
and therefore not subject to scaling. The loop-based delay continues to
use loops_per_jiffy directly, as it should.
Acked-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Commit 455bd4c430 ("ARM: 7668/1: fix memset-related crashes caused by
recent GCC (4.7.2) optimizations") attempted to fix a compliance issue
with the memset return value. However the memset itself became broken
by that patch for misaligned pointers.
This fixes the above by branching over the entry code from the
misaligned fixup code to avoid reloading the original pointer.
Also, because the function entry alignment is wrong in the Thumb mode
compilation, that fixup code is moved to the end.
While at it, the entry instructions are slightly reworked to help dual
issue pipelines.
Signed-off-by: Nicolas Pitre <nico@linaro.org>
Tested-by: Alexander Holler <holler@ahsoftware.de>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Recent GCC versions (e.g. GCC-4.7.2) perform optimizations based on
assumptions about the implementation of memset and similar functions.
The current ARM optimized memset code does not return the value of
its first argument, as is usually expected from standard implementations.
For instance in the following function:
void debug_mutex_lock_common(struct mutex *lock, struct mutex_waiter *waiter)
{
memset(waiter, MUTEX_DEBUG_INIT, sizeof(*waiter));
waiter->magic = waiter;
INIT_LIST_HEAD(&waiter->list);
}
compiled as:
800554d0 <debug_mutex_lock_common>:
800554d0: e92d4008 push {r3, lr}
800554d4: e1a00001 mov r0, r1
800554d8: e3a02010 mov r2, #16 ; 0x10
800554dc: e3a01011 mov r1, #17 ; 0x11
800554e0: eb04426e bl 80165ea0 <memset>
800554e4: e1a03000 mov r3, r0
800554e8: e583000c str r0, [r3, #12]
800554ec: e5830000 str r0, [r3]
800554f0: e5830004 str r0, [r3, #4]
800554f4: e8bd8008 pop {r3, pc}
GCC assumes memset returns the value of pointer 'waiter' in register r0; causing
register/memory corruptions.
This patch fixes the return value of the assembly version of memset.
It adds a 'mov' instruction and merges an additional load+store into
existing load/store instructions.
For ease of review, here is a breakdown of the patch into 4 simple steps:
Step 1
======
Perform the following substitutions:
ip -> r8, then
r0 -> ip,
and insert 'mov ip, r0' as the first statement of the function.
At this point, we have a memset() implementation returning the proper result,
but corrupting r8 on some paths (the ones that were using ip).
Step 2
======
Make sure r8 is saved and restored when (! CALGN(1)+0) == 1:
save r8:
- str lr, [sp, #-4]!
+ stmfd sp!, {r8, lr}
and restore r8 on both exit paths:
- ldmeqfd sp!, {pc} @ Now <64 bytes to go.
+ ldmeqfd sp!, {r8, pc} @ Now <64 bytes to go.
(...)
tst r2, #16
stmneia ip!, {r1, r3, r8, lr}
- ldr lr, [sp], #4
+ ldmfd sp!, {r8, lr}
Step 3
======
Make sure r8 is saved and restored when (! CALGN(1)+0) == 0:
save r8:
- stmfd sp!, {r4-r7, lr}
+ stmfd sp!, {r4-r8, lr}
and restore r8 on both exit paths:
bgt 3b
- ldmeqfd sp!, {r4-r7, pc}
+ ldmeqfd sp!, {r4-r8, pc}
(...)
tst r2, #16
stmneia ip!, {r4-r7}
- ldmfd sp!, {r4-r7, lr}
+ ldmfd sp!, {r4-r8, lr}
Step 4
======
Rewrite register list "r4-r7, r8" as "r4-r8".
Signed-off-by: Ivan Djelic <ivan.djelic@parrot.com>
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Dirk Behme <dirk.behme@gmail.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
When udelay() is implemented using an architected timer, it is wrong
to scale loops_per_jiffy when changing the CPU clock frequency since
the timer clock remains constant.
The lpj should probably become an implementation detail relevant to
the CPU loop based delay routine only and more confined to it. In the
mean time this is the minimal fix needed to have expected delays with
the timer based implementation when cpufreq is also in use.
Reported-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Nicolas Pitre <nico@linaro.org>
Tested-by: Viresh Kumar <viresh.kumar@linaro.org>
Acked-by: Liviu Dudau <Liviu.Dudau@arm.com>
Cc: stable@vger.kernel.org
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
read_current_timer is used by get_cycles since "ARM: 7538/1: delay:
add registration mechanism for delay timer sources", and get_cycles
can be used by device drivers in loadable modules, so it has to
be exported.
Without this patch, building imote2_defconfig fails with
ERROR: "read_current_timer" [crypto/tcrypt.ko] undefined!
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: Jonathan Austin <jonathan.austin@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Russell King <rmk+kernel@arm.linux.org.uk>
The current timer-based delay loop relies on the architected timer to
initiate the switch away from the polling-based implementation. This is
unfortunate for platforms without the architected timers but with a
suitable delay source (that is, constant frequency, always powered-up
and ticking as long as the CPUs are online).
This patch introduces a registration mechanism for the delay timer
(which provides an unconditional read_current_timer implementation) and
updates the architected timer code to use the new interface.
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Jonathan Austin <jonathan.austin@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The delay functions may be called by some platforms between switching to
the timer-based delay loop but before calibration. In this case, the
initial loops_per_jiffy may not be suitable for the timer (although a
compromise may be achievable) and delay times may be considered too
inaccurate.
This patch updates loops_per_jiffy when switching to the timer-based
delay loop so that delays are consistent prior to calibration.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The {get,put}_user macros don't perform range checking on the provided
__user address when !CPU_HAS_DOMAINS.
This patch reworks the out-of-line assembly accessors to check the user
address against a specified limit, returning -EFAULT if is is out of
range.
[will: changed get_user register allocation to match put_user]
[rmk: fixed building on older ARM architectures]
Reported-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Cc: stable@vger.kernel.org
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This partially reverts 357c9c1f07
(ARM: Remove support for ARMv3 ARM610 and ARM710 CPUs).
Although we only support StrongARM on the RiscPC, we need to keep the
ARMv3 user access code for this platform because the bus does not
understand half-word load/stores.
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch allows a timer-based delay implementation to be selected by
switching the delay routines over to use get_cycles, which is
implemented in terms of read_current_timer. This further allows us to
skip the loop calibration and have a consistent delay function in the
face of core frequency scaling.
To avoid the pain of dealing with memory-mapped counters, this
implementation uses the co-processor interface to the architected timers
when they are available. The previous loop-based implementation is
kept around for CPUs without the architected timers and we retain both
the maximum delay (2ms) and the corresponding conversion factors for
determining the number of loops required for a given interval. Since the
indirection of the timer routines will only work when called from C,
the sa1100 sleep routines are modified to branch to the loop-based delay
functions directly.
Tested-by: Shinya Kuribayashi <shinya.kuribayashi.px@renesas.com>
Reviewed-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch implements the word-at-a-time interface for ARM using the
same algorithm as x86. We use the fls macro from ARMv5 onwards, where
we have a clz instruction available which saves us a mov instruction
when targetting Thumb-2. For older CPUs, we use the magic 0x0ff0001
constant. Big-endian configurations make use of the implementation from
asm-generic.
With this implemented, we can replace our byte-at-a-time strnlen_user
and strncpy_from_user functions with the optimised generic versions.
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This patch removes support for ARMv3 CPUs, which haven't worked properly
for quite some time (see the FIXME comment in arch/arm/mm/fault.c). The
only V3 parts left is the cache model for ARMv3, which is needed for some
odd reason by ARM740T CPUs, and being able to build with -march=armv3,
which is required for the RiscPC platform due to its bus structure.
Acked-by: Will Deacon <will.deacon@arm.com>
Acked-by: Jean-Christophe PLAGNIOL-VILLARD <plagnioj@jcrosoft.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
When disabling the MMU, it is necessary to take out a 1:1 identity map
of the reset code so that it can safely be executed with and without
the MMU active. To avoid the situation where the physical address of the
reset code aliases with the virtual address of the active stack (which
cannot be included in the 1:1 mapping), it is desirable to change to a
new stack at a location which is less likely to alias.
This code adds a new lib function, call_with_stack:
void call_with_stack(void (*fn)(void *), void *arg, void *sp);
which changes the stack to point at the sp parameter, before invoking
fn(arg) with the new stack selected.
Reviewed-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Dave Martin <dave.martin@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
The bitops functions (e.g. _test_and_set_bit) on ARM do not have unwind
annotations and therefore the kernel cannot backtrace out of them on a
fatal error (for example, NULL pointer dereference).
This patch annotates the bitops assembly macros with UNWIND annotations
so that we can produce a meaningful backtrace on error. Callers of the
macros are modified to pass their function name as a macro parameter,
enforcing that the macros are used as standalone function implementations.
Acked-by: Dave Martin <dave.martin@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The 64bit division functions never had unwinding annotations
added. This prevents a backtrace from being printed within
the function and if a division by 0 occurs. Add the annotations.
Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
Acked-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Currently, show_regs calls __backtrace which does
nothing if CONFIG_FRAME_POINTER is not set. Switch to
dump_stack which handles both CONFIG_FRAME_POINTER and
CONFIG_ARM_UNWIND correctly.
__backtrace is now superseded by dump_stack in general
and show_regs was the last caller so remove __backtrace
as well.
Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
Acked-by: Nicolas Pitre <nicolas.pitre@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
