If the buffer needing cache invalidation for inbound DMA does start or
end on a cache line aligned address, we need to use the non-destructive
clean&invalidate operation. This issue was introduced by commit
7363590d2c (arm64: Implement coherent DMA API based on swiotlb).
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Jon Medhurst (Tixy) <tixy@linaro.org>
The Kconfig for CONFIG_STRICT_DEVMEM is missing despite being
used in mmap.c. Add it.
Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Recent arm64 builds using CONFIG_ARM64_64K_PAGES are failing with:
arch/arm64/kernel/perf_regs.c: In function ‘perf_reg_abi’:
arch/arm64/kernel/perf_regs.c:41:2: error: implicit declaration of function ‘is_compat_thread’
arch/arm64/kernel/perf_event.c:1398:2: error: unknown type name ‘compat_uptr_t’
This is due to some recent arm64 perf commits with compat support:
commit 23c7d70d55:
ARM64: perf: add support for frame pointer unwinding in compat mode
commit 2ee0d7fd36:
ARM64: perf: add support for perf registers API
Those patches make the arm64 kernel unbuildable if CONFIG_COMPAT is not
defined and CONFIG_ARM64_64K_PAGES depends on !CONFIG_COMPAT. This patch
allows the arm64 kernel to build with and without CONFIG_COMPAT.
Signed-off-by: Mark Salter <msalter@redhat.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This reverts commit 82b2f495fb. The
__boot_cpu_mode variable is flushed in head.S after being written,
therefore the additional cache flushing is no longer required.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
With system caches for the host OS or architected caches for guest OS we
cannot easily guarantee that there are no dirty or stale cache lines for
the areas of memory written by the kernel during boot with the MMU off
(therefore non-cacheable accesses).
This patch adds the necessary cache maintenance during boot and relaxes
the booting requirements.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
The current TCR register setting in arch/arm64/mm/proc.S assumes that
TCR_EL1.TG* fields are one bit wide and bit 31 is RES1 (reserved, set to
1). With the addition of 16K pages (currently unsupported in the
kernel), the TCR_EL1.TG* fields have been extended to two bits. This
patch updates the corresponding Linux definitions and drops the bit 31
setting in proc.S in favour of the new macros.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Joe Sylve <joe.sylve@gmail.com>
Since this macro is identical to pgprot_writecombine() and is only used
in a single place, remove it completely to avoid confusion. On ARMv7+
processors, the coherent DMA mapping must be Normal NonCacheable (a.k.a.
writecombine) to avoid mismatched hardware attribute aliases (with the
kernel linear mapping as Normal Cacheable).
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
DMA_ATTR_WRITE_COMBINE is currently ignored. Set the pgprot
appropriately for non coherent opperations.
Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
The current dma_ops do not specify an mmap function so maping
falls back to the default implementation. There are at least
two issues with using the default implementation:
1) The pgprot is always pgprot_noncached (strongly ordered)
memory even with coherent operations
2) dma_common_mmap calls virt_to_page on the remapped non-coherent
address which leads to invalid memory being mapped.
Fix both these issue by implementing a custom mmap function which
correctly accounts for remapped addresses and sets vm_pg_prot
appropriately.
Signed-off-by: Laura Abbott <lauraa@codeaurora.org>
[catalin.marinas@arm.com: replaced "arm64_" with "__" prefix for consistency]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Without this, the following scenario is incorrectly determined
to be invalid.
addr 0x7f_ffffe000 size 8192 addr_limit 0x80_00000000
This behavior was observed while trying to vmsplice the stack
as part of a CRIU dump of a process on a system started with the
norandmaps kernel parameter.
Signed-off-by: Christopher Covington <cov@codeaurora.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Probably due to rebasing over the lengthy time it took to get the patch
merged commit addea9ef05 (cpufreq: enable ARM drivers on arm64) added
a duplicate Power management options section. Add CPUfreq to the CPU
power management section and remove a duplicate include of the main
power section.
Signed-off-by: Mark Brown <broonie@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Rather than have separate hugetlb and transparent huge page pmd
manipulation functions, re-wire our thp functions to simply call the
pte equivalents.
This allows THP to take advantage of the new PTE_WRITE logic introduced
in:
c2c93e5 arm64: mm: Introduce PTE_WRITE
To represent splitting THPs we use the PTE_SPECIAL bit as this is not
used for pmds.
Signed-off-by: Steve Capper <steve.capper@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
asm-generic offers an atomic-add based rwsem implementation, which
can avoid the need for heavier, spinlock-based synchronisation on the
fast path.
This patch makes use of the optimised implementation for arm64 CPUs.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This enables support for the generic CPU feature modalias implementation that
wires up optional CPU features to udev based module autoprobing.
A file <asm/cpufeature.h> is provided that maps CPU feature numbers to
elf_hwcap bits, which is the standard way on arm64 to advertise optional CPU
features both internally and to user space.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
[catalin.marinas@arm.com: removed unnecessary "!!"]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Make smp_spin_table_cpu_postboot() static, because this function
is used only in this file.
Signed-off-by: Jingoo Han <jg1.han@samsung.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Make local symbols static, because these are used only in this
file.
Signed-off-by: Jingoo Han <jg1.han@samsung.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add support for unwinding using the dwarf information in compat
mode. Using the correct user stack pointer allows perf to record
the frames correctly in the native and compat modes.
Note that although the dwarf frame unwinding works ok using
libunwind in native mode (on ARMv7 & ARMv8), some changes are
required to the libunwind code for the compat mode. Those changes
are posted separately on the libunwind mailing list.
Tested on ARMv8 platform with v8 and compat v7 binaries, the latter
are statically built.
Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
When profiling a 32-bit application, user space callchain unwinding
using the frame pointer is performed in compat mode. The code is taken
over from the AARCH32 code and adapted to work on AARCH64.
Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This patch implements the functions required for the perf registers API,
allowing the perf tool to interface kernel register dumps with libunwind
in order to provide userspace backtracing.
Compat mode is also supported.
Only the general purpose user space registers are exported, i.e.:
PERF_REG_ARM_X0,
...
PERF_REG_ARM_X28,
PERF_REG_ARM_FP,
PERF_REG_ARM_LR,
PERF_REG_ARM_SP,
PERF_REG_ARM_PC
and not the PERF_REG_ARM_V* registers.
Signed-off-by: Jean Pihet <jean.pihet@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
ARMv8 supports a range of physical address bit sizes. The PARange bits
from ID_AA64MMFR0_EL1 register are read during boot-time and the
intermediate physical address size bits are written in the translation
control registers (TCR_EL1 and VTCR_EL2).
There is no change in the VA bits and levels of translation.
Signed-off-by: Radha Mohan Chintakuntla <rchintakuntla@cavium.com>
Reviewed-by: Will Deacon <Will.deacon@arm.com>
Acked-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Special pte mappings are not intended to be executable and do not even
have an associated struct page. This patch ensures that we do not call
__sync_icache_dcache() on such ptes.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Steve Capper <Steve.Capper@arm.com>
Tested-by: Laura Abbott <lauraa@codeaurora.org>
Tested-by: Bharat Bhushan <Bharat.Bhushan@freescale.com>
Cc: <stable@vger.kernel.org>
pgprot_{dmacoherent,writecombine,noncached} don't need to generate
executable mappings with side-effects like __sync_icache_dcache() being
called when the mapping is in user space.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Bharat Bhushan <Bharat.Bhushan@freescale.com>
Tested-by: Laura Abbott <lauraa@codeaurora.org>
Tested-by: Bharat Bhushan <Bharat.Bhushan@freescale.com>
Cc: <stable@vger.kernel.org>
Commit 8adbf57fc4 ("irqchip: gic: use dmb ishst instead of dsb when
raising a softirq") added an explicit dmb(...) call to the GIC driver.
This patch adds a simple dmb() macro to arm64, which expands to a DMB SY
instruction.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add basic CPU topology support to arm64, based on the existing pre-v8
code and some work done by Mark Hambleton. This patch does not
implement any topology discovery support since that should be based on
information from firmware, it merely implements the scaffolding for
integration of topology support in the architecture.
No locking of the topology data is done since it is only modified during
CPU bringup with external serialisation from the SMP code.
The goal is to separate the architecture hookup for providing topology
information from the DT parsing in order to ease review and avoid
blocking the architecture code (which will be built on by other work)
with the DT code review by providing something simple and basic.
Following patches will implement support for interpreting topology
information from MPIDR and for parsing the DT topology bindings for ARM,
similar patches will be needed for ACPI.
Signed-off-by: Mark Brown <broonie@linaro.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
[catalin.marinas@arm.com: removed CONFIG_CPU_TOPOLOGY, always on if SMP]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This adds support for advertising the presence of ARMv8 Crypto
Extensions in the Aarch32 execution state to 32-bit ELF binaries
running in 32-bit compat mode under the arm64 kernel.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add support for the ELF auxv entry AT_HWCAP2 when running 32-bit
ELF binaries in compat mode.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Currently we flush the entire dcache at boot within __cpu_setup, but
this is unnecessary as the booting protocol demands that the dcache is
invalid and off upon entering the kernel. The presence of the cache
flush only serves to hide bugs in bootloaders, and is not safe in the
presence of SMP.
In an SMP boot scenario the CPUs enter coherency outside of the kernel,
and the primary CPU enables its caches before bringing up secondary
CPUs. Therefore if any secondary CPU has an entry in its cache (in
violation of the boot protocol), the primary CPU might snoop it even if
the secondary CPU's cache is disabled. The boot-time cache flush only
serves to hide a firmware bug, and slows down a cpu boot unnecessarily.
This patch removes the unnecessary boot-time cache flush.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Will Deacon <will.deacon@arm.com>
[catalin.marinas@arm.com: make __flush_dcache_all local only]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Enable cpufreq and power kconfig menus on arm64 along with arm cpufreq
drivers. The power menu is needed for OPP support. At least on Calxeda
systems, the same cpufreq driver is used for arm and arm64 based
systems.
Signed-off-by: Rob Herring <rob.herring@calxeda.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Mark Brown <broonie@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
psci_init() is written to return err code if something goes wrong. However,
the single user, setup_arch(), doesn't care about it. Moreover, every error
path is supplied with a clear message which is enough for pleasant debugging.
Signed-off-by: Vladimir Murzin <vladimir.murzin@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Since 652af89979 "arm64: factor out spin-table
boot method" psci prefix's been introduced. We have a common pr_fmt, so clean
them up.
Signed-off-by: Vladimir Murzin <vladimir.murzin@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This patch adds support for DMA API cache maintenance on SoCs without
hardware device cache coherency.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Since arm64 does not support ISA, there is no need for early swiotlb
initialisation. This patch switches the DMA mapping code to
swiotlb_tlb_late_init_with_default_size(). A side effect of this is that
GFP_DMA is used for the swiotlb buffer and devices with a 32-bit
coherent mask are correctly supported.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
On arm64 we do not have two DMA zones, so it does not make sense to
implement ZONE_DMA32. This patch changes ZONE_DMA32 with ZONE_DMA, the
latter covering 32-bit dma address space to honour GFP_DMA allocations.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Remove some unnecessary bits that were apparently carried over from
another architecture's implementation:
- No need to get_page() the vdso text/data - these are part of the
kernel image.
- No need for ClearPageReserved on the vdso text.
- No need to vmap the first text page to check the ELF header - this
can be done through &vdso_start.
Also some minor cleanup:
- Use kcalloc for vdso_pagelist array allocation.
- Don't print on allocation failure, slab/slub will do that for us.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
arm64_swiotlb_alloc/free_coherent name can be misleading
somtimes with CMA support being enabled after this
patch (c2104debc235b745265b64d610237a6833fd53)
Change this name to be more generic:
__dma_alloc/free_coherent
Signed-off-by: Ritesh Harjani <ritesh.harjani@gmail.com>
[catalin.marinas@arm.com: renamed arm64_swiotlb_dma_ops to coherent_swiotlb_dma_ops]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Change the soft_restart() routine to call cpu_reset() at its identity mapped
physical address.
The cpu_reset() routine must be called at its identity mapped physical address
so that when the MMU is turned off the instruction pointer will be at the correct
location in physical memory.
Signed-off-by: Geoff Levand <geoff@infradead.org> for Huawei, Linaro
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This patch changes the idmap page table creation during boot to cover
the whole kernel image, allowing functions like cpu_reset() to be safely
called with the physical address.
This patch also simplifies the create_block_map asm macro to no longer
take an idmap argument and always use the phys/virt/end parameters. For
the idmap case, phys == virt.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Over the past couple of years, the generic mmu_gather gained range
tracking - 597e1c3580 (mm/mmu_gather: enable tlb flush range in generic
mmu_gather), 2b047252d0 (Fix TLB gather virtual address range
invalidation corner cases) - and tlb_fast_mode() has been removed -
29eb77825c (arch, mm: Remove tlb_fast_mode()).
The new mmu_gather structure is now suitable for arm64 and this patch
converts the arch asm/tlb.h to the generic code. One functional
difference is the shift_arg_pages() case where previously the code was
flushing the full mm (no tlb_start_vma call) but now it flushes the
range given to tlb_gather_mmu() (possibly slightly more efficient
previously).
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
The patch moves the PCI I/O space (currently at 64K) before the
earlyprintk mapping and extends it to 16MB.
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
processor debug state PSTATE.D is unmasked in smp call
clear_os_lock for secondary cpus. So debug state is still
masked in normal kernel context. With this patch, unmask
debug state on secondary boot for the cpus in normal kernel
context. Now kgdb tests passed with multicore.
Signed-off-by: Vijaya Kumar K <Vijaya.Kumar@caviumnetworks.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
typecast instruction_pointer macro to unsigned long to
resolve following compiler warnings like
warning: format '%lx' expects argument of type 'long unsigned int',
but argument 2 has type 'u64' [-Wformat]
Signed-off-by: Vijaya Kumar K <Vijaya.Kumar@caviumnetworks.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add KGDB software step debugging support for EL1 debug
in AArch64 mode.
KGDB registers step debug handler with debug monitor.
On receiving 'step' command from GDB tool, target enables
software step debugging and step address is updated in ELR.
Software Step debugging is disabled when 'continue' command
is received
Signed-off-by: Vijaya Kumar K <Vijaya.Kumar@caviumnetworks.com>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add KGDB debug support for kernel debugging.
With this patch, basic KGDB debugging is possible.GDB register
layout is updated and GDB tool can establish connection with
target and can set/clear breakpoints.
Signed-off-by: Vijaya Kumar K <Vijaya.Kumar@caviumnetworks.com>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Add macros to enable and disable to manage PSTATE.D
for debugging. The macros local_dbg_save and local_dbg_restore
are moved to irqflags.h file
KGDB boot tests fail because of PSTATE.D is masked.
unmask it for debugging support
Signed-off-by: Vijaya Kumar K <Vijaya.Kumar@caviumnetworks.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.22 (GNU/Linux)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=+70Q
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
Pull KVM fixes from Paolo Bonzini:
"A small error handling problem and a compile breakage for ARM64"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm:
arm64: KVM: Add VGIC device control for arm64
KVM: return an error code in kvm_vm_ioctl_register_coalesced_mmio()
This fixes the build breakage introduced by
c07a0191ef and adds support for the device
control API and save/restore of the VGIC state for ARMv8.
The defines were simply missing from the arm64 header files and
uaccess.h must be implicitly imported from somewhere else on arm.
Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
FPGA implementations of the Cortex-A57 and Cortex-A53 are now available
in the form of the SMM-A57 and SMM-A53 Soft Macrocell Models (SMMs) for
Versatile Express. As these attach to a Motherboard Express V2M-P1 it
would be useful to have support for some V2M-P1 peripherals enabled by
default.
Additionally a couple of of features have been introduced since the last
defconfig update (CMA, jump labels) that would be good to have enabled
by default to ensure they are build and boot tested.
This patch updates the arm64 defconfig to enable support for these
devices and features. The arm64 Kconfig is modified to select
HAVE_PATA_PLATFORM, which is required to enable support for the
CompactFlash controller on the V2M-P1.
A few options which don't need to appear in defconfig are trimmed:
* BLK_DEV - selected by default
* EXPERIMENTAL - otherwise gone from the kernel
* MII - selected by drivers which require it
* USB_SUPPORT - selected by default
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
cbnz/tbnz don't update the condition flags, so remove the "cc" clobbers
from inline asm blocks that only use these instructions to implement
conditional branches.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Linux requires a number of atomic operations to provide full barrier
semantics, that is no memory accesses after the operation can be
observed before any accesses up to and including the operation in
program order.
On arm64, these operations have been incorrectly implemented as follows:
// A, B, C are independent memory locations
<Access [A]>
// atomic_op (B)
1: ldaxr x0, [B] // Exclusive load with acquire
<op(B)>
stlxr w1, x0, [B] // Exclusive store with release
cbnz w1, 1b
<Access [C]>
The assumption here being that two half barriers are equivalent to a
full barrier, so the only permitted ordering would be A -> B -> C
(where B is the atomic operation involving both a load and a store).
Unfortunately, this is not the case by the letter of the architecture
and, in fact, the accesses to A and C are permitted to pass their
nearest half barrier resulting in orderings such as Bl -> A -> C -> Bs
or Bl -> C -> A -> Bs (where Bl is the load-acquire on B and Bs is the
store-release on B). This is a clear violation of the full barrier
requirement.
The simple way to fix this is to implement the same algorithm as ARMv7
using explicit barriers:
<Access [A]>
// atomic_op (B)
dmb ish // Full barrier
1: ldxr x0, [B] // Exclusive load
<op(B)>
stxr w1, x0, [B] // Exclusive store
cbnz w1, 1b
dmb ish // Full barrier
<Access [C]>
but this has the undesirable effect of introducing *two* full barrier
instructions. A better approach is actually the following, non-intuitive
sequence:
<Access [A]>
// atomic_op (B)
1: ldxr x0, [B] // Exclusive load
<op(B)>
stlxr w1, x0, [B] // Exclusive store with release
cbnz w1, 1b
dmb ish // Full barrier
<Access [C]>
The simple observations here are:
- The dmb ensures that no subsequent accesses (e.g. the access to C)
can enter or pass the atomic sequence.
- The dmb also ensures that no prior accesses (e.g. the access to A)
can pass the atomic sequence.
- Therefore, no prior access can pass a subsequent access, or
vice-versa (i.e. A is strictly ordered before C).
- The stlxr ensures that no prior access can pass the store component
of the atomic operation.
The only tricky part remaining is the ordering between the ldxr and the
access to A, since the absence of the first dmb means that we're now
permitting re-ordering between the ldxr and any prior accesses.
From an (arbitrary) observer's point of view, there are two scenarios:
1. We have observed the ldxr. This means that if we perform a store to
[B], the ldxr will still return older data. If we can observe the
ldxr, then we can potentially observe the permitted re-ordering
with the access to A, which is clearly an issue when compared to
the dmb variant of the code. Thankfully, the exclusive monitor will
save us here since it will be cleared as a result of the store and
the ldxr will retry. Notice that any use of a later memory
observation to imply observation of the ldxr will also imply
observation of the access to A, since the stlxr/dmb ensure strict
ordering.
2. We have not observed the ldxr. This means we can perform a store
and influence the later ldxr. However, that doesn't actually tell
us anything about the access to [A], so we've not lost anything
here either when compared to the dmb variant.
This patch implements this solution for our barriered atomic operations,
ensuring that we satisfy the full barrier requirements where they are
needed.
Cc: <stable@vger.kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
