firefly-linux-kernel-4.4.55.git
8 years agoarm64: add support for kernel ASLR
Ard Biesheuvel [Tue, 26 Jan 2016 13:12:01 +0000 (14:12 +0100)]
arm64: add support for kernel ASLR

This adds support for KASLR is implemented, based on entropy provided by
the bootloader in the /chosen/kaslr-seed DT property. Depending on the size
of the address space (VA_BITS) and the page size, the entropy in the
virtual displacement is up to 13 bits (16k/2 levels) and up to 25 bits (all
4 levels), with the sidenote that displacements that result in the kernel
image straddling a 1GB/32MB/512MB alignment boundary (for 4KB/16KB/64KB
granule kernels, respectively) are not allowed, and will be rounded up to
an acceptable value.

If CONFIG_RANDOMIZE_MODULE_REGION_FULL is enabled, the module region is
randomized independently from the core kernel. This makes it less likely
that the location of core kernel data structures can be determined by an
adversary, but causes all function calls from modules into the core kernel
to be resolved via entries in the module PLTs.

If CONFIG_RANDOMIZE_MODULE_REGION_FULL is not enabled, the module region is
randomized by choosing a page aligned 128 MB region inside the interval
[_etext - 128 MB, _stext + 128 MB). This gives between 10 and 14 bits of
entropy (depending on page size), independently of the kernel randomization,
but still guarantees that modules are within the range of relative branch
and jump instructions (with the caveat that, since the module region is
shared with other uses of the vmalloc area, modules may need to be loaded
further away if the module region is exhausted)

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit f80fb3a3d50843a401dac4b566b3b131da8077a2)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: add support for building vmlinux as a relocatable PIE binary
Ard Biesheuvel [Tue, 26 Jan 2016 08:13:44 +0000 (09:13 +0100)]
arm64: add support for building vmlinux as a relocatable PIE binary

This implements CONFIG_RELOCATABLE, which links the final vmlinux
image with a dynamic relocation section, allowing the early boot code
to perform a relocation to a different virtual address at runtime.

This is a prerequisite for KASLR (CONFIG_RANDOMIZE_BASE).

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 1e48ef7fcc374051730381a2a05da77eb4eafdb0)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: switch to relative exception tables
Ard Biesheuvel [Fri, 1 Jan 2016 14:02:12 +0000 (15:02 +0100)]
arm64: switch to relative exception tables

Instead of using absolute addresses for both the exception location
and the fixup, use offsets relative to the exception table entry values.
Not only does this cut the size of the exception table in half, it is
also a prerequisite for KASLR, since absolute exception table entries
are subject to dynamic relocation, which is incompatible with the sorting
of the exception table that occurs at build time.

This patch also introduces the _ASM_EXTABLE preprocessor macro (which
exists on x86 as well) and its _asm_extable assembly counterpart, as
shorthands to emit exception table entries.

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 6c94f27ac847ff8ef15b3da5b200574923bd6287)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoextable: add support for relative extables to search and sort routines
Ard Biesheuvel [Fri, 1 Jan 2016 11:39:09 +0000 (12:39 +0100)]
extable: add support for relative extables to search and sort routines

This adds support to the generic search_extable() and sort_extable()
implementations for dealing with exception table entries whose fields
contain relative offsets rather than absolute addresses.

Acked-by: Helge Deller <deller@gmx.de>
Acked-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Acked-by: H. Peter Anvin <hpa@linux.intel.com>
Acked-by: Tony Luck <tony.luck@intel.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Acked-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit a272858a3c1ecd4a935ba23c66668f81214bd110)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoscripts/sortextable: add support for ET_DYN binaries
Ard Biesheuvel [Sun, 10 Jan 2016 10:42:28 +0000 (11:42 +0100)]
scripts/sortextable: add support for ET_DYN binaries

Add support to scripts/sortextable for handling relocatable (PIE)
executables, whose ELF type is ET_DYN, not ET_EXEC. Other than adding
support for the new type, no changes are needed.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 7b957b6e603623ef8b2e8222fa94b976df613fa2)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: futex.h: Add missing PAN toggling
James Morse [Tue, 2 Feb 2016 15:53:59 +0000 (15:53 +0000)]
arm64: futex.h: Add missing PAN toggling

futex.h's futex_atomic_cmpxchg_inatomic() does not use the
__futex_atomic_op() macro and needs its own PAN toggling. This was missed
when the feature was implemented.

Fixes: 338d4f49d6f ("arm64: kernel: Add support for Privileged Access Never")
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 811d61e384e24759372bb3f01772f3744b0a8327)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: make asm/elf.h available to asm files
Ard Biesheuvel [Mon, 11 Jan 2016 16:08:26 +0000 (17:08 +0100)]
arm64: make asm/elf.h available to asm files

This reshuffles some code in asm/elf.h and puts a #ifndef __ASSEMBLY__
around its C definitions so that the CPP defines can be used in asm
source files as well.

Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 4a2e034e5cdadde4c712f79bdd57d1455c76a3db)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: avoid dynamic relocations in early boot code
Ard Biesheuvel [Sat, 26 Dec 2015 11:46:40 +0000 (12:46 +0100)]
arm64: avoid dynamic relocations in early boot code

Before implementing KASLR for arm64 by building a self-relocating PIE
executable, we have to ensure that values we use before the relocation
routine is executed are not subject to dynamic relocation themselves.
This applies not only to virtual addresses, but also to values that are
supplied by the linker at build time and relocated using R_AARCH64_ABS64
relocations.

So instead, use assemble time constants, or force the use of static
relocations by folding the constants into the instructions.

Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 2bf31a4a05f5b00f37d65ba029d36a0230286cb7)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: avoid R_AARCH64_ABS64 relocations for Image header fields
Ard Biesheuvel [Sat, 26 Dec 2015 12:48:02 +0000 (13:48 +0100)]
arm64: avoid R_AARCH64_ABS64 relocations for Image header fields

Unfortunately, the current way of using the linker to emit build time
constants into the Image header will no longer work once we switch to
the use of PIE executables. The reason is that such constants are emitted
into the binary using R_AARCH64_ABS64 relocations, which are resolved at
runtime, not at build time, and the places targeted by those relocations
will contain zeroes before that.

So refactor the endian swapping linker script constant generation code so
that it emits the upper and lower 32-bit words separately.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 6ad1fe5d9077a1ab40bf74b61994d2e770b00b14)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: add support for module PLTs
Ard Biesheuvel [Tue, 24 Nov 2015 11:37:35 +0000 (12:37 +0100)]
arm64: add support for module PLTs

This adds support for emitting PLTs at module load time for relative
branches that are out of range. This is a prerequisite for KASLR, which
may place the kernel and the modules anywhere in the vmalloc area,
making it more likely that branch target offsets exceed the maximum
range of +/- 128 MB.

In this version, I removed the distinction between relocations against
.init executable sections and ordinary executable sections. The reason
is that it is hardly worth the trouble, given that .init.text usually
does not contain that many far branches, and this version now only
reserves PLT entry space for jump and call relocations against undefined
symbols (since symbols defined in the same module can be assumed to be
within +/- 128 MB)

For example, the mac80211.ko module (which is fairly sizable at ~400 KB)
built with -mcmodel=large gives the following relocation counts:

                    relocs    branches   unique     !local
  .text              3925       3347       518        219
  .init.text           11          8         7          1
  .exit.text            4          4         4          1
  .text.unlikely       81         67        36         17

('unique' means branches to unique type/symbol/addend combos, of which
!local is the subset referring to undefined symbols)

IOW, we are only emitting a single PLT entry for the .init sections, and
we are better off just adding it to the core PLT section instead.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit fd045f6cd98ec4953147b318418bd45e441e52a3)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: move brk immediate argument definitions to separate header
Ard Biesheuvel [Tue, 23 Feb 2016 07:56:45 +0000 (08:56 +0100)]
arm64: move brk immediate argument definitions to separate header

Instead of reversing the header dependency between asm/bug.h and
asm/debug-monitors.h, split off the brk instruction immediate value
defines into a new header asm/brk-imm.h, and include it from both.

This solves the circular dependency issue that prevents BUG() from
being used in some header files, and keeps the definitions together.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit f98deee9a9f8c47d05a0f64d86440882dca772ff)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: use bit ops rather than arithmetic in pa/va translations
Ard Biesheuvel [Mon, 22 Feb 2016 17:46:04 +0000 (18:46 +0100)]
arm64: mm: use bit ops rather than arithmetic in pa/va translations

Since PAGE_OFFSET is chosen such that it cuts the kernel VA space right
in half, and since the size of the kernel VA space itself is always a
power of 2, we can treat PAGE_OFFSET as a bitmask and replace the
additions/subtractions with 'or' and 'and-not' operations.

For the comparison against PAGE_OFFSET, a mov/cmp/branch sequence ends
up getting replaced with a single tbz instruction. For the additions and
subtractions, we save a mov instruction since the mask is folded into the
instruction's immediate field.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 8439e62a15614e8fcd43835d57b7245cd9870dc5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: only perform memstart_addr sanity check if DEBUG_VM
Ard Biesheuvel [Mon, 22 Feb 2016 17:46:03 +0000 (18:46 +0100)]
arm64: mm: only perform memstart_addr sanity check if DEBUG_VM

Checking whether memstart_addr has been assigned every time it is
referenced adds a branch instruction that may hurt performance if
the reference in question occurs on a hot path. So only perform the
check if CONFIG_DEBUG_VM=y.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
[catalin.marinas@arm.com: replaced #ifdef with VM_BUG_ON]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit a92405f082d43267575444a6927085e4c8a69e4e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: User die() instead of panic() in do_page_fault()
Catalin Marinas [Fri, 19 Feb 2016 14:28:58 +0000 (14:28 +0000)]
arm64: User die() instead of panic() in do_page_fault()

The former gives better error reporting on unhandled permission faults
(introduced by the UAO patches).

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 70c8abc28762d04e36c92e07eee2ce6ab41049cb)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: allow kernel Image to be loaded anywhere in physical memory
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:42 +0000 (13:52 +0100)]
arm64: allow kernel Image to be loaded anywhere in physical memory

This relaxes the kernel Image placement requirements, so that it
may be placed at any 2 MB aligned offset in physical memory.

This is accomplished by ignoring PHYS_OFFSET when installing
memblocks, and accounting for the apparent virtual offset of
the kernel Image. As a result, virtual address references
below PAGE_OFFSET are correctly mapped onto physical references
into the kernel Image regardless of where it sits in memory.

Special care needs to be taken for dealing with memory limits passed
via mem=, since the generic implementation clips memory top down, which
may clip the kernel image itself if it is loaded high up in memory. To
deal with this case, we simply add back the memory covering the kernel
image, which may result in more memory to be retained than was passed
as a mem= parameter.

Since mem= should not be considered a production feature, a panic notifier
handler is installed that dumps the memory limit at panic time if one was
set.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit a7f8de168ace487fa7b88cb154e413cf40e87fc6)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: defer __va translation of initrd_start and initrd_end
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:41 +0000 (13:52 +0100)]
arm64: defer __va translation of initrd_start and initrd_end

Before deferring the assignment of memstart_addr in a subsequent patch, to
the moment where all memory has been discovered and possibly clipped based
on the size of the linear region and the presence of a mem= command line
parameter, we need to ensure that memstart_addr is not used to perform __va
translations before it is assigned.

One such use is in the generic early DT discovery of the initrd location,
which is recorded as a virtual address in the globals initrd_start and
initrd_end. So wire up the generic support to declare the initrd addresses,
and implement it without __va() translations, and perform the translation
after memstart_addr has been assigned.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit a89dea585371a9d5d85499db47c93f129be8e0c4)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: move kernel image to base of vmalloc area
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:40 +0000 (13:52 +0100)]
arm64: move kernel image to base of vmalloc area

This moves the module area to right before the vmalloc area, and moves
the kernel image to the base of the vmalloc area. This is an intermediate
step towards implementing KASLR, which allows the kernel image to be
located anywhere in the vmalloc area.

Since other subsystems such as hibernate may still need to refer to the
kernel text or data segments via their linears addresses, both are mapped
in the linear region as well. The linear alias of the text region is
mapped read-only/non-executable to prevent inadvertent modification or
execution.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit f9040773b7bbbd9e98eb6184a263512a7cfc133f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kvm: deal with kernel symbols outside of linear mapping
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:39 +0000 (13:52 +0100)]
arm64: kvm: deal with kernel symbols outside of linear mapping

KVM on arm64 uses a fixed offset between the linear mapping at EL1 and
the HYP mapping at EL2. Before we can move the kernel virtual mapping
out of the linear mapping, we have to make sure that references to kernel
symbols that are accessed via the HYP mapping are translated to their
linear equivalent.

Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit a0bf9776cd0be4490d4675d4108e13379849fc7f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
Conflicts:
skip new funcs create_hyp_mappings(__start_rodata,
in arch/arm/kvm/arm.c and keep funcs in arch/arm64/kvm/hyp.S

8 years agoarm64: decouple early fixmap init from linear mapping
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:38 +0000 (13:52 +0100)]
arm64: decouple early fixmap init from linear mapping

Since the early fixmap page tables are populated using pages that are
part of the static footprint of the kernel, they are covered by the
initial kernel mapping, and we can refer to them without using __va/__pa
translations, which are tied to the linear mapping.

Since the fixmap page tables are disjoint from the kernel mapping up
to the top level pgd entry, we can refer to bm_pte[] directly, and there
is no need to walk the page tables and perform __pa()/__va() translations
at each step.

Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 157962f5a8f236cab898b68bdaa69ce68922f0bf)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: pgtable: implement static [pte|pmd|pud]_offset variants
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:37 +0000 (13:52 +0100)]
arm64: pgtable: implement static [pte|pmd|pud]_offset variants

The page table accessors pte_offset(), pud_offset() and pmd_offset()
rely on __va translations, so they can only be used after the linear
mapping has been installed. For the early fixmap and kasan init routines,
whose page tables are allocated statically in the kernel image, these
functions will return bogus values. So implement pte_offset_kimg(),
pmd_offset_kimg() and pud_offset_kimg(), which can be used instead
before any page tables have been allocated dynamically.

Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 6533945a32c762c5db70d7a3ec251a040b2d9661)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: introduce KIMAGE_VADDR as the virtual base of the kernel region
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:36 +0000 (13:52 +0100)]
arm64: introduce KIMAGE_VADDR as the virtual base of the kernel region

This introduces the preprocessor symbol KIMAGE_VADDR which will serve as
the symbolic virtual base of the kernel region, i.e., the kernel's virtual
offset will be KIMAGE_VADDR + TEXT_OFFSET. For now, we define it as being
equal to PAGE_OFFSET, but in the future, it will be moved below it once
we move the kernel virtual mapping out of the linear mapping.

Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit ab893fb9f1b17f02139bce547bb4b69e96b9ae16)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: add support for ioremap() block mappings
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:35 +0000 (13:52 +0100)]
arm64: add support for ioremap() block mappings

This wires up the existing generic huge-vmap feature, which allows
ioremap() to use PMD or PUD sized block mappings. It also adds support
to the unmap path for dealing with block mappings, which will allow us
to unmap the __init region using unmap_kernel_range() in a subsequent
patch.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 324420bf91f60582bb481133db9547111768ef17)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: prevent potential circular header dependencies in asm/bug.h
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:34 +0000 (13:52 +0100)]
arm64: prevent potential circular header dependencies in asm/bug.h

Currently, using BUG_ON() in header files is cumbersome, due to the fact
that asm/bug.h transitively includes a lot of other header files, resulting
in the actual BUG_ON() invocation appearing before its definition in the
preprocessor input. So let's reverse the #include dependency between
asm/bug.h and asm/debug-monitors.h, by moving the definition of BUG_BRK_IMM
from the latter to the former. Also fix up one user of asm/debug-monitors.h
which relied on a transitive include.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 03336b1df9929e5d9c28fd9768948b6151cb046c)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
Conflicts:
skip arch/arm64/kvm/hyp/debug-sr.c

8 years agoof/fdt: factor out assignment of initrd_start/initrd_end
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:33 +0000 (13:52 +0100)]
of/fdt: factor out assignment of initrd_start/initrd_end

Since architectures may not yet have their linear mapping up and running
when the initrd address is discovered from the DT, factor out the
assignment of initrd_start and initrd_end, so that an architecture can
override it and use the translation it needs.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Rob Herring <robh@kernel.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 369bc9abf22bf026e8645a4dd746b90649a2f6ee)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoof/fdt: make memblock minimum physical address arch configurable
Ard Biesheuvel [Tue, 16 Feb 2016 12:52:32 +0000 (13:52 +0100)]
of/fdt: make memblock minimum physical address arch configurable

By default, early_init_dt_add_memory_arch() ignores memory below
the base of the kernel image since it won't be addressable via the
linear mapping. However, this is not appropriate anymore once we
decouple the kernel text mapping from the linear mapping, so archs
may want to drop the low limit entirely. So allow the minimum to be
overridden by setting MIN_MEMBLOCK_ADDR.

Acked-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Rob Herring <robh@kernel.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 270522a04f7a9911983878fa37da467f9ff1c938)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Remove the get_thread_info() function
Catalin Marinas [Thu, 18 Feb 2016 15:50:04 +0000 (15:50 +0000)]
arm64: Remove the get_thread_info() function

This function was introduced by previous commits implementing UAO.
However, it can be replaced with task_thread_info() in
uao_thread_switch() or get_fs() in do_page_fault() (the latter being
called only on the current context, so no need for using the saved
pt_regs).

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit e950631e84e7e38892ffbeee5e1816b270026b0e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kernel: Don't toggle PAN on systems with UAO
James Morse [Fri, 5 Feb 2016 14:58:50 +0000 (14:58 +0000)]
arm64: kernel: Don't toggle PAN on systems with UAO

If a CPU supports both Privileged Access Never (PAN) and User Access
Override (UAO), we don't need to disable/re-enable PAN round all
copy_to_user() like calls.

UAO alternatives cause these calls to use the 'unprivileged' load/store
instructions, which are overridden to be the privileged kind when
fs==KERNEL_DS.

This patch changes the copy_to_user() calls to have their PAN toggling
depend on a new composite 'feature' ARM64_ALT_PAN_NOT_UAO.

If both features are detected, PAN will be enabled, but the copy_to_user()
alternatives will not be applied. This means PAN will be enabled all the
time for these functions. If only PAN is detected, the toggling will be
enabled as normal.

This will save the time taken to disable/re-enable PAN, and allow us to
catch copy_to_user() accesses that occur with fs==KERNEL_DS.

Futex and swp-emulation code continue to hang their PAN toggling code on
ARM64_HAS_PAN.

Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 705441960033e66b63524521f153fbb28c99ddbd)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: cpufeature: Test 'matches' pointer to find the end of the list
James Morse [Fri, 5 Feb 2016 14:58:49 +0000 (14:58 +0000)]
arm64: cpufeature: Test 'matches' pointer to find the end of the list

CPU feature code uses the desc field as a test to find the end of the list,
this means every entry must have a description. This generates noise for
entries in the list that aren't really features, but combinations of them.
e.g.
> CPU features: detected feature: Privileged Access Never
> CPU features: detected feature: PAN and not UAO

These combination features are needed for corner cases with alternatives,
where cpu features interact.

Change all walkers of the arm64_features[] and arm64_hwcaps[] lists to test
'matches' not 'desc', and only print 'desc' if it is non-NULL.

Signed-off-by: James Morse <james.morse@arm.com>
Reviewed-by : Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 644c2ae198412c956700e55a2acf80b2541f6aa5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kernel: Add support for User Access Override
James Morse [Fri, 5 Feb 2016 14:58:48 +0000 (14:58 +0000)]
arm64: kernel: Add support for User Access Override

'User Access Override' is a new ARMv8.2 feature which allows the
unprivileged load and store instructions to be overridden to behave in
the normal way.

This patch converts {get,put}_user() and friends to use ldtr*/sttr*
instructions - so that they can only access EL0 memory, then enables
UAO when fs==KERNEL_DS so that these functions can access kernel memory.

This allows user space's read/write permissions to be checked against the
page tables, instead of testing addr<USER_DS, then using the kernel's
read/write permissions.

Signed-off-by: James Morse <james.morse@arm.com>
[catalin.marinas@arm.com: move uao_thread_switch() above dsb()]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 57f4959bad0a154aeca125b7d38d1d9471a12422)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: add ARMv8.2 id_aa64mmfr2 boiler plate
James Morse [Fri, 5 Feb 2016 14:58:47 +0000 (14:58 +0000)]
arm64: add ARMv8.2 id_aa64mmfr2 boiler plate

ARMv8.2 adds a new feature register id_aa64mmfr2. This patch adds the
cpu feature boiler plate used by the actual features in later patches.

Signed-off-by: James Morse <james.morse@arm.com>
Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 406e308770a92bd33995b2e5b681e86358328bb0)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: cpufeature: Change read_cpuid() to use sysreg's mrs_s macro
James Morse [Fri, 5 Feb 2016 14:58:46 +0000 (14:58 +0000)]
arm64: cpufeature: Change read_cpuid() to use sysreg's mrs_s macro

Older assemblers may not have support for newer feature registers. To get
round this, sysreg.h provides a 'mrs_s' macro that takes a register
encoding and generates the raw instruction.

Change read_cpuid() to use mrs_s in all cases so that new registers
don't have to be a special case. Including sysreg.h means we need to move
the include and definition of read_cpuid() after the #ifndef __ASSEMBLY__
to avoid syntax errors in vmlinux.lds.

Signed-off-by: James Morse <james.morse@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 0f54b14e76f5302afe164dc911b049b5df836ff5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: use local label prefixes for __reg_num symbols
Ard Biesheuvel [Mon, 15 Feb 2016 08:51:49 +0000 (09:51 +0100)]
arm64: use local label prefixes for __reg_num symbols

The __reg_num_xNN symbols that are used to implement the msr_s and
mrs_s macros are recorded in the ELF metadata of each object file.
This does not affect the size of the final binary, but it does clutter
the output of tools like readelf, i.e.,

  $ readelf -a vmlinux |grep -c __reg_num_x
  50976

So let's use symbols with the .L prefix, these are strictly local,
and don't end up in the object files.

  $ readelf -a vmlinux |grep -c __reg_num_x
  0

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 7abc7d833c9eb16efc8a59239d3771a6e30be367)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: vdso: Mark vDSO code as read-only
David Brown [Wed, 10 Feb 2016 21:52:22 +0000 (13:52 -0800)]
arm64: vdso: Mark vDSO code as read-only

Although the arm64 vDSO is cleanly separated by code/data with the
code being read-only in userspace mappings, the code page is still
writable from the kernel.  There have been exploits (such as
http://itszn.com/blog/?p=21) that take advantage of this on x86 to go
from a bad kernel write to full root.

Prevent this specific exploit on arm64 by putting the vDSO code page
in read-only memory as well.

Before the change:
[    3.138366] vdso: 2 pages (1 code @ ffffffc000a71000, 1 data @ ffffffc000a70000)
---[ Kernel Mapping ]---
0xffffffc000000000-0xffffffc000082000         520K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc000082000-0xffffffc000200000        1528K     ro x  SHD AF            UXN MEM/NORMAL
0xffffffc000200000-0xffffffc000800000           6M     ro x  SHD AF        BLK UXN MEM/NORMAL
0xffffffc000800000-0xffffffc0009b6000        1752K     ro x  SHD AF            UXN MEM/NORMAL
0xffffffc0009b6000-0xffffffc000c00000        2344K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc000c00000-0xffffffc008000000         116M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc00c000000-0xffffffc07f000000        1840M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc800000000-0xffffffc840000000           1G     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc840000000-0xffffffc87ae00000         942M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc87ae00000-0xffffffc87ae70000         448K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87af80000-0xffffffc87af8a000          40K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87af8b000-0xffffffc87b000000         468K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87b000000-0xffffffc87fe00000          78M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc87fe00000-0xffffffc87ff50000        1344K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87ff90000-0xffffffc87ffa0000          64K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87fff0000-0xffffffc880000000          64K     RW NX SHD AF            UXN MEM/NORMAL

After:
[    3.138368] vdso: 2 pages (1 code @ ffffffc0006de000, 1 data @ ffffffc000a74000)
---[ Kernel Mapping ]---
0xffffffc000000000-0xffffffc000082000         520K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc000082000-0xffffffc000200000        1528K     ro x  SHD AF            UXN MEM/NORMAL
0xffffffc000200000-0xffffffc000800000           6M     ro x  SHD AF        BLK UXN MEM/NORMAL
0xffffffc000800000-0xffffffc0009b8000        1760K     ro x  SHD AF            UXN MEM/NORMAL
0xffffffc0009b8000-0xffffffc000c00000        2336K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc000c00000-0xffffffc008000000         116M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc00c000000-0xffffffc07f000000        1840M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc800000000-0xffffffc840000000           1G     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc840000000-0xffffffc87ae00000         942M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc87ae00000-0xffffffc87ae70000         448K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87af80000-0xffffffc87af8a000          40K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87af8b000-0xffffffc87b000000         468K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87b000000-0xffffffc87fe00000          78M     RW NX SHD AF        BLK UXN MEM/NORMAL
0xffffffc87fe00000-0xffffffc87ff50000        1344K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87ff90000-0xffffffc87ffa0000          64K     RW NX SHD AF            UXN MEM/NORMAL
0xffffffc87fff0000-0xffffffc880000000          64K     RW NX SHD AF            UXN MEM/NORMAL

Inspired by https://lkml.org/lkml/2016/1/19/494 based on work by the
PaX Team, Brad Spengler, and Kees Cook.

Signed-off-by: David Brown <david.brown@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
[catalin.marinas@arm.com: removed superfluous __PAGE_ALIGNED_DATA]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 88d8a7994e564d209d4b2583496631c2357d386b)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ubsan: select ARCH_HAS_UBSAN_SANITIZE_ALL
Yang Shi [Fri, 5 Feb 2016 23:50:18 +0000 (15:50 -0800)]
arm64: ubsan: select ARCH_HAS_UBSAN_SANITIZE_ALL

To enable UBSAN on arm64, ARCH_HAS_UBSAN_SANITIZE_ALL need to be selected.

Basic kernel bootup test is passed on arm64 with CONFIG_UBSAN_SANITIZE_ALL
enabled.

Signed-off-by: Yang Shi <yang.shi@linaro.org>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit f0b7f8a4b44657386273a67179dd901c81cd11a6)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ptdump: Indicate whether memory should be faulting
Laura Abbott [Sat, 6 Feb 2016 00:24:48 +0000 (16:24 -0800)]
arm64: ptdump: Indicate whether memory should be faulting

With CONFIG_DEBUG_PAGEALLOC, pages do not have the valid bit
set when free in the buddy allocator. Add an indiciation to
the page table dumping code that the valid bit is not set,
'F' for fault, to make this easier to understand.

Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Laura Abbott <labbott@fedoraproject.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit d7e9d59494a9a5d83274f5af2148b82ca22dff3f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Add support for ARCH_SUPPORTS_DEBUG_PAGEALLOC
Laura Abbott [Sat, 6 Feb 2016 00:24:47 +0000 (16:24 -0800)]
arm64: Add support for ARCH_SUPPORTS_DEBUG_PAGEALLOC

ARCH_SUPPORTS_DEBUG_PAGEALLOC provides a hook to map and unmap
pages for debugging purposes. This requires memory be mapped
with PAGE_SIZE mappings since breaking down larger mappings
at runtime will lead to TLB conflicts. Check if debug_pagealloc
is enabled at runtime and if so, map everyting with PAGE_SIZE
pages. Implement the functions to actually map/unmap the
pages at runtime.

Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Laura Abbott <labbott@fedoraproject.org>
[catalin.marinas@arm.com: static annotation block_mappings_allowed() and #ifdef]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 83863f25e4b8214e994ef8b5647aad614d74b45d)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Drop alloc function from create_mapping
Laura Abbott [Sat, 6 Feb 2016 00:24:46 +0000 (16:24 -0800)]
arm64: Drop alloc function from create_mapping

create_mapping is only used in fixmap_remap_fdt. All the create_mapping
calls need to happen on existing translation table pages without
additional allocations. Rather than have an alloc function be called
and fail, just set it to NULL and catch its use. Also change
the name to create_mapping_noalloc to better capture what exactly is
going on.

Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Laura Abbott <labbott@fedoraproject.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 132233a759580f5ce9b1bfaac9073e47d03c460d)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: prefetch: add missing #include for spin_lock_prefetch
Will Deacon [Wed, 10 Feb 2016 10:07:30 +0000 (10:07 +0000)]
arm64: prefetch: add missing #include for spin_lock_prefetch

As of 52e662326e1e ("arm64: prefetch: don't provide spin_lock_prefetch
with LSE"), spin_lock_prefetch is patched at runtime when the LSE atomics
are in use. This relies on the ARM64_LSE_ATOMIC_INSN macro to drive
the alternatives framework, but that macro is only available via
asm/lse.h, which isn't explicitly included in processor.h. Consequently,
drivers can run into build failures such as:

   In file included from include/linux/prefetch.h:14:0,
                    from drivers/net/ethernet/intel/i40e/i40e_txrx.c:27:
   arch/arm64/include/asm/processor.h: In function 'spin_lock_prefetch':
   arch/arm64/include/asm/processor.h:183:15: error: expected string literal before 'ARM64_LSE_ATOMIC_INSN'
     asm volatile(ARM64_LSE_ATOMIC_INSN(

This patch add the missing include and gets things building again.

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit afb83cc3f0e4f86ea0e1cc3db7a90f58f1abd4d5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: lib: patch in prfm for copy_page if requested
Andrew Pinski [Tue, 2 Feb 2016 12:46:26 +0000 (12:46 +0000)]
arm64: lib: patch in prfm for copy_page if requested

On ThunderX T88 pass 1 and pass 2, there is no hardware prefetching so
we need to patch in explicit software prefetching instructions

Prefetching improves this code by 60% over the original code and 2x
over the code without prefetching for the affected hardware using the
benchmark code at https://github.com/apinski-cavium/copy_page_benchmark

Signed-off-by: Andrew Pinski <apinski@cavium.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Tested-by: Andrew Pinski <apinski@cavium.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 60e0a09db24adc8809696307e5d97cc4ba7cb3e0)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: lib: improve copy_page to deal with 128 bytes at a time
Will Deacon [Tue, 2 Feb 2016 12:46:25 +0000 (12:46 +0000)]
arm64: lib: improve copy_page to deal with 128 bytes at a time

We want to avoid lots of different copy_page implementations, settling
for something that is "good enough" everywhere and hopefully easy to
understand and maintain whilst we're at it.

This patch reworks our copy_page implementation based on discussions
with Cavium on the list and benchmarking on Cortex-A processors so that:

  - The loop is unrolled to copy 128 bytes per iteration

  - The reads are offset so that we read from the next 128-byte block
    in the same iteration that we store the previous block

  - Explicit prefetch instructions are removed for now, since they hurt
    performance on CPUs with hardware prefetching

  - The loop exit condition is calculated at the start of the loop

Signed-off-by: Will Deacon <will.deacon@arm.com>
Tested-by: Andrew Pinski <apinski@cavium.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 223e23e8aa26b0bb62c597637e77295e14f6a62c)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: prefetch: add alternative pattern for CPUs without a prefetcher
Will Deacon [Tue, 2 Feb 2016 12:46:24 +0000 (12:46 +0000)]
arm64: prefetch: add alternative pattern for CPUs without a prefetcher

Most CPUs have a hardware prefetcher which generally performs better
without explicit prefetch instructions issued by software, however
some CPUs (e.g. Cavium ThunderX) rely solely on explicit prefetch
instructions.

This patch adds an alternative pattern (ARM64_HAS_NO_HW_PREFETCH) to
allow our library code to make use of explicit prefetch instructions
during things like copy routines only when the CPU does not have the
capability to perform the prefetching itself.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Tested-by: Andrew Pinski <apinski@cavium.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit d5370f754875460662abe8561388e019d90dd0c4)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: prefetch: don't provide spin_lock_prefetch with LSE
Will Deacon [Tue, 2 Feb 2016 12:46:23 +0000 (12:46 +0000)]
arm64: prefetch: don't provide spin_lock_prefetch with LSE

The LSE atomics rely on us not dirtying data at L1 if we can avoid it,
otherwise many of the potential scalability benefits are lost.

This patch replaces spin_lock_prefetch with a nop when the LSE atomics
are in use, so that users don't shoot themselves in the foot by causing
needless coherence traffic at L1.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Tested-by: Andrew Pinski <apinski@cavium.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit cd5e10bdf3795d22f10787bb1991c43798c885d5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: allow vmalloc regions to be set with set_memory_*
Ard Biesheuvel [Wed, 27 Jan 2016 09:50:19 +0000 (10:50 +0100)]
arm64: allow vmalloc regions to be set with set_memory_*

The range of set_memory_* is currently restricted to the module address
range because of difficulties in breaking down larger block sizes.
vmalloc maps PAGE_SIZE pages so it is safe to use as well. Update the
function ranges and add a comment explaining why the range is restricted
the way it is.

Suggested-by: Laura Abbott <labbott@fedoraproject.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 95f5c80050ad723163aa80dc8bffd48ef4afc6d5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kernel: implement ACPI parking protocol
Lorenzo Pieralisi [Tue, 26 Jan 2016 11:10:38 +0000 (11:10 +0000)]
arm64: kernel: implement ACPI parking protocol

The SBBR and ACPI specifications allow ACPI based systems that do not
implement PSCI (eg systems with no EL3) to boot through the ACPI parking
protocol specification[1].

This patch implements the ACPI parking protocol CPU operations, and adds
code that eases parsing the parking protocol data structures to the
ARM64 SMP initializion carried out at the same time as cpus enumeration.

To wake-up the CPUs from the parked state, this patch implements a
wakeup IPI for ARM64 (ie arch_send_wakeup_ipi_mask()) that mirrors the
ARM one, so that a specific IPI is sent for wake-up purpose in order
to distinguish it from other IPI sources.

Given the current ACPI MADT parsing API, the patch implements a glue
layer that helps passing MADT GICC data structure from SMP initialization
code to the parking protocol implementation somewhat overriding the CPU
operations interfaces. This to avoid creating a completely trasparent
DT/ACPI CPU operations layer that would require creating opaque
structure handling for CPUs data (DT represents CPU through DT nodes, ACPI
through static MADT table entries), which seems overkill given that ACPI
on ARM64 mandates only two booting protocols (PSCI and parking protocol),
so there is no need for further protocol additions.

Based on the original work by Mark Salter <msalter@redhat.com>

[1] https://acpica.org/sites/acpica/files/MP%20Startup%20for%20ARM%20platforms.docx

Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Tested-by: Loc Ho <lho@apm.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Hanjun Guo <hanjun.guo@linaro.org>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Mark Salter <msalter@redhat.com>
Cc: Al Stone <ahs3@redhat.com>
[catalin.marinas@arm.com: Added WARN_ONCE(!acpi_parking_protocol_valid() on the IPI]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 5e89c55e4ed81d7abb1ce8828db35fa389dc0e90)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: create new fine-grained mappings at boot
Mark Rutland [Mon, 25 Jan 2016 11:45:12 +0000 (11:45 +0000)]
arm64: mm: create new fine-grained mappings at boot

At boot we may change the granularity of the tables mapping the kernel
(by splitting or making sections). This may happen when we create the
linear mapping (in __map_memblock), or at any point we try to apply
fine-grained permissions to the kernel (e.g. fixup_executable,
mark_rodata_ro, fixup_init).

Changing the active page tables in this manner may result in multiple
entries for the same address being allocated into TLBs, risking problems
such as TLB conflict aborts or issues derived from the amalgamation of
TLB entries. Generally, a break-before-make (BBM) approach is necessary
to avoid conflicts, but we cannot do this for the kernel tables as it
risks unmapping text or data being used to do so.

Instead, we can create a new set of tables from scratch in the safety of
the existing mappings, and subsequently migrate over to these using the
new cpu_replace_ttbr1 helper, which avoids the two sets of tables being
active simultaneously.

To avoid issues when we later modify permissions of the page tables
(e.g. in fixup_init), we must create the page tables at a granularity
such that later modification does not result in splitting of tables.

This patch applies this strategy, creating a new set of fine-grained
page tables from scratch, and safely migrating to them. The existing
fixmap and kasan shadow page tables are reused in the new fine-grained
tables.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 068a17a5805dfbca4bbf03e664ca6b19709cc7a8)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ensure _stext and _etext are page-aligned
Mark Rutland [Mon, 25 Jan 2016 11:45:11 +0000 (11:45 +0000)]
arm64: ensure _stext and _etext are page-aligned

Currently we have separate ALIGN_DEBUG_RO{,_MIN} directives to align
_etext and __init_begin. While we ensure that __init_begin is
page-aligned, we do not provide the same guarantee for _etext. This is
not problematic currently as the alignment of __init_begin is sufficient
to prevent issues when we modify permissions.

Subsequent patches will assume page alignment of segments of the kernel
we wish to map with different permissions. To ensure this, move _etext
after the ALIGN_DEBUG_RO_MIN for the init section. This renders the
prior ALIGN_DEBUG_RO irrelevant, and hence it is removed. Likewise,
upgrade to ALIGN_DEBUG_RO_MIN(PAGE_SIZE) for _stext.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit fca082bfb543ccaaff864fc0892379ccaa1711cd)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: allow passing a pgdir to alloc_init_*
Mark Rutland [Mon, 25 Jan 2016 11:45:10 +0000 (11:45 +0000)]
arm64: mm: allow passing a pgdir to alloc_init_*

To allow us to initialise pgdirs which are fixmapped, allow explicitly
passing a pgdir rather than an mm. A new __create_pgd_mapping function
is added for this, with existing __create_mapping callers migrated to
this.

The mm argument was previously only used at the top level. Now that it
is redundant at all levels, it is removed. To indicate its new found
similarity to alloc_init_{pud,pmd,pte}, __create_mapping is renamed to
init_pgd.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 11509a306bb6ea595878b2d246d2d56b1783e040)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: allocate pagetables anywhere
Mark Rutland [Mon, 25 Jan 2016 11:45:09 +0000 (11:45 +0000)]
arm64: mm: allocate pagetables anywhere

Now that create_mapping uses fixmap slots to modify pte, pmd, and pud
entries, we can access page tables anywhere in physical memory,
regardless of the extent of the linear mapping.

Given that, we no longer need to limit memblock allocations during page
table creation, and can leave the limit as its default
MEMBLOCK_ALLOC_ANYWHERE.

We never add memory which will fall outside of the linear map range
given phys_offset and MAX_MEMBLOCK_ADDR are configured appropriately, so
any tables we create will fall in the linear map of the final tables.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit cdef5f6e9e0e5ee397759b664a9f875ff59ccf01)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: use fixmap when creating page tables
Mark Rutland [Mon, 25 Jan 2016 11:45:08 +0000 (11:45 +0000)]
arm64: mm: use fixmap when creating page tables

As a preparatory step to allow us to allocate early page tables from
unmapped memory using memblock_alloc, modify the __create_mapping
callees to map and unmap the tables they modify using fixmap entries.

All but the top-level pgd initialisation is performed via the fixmap.
Subsequent patches will inject the pgd physical address, and migrate to
using the FIX_PGD slot.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit f4710445458c0a1bd1c3c014ada2e7d7dc7b882f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: add functions to walk tables in fixmap
Mark Rutland [Mon, 25 Jan 2016 11:45:07 +0000 (11:45 +0000)]
arm64: mm: add functions to walk tables in fixmap

As a preparatory step to allow us to allocate early page tables from
unmapped memory using memblock_alloc, add new p??_{set,clear}_fixmap*
functions which can be used to walk page tables outside of the linear
mapping by using fixmap slots.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 961faac114819a01e627fe9c9c82b830bb3849d4)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: add __{pud,pgd}_populate
Mark Rutland [Mon, 25 Jan 2016 11:45:06 +0000 (11:45 +0000)]
arm64: mm: add __{pud,pgd}_populate

We currently have __pmd_populate for creating a pmd table entry given
the physical address of a pte, but don't have equivalents for the pud or
pgd levels of table.

To enable us to manipulate tables which are mapped outside of the linear
mapping (where we have a PA, but not a linear map VA), it is useful to
have these functions.

This patch adds __{pud,pgd}_populate. As these should not be called when
the kernel uses folded {pmd,pud}s, in these cases they expand to
BUILD_BUG(). So long as the appropriate checks are made on the {pud,pgd}
entry prior to attempting population, these should be optimized out at
compile time.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 1e531cce68c92b46c7d29f36a72f9a3e5886678f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: avoid redundant __pa(__va(x))
Mark Rutland [Mon, 25 Jan 2016 11:45:05 +0000 (11:45 +0000)]
arm64: mm: avoid redundant __pa(__va(x))

When we "upgrade" to a section mapping, we free any table we made
redundant by giving it back to memblock. To get the PA, we acquire the
physical address and convert this to a VA, then subsequently convert
this back to a PA.

This works currently, but will not work if the tables are not accessed
via linear map VAs (e.g. is we use fixmap slots).

This patch uses {pmd,pud}_page_paddr to acquire the PA. This avoids the
__pa(__va()) round trip, saving some work and avoiding reliance on the
linear mapping.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 316b39db06718d59d82736df9fc65cf05b467cc7)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: add functions to walk page tables by PA
Mark Rutland [Mon, 25 Jan 2016 11:45:04 +0000 (11:45 +0000)]
arm64: mm: add functions to walk page tables by PA

To allow us to walk tables allocated into the fixmap, we need to acquire
the physical address of a page, rather than the virtual address in the
linear map.

This patch adds new p??_page_paddr and p??_offset_phys functions to
acquire the physical address of a next-level table, and changes
p??_offset* into macros which simply convert this to a linear map VA.
This renders p??_page_vaddr unused, and hence they are removed.

At the pgd level, a new pgd_offset_raw function is added to find the
relevant PGD entry given the base of a PGD and a virtual address.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit dca56dca7124709f3dfca81afe61b4d98eb9cacf)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: move pte_* macros
Mark Rutland [Mon, 25 Jan 2016 11:45:03 +0000 (11:45 +0000)]
arm64: mm: move pte_* macros

For pmd, pud, and pgd levels of table, functions including p?d_index and
p?d_offset are defined after the p?d_page_vaddr function for the
immediately higher level of table.

The pte functions however are defined much earlier, even though several
rely on the later definition of pmd_page_vaddr. While this isn't
currently a problem as these are macros, it prevents the logical
grouping of later C functions (which cannot rely on prototypes for
functions not yet defined).

Move these definitions after pmd_page_vaddr, for consistency with the
placement of these functions for other levels of table.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 053520f7d3923cc6d37afb28f9887cb1e7d77454)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kasan: avoid TLB conflicts
Mark Rutland [Mon, 25 Jan 2016 11:45:02 +0000 (11:45 +0000)]
arm64: kasan: avoid TLB conflicts

The page table modification performed during the KASAN init risks the
allocation of conflicting TLB entries, as it swaps a set of valid global
entries for another without suitable TLB maintenance.

The presence of conflicting TLB entries can result in the delivery of
synchronous TLB conflict aborts, or may result in the use of erroneous
data being returned in response to a TLB lookup. This can affect
explicit data accesses from software as well as translations performed
asynchronously (e.g. as part of page table walks or speculative I-cache
fetches), and can therefore result in a wide variety of problems.

To avoid this, use cpu_replace_ttbr1 to swap the page tables. This
ensures that when the new tables are installed there are no stale
entries from the old tables which may conflict. As all updates are made
to the tables while they are not active, the updates themselves are
safe.

At the same time, add the missing barrier to ensure that the tmp_pg_dir
entries updated via memcpy are visible to the page table walkers at the
point the tmp_pg_dir is installed. All other page table updates made as
part of KASAN initialisation have the requisite barriers due to the use
of the standard page table accessors.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit c1a88e9124a499939ebd8069d5e4d3937f019157)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: add code to safely replace TTBR1_EL1
Mark Rutland [Mon, 25 Jan 2016 11:45:01 +0000 (11:45 +0000)]
arm64: mm: add code to safely replace TTBR1_EL1

If page tables are modified without suitable TLB maintenance, the ARM
architecture permits multiple TLB entries to be allocated for the same
VA. When this occurs, it is permitted that TLB conflict aborts are
raised in response to synchronous data/instruction accesses, and/or and
amalgamation of the TLB entries may be used as a result of a TLB lookup.

The presence of conflicting TLB entries may result in a variety of
behaviours detrimental to the system (e.g. erroneous physical addresses
may be used by I-cache fetches and/or page table walks). Some of these
cases may result in unexpected changes of hardware state, and/or result
in the (asynchronous) delivery of SError.

To avoid these issues, we must avoid situations where conflicting
entries may be allocated into TLBs. For user and module mappings we can
follow a strict break-before-make approach, but this cannot work for
modifications to the swapper page tables that cover the kernel text and
data.

Instead, this patch adds code which is intended to be executed from the
idmap, which can safely unmap the swapper page tables as it only
requires the idmap to be active. This enables us to uninstall the active
TTBR1_EL1 entry, invalidate TLBs, then install a new TTBR1_EL1 entry
without potentially unmapping code or data required for the sequence.
This avoids the risk of conflict, but requires that updates are staged
in a copy of the swapper page tables prior to being installed.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 50e1881ddde2a986c7d0d2150985239e5e3d7d96)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: add function to install the idmap
Mark Rutland [Mon, 25 Jan 2016 11:45:00 +0000 (11:45 +0000)]
arm64: add function to install the idmap

In some cases (e.g. when making invasive changes to the kernel page
tables) we will need to execute code from the idmap.

Add a new helper which may be used to install the idmap, complementing
the existing cpu_uninstall_idmap.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 609116d202a8c5fd3fe393eb85373cbee906df68)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: unmap idmap earlier
Mark Rutland [Mon, 25 Jan 2016 11:44:59 +0000 (11:44 +0000)]
arm64: unmap idmap earlier

During boot we leave the idmap in place until paging_init, as we
previously had to wait for the zero page to become allocated and
accessible.

Now that we have a statically-allocated zero page, we can uninstall the
idmap much earlier in the boot process, making it far easier to spot
accidental use of physical addresses. This also brings the cold boot
path in line with the secondary boot path.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 86ccce896cb0aa800a7a6dcd29b41ffc4eeb1a75)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: unify idmap removal
Mark Rutland [Mon, 25 Jan 2016 11:44:58 +0000 (11:44 +0000)]
arm64: unify idmap removal

We currently open-code the removal of the idmap and restoration of the
current task's MMU state in a few places.

Before introducing yet more copies of this sequence, unify these to call
a new helper, cpu_uninstall_idmap.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 9e8e865bbe294a69666a1996bda3e87825b258c0)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: place empty_zero_page in bss
Mark Rutland [Mon, 25 Jan 2016 11:44:57 +0000 (11:44 +0000)]
arm64: mm: place empty_zero_page in bss

Currently the zero page is set up in paging_init, and thus we cannot use
the zero page earlier. We use the zero page as a reserved TTBR value
from which no TLB entries may be allocated (e.g. when uninstalling the
idmap). To enable such usage earlier (as may be required for invasive
changes to the kernel page tables), and to minimise the time that the
idmap is active, we need to be able to use the zero page before
paging_init.

This patch follows the example set by x86, by allocating the zero page
at compile time, in .bss. This means that the zero page itself is
available immediately upon entry to start_kernel (as we zero .bss before
this), and also means that the zero page takes up no space in the raw
Image binary. The associated struct page is allocated in bootmem_init,
and remains unavailable until this time.

Outside of arch code, the only users of empty_zero_page assume that the
empty_zero_page symbol refers to the zeroed memory itself, and that
ZERO_PAGE(x) must be used to acquire the associated struct page,
following the example of x86. This patch also brings arm64 inline with
these assumptions.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 5227cfa71f9e8574373f4d0e9e754942d76cdf67)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: specialise pagetable allocators
Mark Rutland [Mon, 25 Jan 2016 11:44:56 +0000 (11:44 +0000)]
arm64: mm: specialise pagetable allocators

We pass a size parameter to early_alloc and late_alloc, but these are
only ever used to allocate single pages. In late_alloc we always
allocate a single page.

Both allocators provide us with zeroed pages (such that all entries are
invalid), but we have no barriers between allocating a page and adding
that page to existing (live) tables. A concurrent page table walk may
see stale data, leading to a number of issues.

This patch specialises the two allocators for page tables. The size
parameter is removed and the necessary dsb(ishst) is folded into each.
To make it clear that the functions are intended for use for page table
allocation, they are renamed to {early,late}_pgtable_alloc, with the
related function pointed renamed to pgtable_alloc.

As the dsb(ishst) is now in the allocator, the existing barrier for the
zero page is redundant and thus is removed. The previously missing
include of barrier.h is added.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 21ab99c289d350f4ae454bc069870009db6df20e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoasm-generic: Fix local variable shadow in __set_fixmap_offset
Mark Rutland [Mon, 25 Jan 2016 11:44:55 +0000 (11:44 +0000)]
asm-generic: Fix local variable shadow in __set_fixmap_offset

Currently __set_fixmap_offset is a macro function which has a local
variable called 'addr'. If a caller passes a 'phys' parameter which is
derived from a variable also called 'addr', the local variable will
shadow this, and the compiler will complain about the use of an
uninitialized variable. To avoid the issue with namespace clashes,
'addr' is prefixed with a liberal sprinkling of underscores.

Turning __set_fixmap_offset into a static inline breaks the build for
several architectures. Fixing this properly requires updates to a number
of architectures to make them agree on the prototype of __set_fixmap (it
could be done as a subsequent patch series).

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Arnd Bergmann <arnd@arndb.de>
[catalin.marinas@arm.com: squashed the original function patch and macro fixup]
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
(cherry picked from commit 3694bd76781b76c4f8d2ecd85018feeb1609f0e5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoEliminate the .eh_frame sections from the aarch64 vmlinux and kernel modules
William Cohen [Fri, 22 Jan 2016 03:56:26 +0000 (22:56 -0500)]
Eliminate the .eh_frame sections from the aarch64 vmlinux and kernel modules

By default the aarch64 gcc generates .eh_frame sections.  Unlike
.debug_frame sections, the .eh_frame sections are loaded into memory
when the associated code is loaded.  On an example kernel being built
with this default the .eh_frame section in vmlinux used an extra 1.7MB
of memory.  The x86 disables the creation of the .eh_frame section.
The aarch64 should probably do the same to save some memory.

Signed-off-by: William Cohen <wcohen@redhat.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 728dabd6d1751cf5e0f8e0535891393da62396e9)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
Conflicts:
pick 67dfa1751 arm64: errata: Add -mpc-relative-literal-loads
in arch/arm64/Makefile

8 years agoarm64: Fix an enum typo in mm/dump.c
Masanari Iida [Sun, 24 Jan 2016 06:24:12 +0000 (15:24 +0900)]
arm64: Fix an enum typo in mm/dump.c

This patch fixes a typo in mm/dump.c:
"MODUELS_END_NR" should be "MODULES_END_NR".

Signed-off-by: Masanari Iida <standby24x7@gmail.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit b3122023df935cf14bf951da98ca598d71b9f826)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: kasan: ensure that the KASAN zero page is mapped read-only
Ard Biesheuvel [Mon, 11 Jan 2016 13:50:21 +0000 (14:50 +0100)]
arm64: kasan: ensure that the KASAN zero page is mapped read-only

When switching from the early KASAN shadow region, which maps the
entire shadow space read-write, to the permanent KASAN shadow region,
which uses a zero page to shadow regions that are not subject to
instrumentation, the lowest level table kasan_zero_pte[] may be
reused unmodified, which means that the mappings of the zero page
that it contains will still be read-write.

So update it explicitly to map the zero page read only when we
activate the permanent mapping.

Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 7b1af9795773d745c2a8c7d4ca5f2936e8b6adfb)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarch/arm64/include/asm/pgtable.h: add pmd_mkclean for THP
Minchan Kim [Sat, 16 Jan 2016 00:55:37 +0000 (16:55 -0800)]
arch/arm64/include/asm/pgtable.h: add pmd_mkclean for THP

MADV_FREE needs pmd_dirty and pmd_mkclean for detecting recent overwrite
of the contents since MADV_FREE syscall is called for THP page.

This patch adds pmd_mkclean for THP page MADV_FREE support.

Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
(cherry picked from commit 05ee26d9e7e29ab026995eab79be3c6e8351908c)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: hide __efistub_ aliases from kallsyms
Ard Biesheuvel [Fri, 15 Jan 2016 12:28:57 +0000 (13:28 +0100)]
arm64: hide __efistub_ aliases from kallsyms

Commit e8f3010f7326 ("arm64/efi: isolate EFI stub from the kernel
proper") isolated the EFI stub code from the kernel proper by prefixing
all of its symbols with __efistub_, and selectively allowing access to
core kernel symbols from the stub by emitting __efistub_ aliases for
functions and variables that the stub can access legally.

As an unintended side effect, these aliases are emitted into the
kallsyms symbol table, which means they may turn up in backtraces,
e.g.,

  ...
  PC is at __efistub_memset+0x108/0x200
  LR is at fixup_init+0x3c/0x48
  ...
  [<ffffff8008328608>] __efistub_memset+0x108/0x200
  [<ffffff8008094dcc>] free_initmem+0x2c/0x40
  [<ffffff8008645198>] kernel_init+0x20/0xe0
  [<ffffff8008085cd0>] ret_from_fork+0x10/0x40

The backtrace in question has nothing to do with the EFI stub, but
simply returns one of the several aliases of memset() that have been
recorded in the kallsyms table. This is undesirable, since it may
suggest to people who are not aware of this that the issue they are
seeing is somehow EFI related.

So hide the __efistub_ aliases from kallsyms, by emitting them as
absolute linker symbols explicitly. The distinction between those
and section relative symbols is completely irrelevant to these
definitions, and to the final link we are performing when these
definitions are being taken into account (the distinction is only
relevant to symbols defined inside a section definition when performing
a partial link), and so the resulting values are identical to the
original ones. Since absolute symbols are ignored by kallsyms, this
will result in these values to be omitted from its symbol table.

After this patch, the backtrace generated from the same address looks
like this:
  ...
  PC is at __memset+0x108/0x200
  LR is at fixup_init+0x3c/0x48
  ...
  [<ffffff8008328608>] __memset+0x108/0x200
  [<ffffff8008094dcc>] free_initmem+0x2c/0x40
  [<ffffff8008645198>] kernel_init+0x20/0xe0
  [<ffffff8008085cd0>] ret_from_fork+0x10/0x40

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 75feee3d9d51775072d3a04f47d4a439a4c4590e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: head.S: use memset to clear BSS
Mark Rutland [Wed, 6 Jan 2016 11:05:27 +0000 (11:05 +0000)]
arm64: head.S: use memset to clear BSS

Currently we use an open-coded memzero to clear the BSS. As it is a
trivial implementation, it is sub-optimal.

Our optimised memset doesn't use the stack, is position-independent, and
for the memzero case can use of DC ZVA to clear large blocks
efficiently. In __mmap_switched the MMU is on and there are no live
caller-saved registers, so we can safely call an uninstrumented memset.

This patch changes __mmap_switched to use memset when clearing the BSS.
We use the __pi_memset alias so as to avoid any instrumentation in all
kernel configurations.

Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 2a803c4db615d85126c5c7afd5849a3cfde71422)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoefi: stub: define DISABLE_BRANCH_PROFILING for all architectures
Ard Biesheuvel [Wed, 23 Dec 2015 09:29:28 +0000 (10:29 +0100)]
efi: stub: define DISABLE_BRANCH_PROFILING for all architectures

This moves the DISABLE_BRANCH_PROFILING define from the x86 specific
to the general CFLAGS definition for the stub. This fixes build errors
when building for arm64 with CONFIG_PROFILE_ALL_BRANCHES_ENABLED.

Reviewed-by: Matt Fleming <matt@codeblueprint.co.uk>
Reported-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit b523e185bba36164ca48a190f5468c140d815414)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: entry: remove pointless SPSR mode check
Mark Rutland [Tue, 5 Jan 2016 17:33:34 +0000 (17:33 +0000)]
arm64: entry: remove pointless SPSR mode check

In work_pending, we may skip work if the stacked SPSR value represents
anything other than an EL0 context. We then immediately invoke the
kernel_exit 0 macro as part of ret_to_user, assuming a return to EL0.
This is somewhat confusing.

We use work_pending as part of the ret_to_user/ret_fast_syscall state
machine. We only use ret_fast_syscall in the return from an SVC issued
from EL0. We use ret_to_user for return from EL0 exception handlers and
also for return from ret_from_fork in the case the task was not a kernel
thread (i.e. it is a user task).

Thus in all cases the stacked SPSR value must represent an EL0 context,
and the check is redundant. This patch removes it, along with the now
unused no_work_pending label.

Cc: Chris Metcalf <cmetcalf@ezchip.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit ee03353bc04f8e460cc4e3da80d9721d9ecb89f1)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: move pgd_cache initialisation to pgtable_cache_init
Will Deacon [Tue, 5 Jan 2016 15:36:59 +0000 (15:36 +0000)]
arm64: mm: move pgd_cache initialisation to pgtable_cache_init

Initialising the suppport for EFI runtime services requires us to
allocate a pgd off the back of an early_initcall. On systems where the
PGD_SIZE is smaller than PAGE_SIZE (e.g. 64k pages and 48-bit VA), the
pgd_cache isn't initialised at this stage, and we panic with a NULL
dereference during boot:

  Unable to handle kernel NULL pointer dereference at virtual address 00000000

  __create_mapping.isra.5+0x84/0x350
  create_pgd_mapping+0x20/0x28
  efi_create_mapping+0x5c/0x6c
  arm_enable_runtime_services+0x154/0x1e4
  do_one_initcall+0x8c/0x190
  kernel_init_freeable+0x84/0x1ec
  kernel_init+0x10/0xe0
  ret_from_fork+0x10/0x50

This patch fixes the problem by initialising the pgd_cache earlier, in
the pgtable_cache_init callback, which sounds suspiciously like what it
was intended for.

Reported-by: Dennis Chen <dennis.chen@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 39b5be9b4233a9f212b98242bddf008f379b5122)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: module: avoid undefined shift behavior in reloc_data()
Ard Biesheuvel [Tue, 5 Jan 2016 09:18:52 +0000 (10:18 +0100)]
arm64: module: avoid undefined shift behavior in reloc_data()

Compilers may engage the improbability drive when encountering shifts
by a distance that is a multiple of the size of the operand type. Since
the required bounds check is very simple here, we can get rid of all the
fuzzy masking, shifting and comparing, and use the documented bounds
directly.

Reported-by: David Binderman <dcb314@hotmail.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit f930896967fa3f9ab16a6f87267b92798308d48f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: module: fix relocation of movz instruction with negative immediate
Ard Biesheuvel [Tue, 5 Jan 2016 09:18:51 +0000 (10:18 +0100)]
arm64: module: fix relocation of movz instruction with negative immediate

The test whether a movz instruction with a signed immediate should be
turned into a movn instruction (i.e., when the immediate is negative)
is flawed, since the value of imm is always positive. Also, the
subsequent bounds check is incorrect since the limit update never
executes, due to the fact that the imm_type comparison will always be
false for negative signed immediates.

Let's fix this by performing the sign test on sval directly, and
replacing the bounds check with a simple comparison against U16_MAX.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
[will: tidied up use of sval, renamed MOVK enum value to MOVKZ]
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit b24a557527f97ad88619d5bd4c8017c635056d69)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: traps: address fallout from printk -> pr_* conversion
Will Deacon [Mon, 21 Dec 2015 16:44:27 +0000 (16:44 +0000)]
arm64: traps: address fallout from printk -> pr_* conversion

Commit ac7b406c1a9d ("arm64: Use pr_* instead of printk") was a fairly
mindless s/printk/pr_*/ change driven by a complaint from checkpatch.

As is usual with such changes, this has led to some odd behaviour on
arm64:

  * syslog now picks up the "pr_emerg" line from dump_backtrace, but not
    the actual trace, which leads to a bunch of "kernel:Call trace:"
    lines in the log

  * __{pte,pmd,pgd}_error print at KERN_CRIT, as opposed to KERN_ERR
    which is used by other architectures.

This patch restores the original printk behaviour for dump_backtrace
and downgrade the pgtable error macros to KERN_ERR.

Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit c9cd0ed925c0b927283d4739bfe689eb9d1e9dfd)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ftrace: fix a stack tracer's output under function graph tracer
AKASHI Takahiro [Tue, 15 Dec 2015 08:33:41 +0000 (17:33 +0900)]
arm64: ftrace: fix a stack tracer's output under function graph tracer

Function graph tracer modifies a return address (LR) in a stack frame
to hook a function return. This will result in many useless entries
(return_to_handler) showing up in
 a) a stack tracer's output
 b) perf call graph (with perf record -g)
 c) dump_backtrace (at panic et al.)

For example, in case of a),
  $ echo function_graph > /sys/kernel/debug/tracing/current_tracer
  $ echo 1 > /proc/sys/kernel/stack_trace_enabled
  $ cat /sys/kernel/debug/tracing/stack_trace
        Depth    Size   Location    (54 entries)
        -----    ----   --------
  0)     4504      16   gic_raise_softirq+0x28/0x150
  1)     4488      80   smp_cross_call+0x38/0xb8
  2)     4408      48   return_to_handler+0x0/0x40
  3)     4360      32   return_to_handler+0x0/0x40
  ...

In case of b),
  $ echo function_graph > /sys/kernel/debug/tracing/current_tracer
  $ perf record -e mem:XXX:x -ag -- sleep 10
  $ perf report
                  ...
                  |          |          |--0.22%-- 0x550f8
                  |          |          |          0x10888
                  |          |          |          el0_svc_naked
                  |          |          |          sys_openat
                  |          |          |          return_to_handler
                  |          |          |          return_to_handler
                  ...

In case of c),
  $ echo function_graph > /sys/kernel/debug/tracing/current_tracer
  $ echo c > /proc/sysrq-trigger
  ...
  Call trace:
  [<ffffffc00044d3ac>] sysrq_handle_crash+0x24/0x30
  [<ffffffc000092250>] return_to_handler+0x0/0x40
  [<ffffffc000092250>] return_to_handler+0x0/0x40
  ...

This patch replaces such entries with real addresses preserved in
current->ret_stack[] at unwind_frame(). This way, we can cover all
the cases.

Reviewed-by: Jungseok Lee <jungseoklee85@gmail.com>
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
[will: fixed minor context changes conflicting with irq stack bits]
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 20380bb390a443b2c5c8800cec59743faf8151b4)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: pass a task parameter to unwind_frame()
AKASHI Takahiro [Tue, 15 Dec 2015 08:33:40 +0000 (17:33 +0900)]
arm64: pass a task parameter to unwind_frame()

Function graph tracer modifies a return address (LR) in a stack frame
to hook a function's return. This will result in many useless entries
(return_to_handler) showing up in a call stack list.
We will fix this problem in a later patch ("arm64: ftrace: fix a stack
tracer's output under function graph tracer"). But since real return
addresses are saved in ret_stack[] array in struct task_struct,
unwind functions need to be notified of, in addition to a stack pointer
address, which task is being traced in order to find out real return
addresses.

This patch extends unwind functions' interfaces by adding an extra
argument of a pointer to task_struct.

Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit fe13f95b720075327a761fe6ddb45b0c90cab504)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ftrace: modify a stack frame in a safe way
AKASHI Takahiro [Tue, 15 Dec 2015 08:33:39 +0000 (17:33 +0900)]
arm64: ftrace: modify a stack frame in a safe way

Function graph tracer modifies a return address (LR) in a stack frame by
calling ftrace_prepare_return() in a traced function's function prologue.
The current code does this modification before preserving an original
address at ftrace_push_return_trace() and there is always a small window
of inconsistency when an interrupt occurs.

This doesn't matter, as far as an interrupt stack is introduced, because
stack tracer won't be invoked in an interrupt context. But it would be
better to proactively minimize such a window by moving the LR modification
after ftrace_push_return_trace().

Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 79fdee9b6355c9720f14717e1ad66af51bb331b5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: remove irq_count and do_softirq_own_stack()
James Morse [Fri, 18 Dec 2015 16:01:47 +0000 (16:01 +0000)]
arm64: remove irq_count and do_softirq_own_stack()

sysrq_handle_reboot() re-enables interrupts while on the irq stack. The
irq_stack implementation wrongly assumed this would only ever happen
via the softirq path, allowing it to update irq_count late, in
do_softirq_own_stack().

This means if an irq occurs in sysrq_handle_reboot(), during
emergency_restart() the stack will be corrupted, as irq_count wasn't
updated.

Lose the optimisation, and instead of moving the adding/subtracting of
irq_count into irq_stack_entry/irq_stack_exit, remove it, and compare
sp_el0 (struct thread_info) with sp & ~(THREAD_SIZE - 1). This tells us
if we are on a task stack, if so, we can safely switch to the irq stack.
Finally, remove do_softirq_own_stack(), we don't need it anymore.

Reported-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: James Morse <james.morse@arm.com>
[will: use get_thread_info macro]
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit d224a69e3d80fe08f285d1f41d21b590bae4fa9f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: hugetlb: add support for PTE contiguous bit
David Woods [Thu, 17 Dec 2015 19:31:26 +0000 (14:31 -0500)]
arm64: hugetlb: add support for PTE contiguous bit

The arm64 MMU supports a Contiguous bit which is a hint that the TTE
is one of a set of contiguous entries which can be cached in a single
TLB entry.  Supporting this bit adds new intermediate huge page sizes.

The set of huge page sizes available depends on the base page size.
Without using contiguous pages the huge page sizes are as follows.

 4KB:   2MB  1GB
64KB: 512MB

With a 4KB granule, the contiguous bit groups together sets of 16 pages
and with a 64KB granule it groups sets of 32 pages.  This enables two new
huge page sizes in each case, so that the full set of available sizes
is as follows.

 4KB:  64KB   2MB  32MB  1GB
64KB:   2MB 512MB  16GB

If a 16KB granule is used then the contiguous bit groups 128 pages
at the PTE level and 32 pages at the PMD level.

If the base page size is set to 64KB then 2MB pages are enabled by
default.  It is possible in the future to make 2MB the default huge
page size for both 4KB and 64KB granules.

Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Reviewed-by: Steve Capper <steve.capper@linaro.org>
Signed-off-by: David Woods <dwoods@ezchip.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 66b3923a1a0f77a563b43f43f6ad091354abbfe9)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Use PoU cache instr for I/D coherency
Ashok Kumar [Thu, 17 Dec 2015 09:38:32 +0000 (01:38 -0800)]
arm64: Use PoU cache instr for I/D coherency

In systems with three levels of cache(PoU at L1 and PoC at L3),
PoC cache flush instructions flushes L2 and L3 caches which could affect
performance.
For cache flushes for I and D coherency, PoU should suffice.
So changing all I and D coherency related cache flushes to PoU.

Introduced a new __clean_dcache_area_pou API for dcache flush till PoU
and provided a common macro for __flush_dcache_area and
__clean_dcache_area_pou.

Also, now in __sync_icache_dcache, icache invalidation for non-aliasing
VIPT icache is done only for that particular page instead of the earlier
__flush_icache_all.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Ashok Kumar <ashoks@broadcom.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 0a28714c53fd4f7aea709be7577dfbe0095c8c3e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
Conflicts:
included reset_pmuserenr_el0 in arch/arm64/mm/proc-macros.S

8 years agoarm64: Defer dcache flush in __cpu_copy_user_page
Ashok Kumar [Thu, 17 Dec 2015 09:38:31 +0000 (01:38 -0800)]
arm64: Defer dcache flush in __cpu_copy_user_page

Defer dcache flushing to __sync_icache_dcache by calling
flush_dcache_page which clears PG_dcache_clean flag.

Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Ashok Kumar <ashoks@broadcom.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit e6b1185f77351aa154e63bd54b05d07ff99d4ffa)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: reduce stack use in irq_handler
James Morse [Tue, 15 Dec 2015 11:21:25 +0000 (11:21 +0000)]
arm64: reduce stack use in irq_handler

The code for switching to irq_stack stores three pieces of information on
the stack, fp+lr, as a fake stack frame (that lets us walk back onto the
interrupted tasks stack frame), and the address of the struct pt_regs that
contains the register values from kernel entry. (which dump_backtrace()
will print in any stack trace).

To reduce this, we store fp, and the pointer to the struct pt_regs.
unwind_frame() can recognise this as the irq_stack dummy frame, (as it only
appears at the top of the irq_stack), and use the struct pt_regs values
to find the missing interrupted link-register.

Suggested-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 971c67ce37cfeeaf560e792a2c3bc21d8b67163a)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Documentation: add list of software workarounds for errata
Will Deacon [Tue, 17 Nov 2015 14:45:47 +0000 (14:45 +0000)]
arm64: Documentation: add list of software workarounds for errata

It's not immediately obvious which hardware errata are worked around in
the Linux kernel for an arbitrary kernel tree, so add a file to keep
track of what we're working around.

Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 9cb9c9e5ba8453537e8e645318edf231fe54eaf9)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: place __cpu_setup in .text
Mark Rutland [Fri, 11 Dec 2015 11:04:31 +0000 (11:04 +0000)]
arm64: mm: place __cpu_setup in .text

We drop __cpu_setup in .text.init, which ends up being part of .text.
The .text.init section was a legacy section name which has been unused
elsewhere for a long time.

The ".text.init" name is misleading if read as a synonym for
".init.text". Any CPU may execute __cpu_setup before turning the MMU on,
so it should simply live in .text.

Remove the pointless section assignment. This will leave __cpu_setup in
the .text section.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit f00083cae331e5d3eecade6b4fdc35d0825e73ef)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: cmpxchg: Don't incldue linux/mmdebug.h
Mark Brown [Thu, 10 Dec 2015 16:54:32 +0000 (16:54 +0000)]
arm64: cmpxchg: Don't incldue linux/mmdebug.h

The arm64 asm/cmpxchg.h includes linux/mmdebug.h but doesn't so far as I
can tell actually use anything from it.  Removing the inclusion reduces
spurious header dependency rebuilds and also avoids issues with
recursive inclusions of headers causing build breaks due to attempts to
use things before they are defined if linux/mmdebug.h starts pulling in
more low level headers.

Such errors have happened in -next recently, for example:

In file included from include/linux/completion.h:11:0,
                 from include/linux/rcupdate.h:43,
                 from include/linux/tracepoint.h:19,
                 from include/linux/mmdebug.h:6,
                 from ./arch/arm64/include/asm/cmpxchg.h:22,
                 from ./arch/arm64/include/asm/atomic.h:41,
                 from include/linux/atomic.h:4,
                 from include/linux/spinlock.h:406,
                 from include/linux/seqlock.h:35,
                 from include/linux/time.h:5,
                 from include/uapi/linux/timex.h:56,
                 from include/linux/timex.h:56,
                 from include/linux/sched.h:19,
                 from arch/arm64/kernel/asm-offsets.c:21:
include/linux/wait.h: In function 'wait_on_atomic_t':
include/linux/wait.h:1218:2: error: implicit declaration of function 'atomic_read' [-Werror=implicit-function-declaration]
 if (atomic_read(val) == 0)

Signed-off-by: Mark Brown <broonie@kernel.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 4a6ccf30263f4e265c0f171561bf4c40bed5f273)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: fold alternatives into .init
Mark Rutland [Wed, 9 Dec 2015 12:44:38 +0000 (12:44 +0000)]
arm64: mm: fold alternatives into .init

Currently we treat the alternatives separately from other data that's
only used during initialisation, using separate .altinstructions and
.altinstr_replacement linker sections. These are freed for general
allocation separately from .init*. This is problematic as:

* We do not remove execute permissions, as we do for .init, leaving the
  memory executable.

* We pad between them, making the kernel Image bianry up to PAGE_SIZE
  bytes larger than necessary.

This patch moves the two sections into the contiguous region used for
.init*. This saves some memory, ensures that we remove execute
permissions, and allows us to remove some code made redundant by this
reorganisation.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Andre Przywara <andre.przywara@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 9aa4ec1571da62366cfddc20f3b923609604fe63)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Remove redundant padding from linker script
Mark Rutland [Wed, 9 Dec 2015 12:44:37 +0000 (12:44 +0000)]
arm64: Remove redundant padding from linker script

Currently we place an ALIGN_DEBUG_RO between text and data for the .text
and .init sections, and depending on configuration each of these may
result in up to SECTION_SIZE bytes worth of padding (for
DEBUG_RODATA_ALIGN).

We make no distinction between the text and data in each of these
sections at any point when creating the initial page tables in head.S.
We also make no distinction when modifying the tables; __map_memblock,
fixup_executable, mark_rodata_ro, and fixup_init only work at section
granularity. Thus this padding is unnecessary.

For the spit between init text and data we impose a minimum alignment of
16 bytes, but this is also unnecessary. The init data is output
immediately after the padding before any symbols are defined, so this is
not required to keep a symbol for linker a section array correctly
associated with the data. Any objects within the section will be given
at least their usual alignment regardless.

This patch removes the redundant padding.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 5b28cd9d084eca8ddc46270d2720305bfd40e348)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: mm: remove pointless PAGE_MASKing
Mark Rutland [Wed, 9 Dec 2015 12:44:36 +0000 (12:44 +0000)]
arm64: mm: remove pointless PAGE_MASKing

As pgd_offset{,_k} shift the input address by PGDIR_SHIFT, the sub-page
bits will always be shifted out. There is no need to apply PAGE_MASK
before this.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Jeremy Linton <jeremy.linton@arm.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit e2c30ee320eb96304896c7ab84499e5bc5e5fb6e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: don't call C code with el0's fp register
James Morse [Thu, 10 Dec 2015 10:22:41 +0000 (10:22 +0000)]
arm64: don't call C code with el0's fp register

On entry from el0, we save all the registers on the kernel stack, and
restore them before returning. x29 remains unchanged when we call out
to C code, which will store x29 as the frame-pointer on the stack.

Instead, write 0 into x29 after entry from el0, to avoid any risk of
tracing into user space.

Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 49003a8d6b35e128ef5e51433e60e783a46fbe5f)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: when walking onto the task stack, check sp & fp are in current->stack
James Morse [Thu, 10 Dec 2015 10:22:40 +0000 (10:22 +0000)]
arm64: when walking onto the task stack, check sp & fp are in current->stack

When unwind_frame() reaches the bottom of the irq_stack, the last fp
points to the original task stack. unwind_frame() uses
IRQ_STACK_TO_TASK_STACK() to find the sp value. If either values is
wrong, we may end up walking a corrupt stack.

Check these values are sane by testing if they are both on the stack
pointed to by current->stack.

Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 1ffe199b1c9b72a8e752a9ae2a7af10128ab2ca1)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Add this_cpu_ptr() assembler macro for use in entry.S
James Morse [Thu, 10 Dec 2015 10:22:39 +0000 (10:22 +0000)]
arm64: Add this_cpu_ptr() assembler macro for use in entry.S

irq_stack is a per_cpu variable, that needs to be access from entry.S.
Use an assembler macro instead of the unreadable details.

Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit aa4d5d3cbc258c355151a3903211b27359390ec5)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: irq: fix walking from irq stack to task stack
Will Deacon [Wed, 9 Dec 2015 13:58:42 +0000 (13:58 +0000)]
arm64: irq: fix walking from irq stack to task stack

Running with CONFIG_DEBUG_SPINLOCK=y can trigger a BUG with the new IRQ
stack code:

  BUG: spinlock lockup suspected on CPU#1

This is due to the IRQ_STACK_TO_TASK_STACK macro incorrectly retrieving
the task stack pointer stashed at the top of the IRQ stack.

Sayeth James:

| Yup, this is what is happening. Its an off-by-one due to broken
| thinking about how the stack works. My broken thinking was:
|
| >   top ------------
| >       | dummy_lr | <- irq_stack_ptr
| >       ------------
| >       |   x29    |
| >       ------------
| >       |   x19    | <- irq_stack_ptr - 0x10
| >       ------------
| >       |   xzr    |
| >       ------------
|
| But the stack-pointer is decreased before use. So it actually looks
| like this:
|
| >       ------------
| >       |          |  <- irq_stack_ptr
| >   top ------------
| >       | dummy_lr |
| >       ------------
| >       |   x29    | <- irq_stack_ptr - 0x10
| >       ------------
| >       |   x19    |
| >       ------------
| >       |   xzr    | <- irq_stack_ptr - 0x20
| >       ------------
|
| The value being used as the original stack is x29, which in all the
| tests is sp but without the current frames data, hence there are no
| missing frames in the output.
|
| Jungseok Lee picked it up with a 32bit user space because aarch32
| can't use x29, so it remains 0 forever. The fix he posted is correct.

This patch fixes the macro and adds some of this wisdom to a comment,
so that the layout of the IRQ stack is well understood.

Cc: James Morse <james.morse@arm.com>
Reported-by: Jungseok Lee <jungseoklee85@gmail.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 7596abf2e5661d52c4f414f37addeed54e098880)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Add do_softirq_own_stack() and enable irq_stacks
James Morse [Fri, 4 Dec 2015 11:02:27 +0000 (11:02 +0000)]
arm64: Add do_softirq_own_stack() and enable irq_stacks

entry.S is modified to switch to the per_cpu irq_stack during el{0,1}_irq.
irq_count is used to detect recursive interrupts on the irq_stack, it is
updated late by do_softirq_own_stack(), when called on the irq_stack, before
__do_softirq() re-enables interrupts to process softirqs.

do_softirq_own_stack() is added by this patch, but does not yet switch
stack.

This patch adds the dummy stack frame and data needed by the previous
stack tracing patches.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 8e23dacd12a48e58125b84c817da50850b73280a)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Modify stack trace and dump for use with irq_stack
AKASHI Takahiro [Fri, 4 Dec 2015 11:02:26 +0000 (11:02 +0000)]
arm64: Modify stack trace and dump for use with irq_stack

This patch allows unwind_frame() to traverse from interrupt stack to task
stack correctly. It requires data from a dummy stack frame, created
during irq_stack_entry(), added by a later patch.

A similar approach is taken to modify dump_backtrace(), which expects to
find struct pt_regs underneath any call to functions marked __exception.
When on an irq_stack, the struct pt_regs is stored on the old task stack,
the location of which is stored in the dummy stack frame.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
[james.morse: merged two patches, reworked for per_cpu irq_stacks, and
 no alignment guarantees, added irq_stack definitions]
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 132cd887b5c54758d04bf25c52fa48f45e843a30)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Store struct thread_info in sp_el0
Jungseok Lee [Fri, 4 Dec 2015 11:02:25 +0000 (11:02 +0000)]
arm64: Store struct thread_info in sp_el0

There is need for figuring out how to manage struct thread_info data when
IRQ stack is introduced. struct thread_info information should be copied
to IRQ stack under the current thread_info calculation logic whenever
context switching is invoked. This is too expensive to keep supporting
the approach.

Instead, this patch pays attention to sp_el0 which is an unused scratch
register in EL1 context. sp_el0 utilization not only simplifies the
management, but also prevents text section size from being increased
largely due to static allocated IRQ stack as removing masking operation
using THREAD_SIZE in many places.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Jungseok Lee <jungseoklee85@gmail.com>
Signed-off-by: James Morse <james.morse@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 6cdf9c7ca687e01840d0215437620a20263012fc)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: Add trace_hardirqs_off annotation in ret_to_user
Catalin Marinas [Fri, 4 Dec 2015 12:42:29 +0000 (12:42 +0000)]
arm64: Add trace_hardirqs_off annotation in ret_to_user

When a kernel is built with CONFIG_TRACE_IRQFLAGS the following warning
is produced when entering userspace for the first time:

  WARNING: at /work/Linux/linux-2.6-aarch64/kernel/locking/lockdep.c:3519
  Modules linked in:
  CPU: 1 PID: 1 Comm: systemd Not tainted 4.4.0-rc3+ #639
  Hardware name: Juno (DT)
  task: ffffffc9768a0000 ti: ffffffc9768a8000 task.ti: ffffffc9768a8000
  PC is at check_flags.part.22+0x19c/0x1a8
  LR is at check_flags.part.22+0x19c/0x1a8
  pc : [<ffffffc0000fba6c>] lr : [<ffffffc0000fba6c>] pstate: 600001c5
  sp : ffffffc9768abe10
  x29: ffffffc9768abe10 x28: ffffffc9768a8000
  x27: 0000000000000000 x26: 0000000000000001
  x25: 00000000000000a6 x24: ffffffc00064be6c
  x23: ffffffc0009f249e x22: ffffffc9768a0000
  x21: ffffffc97fea5480 x20: 00000000000001c0
  x19: ffffffc00169a000 x18: 0000005558cc7b58
  x17: 0000007fb78e3180 x16: 0000005558d2e238
  x15: ffffffffffffffff x14: 0ffffffffffffffd
  x13: 0000000000000008 x12: 0101010101010101
  x11: 7f7f7f7f7f7f7f7f x10: fefefefefefeff63
  x9 : 7f7f7f7f7f7f7f7f x8 : 6e655f7371726964
  x7 : 0000000000000001 x6 : ffffffc0001079c4
  x5 : 0000000000000000 x4 : 0000000000000001
  x3 : ffffffc001698438 x2 : 0000000000000000
  x1 : ffffffc9768a0000 x0 : 000000000000002e
  Call trace:
  [<ffffffc0000fba6c>] check_flags.part.22+0x19c/0x1a8
  [<ffffffc0000fc440>] lock_is_held+0x80/0x98
  [<ffffffc00064bafc>] __schedule+0x404/0x730
  [<ffffffc00064be6c>] schedule+0x44/0xb8
  [<ffffffc000085bb0>] ret_to_user+0x0/0x24
  possible reason: unannotated irqs-off.
  irq event stamp: 502169
  hardirqs last  enabled at (502169): [<ffffffc000085a98>] el0_irq_naked+0x1c/0x24
  hardirqs last disabled at (502167): [<ffffffc0000bb3bc>] __do_softirq+0x17c/0x298
  softirqs last  enabled at (502168): [<ffffffc0000bb43c>] __do_softirq+0x1fc/0x298
  softirqs last disabled at (502143): [<ffffffc0000bb830>] irq_exit+0xa0/0xf0

This happens because we disable interrupts in ret_to_user before calling
schedule() in work_resched. This patch adds the necessary
trace_hardirqs_off annotation.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Reported-by: Mark Rutland <mark.rutland@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit db3899a6477a4dccd26cbfb7f408b6be2cc068e0)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ftrace: fix the comments for ftrace_modify_code
Li Bin [Fri, 4 Dec 2015 03:38:40 +0000 (11:38 +0800)]
arm64: ftrace: fix the comments for ftrace_modify_code

There is no need to worry about module and __init text disappearing
case, because that ftrace has a module notifier that is called when
a module is being unloaded and before the text goes away and this
code grabs the ftrace_lock mutex and removes the module functions
from the ftrace list, such that it will no longer do any
modifications to that module's text, the update to make functions
be traced or not is done under the ftrace_lock mutex as well.
And by now, __init section codes should not been modified
by ftrace, because it is black listed in recordmcount.c and
ignored by ftrace.

Suggested-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Li Bin <huawei.libin@huawei.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 004ab584e028093996cf5b8e220b8bc50c5111cf)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: ftrace: stop using kstop_machine to enable/disable tracing
Li Bin [Fri, 4 Dec 2015 03:38:39 +0000 (11:38 +0800)]
arm64: ftrace: stop using kstop_machine to enable/disable tracing

For ftrace on arm64, kstop_machine which is hugely disruptive
to a running system is not needed to convert nops to ftrace calls
or back, because that to be modified instrucions, that NOP, B or BL,
are all safe instructions which called "concurrent modification
and execution of instructions", that can be executed by one
thread of execution as they are being modified by another thread
of execution without requiring explicit synchronization.

Signed-off-by: Li Bin <huawei.libin@huawei.com>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 81a6a146e88eca5d6726569779778d61489d85aa)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: spinlock: serialise spin_unlock_wait against concurrent lockers
Will Deacon [Thu, 19 Nov 2015 17:48:31 +0000 (17:48 +0000)]
arm64: spinlock: serialise spin_unlock_wait against concurrent lockers

Boqun Feng reported a rather nasty ordering issue with spin_unlock_wait
on architectures implementing spin_lock with LL/SC sequences and acquire
semantics:

 | CPU 1                   CPU 2                     CPU 3
 | ==================      ====================      ==============
 |                                                   spin_unlock(&lock);
 |                         spin_lock(&lock):
 |                           r1 = *lock; // r1 == 0;
 |                         o = READ_ONCE(object); // reordered here
 | object = NULL;
 | smp_mb();
 | spin_unlock_wait(&lock);
 |                           *lock = 1;
 | smp_mb();
 | o->dead = true;
 |                         if (o) // true
 |                           BUG_ON(o->dead); // true!!

The crux of the problem is that spin_unlock_wait(&lock) can return on
CPU 1 whilst CPU 2 is in the process of taking the lock. This can be
resolved by upgrading spin_unlock_wait to a LOCK operation, forcing it
to serialise against a concurrent locker and giving it acquire semantics
in the process (although it is not at all clear whether this is needed -
different callers seem to assume different things about the barrier
semantics and architectures are similarly disjoint in their
implementations of the macro).

This patch implements spin_unlock_wait using an LL/SC sequence with
acquire semantics on arm64. For v8.1 systems with the LSE atomics, the
exclusive writeback is omitted, since the spin_lock operation is
indivisible and no intermediate state can be observed.

Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit d86b8da04dfa4771a68bdbad6c424d40f22f0d14)
Signed-off-by: Alex Shi <alex.shi@linaro.org>
8 years agoarm64: enable HAVE_IRQ_TIME_ACCOUNTING
Will Deacon [Mon, 23 Nov 2015 15:12:59 +0000 (15:12 +0000)]
arm64: enable HAVE_IRQ_TIME_ACCOUNTING

arm64 relies on the arm_arch_timer for sched_clock, so we can select
HAVE_IRQ_TIME_ACCOUNTING and have the core sched-clock code enable the
feature at runtime based on the rate.

Reported-by: Mario Smarduch <m.smarduch@samsung.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
(cherry picked from commit 24da208db32ee1e4757ceaba898c47add8e5361e)
Signed-off-by: Alex Shi <alex.shi@linaro.org>