this is possible with the new build system that allows src/*/$(ARCH)/*
files which do not shadow a file in the parent directory, and yields a
more logical organization. eventually it will be possible to remove
arch/*/src from the build system.
switch to ll/sc model so that new atomic.h can provide optimized
versions of all the atomic primitives without needing an ll/sc loop
written in asm for each one.
all isa levels which use ldrex/strex now use the inline ll/sc model
even if the type of barrier to use is not known until runtime (v6).
the cas model is only used for arm v5 and earlier, and it has been
optimized to make the call via inline asm with custom constraints
rather than as a C function call.
sh needs runtime-selected atomic backends since there are a number of
supported models that use non-forwards-compatible (non-smp-compatible)
atomic mechanisms. previously, the code paths for this were highly
inefficient since they involved C function calls with multiple
branches in the callee and heavy spills in the caller. the new code
performs calls the runtime-selected asm fragment from inline asm with
extremely minimal clobbers, rather than using a function call.
for the sh4a case where the atomic mechanism is known and there is no
forward-compatibility issue, the movli.l and movco.l instructions are
provided as a_ll and a_sc, allowing the new shared atomic.h to
generate efficient inline versions of all the basic atomic operations
without needing a cas loop.
rather than having each arch provide its own atomic.h, there is a new
shared atomic.h in src/internal which pulls arch-specific definitions
from arc/$(ARCH)/atomic_arch.h. the latter can be extremely minimal,
defining only a_cas or new ll/sc type primitives which the shared
atomic.h will use to construct everything else.
this commit avoids making heavy changes to the individual archs'
atomic implementations. definitions which are identical or
near-identical to what the new shared atomic.h would produce have been
removed, but otherwise the changes made are just hooking up the
arch-specific files to the new infrastructure. major changes to take
advantage of the new system will come in subsequent commits.
commit 2f853dd6b9 failed to replicate
the old makefile logic that caused arch/arm/src/arm/atomics.s to be
built. since this was the only .s file under arch/*/src, rather than
trying to reproduce the old logic, I'm just moving it up a level and
adjusting the glob pattern in the makefile to catch it. eventually
arch/*/src will probably be removed in favor of moving all these files
to appropriate src/*/$(ARCH) locations.
the __SOFTFP__ macro which was wrongly being used does not reflect the
ABI (arm vs armhf) but just the availability of floating point
instructions/registers, so -mfloat-abi=softfp was wrongly being
treated as armhf. __ARM_PCS_VFP is the correct predefined macro to
check for the armhf EABI variant. this macro usage was corrected for
the build process in commit 4918c2bb20
but reloc.h was apparently overlooked at the time.
apparently the .gpword directive does not work reliably with local
text labels; values produced were offset by 64k from the correct
value, resulting in incorrect computation of the got pointer at
runtime. instead, use an external label so that the assembler does not
munge the relocation; the linker will then get it right.
commit 6fef8cafbd exposed this issue by
removing the old, non-PIE-compatible handwritten crt1.s, which was not
affected. presumably mips PIE executables (using Scrt1.o produced from
crt_arch.h) were already affected at the time.
at least gcc 4.7 claims c++11 support but does not accept the alignas
keyword, causing breakage when stddef.h is included in c++11 mode.
instead, prefer using __attribute__((__aligned__)) on any compiler
with GNU extensions, and only use the alignas keyword as a fallback
for other C++ compilers.
C code should not be affected by this patch.
commit 8a8fdf6398 was intended to remove
all such usage, but these arch-specific files were overlooked, leading
to inconsistent declarations and definitions.
on linux/nommu, non-writable private mappings of files may actually
use memory shared with other processes or the fs cache. the old nommu
loader code (used when mmap with MAP_FIXED fails) simply wrote over
top of the original file mapping, possibly clobbering this shared
memory. no such breakage was observed in practice, but it should have
been possible.
the new code starts by mapping anonymous writable memory on archs that
might support nommu, then maps load segments over top of it, falling
back to read if MAP_FIXED fails. we use an anonymous map rather than a
writable file map to avoid reading more data from disk than needed.
since pages cannot be loaded lazily on fault, in case of large
data/bss, mapping the full file may read a lot of data that will
subsequently be thrown away when processing additional LOAD segments.
as a result, we cannot skip the first LOAD segment when operating in
this mode.
these changes affect only non-FDPIC nommu support.
these files are all accepted as legacy arm syntax when producing arm
code, but legacy syntax cannot be used for producing thumb2 with
access to the full ISA. even after switching to UAL, some asm source
files contain instructions which are not valid in thumb mode, so these
will need to be addressed separately.
the idea of the three-instruction sequence being removed was to be
able to return to thumb code when used on armv4t+ from a thumb caller,
but also to be able to run on armv4 without the bx instruction
available (in which case the low bit of lr would always be 0).
however, without compiler support for generating such a sequence from
C code, which does not exist and which there is unlikely to be
interest in implementing, there is little point in having it in the
asm, and it would likely be easier to add pre-armv4t support via
enhanced linker handling of R_ARM_V4BX than at the compiler level.
removing this code simplifies adding support for building libc in
thumb2-only form (for cortex-m).
this assumption is borderline-unsafe to begin with, and fails badly
with -ffunction-sections since the linker can move the callee
arbitrarily far away when it lies in a different section.
using the actual mcontext_t definition rather than an overlaid pointer
array both improves correctness/readability and eliminates some ugly
hacks for archs with 64-bit registers bit 32-bit program counter.
also fix UB due to comparison of pointers not in a common array
object.
other archs use asm for the thread pointer load, so making that asm
volatile is sufficient to inform the compiler that it has a "side
effect" (crashing or giving the wrong result if the thread pointer was
not yet initialized) that prevents reordering. however, powerpc and
or1k have dedicated general purpose registers for the thread pointer
and did not need to use any asm to access it; instead, "local register
variables with a specified register" were used. however, there is no
specification for ordering constraints on this type of usage, and
presumably use of the thread pointer could be reordered across its
initialization.
to impose an ordering, I have added empty volatile asm blocks that
produce the "local register variable with a specified register" as
an output constraint.
this builds on commits a603a75a72 and
0ba35d69c0 to ensure that a compiler
cannot conclude that it's valid to reorder the asm to a point before
the thread pointer is set up, or to treat the inline function as if it
were declared with attribute((const)).
other archs already use volatile asm for thread pointer loading.
the restorer function pointer provided in the kernel sigaction
structure is interpreted by the kernel as a raw code address, not a
function descriptor.
this commit moves the declarations of the __restore and __restore_rt
symbols to ksigaction.h so that arch versions of the file can override
them, and introduces a version for sh which declares them as objects
rather than functions.
an alternate solution would have been defining SA_RESTORER to 0 so
that the functions are not used, but this both requires executable
stack (since the sh kernel does not have a vdso page with permanent
restorer functions) and crashes on qemu user-level emulation.
the entry point code supports being loaded by a loader which is not
fdpic-aware (in practice, either kernel with mmu or qemu without fdpic
support). this mostly just works, but signal handling will wrongly use
a function descriptor address as a code address if the personality is
not adjusted to fdpic.
ideally this code could be placed with sigaction so that it's not
needed except if/when a signal handler is installed. however,
personality is incorrectly maintained per-thread by the kernel, rather
than per-process, so it's necessary to correct the personality before
any threads are started. also, in order to skip the personality
syscall when an fdpic-aware loader is used, we need to be able to
detect how the program was loaded, and this information is only
readily available at the entry point.
previously, the normal ELF library loading code was used even for
fdpic, so only the kernel-loaded dynamic linker and main app could
benefit from separate placement of segments and shared text.
the __fdpic_fixup code is not needed for ET_DYN executables, which
instead use reloctions, so we can omit it from the dynamic linker and
static-pie entry point and save some code size.
the C implementation of __unmapself used for potentially-nommu sh
assumed CRTJMP takes a function descriptor rather than a code address;
however, the actual dynamic linker needs a code address, and so commit
7a9669e977 changed the definition of the
macro in reloc.h. this commit puts the old macro back in a place where
it only affects __unmapself.
this is an ugly workaround and should be cleaned up at some point, but
at least it's well isolated.
at this point not all functionality is complete. the dynamic linker
itself, and main app if it is also loaded by the kernel, take
advantage of fdpic and do not need constant displacement between
segments, but additional libraries loaded by the dynamic linker follow
normal ELF semantics for mapping still. this fully works, but does not
admit shared text on nommu.
in terms of actual functional correctness, dlsym's results are
presently incorrect for function symbols, RTLD_NEXT fails to identify
the caller correctly, and dladdr fails almost entirely.
with the dynamic linker entry point working, support for static pie is
automatically included, but linking the main application as ET_DYN
(pie) probably does not make sense for fdpic anyway. ET_EXEC is
equally relocatable but more efficient at representing relocations.
previously, the call into stage 2 was made by looking up the symbol
name "__dls2" (which was chosen short to be easy to look up) from the
dynamic symbol table. this was no problem for the dynamic linker,
since it always exports all its symbols. in the case of the static pie
entry point, however, the dynamic symbol table does not contain the
necessary symbol unless -rdynamic/-E was used when linking. this
linking requirement is a major obstacle both to practical use of
static-pie as a nommu binary format (since it greatly enlarges the
file) and to upstream toolchain support for static-pie (adding -E to
default linking specs is not reasonable).
this patch replaces the runtime symbolic lookup with a link-time
lookup via an inline asm fragment, which reloc.h is responsible for
providing. in this initial commit, the asm is provided only for i386,
and the old lookup code is left in place as a fallback for archs that
have not yet transitioned.
modifying crt_arch.h to pass the stage-2 function pointer as an
argument was considered as an alternative, but such an approach would
not be compatible with fdpic, where it's impossible to compute
function pointers without already having performed relocations. it was
also deemed desirable to keep crt_arch.h as simple/minimal as
possible.
in principle, archs with pc-relative or got-relative addressing of
static variables could instead load the stage-2 function pointer from
a static volatile object. that does not work for fdpic, and is not
safe against reordering on mips-like archs that use got slots even for
static functions, but it's a valid on i386 and many others, and could
provide a reasonable default implementation in the future.
with this commit it should be possible to produce a working
static-linked fdpic libc and application binaries for sh.
the changes in reloc.h are largely unused at this point since dynamic
linking is not supported, but the CRTJMP macro is used one place
outside of dynamic linking, in __unmapself.
this version of the entry point is only suitable for static linking in
ET_EXEC form. neither dynamic linking nor pie is supported yet. at
some point in the future the fdpic and non-fdpic versions of this code
may be unified but for now it's easiest to work with them separately.
clone calls back to a function pointer provided by the caller, which
will actually be a pointer to a function descriptor on fdpic. the
obvious solution is to have a separate version of clone for fdpic, but
I have taken a simpler approach to go around the problem. instead of
calling the pointed-to function from asm, a direct call is made to an
internal C function which then calls the pointed-to function. this
lets the C compiler generate the appropriate calling convention for an
indirect call with no need for ABI-specific assembly.
this error was only found by reading the code, but it seems to have
been causing gcc to produce wrong code in malloc: the same register
was used for the output and the high word of the input. in principle
this could have caused an infinite loop searching for an available
bin, but in practice most x86 models seem to implement the "undefined"
result of the bsf instruction as "unchanged".
these functions are part of the ARM EABI, meaning compilers may
generate references to them. known versions of gcc do not use them,
but llvm does. they are not provided by libgcc, and the de facto
standard seems to be that libc provides them.
commit 3c43c0761e fixed missing
synchronization in the atomic store operation for i386 and x86_64, but
opted to use mfence for the barrier on x86_64 where it's always
available. however, in practice mfence is significantly slower than
the barrier approach used on i386 (a nop-like lock orl operation).
this commit changes x86_64 (and x32) to use the faster barrier.
On 32bit systems long long arguments are passed in a special way
to some syscalls; this accidentally got copied to the AArch64 port.
The following interfaces were broken: fallocate, fanotify, ftruncate,
posix_fadvise, posix_fallocate, pread, pwrite, readahead,
sync_file_range, truncate.
despite being strongly ordered, the x86 memory model does not preclude
reordering of loads across earlier stores. while a plain store
suffices as a release barrier, we actually need a full barrier, since
users of a_store subsequently load a waiter count to determine whether
to issue a futex wait, and using a stale count will result in soft
(fail-to-wake) deadlocks. these deadlocks were observed in malloc and
possible with stdio locks and other libc-internal locking.
on i386, an atomic operation on the caller's stack is used as the
barrier rather than performing the store itself using xchg; this
avoids the need to read the cache line on which the store is being
performed. mfence is used on x86_64 where it's always available, and
could be used on i386 with the appropriate cpu model checks if it's
shown to perform better.
the TLS ABI spec for mips, powerpc, and some other (presently
unsupported) RISC archs has the return value of __tls_get_addr offset
by +0x8000 and the result of DTPOFF relocations offset by -0x8000. I
had previously assumed this part of the ABI was actually just an
implementation detail, since the adjustments cancel out. however, when
the local dynamic model is used for accessing TLS that's known to be
in the same DSO, either of the following may happen:
1. the -0x8000 offset may already be applied to the argument structure
passed to __tls_get_addr at ld time, without any opportunity for
runtime relocations.
2. __tls_get_addr may be used with a zero offset argument to obtain a
base address for the module's TLS, to which the caller then applies
immediate offsets for individual objects accessed using the local
dynamic model. since the immediate offsets have the -0x8000 adjustment
applied to them, the base address they use needs to include the
+0x8000 offset.
it would be possible, but more complex, to store the pointers in the
dtv[] array with the +0x8000 offset pre-applied, to avoid the runtime
cost of adding 0x8000 on each call to __tls_get_addr. this change
could be made later if measurements show that it would help.
nominally the low bits of the trap number on sh are the number of
syscall arguments, but they have never been used by the kernel, and
some code making syscalls does not even know the number of arguments
and needs to pass an arbitrary high number anyway.
sh3/sh4 traditionally used the trap range 16-31 for syscalls, but part
of this range overlapped with hardware exceptions/interrupts on sh2
hardware, so an incompatible range 32-47 was chosen for sh2.
using trap number 31 everywhere, since it's in the existing sh3/sh4
range and does not conflict with sh2 hardware, is a proposed
unification of the kernel syscall convention that will allow binaries
to be shared between sh2 and sh3/sh4. if this is not accepted into the
kernel, we can refit the sh2 target with runtime selection mechanisms
for the trap number, but doing so would be invasive and would entail
non-trivial overhead.
due to the way the interrupt and syscall trap mechanism works,
userspace on sh2 must never set the stack pointer to an invalid value.
thus, the approach used on most archs, where __unmapself executes with
no stack for the interval between SYS_munmap and SYS_exit, is not
viable on sh2.
in order not to pessimize sh3/sh4, the sh asm version of __unmapself
is not removed. instead it's renamed and redirected through code that
calls either the generic (safe) __unmapself or the sh3/sh4 asm,
depending on compile-time and run-time conditions.
the sh2 target is being considered an ISA subset of sh3/sh4, in the
sense that binaries built for sh2 are intended to be usable on later
cpu models/kernels with mmu support. so rather than hard-coding
sh2-specific atomics, the runtime atomic selection mechanisms that was
already in place has been extended to add sh2 atomics.
at this time, the sh2 atomics are not SMP-compatible; since the ISA
lacks actual atomic operations, the new code instead masks interrupts
for the duration of the atomic operation, producing an atomic result
on single-core. this is only possible because the kernel/hardware does
not impose protections against userspace doing so. additional changes
will be needed to support future SMP systems.
care has been taken to avoid producing significant additional code
size in the case where it's known at compile-time that the target is
not sh2 and does not need sh2-specific code.
the instruction used to align the stack, "and $sp, $sp, -8", does not
actually exist; it's expanded to 2 instructions using the 'at'
(assembler temporary) register, and thus cannot be used in a branch
delay slot. since alignment mod 16 commutes with subtracting 8, simply
swapping these two operations fixes the problem.
crt1.o was not affected because it's still being generated from a
dedicated asm source file. dlstart.lo was not affected because the
stack pointer it receives is already aligned by the kernel. but
Scrt1.o was affected in cases where the dynamic linker gave it a
misaligned stack pointer.
Rich Felker
Felix Janda
Timo Teräs
Szabolcs Nagy
Roman Yeryomin
Felix Fietkau