| Age | Commit message (Collapse) | Author |
|---|
|
|
|
Signed-off-by: Dmitry Chestnykh <dm.chestnykh@gmail.com>
|
|
Add acquire/release variants for atomic functions cmpxchg/xchg and
provide a memory barrier after/before exchange. For cmpxchg use compiler
builtins. For xchg functions add memory barrier explicitly.
These barriers are required to keep memory consistency of ARCv3 CPU
cores in SMP.
For ARC700 barriers are not required and the compiler doesn't provide
_atomic_compare_exchange*, use current asm insertion without
acquire/release variants for ARC700.
Signed-off-by: Pavel Kozlov <pavel.kozlov@synopsys.com>
|
|
Add ability to use optimized versions of string functions for ARCv3 32-bit
CPUs with UCLIBC_HAS_STRING_ARCH_OPT option. Add optimized
memcpy/memset/memcmp code for ARCv3 CPUs based on the code from newlib
and adapt for ARCv3 existed optimized strchr/strcmp/strcpy/strlen.
Link to the Synopsys newlib repo with code for ARCv3 on GitHub:
https://github.com/foss-for-synopsys-dwc-arc-processors/newlib
Signed-off-by: Pavel Kozlov <pavel.kozlov@synopsys.com>
|
|
New ARCv3 ISA includes both 64-bit and 32-bit CPU family.
This patch adds support for 32-bit ARCv3 HS5x processors.
Signed-off-by: Sergey Matyukevich <sergey.matyukevich@synopsys.com>
Signed-off-by: Pavel Kozlov <pavel.kozlov@synopsys.com>
|
|
Add a header file with assembler macros to be able to handle in one
place the differences between ARCv2 and ARCv3 ISAs. It is a preparatory
step before the introduction of support for ARCv3 CPUs.
Signed-off-by: Sergey Matyukevich <sergey.matyukevich@synopsys.com>
Signed-off-by: Pavel Kozlov <pavel.kozlov@synopsys.com>
|
|
When MD_FALLBACK_FRAME_STATE_FOR handler in libgcc unwind code checks
the second instruction opcode in __default_rt_sa_restorer function,
it expects to see the following values for ARC cores:
- 0x7ee0781e for ARCv2 LE
- 0x003f226f for ARC700 LE
ARC700 value correspond to trap0 instruction. ARCv2 value corresponds
to the following code:
traps_0
j_s [blink]
However, unlike glibc, uClibc implementation of __default_rt_sa_restorer
for ARC does not have that jump. Hence libgcc unwind code is not able
to recognize signal frame correctly on ARCv2 and completes too early.
This change fixes libgcc unwinding over signal frame on ARCv2 adding
missing jump to __default_rt_sa_restorer.
Signed-off-by: Sergey Matyukevich <sergey.matyukevich@synopsys.com>
|
|
Since Linux 3.11, O_TMPFILE allows to create unnamed files that can be
linked later on. It is internally defined as (O_TMPFILE | O_DIRECTORY)
to make it fail on old kernels.
Copying definitions from glibc for O_TMPFILE is not enough to support
O_TMPFILE; The open() wrapper also need to pass the mode when the flag
contains O_TMPFILE, otherwise, it will pass mode 000 which will succeed
but yield unexpected results.
openat() is curiously not affected since it passes the mode
unconditionally..
Signed-off-by: Nicolas Cavallari <nicolas.cavallari@green-communications.fr>
|
|
The F_DUPFD_CLOEXEC flag was added in POSIX 2008.09.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
|
|
Similar to glibc commit
https://sourceware.org/git/?p=glibc.git;a=commit;h=6bbfc5c09fc5b5e3d4a0cddbbd4e2e457767dae7
we need to handle Linux kernel change, which removed stat64 family from default syscall set.
Signed-off-by: Waldemar Brodkorb <wbx@openadk.org>
Signed-off-by: Waldemar Brodkorb <wbrodkorb@conet.de>
|
|
The return type of syscall() is long so __syscall_error, which is jumped
to by syscall handlers to stash an error number into errno, must return
long too otherwhise it returs 4294967295L instead of -1L. For example,
syscall for x86_64 is defined in libc/sysdeps/linux/x86_64/syscall.S as
syscall:
movq %rdi, %rax /* Syscall number -> rax. */
movq %rsi, %rdi /* shift arg1 - arg5. */
movq %rdx, %rsi
movq %rcx, %rdx
movq %r8, %r10
movq %r9, %r8
movq 8(%rsp),%r9 /* arg6 is on the stack. */
syscall /* Do the system call. */
cmpq $-4095, %rax /* Check %rax for error. */
jae __syscall_error /* Branch forward if it failed. */
ret /* Return to caller. */
In libc/sysdeps/linux/x86_64/__syscall_error.c, __syscall_error is
defined as
int __syscall_error(void) attribute_hidden;
int __syscall_error(void)
{
register int err_no __asm__ ("%rcx");
__asm__ ("mov %rax, %rcx\n\t"
"neg %rcx");
__set_errno(err_no);
return -1;
}
So __syscall_error returns -1 as a 32-bit int in a 64-bit register, %rax
(0x00000000ffffffff, whose decimal value is decimal 4294967295) and a
test like this always returns false:
if (syscall(number, ...) == -1)
foo();
Fix the error by making __syscall_error return a long, like syscall().
The problem can be circumvented by the caller by coercing the returned
value to int before comparing it to -1:
if ((int) syscall(number, ...) == -1)
foo();
The same problem probably occurs on other 64-bit systems but so far only
x86_64 was tested, so this change must be considered experimental.
Signed-off-by: Carlos Santos <unixmania@gmail.com>
|
|
An issue has been found with current implementation of signal restorer
function in uClibc and how GDB handles it. When debugger information is not
present, everything worked fine, because GDB would use a built-in logic to
determine if function is a signal restorer. However when debugging information
is present, debugger would rely solely on it and wouldn't use ARC-specific
functions to detect signal handler frames. Because debug information for signal restorer is generated
completely by the compiler, it lacks a marker, that identifies this as a
signal frame that requires special handling. While it is possible to insert
that marker via inline assembly, that still doesn't solve the whole problem,
because some other expectations are not met by the debug information - there
is no "nop" in front of the function, needed to fool debugger into thinking
that this was a function call, and references to previous frame information
need to be described manually. The simplest way to fix the problem is just
to make sure that signal restorer function will not have any debug function
at all, which can be done by writing it in assembly.
Alternative, more complex, solution, where debug information for signal
frame is manually defined can be found in
glibc/sysdeps/unix/sysv/linux/x86_64/sigaction.c [1].
[1] https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/x86_64/sigaction.c;hb=HEAD
Signed-off-by: Anton Kolesov <Anton.Kolesov@synopsys.com>
|
|
Sync with GNU C library. Found while trying to compile
linux-rdma to uClibc-ng.
Reported-by: Alexey Brodkin <Alexey.Brodkin@synopsys.com>
|
|
Follow GNU C Library from c579f48edba88380635ab98cb612030e3ed8691e
and remove the PID caching. These simplifies the architecture specific
assembly code.
The run of the test suite found no regressions, it even solves
some of the test failures for x86/x86_64/sparc.
Reviewed-by: Max Filippov <jcmvbkbc@gmail.com>
Reviewed-by: Vineet Gupta <Vineet.Gupta1@synopsys.com>
Acked-by: Matthew Fortune <Matthew.Fortune@imgtec.com>
Acked-by: Alexey Brodkin <Alexey.Brodkin@synopsys.com>
|
|
For hardware configurations lacking LLOCK/SCOND (say ARC750),
use a syscall to atomically do the cmpxchg.
This is costly and painful, but really the only way out.
Note that kenrel only guarantees this to work in a UP configuraion
Reported-by: Avinash Patil <avinashp@quantenna.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
ARC EX instruction maps directly to this primitive, thus helps elide the
llock/scond based retry loop where possible.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
Cc: Anton Kolesov <akolesov@synopsys.com>
Cc: linux-snps-arc@lists.infradead.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
The syscall ABI includes the gcc functional calling ABI since a syscall
implies userland caller and kernel callee.
The current gcc ABI (v3) for ARCv2 ISA required 64-bit data be passed in
even-odd register pairs, (potentially punching reg holes when passing such
values as args). This was partly driven by the fact that the double-word
LDD/STD instructions in ARCv2 expect the register alignment and thus gcc
forcing this avoids extra MOV at the cost of a few unused register (which we
have plenty anyways).
This however was rejected as part of upstreaming gcc port to HS. So the new
ABI v4 doesn't enforce the even-odd reg restriction.
Do note that for ARCompact ISA builds v3 and v4 are practically the same in
terms of gcc code generation.
This change is dormant for now (gcc 4.8.x based tools) and will only kick
in with switch to gcc 6.x based tools.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
crt1.S needed to use a got relative reference.
Libraries like pthreads define a _init and unless it is GOT, the
linker will fail because it will try to create a dynamic reloc on
.text section, more precisely on __start.
Signed-off-by: Cupertino Miranda <cmiranda@synopsys.com>
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: Vlad Zakharov <vzakhar@synopsys.com>
Cc: Vineet Gupta <vgupta@synopsys.com>
|
|
This change was inspired by similar change in glibc:
https://sourceware.org/git/?p=glibc.git;a=commit;h=0cb313f7cb0e418b3d56f3a2ac69790522ab825d
Current Linux kernel requires CLONE_VM to be set with CLONE_THREAD
otherwise returning -EINVAL, see man clone2. This means we don't need
to check for both CLONE_THREAD and CLONE_VM instead we may simplify
code a lot and just check 1 bit (CLONE_THREAD).
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
|
|
For thread group case (CLONE_THREAD), the cached PID of new process/thread
need not be reset. The old logic to decide that was flawed as it would be
true only for exact combination of CLONE_THREAD + _VM, but would fail for
CLONE_THREAD + _VM + _xyz.
More detailed tear-down of current and new code below.
Current implementation is:
--------------------->8--------------------
; r12 contains clone flags
mov_s r2, CLONE_THREAD_N_VM; r2 contains bit mask
and_s r2, r2, r12 ; r2 contains bit mask AND clone flags
; but r12 still contains the same flags
brne r2, r12, .Lgo_thread ; here we compare modified mask with
; flags as they were and skip pthread TID/PID
; setup if r2 != r12 which happens all
; the time except clone flags were
; exactly CLONE_THREAD | CLONE_VM
--------------------->8--------------------
New implementation is:
--------------------->8--------------------
; r12 contains clone flags
mov_s r2, CLONE_THREAD_N_VM; r2 contains bit mask
and_s r12, r12, r2 ; r12 contains clone flags AND bit mask
; i.e. we did mask all flags except
; CLONE_THREAD and CLONE_VM
breq r2, r12, .Lgo_thread ; here we compare masked flags with
; target mask and if they match we skip
; pthread TID/PID setup
--------------------->8--------------------
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
|
|
Caught by tst-getpid1 test from uClibc's test-suite.
It looks like original implementation was not correct.
The code in question is supposed to recover PID of the new
thread. And by no means that could happen with clone() syscall
while getpid() does exactly this.
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Acked-by: Vineet Gupta <vgupta@synopsys.com>
Reported-by: Eugeniy Paltsev <paltsev@synopsys.com>
|
|
Reviewed-by: Max Filippov <jcmvbkbc@gmail.com>
|
|
As this is only implemented for a few architecture and not well
tested, just remove it.
Reviewed-by: Max Filippov <jcmvbkbc@gmail.com>
|
|
Currently crt1 takes address of functions (main,_init,_fini) directly
which doesn't generate truely position independent code, but zero based
values instead. e.g.
| __start:
| ...
| add_s r2,sp,0x4
| mov_s r0, main
generates to
| 000156ec <__start>:
| ...
| 156f4: add_s r2,sp,0x4
| 156f6: mov_s r0,0x15f7c
| ...
| 00015f7c <main>:
| 15f7c: push_s blink
This works just fine for the normal (non PIE) dynamic executables since
they are loaded at address 0. However this is not true for PIE
executables. So for Scrt1 we use a true position independent way when
taking function addresses.
Cc: uclibc@uclibc.org <uclibc@uclibc.org>
Cc: devel@uclibc-ng.org <devel@uclibc-ng.org>
Cc: Cupertino Miranda <cmiranda@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
Linux Standard Base specifies section .note.ABI-tag that can be considered
as a marker for ELF files targeted to Linux systems. See
https://refspecs.linuxfoundation.org/LSB_1.2.0/gLSB/noteabitag.html
This section, for example, is used by the GDB to identify Linux ELFs as
compared to baremetal ELFs that do not have this section.
Signed-off-by: Anton Kolesov <Anton.Kolesov@synopsys.com>
Cc: Vineet Gupta <Vineet.Gupta1@synopsys.com>
|
|
Only alpha, hppa and sparc need non-default value.
|
|
.globl can be used for every architecture so remove the define.
Sync with GNU C library.
|
|
It's even no longer required for non-ported ppc64
architecture. Sync with GNU C library.
This simplify the macros in include/libc-symbols.h.
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
|
|
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
[updated changelog]
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
when building uClibc with -O0 (DODEBUG build) the default sigrestorer
had some extra glue code generated for stack manipulation which was
messing up resume from signal path.
So annotate the function with -Os so that gcc would only generate the
bare min 2 instruction TRAP sequence
Reported-and-Debugged-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
This is next gen Instruction Set Architecture from Synopsys and basis
for the ARC HS family of processors.
http://www.synopsys.com/dw/ipdir.php?ds=arc-hs38-processor&elq_mid=5732&elq_cid=458802
http://www.synopsys.com/IP/ProcessorIP/ARCProcessors/arc-hs/Pages/default.aspx
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Not relevant anymore since commit e8cc14e59ed3f66b84e,
"libc: rename TRUNCATE64_HAS_4_ARGS to SYSCALL_ALIGN_64BIT"
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
ARC CPU may have MMU page size of 4/8(default)/16k.
uClibc needs to have page size configured accodring to HW it will be run on.
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Khem Raj <raj.khem@gmail.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Typically kernel would set the TP register of newly created thread in
clone syscall with CLONE_SETTLS.
However this implies that:
* kernel knows of the exact TP register (which is already different for
ARCompact and ARCv2) and is a detail which becomes part of the ABI
* kernel also needs to handle the unlikely but possible case of !TLS
userspace (e.g. Android libc) case where TP reg is not reserved and
thus need not be mucked with (using ELF personality based detection)
It is better to confine this detail to userspace runtime and set TP reg
right after clone syscall returns.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Native GDB requires elf_fpregset_t to be defined, even if floating point
registers are absent. In this case we use empty structure. Additionally we
redefined prfpregset_t to use this new type.
And there has been an error: it was defined that user_regs_struct has 20
registers, while it has 40.
Signed-off-by: Anton Kolesov <Anton.Kolesov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
This showed up due to longstanding test/unistd/clone failure where post
clone, the callback was not exiting and rather falling thru in into main
program.
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
- Use syscalls-common.h vers of INTERNAL_SYSCALL / INLINE_SYSCALL_NOERR
- INLINE_SYSCALL takes syscall name (e.g. write)
- {INLINE,INTERNAL}_SYSCALL_NCS macros take syscall num (__NR_write)
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
Anton reported:
------------------->8---------------------
Package iperf cannot be built in Buildroot due to following error:
/home/akolesov/env/autobuild/tmp/host/usr/bin/arc-buildroot-linux-uclibc-g++
-Wall -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64
-pipe -Os --static -o iperf -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE
-D_FILE_OFFSET_BITS=64 -pipe -Os -pthread -DHAVE_CONFIG_H Client.o
Extractor.o Launch.o List.o Listener.o Locale.o PerfSocket.o ReportCSV.o
ReportDefault.o Reporter.o Server.o Settings.o SocketAddr.o gnu_getopt.o
gnu_getopt_long.o main.o service.o sockets.o stdio.o tcp_window_size.o
../compat/libcompat.a
/home/akolesov/env/autobuild/tmp/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/libc.a(vfork.os):
In function `vfork':
(.text+0x0): multiple definition of `vfork'
/home/akolesov/env/autobuild/tmp/host/usr/arc-buildroot-linux-uclibc/sysroot/usr/lib/libpthread.a(ptfork.os):ptfork.c:(.text+0xc0):
first defined here
collect2: error: ld returned 1 exit status
------------------->8---------------------
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Cc: Anton Kolesov <Anton.Kolesov@synopsys.com>
Cc: Francois Bedard <Francois.Bedard@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
crt1.S -> include sysdep.h -> include sysnum.h
In multistage gcc build, Buildroot happens to build crt1.S before
sysnum.h is generated. So break the above include chain and opencode the
ENTRY/END macros.
Reported-by: Anton Kolesov <akolesov@synopsys.com>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
For this port, I would like to give due credit to:
- Folks from Codito technologies (Sameer, Amit, Kanika, Ramana,...)
who did the very first port
- ARC UK from 2007-2009 (Joern, Irfan, Khurram, Phil...
- Late Brendan Kehoe (may he RIP)
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
