On systems without enough random-access memory, stdlib/test-bz22786
will go deeply into swap and time out, even with a substantial
TIMEOUTFACTOR. This commit adds a facility to construct repeating
strings with alias mappings, so that the requirement for physical
memory, and uses it in stdlib/test-bz22786.
Mark the ra register as undefined in _start, so that unwinding through
main works correctly. Also, don't use a tail call so that ra points after
the call to __libc_start_main, not after the previous call.
Now that we require Python 3.4 or later, Python code creating
temporary directories can use tempfile.TemporaryDirectory in "with" to
have the directory deleted automatically instead of needing to use
try/finally to handle removing a directory created with
tempfile.mkdtemp. This patch does so in conform/glibcconform.py.
Tested for x86_64.
* conform/glibcconform.py: Do not import shutil.
(list_exported_functions): Use tempfile.TemporaryDirectory instead
of mkdtemp.
This patch makes Python 3.4 or later a required tool for building
glibc, so allowing changes of awk, perl etc. code used in the build
and test to Python code without any such changes needing makefile
conditionals or to handle older Python versions.
This patch makes the configure test for Python check the version and
give an error if Python is missing or too old, and removes makefile
conditionals that are no longer needed. It does not itself convert
any code from another language to Python, and does not remove any
compatibility with older Python versions from existing scripts.
Tested for x86_64.
* configure.ac (PYTHON_PROG): Use AC_CHECK_PROG_VER. Set
critic_missing for versions before 3.4.
* configure: Regenerated.
* manual/install.texi (Tools for Compilation): Document
requirement for Python to build glibc.
* INSTALL: Regenerated.
* Rules [PYTHON]: Make code unconditional.
* benchtests/Makefile [PYTHON]: Likewise.
* conform/Makefile [PYTHON]: Likewise.
* manual/Makefile [PYTHON]: Likewise.
* math/Makefile [PYTHON]: Likewise.
* sysdeps/mach/hurd/i386/intr-msg.h (INTR_MSG_TRAP): Make
_hurd_intr_rpc_msg_about_to global point to start of controlled
assembly snippet. Make it check canceled flag.
* hurd/hurdsig.c (_hurdsig_abort_rpcs): Only mutate thread if it passed
the _hurd_intr_rpc_msg_about_to point.
* hurd/intr-msg.c (_hurd_intr_rpc_mach_msg): Remove comment on mutation
issue, remove cancel flag check.
* sysdeps/mach/hurd/i386/intr-msg.h (INTR_MSG_TRAP): Set
intr_port to 0 when canceled.
* hurd/intr-msg.c (_hurd_intr_rpc_mach_msg): Pass intr_port
address.
* sysdeps/mach/hurd/i386/intr-msg.h (INTR_MSG_TRAP): Make
_hurd_intr_rpc_msg_about_to global point to start of controlled
assembly snippet. Make it check canceled flag again.
* hurd/hurdsig.c (_hurdsig_abort_rpcs): Only mutate thread if it passed the
_hurd_intr_rpc_msg_about_to point.
* hurd/intr-msg.c (_hurd_intr_rpc_mach_msg): Remove comment on mutation
issue.
since we do not actually know whether the RPC was completed or not,
which makes a huge difference for e.g. write(), so better really error
out than letting caller think that the RPC did not happen.
* hurd/intr-msg.c (_hurd_intr_rpc_mach_msg): When the server does not
answer to interrupt_operation, return EIO instead of EINTR.
Seeing a server not able to get interrupted for 3s is not so surprising when
e.g. a lot of writes are happening. 1 minute allows to actually notice the
issue and be able to debug it.
* hurd/hurdsig.c (_hurd_interrupted_rpc_timeout): Set to 60000.
Since we have consensus on requiring Python 3.4 or later to build
glibc, it follows that compatibility with older Python versions is
also no longer relevant to auxiliary Python scripts for use in glibc
development. This patch removes such compatibility code from
build-many-glibcs.py (compatibility code needed for 3.4, which lacks
the newer subprocess interface, is kept). Because
build-many-glibcs.py is not itself called from the glibc build system,
this patch is independent of the configure checks for having a
new-enough Python version, which are only relevant to uses of Python
from the main build and test process.
Tested with build-many-glibcs.py building glibc for aarch64-linux-gnu
(with Python 3.4 to make sure that still works).
* scripts/build-many-glibcs.py: Remove compatibility for missing
os.cpu_count and re.fullmatch.
When new symbol versions were introduced without SVID compatible
error handling the exp2f, log2f and powf symbols were accidentally
removed from the ia64 lim.a. The regression was introduced by
the commits
f5f0f52651
New expf and exp2f version without SVID compat wrapper
72d3d28108
New symbol version for logf, log2f and powf without SVID compat
With WEAK_LIBM_ENTRY(foo), there is a hidden __foo and weak foo
symbol definition in both SHARED and !SHARED build.
[BZ #23822]
* sysdeps/ia64/fpu/e_exp2f.S (exp2f): Use WEAK_LIBM_ENTRY.
* sysdeps/ia64/fpu/e_log2f.S (log2f): Likewise.
* sysdeps/ia64/fpu/e_exp2f.S (powf): Likewise.
This patch adds the IN_MASK_CREATE macro from Linux 4.19 to
sys/inotify.h.
Tested for x86_64.
* sysdeps/unix/sysv/linux/sys/inotify.h (IN_MASK_CREATE): New
macro.
This patch adds NT_MIPS_DSP and NT_MIPS_FP_MODE from Linux 4.19 to
elf.h.
Tested for x86_64.
* elf/elf.h (NT_MIPS_DSP): New macro.
(NT_MIPS_FP_MODE): Likewise.
This patch extends gen-libm-test.py to generate the ulps table for the
manual, so meaning there is only a single ulps file parser needed and
another Perl script is eliminated. As with the introduction of
gen-libm-test.py, this is designed to generate exactly the same
libm-err.texi as libm-err-tab.pl did. (gen-libm-test.py is still
shorter in lines than the old gen-libm-test.pl even after this patch.)
Note that this introduces a Python dependency for building the manual,
which is thus noted in install.texi and NEWS.
Tested building html / info / pdf versions of the manual.
* math/gen-libm-test.py: Import os.
(ALL_FLOATS_MANUAL): New constant.
(ALL_FLOATS_SUFFIX): Likewise.
(Ulps.all_functions): New function.
(real_all_ulps): Likewise.
(generate_err_table_sub): Likewise.
(generate_err_table): Likewise.
(main): Handle -s and -m options.
* manual/libm-err-tab.pl: Remove.
* manual/Makefile ($(objpfx)stamp-libm-err): Use gen-libm-test.py
instead of libm-err-tab.pl.
[$(PERL) != no]: Change condition to [$(if $(PYTHON),$(PERL),no)
!= no].
* manual/install.texi (Tools for Compilation): Document
requirement for Python to build manual.
* INSTALL: Regenerated.
glibc support for 64-bit time_t on 32-bit architectures
will involve:
- Using 64-bit times inside glibc, with conversions
to and from 32-bit times taking place as necessary
for interfaces using such times.
- Adding 64-bit-time support in the glibc public API.
This support should be dynamic, i.e. glibc should
provide both 32-bit and 64-bit implementations and
let user code choose at compile time whether to use
the 32-bit or 64-bit interfaces.
This requires a glibc-internal name for a type for times
that are always 64-bit.
Based on __TIMESIZE, a new macro is defined, __TIME64_T_TYPE,
which is always the right __*_T_TYPE to hold a 64-bit-time.
__TIME64_T_TYPE equals __TIME_T_TYPE if __TIMESIZE equals 64
and equals __SQUAD_T_TYPE otherwise.
__time64_t can then replace uses of internal_time_t.
This patch was tested by running 'make check' on branch
master then applying this patch and its predecessor and
running 'make check' again, and checking that both 'make
check' yield identical results. This was done on
x86_64-linux-gnu and i686-linux-gnu.
* bits/time64.h: New file.
* include/time.h: Replace internal_time_t with __time64_t.
* posix/bits/types (__time64_t): Add.
* stdlib/Makefile: Add bits/time64.h to includes.
* time/tzfile.c: Replace internal_time_t with __time64_t.
To determine whether the default time_t interfaces are 32-bit
and so need conversions, or are 64-bit and so are compatible
with the internal 64-bit type without conversions, a macro
giving the size of the default time_t is also required.
This macro is called __TIMESIZE.
This macro can then be used instead of __WORDSIZE in msq-pad.h
and shm-pad.h files, which in turn allows removing their x86
variants, and in sem-pad.h files but keeping the x86 variant.
This patch was tested by running 'make check' on branch master
then applying this patch and running 'make check' again, and
checking that both 'make check' yield identical results.
This was done on x86_64-linux-gnu and i686-linux-gnu.
* bits/timesize.h: New file.
* stdlib/Makefile (headers): Add bits/timesize.h.
* sysdeps/unix/sysv/linux/bits/msq-pad.h
(__MSQ_PAD_AFTER_TIME): Use __TIMESIZE instead of __WORDSIZE.
* sysdeps/unix/sysv/linux/bits/sem-pad.h
(__SEM_PAD_AFTER_TIME): Likewise.
* sysdeps/unix/sysv/linux/bits/shm-pad.h
(__SHM_PAD_AFTER_TIME): Likewise.
* sysdeps/unix/sysv/linux/hppa/bits/msq-pad.h
(__MSQ_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/hppa/bits/sem-pad.h
(__SEM_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/hppa/bits/shm-pad.h
(__SHM_PAD_BEFORE_TIME, __SHM_PAD_BETWEEN_TIME_AND_SEGSZ): Likewise.
* sysdeps/unix/sysv/linux/mips/bits/msq-pad.h
(__MSQ_PAD_AFTER_TIME, __MSQ_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/msq-pad.h
(__MSQ_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/sem-pad.h
(__SEM_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/shm-pad.h
(__SHM_PAD_BEFORE_TIME, __SHM_PAD_BETWEEN_TIME_AND_SEGSZ): Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/msq-pad.h
(__MSQ_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/sem-pad.h
(__SEM_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/shm-pad.h
(__SHM_PAD_BEFORE_TIME): Likewise.
* sysdeps/unix/sysv/linux/x86/bits/msq-pad.h: Delete file.
* sysdeps/unix/sysv/linux/x86/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/x86/bits/timesize.h: New file.
RDTSCP waits until all previous instructions have executed and all
previous loads are globally visible before reading the counter. RDTSC
doesn't wait until all previous instructions have been executed before
reading the counter. All x86 processors since 2010 support RDTSCP
instruction. This patch adds RDTSCP support to benchtests.
* benchtests/Makefile (CPPFLAGS-nonlib): Add -DUSE_RDTSCP if
USE_RDTSCP is defined.
* sysdeps/x86/hp-timing.h (HP_TIMING_NOW): Use RDTSCP if
USE_RDTSCP is defined.
Commit 7a16bdbb9f uses IOV_MAX, which is not defined on hurd.
Checked on a build for i686-gnu.
* misc/tst-preadvwritev2-common.c (IOV_MAX): Define if not
defined.
Th commit 'Disable TSX on some Haswell processors.' (2702856bf4) changed the
default flags for Haswell models. Previously, new models were handled by the
default switch path, which assumed a Core i3/i5/i7 if AVX is available. After
the patch, Haswell models (0x3f, 0x3c, 0x45, 0x46) do not set the flags
Fast_Rep_String, Fast_Unaligned_Load, Fast_Unaligned_Copy, and
Prefer_PMINUB_for_stringop (only the TSX one).
This patch fixes it by disentangle the TSX flag handling from the memory
optimization ones. The strstr case cited on patch now selects the
__strstr_sse2_unaligned as expected for the Haswell cpu.
Checked on x86_64-linux-gnu.
[BZ #23709]
* sysdeps/x86/cpu-features.c (init_cpu_features): Set TSX bits
independently of other flags.
Linux 4.19 does not add any new syscalls (some existing ones are added
to more architectures); this patch updates the version number in
syscall-names.list to reflect that it's still current for 4.19.
Tested with build-many-glibcs.py.
* sysdeps/unix/sysv/linux/syscall-names.list: Update kernel
version to 4.19.
glibc does:
/* There should be no difference between the UTF-32 handling required
by c32rtomb and the wchar_t handling which has long since been
implemented in wcrtomb. */
weak_alias (__wcrtomb, c32rtomb)
/* There should be no difference between the UTF-32 handling required
by mbrtoc32 and the wchar_t handling which has long since been
implemented in mbrtowc. */
weak_alias (__mbrtowc, mbrtoc32)
The reasoning in those comments to justify those aliases is incorrect:
ISO C requires that, for the case of a NULL mbstate_t* being passed,
each function has its *own* internal static mbstate_t. Thus a program
must be able to use both wcrtomb and c32rtomb at the same time with
each keeping its own separate state, and likewise for mbrtowc and
mbrtoc32.
This patch duly sets up separarate char32_t function that wrap the
wchar_t ones. Note that the included test only covers the mbrtoc32 /
mbrtowc pair. While I think the change made is logically correct for
c32rtomb / wcrtomb as well, I'm not sure we have a locale with a
suitable state-dependent multibyte encoding for testing that part of
the change.
Tested for x86_64.
[BZ #23793]
* wcsmbs/c32rtomb.c: New file.
* wcsmbs/mbrtoc32.c: Likewise.
* wcsmbs/tst-c32-state.c: Likewise.
* wcsmbs/mbrtowc.c (mbrtoc32): Do not define as alias.
* wcsmbs/wcrtomb.c (c32rtomb): Likewise.
* wcsmbs/Makefile (routines): Add mbrtoc32 and c32rtomb.
(tests): Add tst-c32-state.
[$(run-built-tests) = yes] ($(objpfx)tst-c32-state.out): Depend on
$(gen-locales).
Use __builtin_ia32_rdtsc directly since including <x86intrin.h> makes
building glibc very slow. On Intel Core i5-6260U, this patch reduces
x86-64 build time from 8 minutes 33 seconds to 3 minutes 48 seconds
with "make -j4" and GCC 8.2.1.
* sysdeps/x86/hp-timing.h: Don't include <x86intrin.h>.
(HP_TIMING_NOW): Replace _rdtsc with __builtin_ia32_rdtsc.
The c16rtomb implementation has:
// XXX The ISO C 11 spec I have does not say anything about handling
// XXX surrogates in this interface.
The DR#488 resolution, as applied to C2X, requires surrogate pairs to
be handled here (so the first call returns 0 and stores the high
surrogate in the mbstate_t, while the second call combines the
surrogates, produces a multibyte character and returns the number of
bytes written). This patch implements that. (mbrtoc16 already
handled producing surrogates as output.)
Tested for x86_64.
[BZ #23794]
* wcsmbs/c16rtomb.c (c16rtomb): Save first character of surrogate
pair and return 0 in that case, and use saved character to
interpret following character.
* wcsmbs/tst-c16-surrogate.c: New file.
* wcsmbs/Makefile (tests): Add tst-c16-surrogate.c.
[$(run-built-tests) = yes] ($(objpfx)tst-c16-surrogate.out):
Depend on $(gen-locales)
Since _rdtsc intrinsic is supported in GCC 4.9, we can use it for
HP_TIMING_NOW. This patch
1. Create x86 hp-timing.h to replace i686 and x86_64 hp-timing.h.
2. Move MINIMUM_ISA from init-arch.h to isa.h so that x86 hp-timing.h
can check minimum x86 ISA to decide if _rdtsc can be used.
NB: Checking if __i686__ isn't sufficient since __i686__ may not be
defined when building for i686 class processors.
* sysdeps/i386/init-arch.h: Removed.
* sysdeps/i386/i586/init-arch.h: Likewise.
* sysdeps/i386/i686/init-arch.h: Likewise.
* sysdeps/i386/i686/hp-timing.h: Likewise.
* sysdeps/x86_64/hp-timing.h: Likewise.
* sysdeps/i386/isa.h: New file.
* sysdeps/i386/i586/isa.h: Likewise.
* sysdeps/i386/i686/isa.h: Likewise.
* sysdeps/x86_64/isa.h: Likewise.
* sysdeps/x86/hp-timing.h: New file.
* sysdeps/x86/init-arch.h: Include <isa.h>.
C99 wrongly specified a divide-by-zero exception for pow(+/- 0, -Inf);
C11 made it optional after this was pointed out, and the permission
for this exception has been removed in the current C2x draft. This
patch makes the glibc pow tests reflect the stricter requirement
(which follows the normal IEEE rules that a divide-by-zero exception
is for the case of exact infinite results from *finite* operands, not
for such results when any operand is infinite).
Tested for x86_64 and x86. (If any other pow implementation in glibc,
not exercised on those architectures, turns out to fail the stricter
test, it should be fixed to avoid the exception in this case.)
* math/libm-test-pow.inc (pow_test_data): Do not allow
divide-by-zero exception for pow(+/- 0, -Inf).
Job control was made mandatory in POSIX.1-2001: compare
<http://pubs.opengroup.org/onlinepubs/7990989775/xsh/unistd.h.html> with
<http://pubs.opengroup.org/onlinepubs/009695399/basedefs/unistd.h.html>.
Seventeen years later, we need not devote an entire manual @node to
warning people that this was once an optional POSIX feature.
* manual/job.texi (Job Control is Optional): Remove node, as
job control has not been optional in quite some time.
(Job Control): Mention briefly that systems older than
POSIX.1-2001 might not support job control.
* manual/conf.texi (_POSIX_JOB_CONTROL): Will always be
defined on systems conforming to POSIX.1-2001.
In iconv/gconv_conf.c, __gconv_get_path unnecessarily obtains a lock when
populating the array pointed to by __gconv_path_elem. The locking is not
necessary because all calls to __gconv_read_conf (which in turn calls
__gconv_get_path) are serialized using __libc_once.
This patch:
- removes all locking in __gconv_get_path;
- replaces all explicitly serialized __gconv_read_conf calls with calls to
__gconv_load_conf, a new wrapper that is serialized internally;
- adds a new test, iconv/tst-iconv_mt.c, to exercise iconv initialization,
usage, and cleanup in a multi-threaded program;
- indents __gconv_get_path correctly, removing tab characters (which makes
the patch look a little bigger than it really is).
After removing the unnecessary locking, it was confirmed that the test case
fails if the relevant __libc_once is removed. Additionally, four localedata
and iconvdata tests also fail. This gives confidence that the testsuite
sufficiently guards against some regressions relating to multi-threading
with iconv.
Tested on x86_64 and i686.
After my patch to move SHMLBA to its own header, the bits/shm.h
headers for architectures using the Linux kernel still vary in a few
ways: the use of __syscall_ulong_t; whether padding for 32-bit systems
is present before or after time fields, or missing altogether (mips,
x32); whether shm_segsz is before or after the time fields; whether,
if after time fields, there is extra padding before shm_segsz.
This patch arranges for a single header to be used. __syscall_ulong_t
is safe to use everywhere, while bits/shm-pad.h is added with new
macros __SHM_PAD_AFTER_TIME, __SHM_PAD_BEFORE_TIME,
__SHM_SEGSZ_AFTER_TIME and __SHM_PAD_BETWEEN_TIME_AND_SEGSZ to
describe the differences.
Tested for x86_64 and x86, and with build-many-glibcs.py.
* sysdeps/unix/sysv/linux/Makefile (sysdep_headers): Add
bits/shm-pad.h.
* sysdeps/unix/sysv/linux/bits/shm.h: Include <bits/shm-pad.h>.
(shmatt_t): Define as __syscall_ulong_t.
(__SHM_PAD_TIME): New macro, depending on [__SHM_PAD_BEFORE_TIME]
and [__SHM_PAD_AFTER_TIME].
(struct shmid_ds): Define time fields using __SHM_PAD_TIME.
Define shm_segsz and associated padding based on
[__SHM_SEGSZ_AFTER_TIME] and [__SHM_PAD_BETWEEN_TIME_AND_SEGSZ].
Use __syscall_ulong_t instead of unsigned long int.
[__USE_MISC] (struct shminfo): Use __syscall_ulong_t instead of
unsigned long int.
[__USE_MISC] (struct shm_info): Likewise.
* sysdeps/unix/sysv/linux/bits/shm-pad.h: New file.
* sysdeps/unix/sysv/linux/hppa/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/mips/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/x86/bits/shm-pad.h: Likewise.
* sysdeps/unix/sysv/linux/hppa/bits/shm.h: Remove.
* sysdeps/unix/sysv/linux/mips/bits/shm.h: Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/shm.h: Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/shm.h: Likewise.
* sysdeps/unix/sysv/linux/x86/bits/shm.h: Likewise.
One difference between bits/shm.h headers for architectures using the
Linux kernel is the definition of SHMLBA. This was noted in
<https://sourceware.org/ml/libc-alpha/2018-09/msg00175.html> as a
reason why even a new architecture (C-SKY) might need its own
bits/shm.h; thus, splitting it out of bits/shm.h can allow less
duplication of headers for new architectures.
This patch moves that definition to its own header, bits/shmlba.h, to
allow more sharing of headers between architectures. That move allows
the arm, ia64 and sh variants of bits/shm.h to be removed, as they had
no other significant differences from the generic bits/shm.h; powerpc
and x86 have their own bits/shm.h but do not need to get their own
bits/shmlba.h because they use the same SHMLBA as the generic header.
Other architectures with their own bits/shm.h get their own
bits/shmlba.h without being able to remove their own bits/shm.h until
the generic one has been adapted to be able to handle more
architectures (where, in addition to the differences seen for
bits/msq.h and bits/sem.h, the position of shm_segsz in struct
shmid_ds also depends on the architecture).
Tested for x86_64 and x86, and with build-many-glibcs.py.
* sysdeps/unix/sysv/linux/Makefile (sysdep_headers): Add
bits/shmlba.h.
* sysdeps/unix/sysv/linux/bits/shm.h: Include <bits/shmlba.h>.
(SHMLBA): Remove macro.
(__getpagesize): Remove function declaration.
* sysdeps/unix/sysv/linux/hppa/bits/shm.h: Include
<bits/shmlba.h>.
(SHMLBA): Remove macro.
* sysdeps/unix/sysv/linux/mips/bits/shm.h: Include
<bits/shmlba.h>.
(SHMLBA): Remove macro.
* sysdeps/unix/sysv/linux/powerpc/bits/shm.h: Include
<bits/shmlba.h>.
(SHMLBA): Remove macro.
(__getpagesize): Remove function declaration.
* sysdeps/unix/sysv/linux/sparc/bits/shm.h: Include
<bits/shmlba.h>.
(SHMLBA): Remove macro.
(__getshmlba): Remove function declaration.
* sysdeps/unix/sysv/linux/x86/bits/shm.h: Include <bits/shmlba.h>.
(SHMLBA): Remove macro.
(__getpagesize): Remove function declaration.
* sysdeps/unix/sysv/linux/arm/bits/shm.h: Remove file.
* sysdeps/unix/sysv/linux/ia64/bits/shm.h: Likewise.
* sysdeps/unix/sysv/linux/sh/bits/shm.h: Likewise.
* sysdeps/unix/sysv/linux/bits/shmlba.h: New file.
* sysdeps/unix/sysv/linux/arm/bits/shmlba.h: Likewise.
* sysdeps/unix/sysv/linux/hppa/bits/shmlba.h: Likewise.
* sysdeps/unix/sysv/linux/ia64/bits/shmlba.h: Likewise.
* sysdeps/unix/sysv/linux/mips/bits/shmlba.h: Likewise.
* sysdeps/unix/sysv/linux/sh/bits/shmlba.h: Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/shmlba.h: Likewise.
The race leads either to pthread_mutex_destroy returning EBUSY
or triggering an assertion (See description in bugzilla).
This patch is fixing the race by ensuring that the elision path is
used in all cases if elision is enabled by the GLIBC_TUNABLES framework.
The __kind variable in struct __pthread_mutex_s is accessed concurrently.
Therefore we are now using the atomic macros.
The new testcase tst-mutex10 is triggering the race on s390x and intel.
Presumably also on power, but I don't have access to a power machine
with lock-elision. At least the code for power is the same as on the other
two architectures.
ChangeLog:
[BZ #23275]
* nptl/tst-mutex10.c: New File.
* nptl/Makefile (tests): Add tst-mutex10.
(tst-mutex10-ENV): New variable.
* sysdeps/unix/sysv/linux/s390/force-elision.h: (FORCE_ELISION):
Ensure that elision path is used if elision is available.
* sysdeps/unix/sysv/linux/powerpc/force-elision.h (FORCE_ELISION):
Likewise.
* sysdeps/unix/sysv/linux/x86/force-elision.h: (FORCE_ELISION):
Likewise.
* nptl/pthreadP.h (PTHREAD_MUTEX_TYPE, PTHREAD_MUTEX_TYPE_ELISION)
(PTHREAD_MUTEX_PSHARED): Use atomic_load_relaxed.
* nptl/pthread_mutex_consistent.c (pthread_mutex_consistent): Likewise.
* nptl/pthread_mutex_getprioceiling.c (pthread_mutex_getprioceiling):
Likewise.
* nptl/pthread_mutex_lock.c (__pthread_mutex_lock_full)
(__pthread_mutex_cond_lock_adjust): Likewise.
* nptl/pthread_mutex_setprioceiling.c (pthread_mutex_setprioceiling):
Likewise.
* nptl/pthread_mutex_timedlock.c (__pthread_mutex_timedlock): Likewise.
* nptl/pthread_mutex_trylock.c (__pthread_mutex_trylock): Likewise.
* nptl/pthread_mutex_unlock.c (__pthread_mutex_unlock_full): Likewise.
* sysdeps/nptl/bits/thread-shared-types.h (struct __pthread_mutex_s):
Add comments.
* nptl/pthread_mutex_destroy.c (__pthread_mutex_destroy):
Use atomic_load_relaxed and atomic_store_relaxed.
* nptl/pthread_mutex_init.c (__pthread_mutex_init):
Use atomic_store_relaxed.
Since aligned loads and stores are huge performance
advantage the implementation always tries to do aligned
access. Among the cases when src and dst addresses are
aligned or unaligned evenly there are cases of not evenly
unaligned src and dst. For such cases (if the length is
big enough) ext instruction is used to merge-and-shift
two memory chunks loaded from two adjacent aligned
locations and then the adjusted chunk gets stored to
aligned address.
Performance gain against the current T2 implementation:
memcpy-large: 65K-32M: +40% - +10%
memcpy-walk: 128-32M: +20% - +2%
