This patch adds some randomly-generated tests of clog and clog10 that
are observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of clog and clog10.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds some randomly-generated tests of atanh that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of atanh.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds some randomly-generated tests of atan that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of atan.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds some randomly-generated tests of cbrt that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of cbrt.
* math/auto-libm-test-out: Regenerated.
* sysdeps/x86_64/fpu/libm-test-ulps: Update.
This patch adds some randomly-generated tests of cabs that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of cabs.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
The dbl-64 implementation of atan2 does computations that expect to
run in round-to-nearest mode, and in other modes the errors can
accumulate to more than the maximum accepted 9ulp. This patch makes
it use FE_TONEAREST internally, similar to other functions with such
issues. Tests that previously produced large errors are added for
atan2 and the closely related carg, clog and clog10 functions.
Tested for x86_64 and x86 and ulps updated accordingly.
[BZ #18210]
[BZ #18211]
* sysdeps/ieee754/dbl-64/e_atan2.c: Include <fenv.h>.
(__ieee754_atan2): Set FE_TONEAREST mode for internal
computations.
* math/auto-libm-test-in: Add more tests of atan2, carg, clog and
clog10.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
The dbl-64 implementation of atan does computations that expect to run
in round-to-nearest mode, and in other modes the errors can accumulate
to more than the maximum accepted 9ulp. This patch makes it use
FE_TONEAREST internally, similar to other functions with such issues.
Tested for x86_64 and x86; no ulps updates needed.
[BZ #18197]
* sysdeps/ieee754/dbl-64/s_atan.c: Include <fenv.h>.
(atan): Set FE_TONEAREST mode for internal computations.
* math/auto-libm-test-in: Add more tests of atan.
* math/auto-libm-test-out: Regenerated.
On Alpha and IA-64, fexcept_t is unsigned long. But all the values
fit within an int, so the cast is ok for printing. All other hosts
use unsigned int or unsigned short already.
Trimming heaps is a balance between saving memory and the system overhead
required to update page tables and discard allocated pages. The malloc
option M_TRIM_THRESHOLD is a tunable that users are meant to use to decide
where this balance point is but it is only applied to the main arena.
For scalability reasons, glibc malloc has per-thread heaps but these are
shrunk with madvise() if there is one page free at the top of the heap.
In some circumstances this can lead to high system overhead if a thread
has a control flow like
while (data_to_process) {
buf = malloc(large_size);
do_stuff();
free(buf);
}
For a large size, the free() will call madvise (pagetable teardown, page
free and TLB flush) every time followed immediately by a malloc (fault,
kernel page alloc, zeroing and charge accounting). The kernel overhead
can dominate such a workload.
This patch allows the user to tune when madvise gets called by applying
the trim threshold to the per-thread heaps and using similar logic to the
main arena when deciding whether to shrink. Alternatively if the dynamic
brk/mmap threshold gets adjusted then the new values will be obeyed by
the per-thread heaps.
Bug 17195 was a test case motivated by a problem encountered in scientific
applications written in python that performance badly due to high page fault
overhead. The basic operation of such a program was posted by Julian Taylor
https://sourceware.org/ml/libc-alpha/2015-02/msg00373.html
With this patch applied, the overhead is eliminated. All numbers in this
report are in seconds and were recorded by running Julian's program 30
times.
pyarray
glibc madvise
2.21 v2
System min 1.81 ( 0.00%) 0.00 (100.00%)
System mean 1.93 ( 0.00%) 0.02 ( 99.20%)
System stddev 0.06 ( 0.00%) 0.01 ( 88.99%)
System max 2.06 ( 0.00%) 0.03 ( 98.54%)
Elapsed min 3.26 ( 0.00%) 2.37 ( 27.30%)
Elapsed mean 3.39 ( 0.00%) 2.41 ( 28.84%)
Elapsed stddev 0.14 ( 0.00%) 0.02 ( 82.73%)
Elapsed max 4.05 ( 0.00%) 2.47 ( 39.01%)
glibc madvise
2.21 v2
User 141.86 142.28
System 57.94 0.60
Elapsed 102.02 72.66
Note that almost a minutes worth of system time is eliminted and the
program completes 28% faster on average.
To illustrate the problem without python this is a basic test-case for
the worst case scenario where every free is a madvise followed by a an alloc
/* gcc bench-free.c -lpthread -o bench-free */
static int num = 1024;
void __attribute__((noinline,noclone)) dostuff (void *p)
{
}
void *worker (void *data)
{
int i;
for (i = num; i--;)
{
void *m = malloc (48*4096);
dostuff (m);
free (m);
}
return NULL;
}
int main()
{
int i;
pthread_t t;
void *ret;
if (pthread_create (&t, NULL, worker, NULL))
exit (2);
if (pthread_join (t, &ret))
exit (3);
return 0;
}
Before the patch, this resulted in 1024 calls to madvise. With the patch applied,
madvise is called twice because the default trim threshold is high enough to avoid
this.
This a more complex case where there is a mix of frees. It's simply a different worker
function for the test case above
void *worker (void *data)
{
int i;
int j = 0;
void *free_index[num];
for (i = num; i--;)
{
void *m = malloc ((i % 58) *4096);
dostuff (m);
if (i % 2 == 0) {
free (m);
} else {
free_index[j++] = m;
}
}
for (; j >= 0; j--)
{
free(free_index[j]);
}
return NULL;
}
glibc 2.21 calls malloc 90305 times but with the patch applied, it's
called 13438. Increasing the trim threshold will decrease the number of
times it's called with the option of eliminating the overhead.
ebizzy is meant to generate a workload resembling common web application
server workloads. It is threaded with a large working set that at its core
has an allocation, do_stuff, free loop that also hits this case. The primary
metric of the benchmark is records processed per second. This is running on
my desktop which is a single socket machine with an I7-4770 and 8 cores.
Each thread count was run for 30 seconds. It was only run once as the
performance difference is so high that the variation is insignificant.
glibc 2.21 patch
threads 1 10230 44114
threads 2 19153 84925
threads 4 34295 134569
threads 8 51007 183387
Note that the saving happens to be a concidence as the size allocated
by ebizzy was less than the default threshold. If a different number of
chunks were specified then it may also be necessary to tune the threshold
to compensate
This is roughly quadrupling the performance of this benchmark. The difference in
system CPU usage illustrates why.
ebizzy running 1 thread with glibc 2.21
10230 records/s 306904
real 30.00 s
user 7.47 s
sys 22.49 s
22.49 seconds was spent in the kernel for a workload runinng 30 seconds. With the
patch applied
ebizzy running 1 thread with patch applied
44126 records/s 1323792
real 30.00 s
user 29.97 s
sys 0.00 s
system CPU usage was zero with the patch applied. strace shows that glibc
running this workload calls madvise approximately 9000 times a second. With
the patch applied madvise was called twice during the workload (or 0.06
times per second).
2015-02-10 Mel Gorman <mgorman@suse.de>
[BZ #17195]
* malloc/arena.c (free): Apply trim threshold to per-thread heaps
as well as the main arena.
Silvermont and Knights Landing have a modular system design with two cores
sharing an L2 cache. If more than 2 cores are detected to shared L2 cache,
it should be adjusted for Silvermont and Knights Landing.
[BZ #18185]
* sysdeps/x86_64/cacheinfo.c (init_cacheinfo): Limit threads
sharing L2 cache to 2 for Silvermont/Knights Landing.
Linkers in some versions of binutils 2.25 and 2.26 don't support protected
data symbol with error messsage like:
/usr/bin/ld: copy reloc against protected `bar' is invalid
/usr/bin/ld: failed to set dynamic section sizes: Bad value
We check if linker supports copy reloc against protected data symbol to
avoid running the test if linker is broken.
[BZ #17711]
* config.make.in (have-protected-data): New.
* configure.ac: Check linker support for protected data symbol.
* configure: Regenerated.
* elf/Makefile (modules-names): Add tst-protected1moda and
tst-protected1modb if $(have-protected-data) is yes.
(tests): Add tst-protected1a and tst-protected1b if
$(have-protected-data) is yes.
($(objpfx)tst-protected1a): New.
($(objpfx)tst-protected1b): Likewise.
(tst-protected1modb.so-no-z-defs): Likewise.
* elf/tst-protected1a.c: New file.
* elf/tst-protected1b.c: Likewise.
* elf/tst-protected1mod.h: Likewise.
* elf/tst-protected1moda.c: Likewise.
* elf/tst-protected1modb.c: Likewise.
With copy relocation, address of protected data defined in the shared
library may be external. When there is a relocation against the
protected data symbol within the shared library, we need to check if we
should skip the definition in the executable copied from the protected
data. This patch adds ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA and defines
it for x86. If ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA isn't 0, do_lookup_x
will skip the data definition in the executable from copy reloc.
[BZ #17711]
* elf/dl-lookup.c (do_lookup_x): When UNDEF_MAP is NULL, which
indicates it is called from do_lookup_x on relocation against
protected data, skip the data definion in the executable from
copy reloc.
(_dl_lookup_symbol_x): Pass ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA,
instead of ELF_RTYPE_CLASS_PLT, to do_lookup_x for
EXTERN_PROTECTED_DATA relocation against STT_OBJECT symbol.
* sysdeps/generic/ldsodefs.h * (ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA):
New. Defined to 4 if DL_EXTERN_PROTECTED_DATA is defined,
otherwise to 0.
* sysdeps/i386/dl-lookupcfg.h (DL_EXTERN_PROTECTED_DATA): New.
* sysdeps/i386/dl-machine.h (elf_machine_type_class): Set class
to ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA for R_386_GLOB_DAT.
* sysdeps/x86_64/dl-lookupcfg.h (DL_EXTERN_PROTECTED_DATA): New.
* sysdeps/x86_64/dl-machine.h (elf_machine_type_class): Set class
to ELF_RTYPE_CLASS_EXTERN_PROTECTED_DATA for R_X86_64_GLOB_DAT.
IFUNC is difficult to correctly implement on any target needing a GOT
to support position independent code, due to the dependency on order
of dynamic relocations. ld.so should be changed to apply IFUNC
relocations last, globally, because without that it is actually
impossible to write an IFUNC resolver in C that works in all
situations. Case in point, vfork in libpthread.so is an IFUNC with
the resolver returning &__libc_vfork. (system and fork are similar.)
If another shared library, libA say, uses vfork then it is quite
possible that libpthread.so hasn't been dynamically relocated before
the unfortunate libA is dynamically relocated. In that case the GOT
entry for &__libc_vfork is still zero, so the IFUNC resolver returns
NULL. LD_BIND_NOW=1 results in libA PLT dynamic relocations being
applied using this NULL value and ld.so segfaults.
This patch hardens ld.so to not segfault on a NULL from an IFUNC
resolver. It also fixes a problem with undefined weak. If you leave
the plt entry as-is for undefined weak then if the entry is ever
called it will loop in ld.so rather than segfaulting.
* sysdeps/powerpc/powerpc64/dl-machine.h (elf_machine_fixup_plt):
Don't segfault if ifunc resolver returns a NULL. Do set plt to
zero for undefined weak.
(elf_machine_plt_conflict): Similarly.
This patch adds some randomly-generated tests of acosh, asinh and
atanh that are observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of acosh, asinh and
atanh.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds a randomly-generated test of asin that is observed to
increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add another test of asin.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
In the course of the work on six-argument syscalls I noticed that the
i386 lowlevellock.h contained some unused macro definitions (already
unused before my patch). This patch removes them.
Tested for x86 that installed stripped shared libraries are unchanged
by this patch.
* sysdeps/unix/sysv/linux/i386/lowlevellock.h (LLL_EBX_LOAD):
Remove macro.
(LLL_EBX_REG): Likewise.
(LLL_ENTER_KERNEL): Likewise.
This patch adds some randomly-generated tests of asin that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of asin.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch follows the approach outlined in
<https://sourceware.org/ml/libc-alpha/2015-03/msg00656.html> to
support six-argument syscalls from INTERNAL_SYSCALL for 32-bit x86,
making them call a function __libc_do_syscall that takes the syscall
number and three syscall arguments in the registers in which the
kernel expects them, along with a pointer to a structure containing
the other three arguments.
In turn, this allows the generic lowlevellock-futex.h to be used on
32-bit x86, so supporting lll_futex_timed_wait_bitset (and so allowing
FUTEX_CLOCK_REALTIME to be used in various cases, so fixing bug 18138
for 32-bit x86 and leaving hppa as the only architecture missing
lll_futex_timed_wait_bitset). The change to lowlevellock.h's
definition of SYS_futex is because the generic lowlevelloc-futex.h
ends up bringing in bits/syscall.h which defines SYS_futex to
__NR_futex, so resulting in redefinition errors. The revised
definition in lowlevellock.h is in line with what the x86_64 version
does.
__libc_do_syscall is only needed in libpthread at present (meaning
nothing special needs to be done to make it shared-only in most
libraries containing it, static in libc only, as on ARM).
Tested for 32-bit x86, with the glibc testsuite and with the test in
bug 18138. The failures seen
FAIL: nptl/tst-cleanupx4
FAIL: rt/tst-cpuclock2
are pre-existing.
[BZ #18138]
* sysdeps/unix/sysv/linux/i386/sysdep.h (struct
libc_do_syscall_args): New structure.
(INTERNAL_SYSCALL_MAIN_0): New macro.
(INTERNAL_SYSCALL_MAIN_1): Likewise.
(INTERNAL_SYSCALL_MAIN_2): Likewise.
(INTERNAL_SYSCALL_MAIN_3): Likewise.
(INTERNAL_SYSCALL_MAIN_4): Likewise.
(INTERNAL_SYSCALL_MAIN_5): Likewise.
(INTERNAL_SYSCALL_MAIN_6): Likewise. Call __libc_do_syscall.
(INTERNAL_SYSCALL): Define to use INTERNAL_SYSCALL_MAIN_##nr.
Replace conditional definitions by conditional definitions of ....
(INTERNAL_SYSCALL_MAIN_INLINE): ... this. New macro.
* sysdeps/unix/sysv/linux/i386/libc-do-syscall.S: New file.
* sysdeps/unix/sysv/linux/i386/Makefile [$(subdir) = nptl]
(libpthread-sysdep_routines): Add libc-do-syscall.
* sysdeps/unix/sysv/linux/i386/lowlevellock-futex.h: Remove file.
* sysdeps/unix/sysv/linux/i386/lowlevellock.h (SYS_futex): Define
to __NR_futex not 240.
This patch is glibc support for a PowerPC TLS optimization, inspired
by Alexandre Oliva's TLS optimization for other processors,
http://www.lsd.ic.unicamp.br/~oliva/writeups/TLS/RFC-TLSDESC-x86.txt
In essence, this optimization uses a zero module id in the tls_index
GOT entry to indicate that a TLS variable is allocated space in the
static TLS area. A special plt call linker stub for __tls_get_addr
checks for such a tls_index and if found, returns the offset
immediately. The linker communicates the fact that the special
__tls_get_addr stub is used by setting a bit in the dynamic tag
DT_PPC64_OPT/DT_PPC_OPT. glibc communicates to the linker that this
optimization is available by the presence of __tls_get_addr_opt.
tst-tlsmod2.so is built with -Wl,--no-tls-get-addr-optimize for
tst-tls-dlinfo, which otherwise would fail since it tests that no
static tls is allocated. The ld option --no-tls-get-addr-optimize has
been available since binutils-2.20 so doesn't need a configure test.
* NEWS: Advertise TLS optimization.
* elf/elf.h (R_PPC_TLSGD, R_PPC_TLSLD, DT_PPC_OPT, PPC_OPT_TLS): Define.
(DT_PPC_NUM): Increment.
* elf/dynamic-link.h (HAVE_STATIC_TLS): Define.
(CHECK_STATIC_TLS): Use here.
* sysdeps/powerpc/powerpc32/dl-machine.h (elf_machine_rela): Optimize
TLS descriptors.
* sysdeps/powerpc/powerpc64/dl-machine.h (elf_machine_rela): Likewise.
* sysdeps/powerpc/dl-tls.c: New file.
* sysdeps/powerpc/Versions: Add __tls_get_addr_opt.
* sysdeps/powerpc/tst-tlsopt-powerpc.c: New tls test.
* sysdeps/unix/sysv/linux/powerpc/Makefile: Add new test.
Build tst-tlsmod2.so with --no-tls-get-addr-optimize.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/ld.abilist: Update.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/ld.abilist: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/ld-le.abilist: Likewise.
This feature doesn't depend on the linker, as can be seen from the
actual test. It's a compiler feature.
* sysdeps/powerpc/powerpc64/configure.ac: Correct "linker support
for overlapping .opd entries" to "support...".
* sysdeps/powerpc/powerpc64/configure: Regenerate
This patch adds some randomly-generated tests of acos that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of acos.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds some randomly-generated tests of expm1 that are
observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of expm1.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch adds some randomly-generated tests of cosh and sinh that
are observed to increase ulps on x86_64.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of cosh and sinh.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
The x86_64 and x86 libm-test-ulps files hadn't been regenerated from
scratch for some time, as evidenced by the presence of entries for
*_tonearest functions (those tests duplicated the
default-rounding-mode tests, and such duplicates are no longer run).
The aarch64, alpha, hppa, ia64, m68k, microblaze, powerpc, s390, sh,
sparc, tile files similarly could do with from-scratch regeneration as
evidenced by the presence of such entries. (Truncate the existing
file then run "make regen-ulps" and move the resulting file into
place.)
This patch regenerates the x86_64 and x86 files from scratch. It's
likely some of the reduced / removed ulps will need restoring because
they appear on processors or compiler versions other than the one I
tested on, but in such cases I'd like to first see if I can generate
new tests that show such ulps on the Intel processor I'm testing on,
to reduce the effects from different people using different processors
and compilers to regenerate the ulps.
* sysdeps/i386/fpu/libm-test-ulps: Regenerated.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
In testing for x86_64 on an AMD processor, I observed libm test
failures of the form:
testing long double (without inline functions)
Failure: Test: log2_downward (0x2.b7e151628aed4p+0)
Result:
is: 1.44269504088896356633e+00 0xb.8aa3b295c17f67600000p-3
should be: 1.44269504088896356622e+00 0xb.8aa3b295c17f67500000p-3
difference: 1.08420217248550443400e-19 0x8.00000000000000000000p-66
ulp : 1.0000
max.ulp : 0.0000
Maximal error of `log2_downward'
is : 1 ulp
accepted: 0 ulp
These issues arise because the maximum ulps when regenerating on one
processor are not the same as on another processor, so regeneration on
several processors may be needed when updating libm-test-ulps to avoid
failures for some users testing glibc - but such regeneration on
multiple processors is inconvenient. Causes can be: on x86 and, for
x86_64, for long double, variation in results of x87 instructions for
transcendental operations between processors; on x86, variation in
compiler excess precision between compiler versions and
configurations; on any processor where the compiler may contract
expressions using fused multiply-add, variation in what contraction
occurs.
Although it's hard to be sure libm-test-ulps covers all ulps that may
be seen in any configuration for the given architecture, in practice
it helps simply to add wider test coverage to make it more likely
that, when testing on one processor, the ulps seen are the biggest
that can be seen for that function on that processor, and hopefully
they are also the biggest that can be seen for that function in other
configurations for that architecture. Thus, this patch adds some
tests of log2 that increase the ulps I see on x86_64 on an Intel
processor, so that hopefully future from-scratch regenerations on that
processor will produce ulps big enough not to have errors from testing
on AMD processors. These tests were found by randomly generating
inputs and seeing what produced ulps larger than those currently in
libm-test-ulps. Of course such increases also improve the accuracy of
the empirical table of known ulps generated from libm-test-ulps files
that goes in the manual.
Tested for x86_64 and x86 and ulps updated accordingly.
* math/auto-libm-test-in: Add more tests of log2.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
extend_alloca was used to emulate VLA deallocation. The new version
also handles the res == 0 corner case more explicitly, by returning 0
instead of the (potentially undefined, but usually zero) system call
error.
In bc0cdc498 the configure check for HAVE_ASM_PPC_REL16 was removed
on the grounds that the minimum binutils supports rel16 relocs. This
is true, but not all references to HAVE_ASM_PPC_REL16 in the sources
were removed.
* config.h.in: Remove HAVE_ASM_PPC_REL16.
* sysdeps/powerpc/powerpc32/tls-macros.h: Remove HAVE_ASM_PPC_REL16
and false branch of conditional.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/swapcontext-common.S:
Likewise.
