Detecting the cpus based on sysconf of the online CPUs can significantly
over estimate the number of cpus available.
Wheither via numactl, cgroups, taskset, systemd constraints, docker
containers and probably other mechanisms, the number of threads mysqld
can be run on can be quite less.
As such we use the pthread_getaffinity_np function on Linux and FreeBSD
(identical API) to get the number of CPUs.
The number of CPUs is the default for the thread_pool_size and a too
high default will resulting in large memory usage and high context
switching overhead.
Closes PR #922
The fix consists of three commits backported from 10.3:
1) Cleanup isnan() portability checks
(cherry picked from commit 7ffd7fe962)
2) Cleanup isinf() portability checks
Original problem reported by Wlad: re-compilation of 10.3 on top of 10.2
build would cache undefined HAVE_ISINF from 10.2, whereas it is expected
to be 1 in 10.3.
std::isinf() seem to be available on all supported platforms.
(cherry picked from commit bc469a0bdf)
3) Use std::isfinite in C++ code
This is addition to parent revision fixing build failures.
(cherry picked from commit 54999f4e75)
On clang, use __builtin_readcyclecounter() when available.
Hinted by Sergey Vojtovich. (This may lead to runtime failure
on ARM systems. The hardware should be available on ARMv8 (AArch64),
but access to it may require special privileges.)
We remove support for the proprietary Sun Microsystems compiler,
and rely on clang or the __GNUC__ assembler syntax instead.
For now, we retain support for IA-64 (Itanium) and 32-bit SPARC,
even though those platforms are likely no longer widely used.
We remove support for clock_gettime(CLOCK_SGI_CYCLE),
because Silicon Graphics ceased supporting IRIX in December 2013.
This was the only cycle timer interface available for MIPS.
On PowerPC, we rely on the GCC 4.8 __builtin_ppc_get_timebase()
(or clang __builtin_readcyclecounter()), which should be equivalent
to the old assembler code on both 64-bit and 32-bit targets.
Starting with the Intel Skylake microarchitecture, the PAUSE
instruction latency is about 140 clock cycles instead of earlier 10.
On AMD processors, the latency could be 10 or 50 clock cycles,
depending on microarchitecture.
Because of this big range of latency, let us scale the loops around
the PAUSE instruction based on timing results at server startup.
my_cpu_relax_multiplier: New variable: How many times to invoke PAUSE
in a loop. Only defined for IA-32 and AMD64.
my_cpu_init(): Determine with RDTSC the time to run 16 PAUSE instructions
in two unrolled loops according, and based on the quicker of the two
runs, initialize my_cpu_relax_multiplier. This form of calibration was
suggested by Mikhail Sinyavin from Intel.
LF_BACKOFF(), ut_delay(): Use my_cpu_relax_multiplier when available.
ut_delay(): Define inline in my_cpu.h.
UT_COMPILER_BARRIER(): Remove. This does not seem to have any effect,
because in our ut_delay() implementation, no computations are being
performed inside the loop. The purpose of UT_COMPILER_BARRIER() was to
prohibit the compiler from reordering computations. It was not
emitting any code.
For tablespaces that do not reside on spinning storage, it does
not make sense to attempt to write nearby pages when writing out
dirty pages from the InnoDB buffer pool. It is actually detrimental
to performance and to the life span of flash ROM storage.
With this change, MariaDB will detect whether an InnoDB file resides
on solid-state storage. The detection has been implemented for Linux
and Microsoft Windows. For other systems, we will err on the safe side
and assume that files reside on SSD.
As part of this change, we will reduce the number of fstat() calls
when opening data files on POSIX systems and slightly clean up some
file I/O code.
FIXME: os_is_sparse_file_supported() on POSIX works in a destructive
manner. Thus, we can only invoke it when creating files, not when
opening them.
For diagnostics, we introduce the column ON_SSD to the table
INFORMATION_SCHEMA.INNODB_TABLESPACES_SCRUBBING. The table
INNODB_SYS_TABLESPACES might seem more appropriate, but its purpose
is to reflect the contents of the InnoDB system table SYS_TABLESPACES,
which we would like to remove at some point.
On Microsoft Windows, querying StorageDeviceSeekPenaltyProperty
sometimes returns ERROR_GEN_FAILURE instead of ERROR_INVALID_FUNCTION
or ERROR_NOT_SUPPORTED. We will silently ignore also this error,
and assume that the file does not reside on SSD.
On Linux, the detection will be based on the files
/sys/block/*/queue/rotational and /sys/block/*/dev.
Especially for USB storage, it is possible that
/sys/block/*/queue/rotational will wrongly report 1 instead of 0.
fil_node_t::on_ssd: Whether the InnoDB data file resides on
solid-state storage.
fil_system_t::ssd: Collection of Linux block devices that reside on
non-rotational storage.
fil_system_t::create(): Detect ssd on Linux based on the contents
of /sys/block/*/queue/rotational and /sys/block/*/dev.
fil_system_t::is_ssd(dev_t): Determine if a Linux block device is
non-rotational. Partitions will be identified with the containing
block device by assuming that the least significant 4 bits of the
minor number identify a partition, and that the "partition number"
of the entire device is 0.
SIGHUP causes debug info in the error log and reload of
logs/privileges/tables/etc. The server should only do it when
a user intentionally sends SUGHUP, not when a parent terminal gets
disconnected or something.
In particular, not ignoring kernel SIGHUP causes FLUSH PRIVILEGES
at some random point during non-systemd Debian upgrades (Debian
restarts mysqld, debian-start script runs mysql_upgrade in the background,
postinit script ends and kernel sends SIGHUP to all background processes
it has started). And during mysql_upgrade privilege tables aren't
necessarily ready to be reloaded.
main.derived_cond_pushdown: Move all 10.3 tests to the end,
trim trailing white space, and add an "End of 10.3 tests" marker.
Add --sorted_result to tests where the ordering is not deterministic.
main.win_percentile: Add --sorted_result to tests where the
ordering is no longer deterministic.
On some archs uclibc does not provide the ucontext structure despite
providing ucontext.h, for details see
https://git.buildroot.net/buildroot/commit/?id=f1cbfeea95e6287c7a666aafc182ffa318eff262
This patch improves the detection of ucontext by making sure that
HAVE_UCONTEXT_H is only set when makecontext() was found.
Signed-off-by: Bernd Kuhls <bernd.kuhls@t-online.de>
Original problem reported by Wlad: re-compilation of 10.3 on top of 10.2
build would cache undefined HAVE_ISINF from 10.2, whereas it is expected
to be 1 in 10.3.
std::isinf() seem to be available on all supported platforms.
musl ships the header for other purposes, but makecontext is not
implemented. fix the check to detect if makecontext is implemented
before enabling code using it.
Resolving a stacktrace including functions in dynamic libraries requires
us to look inside the libraries for the symbols. Addr2line needs to be
started with the correct binary for each address on the stack. To do this,
figure out which library it is using dladdr, then if the addr2line
binary was started with a different binary, fork it again with the
correct one.
We only have one addr2line process running at any point during the
stacktrace resolving step. The maximum number of forks for addr2line should
generally be around 6.
One for server stacktrace code, one for plugin code, one when going back
into server code, one for pthread library, one for libc, one for the
_start function in the server. More can come up if plugin calls server
function which goes back to a plugin, etc.
