freebsd-nq/sys/dev/hwpmc/hwpmc_logging.c

1286 lines
31 KiB
C
Raw Normal View History

/*-
* SPDX-License-Identifier: BSD-2-Clause-FreeBSD
*
* Copyright (c) 2005-2007 Joseph Koshy
* Copyright (c) 2007 The FreeBSD Foundation
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
* Copyright (c) 2018 Matthew Macy
* All rights reserved.
*
* Portions of this software were developed by A. Joseph Koshy under
* sponsorship from the FreeBSD Foundation and Google, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
*/
/*
* Logging code for hwpmc(4)
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/capsicum.h>
#include <sys/domainset.h>
#include <sys/file.h>
#include <sys/kernel.h>
#include <sys/kthread.h>
#include <sys/lock.h>
#include <sys/module.h>
#include <sys/mutex.h>
#include <sys/pmc.h>
#include <sys/pmckern.h>
#include <sys/pmclog.h>
#include <sys/proc.h>
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#include <sys/sched.h>
#include <sys/signalvar.h>
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#include <sys/smp.h>
#include <sys/syscallsubr.h>
#include <sys/sysctl.h>
#include <sys/systm.h>
#include <sys/uio.h>
#include <sys/unistd.h>
#include <sys/vnode.h>
#if defined(__i386__) || defined(__amd64__)
#include <machine/clock.h>
#endif
#define curdomain PCPU_GET(domain)
/*
* Sysctl tunables
*/
SYSCTL_DECL(_kern_hwpmc);
/*
* kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
*/
static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
&pmclog_buffer_size, 0, "size of log buffers in kilobytes");
/*
* kern.hwpmc.nbuffers_pcpu -- number of global log buffers
*/
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
&pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
/*
* Global log buffer list and associated spin lock.
*/
static struct mtx pmc_kthread_mtx; /* sleep lock */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do { \
(D)->plb_fence = ((char *) (buf)) + 1024*pmclog_buffer_size; \
(D)->plb_base = (D)->plb_ptr = ((char *) (buf)); \
(D)->plb_domain = domain; \
} while (0)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do { \
(D)->plb_ptr = (D)->plb_base; \
} while (0)
/*
* Log file record constructors.
*/
#define _PMCLOG_TO_HEADER(T,L) \
((PMCLOG_HEADER_MAGIC << 24) | \
(PMCLOG_TYPE_ ## T << 16) | \
((L) & 0xFFFF))
/* reserve LEN bytes of space and initialize the entry header */
#define _PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do { \
uint32_t *_le; \
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
int _len = roundup((LEN), sizeof(uint32_t)); \
struct pmclog_header *ph; \
if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
ACTION; \
} \
ph = (struct pmclog_header *)_le; \
ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
ph->pl_tsc = (TSC); \
_le += sizeof(*ph)/4 /* skip over timestamp */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
/* reserve LEN bytes of space and initialize the entry header */
#define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \
uint32_t *_le; \
int _len = roundup((LEN), sizeof(uint32_t)); \
uint64_t tsc; \
struct pmclog_header *ph; \
tsc = pmc_rdtsc(); \
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
spinlock_enter(); \
if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
spinlock_exit(); \
ACTION; \
} \
ph = (struct pmclog_header *)_le; \
ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
ph->pl_tsc = tsc; \
_le += sizeof(*ph)/4 /* skip over timestamp */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_RESERVE_SAFE(P,T,L,TSC) _PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
#define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return)
#define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \
error=ENOMEM;goto error)
#define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0)
#define PMCLOG_EMIT64(V) do { \
*_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \
*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \
} while (0)
/* Emit a string. Caution: does NOT update _le, so needs to be last */
#define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0)
#define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_DESPATCH_SAFE(PO) \
pmclog_release((PO)); \
} while (0)
#define PMCLOG_DESPATCH_SCHED_LOCK(PO) \
pmclog_release_flags((PO), 0); \
} while (0)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_DESPATCH(PO) \
pmclog_release((PO)); \
spinlock_exit(); \
} while (0)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
#define PMCLOG_DESPATCH_SYNC(PO) \
pmclog_schedule_io((PO), 1); \
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
spinlock_exit(); \
} while (0)
#define TSDELTA 4
/*
* Assertions about the log file format.
*/
CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
5*4 + sizeof(uintfptr_t));
2006-03-26 12:20:54 +00:00
CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
5*4 + TSDELTA + sizeof(uintfptr_t));
CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
sizeof(uintfptr_t) + TSDELTA);
CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
sizeof(uintfptr_t));
CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
/*
* Log buffer structure
*/
struct pmclog_buffer {
TAILQ_ENTRY(pmclog_buffer) plb_next;
char *plb_base;
char *plb_ptr;
char *plb_fence;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
uint16_t plb_domain;
} __aligned(CACHE_LINE_SIZE);
/*
* Prototypes
*/
static int pmclog_get_buffer(struct pmc_owner *po);
static void pmclog_loop(void *arg);
static void pmclog_release(struct pmc_owner *po);
static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
static void pmclog_schedule_all(struct pmc_owner *po);
static void pmclog_stop_kthread(struct pmc_owner *po);
/*
* Helper functions
*/
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
static inline void
pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
{
TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
}
static inline void
pmc_plb_rele(struct pmclog_buffer *plb)
{
mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
pmc_plb_rele_unlocked(plb);
mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
}
/*
* Get a log buffer
*/
static int
pmclog_get_buffer(struct pmc_owner *po)
{
struct pmclog_buffer *plb;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
int domain;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
KASSERT(po->po_curbuf[curcpu] == NULL,
("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
domain = curdomain;
MPASS(pmc_dom_hdrs[domain]);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
#ifdef HWPMC_DEBUG
if (plb)
KASSERT(plb->plb_ptr == plb->plb_base &&
plb->plb_base < plb->plb_fence,
("[pmclog,%d] po=%p buffer invariants: ptr=%p "
"base=%p fence=%p", __LINE__, po, plb->plb_ptr,
plb->plb_base, plb->plb_fence));
#endif
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
po->po_curbuf[curcpu] = plb;
/* update stats */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
counter_u64_add(pmc_stats.pm_buffer_requests, 1);
if (plb == NULL)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
return (plb ? 0 : ENOMEM);
}
struct pmclog_proc_init_args {
struct proc *kthr;
struct pmc_owner *po;
bool exit;
bool acted;
};
int
pmclog_proc_create(struct thread *td, void **handlep)
{
struct pmclog_proc_init_args *ia;
int error;
ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
error = kproc_create(pmclog_loop, ia, &ia->kthr,
RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
if (error == 0)
*handlep = ia;
return (error);
}
void
pmclog_proc_ignite(void *handle, struct pmc_owner *po)
{
struct pmclog_proc_init_args *ia;
ia = handle;
mtx_lock(&pmc_kthread_mtx);
MPASS(!ia->acted);
MPASS(ia->po == NULL);
MPASS(!ia->exit);
MPASS(ia->kthr != NULL);
if (po == NULL) {
ia->exit = true;
} else {
ia->po = po;
KASSERT(po->po_kthread == NULL,
("[pmclog,%d] po=%p kthread (%p) already present",
__LINE__, po, po->po_kthread));
po->po_kthread = ia->kthr;
}
wakeup(ia);
while (!ia->acted)
msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
mtx_unlock(&pmc_kthread_mtx);
free(ia, M_TEMP);
}
/*
* Log handler loop.
*
* This function is executed by each pmc owner's helper thread.
*/
static void
pmclog_loop(void *arg)
{
struct pmclog_proc_init_args *ia;
struct pmc_owner *po;
struct pmclog_buffer *lb;
struct proc *p;
struct ucred *ownercred;
struct ucred *mycred;
struct thread *td;
sigset_t unb;
struct uio auio;
struct iovec aiov;
size_t nbytes;
int error;
td = curthread;
SIGEMPTYSET(unb);
SIGADDSET(unb, SIGHUP);
(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
ia = arg;
MPASS(ia->kthr == curproc);
MPASS(!ia->acted);
mtx_lock(&pmc_kthread_mtx);
while (ia->po == NULL && !ia->exit)
msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
if (ia->exit) {
ia->acted = true;
wakeup(ia);
mtx_unlock(&pmc_kthread_mtx);
kproc_exit(0);
}
MPASS(ia->po != NULL);
po = ia->po;
ia->acted = true;
wakeup(ia);
mtx_unlock(&pmc_kthread_mtx);
ia = NULL;
p = po->po_owner;
mycred = td->td_ucred;
PROC_LOCK(p);
ownercred = crhold(p->p_ucred);
PROC_UNLOCK(p);
PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
KASSERT(po->po_kthread == curthread->td_proc,
("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
po, po->po_kthread, curthread->td_proc));
lb = NULL;
/*
* Loop waiting for I/O requests to be added to the owner
* struct's queue. The loop is exited when the log file
* is deconfigured.
*/
mtx_lock(&pmc_kthread_mtx);
for (;;) {
/* check if we've been asked to exit */
if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
break;
if (lb == NULL) { /* look for a fresh buffer to write */
mtx_lock_spin(&po->po_mtx);
if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
mtx_unlock_spin(&po->po_mtx);
/* No more buffers and shutdown required. */
if (po->po_flags & PMC_PO_SHUTDOWN)
break;
2008-11-30 05:10:14 +00:00
(void) msleep(po, &pmc_kthread_mtx, PWAIT,
"pmcloop", 250);
continue;
}
TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
mtx_unlock_spin(&po->po_mtx);
}
mtx_unlock(&pmc_kthread_mtx);
/* process the request */
PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
lb->plb_base, lb->plb_ptr);
/* change our thread's credentials before issuing the I/O */
aiov.iov_base = lb->plb_base;
aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base;
auio.uio_iov = &aiov;
auio.uio_iovcnt = 1;
auio.uio_offset = -1;
auio.uio_resid = nbytes;
auio.uio_rw = UIO_WRITE;
auio.uio_segflg = UIO_SYSSPACE;
auio.uio_td = td;
/* switch thread credentials -- see kern_ktrace.c */
td->td_ucred = ownercred;
error = fo_write(po->po_file, &auio, ownercred, 0, td);
td->td_ucred = mycred;
if (error) {
/* XXX some errors are recoverable */
/* send a SIGIO to the owner and exit */
PROC_LOCK(p);
kern_psignal(p, SIGIO);
PROC_UNLOCK(p);
mtx_lock(&pmc_kthread_mtx);
po->po_error = error; /* save for flush log */
PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
break;
}
mtx_lock(&pmc_kthread_mtx);
/* put the used buffer back into the global pool */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
pmc_plb_rele(lb);
lb = NULL;
}
wakeup_one(po->po_kthread);
po->po_kthread = NULL;
mtx_unlock(&pmc_kthread_mtx);
/* return the current I/O buffer to the global pool */
if (lb) {
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
pmc_plb_rele(lb);
}
/*
* Exit this thread, signalling the waiter
*/
crfree(ownercred);
kproc_exit(0);
}
/*
* Release and log entry and schedule an I/O if needed.
*/
static void
pmclog_release_flags(struct pmc_owner *po, int wakeup)
{
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
struct pmclog_buffer *plb;
plb = po->po_curbuf[curcpu];
KASSERT(plb->plb_ptr >= plb->plb_base,
("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
po, plb->plb_ptr, plb->plb_base));
KASSERT(plb->plb_ptr <= plb->plb_fence,
("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
po, plb->plb_ptr, plb->plb_fence));
/* schedule an I/O if we've filled a buffer */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (plb->plb_ptr >= plb->plb_fence)
pmclog_schedule_io(po, wakeup);
PMCDBG1(LOG,REL,1, "po=%p", po);
}
static void
pmclog_release(struct pmc_owner *po)
{
pmclog_release_flags(po, 1);
}
/*
* Attempt to reserve 'length' bytes of space in an owner's log
* buffer. The function returns a pointer to 'length' bytes of space
* if there was enough space or returns NULL if no space was
* available. Non-null returns do so with the po mutex locked. The
* caller must invoke pmclog_release() on the pmc owner structure
* when done.
*/
static uint32_t *
pmclog_reserve(struct pmc_owner *po, int length)
{
uintptr_t newptr, oldptr __diagused;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
struct pmclog_buffer *plb, **pplb;
PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
KASSERT(length % sizeof(uint32_t) == 0,
("[pmclog,%d] length not a multiple of word size", __LINE__));
/* No more data when shutdown in progress. */
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (po->po_flags & PMC_PO_SHUTDOWN)
return (NULL);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
pplb = &po->po_curbuf[curcpu];
if (*pplb == NULL && pmclog_get_buffer(po) != 0)
goto fail;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
KASSERT(*pplb != NULL,
("[pmclog,%d] po=%p no current buffer", __LINE__, po));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
plb = *pplb;
KASSERT(plb->plb_ptr >= plb->plb_base &&
plb->plb_ptr <= plb->plb_fence,
("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
__LINE__, po, plb->plb_ptr, plb->plb_base,
plb->plb_fence));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
oldptr = (uintptr_t) plb->plb_ptr;
newptr = oldptr + length;
KASSERT(oldptr != (uintptr_t) NULL,
("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
/*
* If we have space in the current buffer, return a pointer to
* available space with the PO structure locked.
*/
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (newptr <= (uintptr_t) plb->plb_fence) {
plb->plb_ptr = (char *) newptr;
goto done;
}
/*
* Otherwise, schedule the current buffer for output and get a
* fresh buffer.
*/
pmclog_schedule_io(po, 0);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (pmclog_get_buffer(po) != 0)
goto fail;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
plb = *pplb;
KASSERT(plb != NULL,
("[pmclog,%d] po=%p no current buffer", __LINE__, po));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
KASSERT(plb->plb_ptr != NULL,
("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
KASSERT(plb->plb_ptr == plb->plb_base &&
plb->plb_ptr <= plb->plb_fence,
("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
__LINE__, po, plb->plb_ptr, plb->plb_base,
plb->plb_fence));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
oldptr = (uintptr_t) plb->plb_ptr;
done:
return ((uint32_t *) oldptr);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
fail:
return (NULL);
}
/*
* Schedule an I/O.
*
* Transfer the current buffer to the helper kthread.
*/
static void
pmclog_schedule_io(struct pmc_owner *po, int wakeup)
{
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
struct pmclog_buffer *plb;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
plb = po->po_curbuf[curcpu];
po->po_curbuf[curcpu] = NULL;
KASSERT(plb != NULL,
("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
KASSERT(plb->plb_ptr >= plb->plb_base,
("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
po, plb->plb_ptr, plb->plb_base));
KASSERT(plb->plb_ptr <= plb->plb_fence,
("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
po, plb->plb_ptr, plb->plb_fence));
PMCDBG1(LOG,SIO, 1, "po=%p", po);
/*
* Add the current buffer to the tail of the buffer list and
* wakeup the helper.
*/
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
mtx_lock_spin(&po->po_mtx);
TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
mtx_unlock_spin(&po->po_mtx);
if (wakeup)
wakeup_one(po);
}
/*
* Stop the helper kthread.
*/
static void
pmclog_stop_kthread(struct pmc_owner *po)
{
mtx_lock(&pmc_kthread_mtx);
po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
if (po->po_kthread != NULL) {
PROC_LOCK(po->po_kthread);
kern_psignal(po->po_kthread, SIGHUP);
PROC_UNLOCK(po->po_kthread);
}
wakeup_one(po);
while (po->po_kthread)
msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
mtx_unlock(&pmc_kthread_mtx);
}
/*
* Public functions
*/
/*
* Configure a log file for pmc owner 'po'.
*
* Parameter 'logfd' is a file handle referencing an open file in the
* owner process. This file needs to have been opened for writing.
*/
int
pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
{
struct proc *p;
struct timespec ts;
int error;
sx_assert(&pmc_sx, SA_XLOCKED);
PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
p = po->po_owner;
/* return EBUSY if a log file was already present */
if (po->po_flags & PMC_PO_OWNS_LOGFILE)
return (EBUSY);
KASSERT(po->po_file == NULL,
("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
po->po_file));
/* get a reference to the file state */
error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
if (error)
goto error;
/* mark process as owning a log file */
po->po_flags |= PMC_PO_OWNS_LOGFILE;
/* mark process as using HWPMCs */
PROC_LOCK(p);
p->p_flag |= P_HWPMC;
PROC_UNLOCK(p);
nanotime(&ts);
/* create a log initialization entry */
PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
sizeof(struct pmclog_initialize));
PMCLOG_EMIT32(PMC_VERSION);
PMCLOG_EMIT32(md->pmd_cputype);
#if defined(__i386__) || defined(__amd64__)
PMCLOG_EMIT64(tsc_freq);
#else
/* other architectures will need to fill this in */
2018-06-07 02:20:27 +00:00
PMCLOG_EMIT32(0);
PMCLOG_EMIT32(0);
#endif
memcpy(_le, &ts, sizeof(ts));
_le += sizeof(ts)/4;
PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
return (0);
error:
KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
"stopped", __LINE__, po));
if (po->po_file)
(void) fdrop(po->po_file, curthread);
po->po_file = NULL; /* clear file and error state */
po->po_error = 0;
po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
return (error);
}
/*
* De-configure a log file. This will throw away any buffers queued
* for this owner process.
*/
int
pmclog_deconfigure_log(struct pmc_owner *po)
{
int error;
struct pmclog_buffer *lb;
struct pmc_binding pb;
PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
return (EINVAL);
KASSERT(po->po_sscount == 0,
("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
KASSERT(po->po_file != NULL,
("[pmclog,%d] po=%p no log file", __LINE__, po));
/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
pmclog_stop_kthread(po);
KASSERT(po->po_kthread == NULL,
("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
/* return all queued log buffers to the global pool */
while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
pmc_plb_rele(lb);
}
pmc_save_cpu_binding(&pb);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
for (int i = 0; i < mp_ncpus; i++) {
pmc_select_cpu(i);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
/* return the 'current' buffer to the global pool */
if ((lb = po->po_curbuf[curcpu]) != NULL) {
PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
pmc_plb_rele(lb);
}
}
pmc_restore_cpu_binding(&pb);
/* drop a reference to the fd */
if (po->po_file != NULL) {
error = fdrop(po->po_file, curthread);
po->po_file = NULL;
} else
error = 0;
po->po_error = 0;
return (error);
}
/*
* Flush a process' log buffer.
*/
int
pmclog_flush(struct pmc_owner *po, int force)
{
int error;
PMCDBG1(LOG,FLS,1, "po=%p", po);
/*
* If there is a pending error recorded by the logger thread,
* return that.
*/
if (po->po_error)
return (po->po_error);
error = 0;
/*
* Check that we do have an active log file.
*/
mtx_lock(&pmc_kthread_mtx);
if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
error = EINVAL;
goto error;
}
pmclog_schedule_all(po);
error:
mtx_unlock(&pmc_kthread_mtx);
return (error);
}
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
static void
pmclog_schedule_one_cond(struct pmc_owner *po)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
{
struct pmclog_buffer *plb;
int cpu;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
spinlock_enter();
cpu = curcpu;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
/* tell hardclock not to run again */
if (PMC_CPU_HAS_SAMPLES(cpu))
PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
plb = po->po_curbuf[cpu];
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (plb && plb->plb_ptr != plb->plb_base)
pmclog_schedule_io(po, 1);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
spinlock_exit();
}
static void
pmclog_schedule_all(struct pmc_owner *po)
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
{
struct pmc_binding pb;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
/*
* Schedule the current buffer if any and not empty.
*/
pmc_save_cpu_binding(&pb);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
for (int i = 0; i < mp_ncpus; i++) {
pmc_select_cpu(i);
pmclog_schedule_one_cond(po);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
}
pmc_restore_cpu_binding(&pb);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
}
int
pmclog_close(struct pmc_owner *po)
{
PMCDBG1(LOG,CLO,1, "po=%p", po);
pmclog_process_closelog(po);
mtx_lock(&pmc_kthread_mtx);
/*
* Initiate shutdown: no new data queued,
* thread will close file on last block.
*/
po->po_flags |= PMC_PO_SHUTDOWN;
/* give time for all to see */
DELAY(50);
/*
* Schedule the current buffer.
*/
pmclog_schedule_all(po);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
wakeup_one(po);
mtx_unlock(&pmc_kthread_mtx);
return (0);
}
void
pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
{
int n, recordlen;
uint32_t flags;
struct pmc_owner *po;
PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
ps->ps_nsamples);
recordlen = offsetof(struct pmclog_callchain, pl_pc) +
ps->ps_nsamples * sizeof(uintfptr_t);
po = pm->pm_owner;
flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen, ps->ps_tsc);
PMCLOG_EMIT32(ps->ps_pid);
PMCLOG_EMIT32(ps->ps_tid);
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(flags);
for (n = 0; n < ps->ps_nsamples; n++)
PMCLOG_EMITADDR(ps->ps_pc[n]);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SAFE(po);
}
void
pmclog_process_closelog(struct pmc_owner *po)
{
PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
}
void
pmclog_process_dropnotify(struct pmc_owner *po)
{
PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
PMCLOG_DESPATCH(po);
}
void
2006-03-26 12:20:54 +00:00
pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
const char *path)
{
int pathlen, recordlen;
2006-03-26 12:20:54 +00:00
KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
pathlen = strlen(path) + 1; /* #bytes for path name */
2006-03-26 12:20:54 +00:00
recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
pathlen;
2006-03-26 12:20:54 +00:00
PMCLOG_RESERVE(po, MAP_IN, recordlen);
PMCLOG_EMIT32(pid);
PMCLOG_EMIT32(0);
2006-03-26 12:20:54 +00:00
PMCLOG_EMITADDR(start);
PMCLOG_EMITSTRING(path,pathlen);
PMCLOG_DESPATCH_SYNC(po);
}
2006-03-26 12:20:54 +00:00
void
pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
uintfptr_t end)
{
KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
PMCLOG_EMIT32(pid);
PMCLOG_EMIT32(0);
2006-03-26 12:20:54 +00:00
PMCLOG_EMITADDR(start);
PMCLOG_EMITADDR(end);
PMCLOG_DESPATCH(po);
}
void
pmclog_process_pmcallocate(struct pmc *pm)
{
struct pmc_owner *po;
struct pmc_soft *ps;
po = pm->pm_owner;
PMCDBG1(LOG,ALL,1, "pm=%p", pm);
if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
PMCLOG_RESERVE(po, PMCALLOCATEDYN,
sizeof(struct pmclog_pmcallocatedyn));
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pm->pm_event);
PMCLOG_EMIT32(pm->pm_flags);
PMCLOG_EMIT32(0);
PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
ps = pmc_soft_ev_acquire(pm->pm_event);
if (ps != NULL)
PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
else
PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
pmc_soft_ev_release(ps);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
} else {
PMCLOG_RESERVE(po, PMCALLOCATE,
sizeof(struct pmclog_pmcallocate));
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pm->pm_event);
PMCLOG_EMIT32(pm->pm_flags);
PMCLOG_EMIT32(0);
PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
}
}
void
pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
{
int pathlen, recordlen;
struct pmc_owner *po;
PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
po = pm->pm_owner;
pathlen = strlen(path) + 1; /* #bytes for the string */
recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
PMCLOG_RESERVE(po, PMCATTACH, recordlen);
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pid);
PMCLOG_EMITSTRING(path, pathlen);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
}
void
pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
{
struct pmc_owner *po;
PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
po = pm->pm_owner;
PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pid);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_DESPATCH_SYNC(po);
}
void
pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
{
if (sync) {
PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
PMCLOG_EMIT32(p->p_pid);
PMCLOG_EMIT32(p->p_flag);
PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
PMCLOG_DESPATCH_SYNC(po);
} else {
PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
PMCLOG_EMIT32(p->p_pid);
PMCLOG_EMIT32(p->p_flag);
PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
PMCLOG_DESPATCH(po);
}
}
/*
* Log a context switch event to the log file.
*/
void
pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
{
struct pmc_owner *po;
KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
v);
po = pm->pm_owner;
PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
PMCLOG_EMIT64(v);
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pp->pp_proc->p_pid);
PMCLOG_EMIT32(td->td_tid);
PMCLOG_EMIT32(0);
PMCLOG_DESPATCH_SCHED_LOCK(po);
}
void
pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
uintfptr_t startaddr, char *path)
{
int pathlen, recordlen;
PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
pathlen = strlen(path) + 1; /* #bytes for the path */
recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
PMCLOG_RESERVE(po, PROCEXEC, recordlen);
PMCLOG_EMIT32(pid);
PMCLOG_EMIT32(pmid);
PMCLOG_EMITADDR(startaddr);
PMCLOG_EMITSTRING(path,pathlen);
PMCLOG_DESPATCH_SYNC(po);
}
/*
* Log a process exit event (and accumulated pmc value) to the log file.
*/
void
pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
{
int ri;
struct pmc_owner *po;
ri = PMC_TO_ROWINDEX(pm);
PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
pp->pp_pmcs[ri].pp_pmcval);
po = pm->pm_owner;
PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
PMCLOG_EMIT32(pm->pm_id);
PMCLOG_EMIT32(pp->pp_proc->p_pid);
PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
PMCLOG_DESPATCH(po);
}
/*
* Log a fork event.
*/
void
pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
{
PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
PMCLOG_EMIT32(oldpid);
PMCLOG_EMIT32(newpid);
PMCLOG_DESPATCH(po);
}
/*
* Log a process exit event of the form suitable for system-wide PMCs.
*/
void
pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
{
PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
PMCLOG_EMIT32(pid);
PMCLOG_DESPATCH(po);
}
void
pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
{
struct proc *p;
p = td->td_proc;
if (sync) {
PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
PMCLOG_EMIT32(td->td_tid);
PMCLOG_EMIT32(p->p_pid);
PMCLOG_EMIT32(p->p_flag);
PMCLOG_EMIT32(0);
PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
PMCLOG_DESPATCH_SYNC(po);
} else {
PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
PMCLOG_EMIT32(td->td_tid);
PMCLOG_EMIT32(p->p_pid);
PMCLOG_EMIT32(p->p_flag);
PMCLOG_EMIT32(0);
PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
PMCLOG_DESPATCH(po);
}
}
void
pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
{
PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
PMCLOG_EMIT32(td->td_tid);
PMCLOG_DESPATCH(po);
}
/*
* Write a user log entry.
*/
int
pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
{
int error;
PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
error = 0;
PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
sizeof(struct pmclog_userdata));
PMCLOG_EMIT32(wl->pm_userdata);
PMCLOG_DESPATCH(po);
error:
return (error);
}
/*
* Initialization.
*
* Create a pool of log buffers and initialize mutexes.
*/
void
pmclog_initialize(void)
{
struct pmclog_buffer *plb;
int domain, ncpus, total;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
(void) printf("hwpmc: tunable logbuffersize=%d must be "
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
"greater than zero and less than or equal to 16MB.\n",
pmclog_buffer_size);
pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
}
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (pmc_nlogbuffers_pcpu <= 0) {
(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
"than zero.\n", pmc_nlogbuffers_pcpu);
pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
}
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
pmc_nlogbuffers_pcpu*pmclog_buffer_size);
pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
}
for (domain = 0; domain < vm_ndomains; domain++) {
ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
total = ncpus * pmc_nlogbuffers_pcpu;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
pmc_dom_hdrs[domain]->pdbh_plbs = plb;
for (; total > 0; total--, plb++) {
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
void *buf;
buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
pmc_plb_rele_unlocked(plb);
}
}
mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
}
/*
* Shutdown logging.
*
* Destroy mutexes and release memory back the to free pool.
*/
void
pmclog_shutdown(void)
{
struct pmclog_buffer *plb;
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
int domain;
mtx_destroy(&pmc_kthread_mtx);
for (domain = 0; domain < vm_ndomains; domain++) {
hwpmc(9): Make pmclog buffer pcpu and update constants On non-trivial SMP systems the contention on the pmc_owner mutex leads to a substantial number of samples captured being from the pmc process itself. This change a) makes buffers larger to avoid contention on the global list b) makes the working sample buffer per cpu. Run pmcstat in the background (default event rate of 64k): pmcstat -S UNHALTED_CORE_CYCLES -O /dev/null sleep 600 & Before: make -j96 buildkernel -s >&/dev/null 3336.68s user 24684.10s system 7442% cpu 6:16.50 total After: make -j96 buildkernel -s >&/dev/null 2697.82s user 1347.35s system 6058% cpu 1:06.77 total For more realistic overhead measurement set the sample rate for ~2khz on a 2.1Ghz processor: pmcstat -n 1050000 -S UNHALTED_CORE_CYCLES -O /dev/null sleep 6000 & Collecting 10 samples of `make -j96 buildkernel` from each: x before + after real time: N Min Max Median Avg Stddev x 10 76.4 127.62 84.845 88.577 15.100031 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence -28.398 +/- 10.0344 -32.0602% +/- 7.69825% (Student's t, pooled s = 10.6794) system time: N Min Max Median Avg Stddev x 10 2277.96 6948.53 2949.47 3341.492 1385.2677 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence -2277.47 +/- 920.425 -68.1574% +/- 8.77623% (Student's t, pooled s = 979.596) x no pmc + pmc running real time: HEAD: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 76.4 127.62 84.845 88.577 15.100031 Difference at 95.0% confidence 29.73 +/- 10.0335 50.5208% +/- 17.0525% (Student's t, pooled s = 10.6785) patched: N Min Max Median Avg Stddev x 10 58.38 59.15 58.86 58.847 0.22504567 + 10 59.71 60.79 60.135 60.179 0.29957192 Difference at 95.0% confidence 1.332 +/- 0.248939 2.2635% +/- 0.426506% (Student's t, pooled s = 0.264942) system time: HEAD: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 2277.96 6948.53 2949.47 3341.492 1385.2677 Difference at 95.0% confidence 2309.97 +/- 920.443 223.937% +/- 89.3039% (Student's t, pooled s = 979.616) patched: N Min Max Median Avg Stddev x 10 1010.15 1073.31 1025.465 1031.524 18.135705 + 10 1038.7 1081.06 1070.555 1064.017 15.85404 Difference at 95.0% confidence 32.493 +/- 16.0042 3.15% +/- 1.5794% (Student's t, pooled s = 17.0331) Reviewed by: jeff@ Approved by: sbruno@ Differential Revision: https://reviews.freebsd.org/D15155
2018-05-12 01:26:34 +00:00
while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
free(plb->plb_base, M_PMC);
}
free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
}
}