freebsd-skq/sys/kern/subr_prof.c
Konstantin Belousov 5c7bebf961 The process spin lock currently has the following distinct uses:
- Threads lifetime cycle, in particular, counting of the threads in
  the process, and interlocking with process mutex and thread lock.
  The main reason of this is that turnstile locks are after thread
  locks, so you e.g. cannot unlock blockable mutex (think process
  mutex) while owning thread lock.

- Virtual and profiling itimers, since the timers activation is done
  from the clock interrupt context.  Replace the p_slock by p_itimmtx
  and PROC_ITIMLOCK().

- Profiling code (profil(2)), for similar reason.  Replace the p_slock
  by p_profmtx and PROC_PROFLOCK().

- Resource usage accounting.  Need for the spinlock there is subtle,
  my understanding is that spinlock blocks context switching for the
  current thread, which prevents td_runtime and similar fields from
  changing (updates are done at the mi_switch()).  Replace the p_slock
  by p_statmtx and PROC_STATLOCK().

The split is done mostly for code clarity, and should not affect
scalability.

Tested by:	pho
Sponsored by:	The FreeBSD Foundation
MFC after:	1 week
2014-11-26 14:10:00 +00:00

544 lines
16 KiB
C

/*-
* Copyright (c) 1982, 1986, 1993
* The Regents of the University of California. All rights reserved.
*
* 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.
* 4. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS 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 REGENTS 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.
*
* @(#)subr_prof.c 8.3 (Berkeley) 9/23/93
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/sysproto.h>
#include <sys/kernel.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/resourcevar.h>
#include <sys/sysctl.h>
#include <machine/cpu.h>
#ifdef GPROF
#include <sys/malloc.h>
#include <sys/gmon.h>
#undef MCOUNT
static MALLOC_DEFINE(M_GPROF, "gprof", "kernel profiling buffer");
static void kmstartup(void *);
SYSINIT(kmem, SI_SUB_KPROF, SI_ORDER_FIRST, kmstartup, NULL);
struct gmonparam _gmonparam = { GMON_PROF_OFF };
#ifdef GUPROF
void
nullfunc_loop_profiled()
{
int i;
for (i = 0; i < CALIB_SCALE; i++)
nullfunc_profiled();
}
#define nullfunc_loop_profiled_end nullfunc_profiled /* XXX */
void
nullfunc_profiled()
{
}
#endif /* GUPROF */
/*
* Update the histograms to support extending the text region arbitrarily.
* This is done slightly naively (no sparse regions), so will waste slight
* amounts of memory, but will overall work nicely enough to allow profiling
* of KLDs.
*/
void
kmupetext(uintfptr_t nhighpc)
{
struct gmonparam np; /* slightly large */
struct gmonparam *p = &_gmonparam;
char *cp;
GIANT_REQUIRED;
bcopy(p, &np, sizeof(*p));
np.highpc = ROUNDUP(nhighpc, HISTFRACTION * sizeof(HISTCOUNTER));
if (np.highpc <= p->highpc)
return;
np.textsize = np.highpc - p->lowpc;
np.kcountsize = np.textsize / HISTFRACTION;
np.hashfraction = HASHFRACTION;
np.fromssize = np.textsize / HASHFRACTION;
np.tolimit = np.textsize * ARCDENSITY / 100;
if (np.tolimit < MINARCS)
np.tolimit = MINARCS;
else if (np.tolimit > MAXARCS)
np.tolimit = MAXARCS;
np.tossize = np.tolimit * sizeof(struct tostruct);
cp = malloc(np.kcountsize + np.fromssize + np.tossize,
M_GPROF, M_WAITOK);
/*
* Check for something else extending highpc while we slept.
*/
if (np.highpc <= p->highpc) {
free(cp, M_GPROF);
return;
}
np.tos = (struct tostruct *)cp;
cp += np.tossize;
np.kcount = (HISTCOUNTER *)cp;
cp += np.kcountsize;
np.froms = (u_short *)cp;
#ifdef GUPROF
/* Reinitialize pointers to overhead counters. */
np.cputime_count = &KCOUNT(&np, PC_TO_I(&np, cputime));
np.mcount_count = &KCOUNT(&np, PC_TO_I(&np, mcount));
np.mexitcount_count = &KCOUNT(&np, PC_TO_I(&np, mexitcount));
#endif
critical_enter();
bcopy(p->tos, np.tos, p->tossize);
bzero((char *)np.tos + p->tossize, np.tossize - p->tossize);
bcopy(p->kcount, np.kcount, p->kcountsize);
bzero((char *)np.kcount + p->kcountsize, np.kcountsize -
p->kcountsize);
bcopy(p->froms, np.froms, p->fromssize);
bzero((char *)np.froms + p->fromssize, np.fromssize - p->fromssize);
cp = (char *)p->tos;
bcopy(&np, p, sizeof(*p));
critical_exit();
free(cp, M_GPROF);
}
static void
kmstartup(dummy)
void *dummy;
{
char *cp;
struct gmonparam *p = &_gmonparam;
#ifdef GUPROF
int cputime_overhead;
int empty_loop_time;
int i;
int mcount_overhead;
int mexitcount_overhead;
int nullfunc_loop_overhead;
int nullfunc_loop_profiled_time;
uintfptr_t tmp_addr;
#endif
/*
* Round lowpc and highpc to multiples of the density we're using
* so the rest of the scaling (here and in gprof) stays in ints.
*/
p->lowpc = ROUNDDOWN((u_long)btext, HISTFRACTION * sizeof(HISTCOUNTER));
p->highpc = ROUNDUP((u_long)etext, HISTFRACTION * sizeof(HISTCOUNTER));
p->textsize = p->highpc - p->lowpc;
printf("Profiling kernel, textsize=%lu [%jx..%jx]\n",
p->textsize, (uintmax_t)p->lowpc, (uintmax_t)p->highpc);
p->kcountsize = p->textsize / HISTFRACTION;
p->hashfraction = HASHFRACTION;
p->fromssize = p->textsize / HASHFRACTION;
p->tolimit = p->textsize * ARCDENSITY / 100;
if (p->tolimit < MINARCS)
p->tolimit = MINARCS;
else if (p->tolimit > MAXARCS)
p->tolimit = MAXARCS;
p->tossize = p->tolimit * sizeof(struct tostruct);
cp = (char *)malloc(p->kcountsize + p->fromssize + p->tossize,
M_GPROF, M_WAITOK | M_ZERO);
p->tos = (struct tostruct *)cp;
cp += p->tossize;
p->kcount = (HISTCOUNTER *)cp;
cp += p->kcountsize;
p->froms = (u_short *)cp;
p->histcounter_type = FUNCTION_ALIGNMENT / HISTFRACTION * NBBY;
#ifdef GUPROF
/* Signed counters. */
p->histcounter_type = -p->histcounter_type;
/* Initialize pointers to overhead counters. */
p->cputime_count = &KCOUNT(p, PC_TO_I(p, cputime));
p->mcount_count = &KCOUNT(p, PC_TO_I(p, mcount));
p->mexitcount_count = &KCOUNT(p, PC_TO_I(p, mexitcount));
/*
* Disable interrupts to avoid interference while we calibrate
* things.
*/
critical_enter();
/*
* Determine overheads.
* XXX this needs to be repeated for each useful timer/counter.
*/
cputime_overhead = 0;
startguprof(p);
for (i = 0; i < CALIB_SCALE; i++)
cputime_overhead += cputime();
empty_loop();
startguprof(p);
empty_loop();
empty_loop_time = cputime();
nullfunc_loop_profiled();
/*
* Start profiling. There won't be any normal function calls since
* interrupts are disabled, but we will call the profiling routines
* directly to determine their overheads.
*/
p->state = GMON_PROF_HIRES;
startguprof(p);
nullfunc_loop_profiled();
startguprof(p);
for (i = 0; i < CALIB_SCALE; i++)
MCOUNT_OVERHEAD(sys_profil);
mcount_overhead = KCOUNT(p, PC_TO_I(p, sys_profil));
startguprof(p);
for (i = 0; i < CALIB_SCALE; i++)
MEXITCOUNT_OVERHEAD();
MEXITCOUNT_OVERHEAD_GETLABEL(tmp_addr);
mexitcount_overhead = KCOUNT(p, PC_TO_I(p, tmp_addr));
p->state = GMON_PROF_OFF;
stopguprof(p);
critical_exit();
nullfunc_loop_profiled_time = 0;
for (tmp_addr = (uintfptr_t)nullfunc_loop_profiled;
tmp_addr < (uintfptr_t)nullfunc_loop_profiled_end;
tmp_addr += HISTFRACTION * sizeof(HISTCOUNTER))
nullfunc_loop_profiled_time += KCOUNT(p, PC_TO_I(p, tmp_addr));
#define CALIB_DOSCALE(count) (((count) + CALIB_SCALE / 3) / CALIB_SCALE)
#define c2n(count, freq) ((int)((count) * 1000000000LL / freq))
printf("cputime %d, empty_loop %d, nullfunc_loop_profiled %d, mcount %d, mexitcount %d\n",
CALIB_DOSCALE(c2n(cputime_overhead, p->profrate)),
CALIB_DOSCALE(c2n(empty_loop_time, p->profrate)),
CALIB_DOSCALE(c2n(nullfunc_loop_profiled_time, p->profrate)),
CALIB_DOSCALE(c2n(mcount_overhead, p->profrate)),
CALIB_DOSCALE(c2n(mexitcount_overhead, p->profrate)));
cputime_overhead -= empty_loop_time;
mcount_overhead -= empty_loop_time;
mexitcount_overhead -= empty_loop_time;
/*-
* Profiling overheads are determined by the times between the
* following events:
* MC1: mcount() is called
* MC2: cputime() (called from mcount()) latches the timer
* MC3: mcount() completes
* ME1: mexitcount() is called
* ME2: cputime() (called from mexitcount()) latches the timer
* ME3: mexitcount() completes.
* The times between the events vary slightly depending on instruction
* combination and cache misses, etc. Attempt to determine the
* minimum times. These can be subtracted from the profiling times
* without much risk of reducing the profiling times below what they
* would be when profiling is not configured. Abbreviate:
* ab = minimum time between MC1 and MC3
* a = minumum time between MC1 and MC2
* b = minimum time between MC2 and MC3
* cd = minimum time between ME1 and ME3
* c = minimum time between ME1 and ME2
* d = minimum time between ME2 and ME3.
* These satisfy the relations:
* ab <= mcount_overhead (just measured)
* a + b <= ab
* cd <= mexitcount_overhead (just measured)
* c + d <= cd
* a + d <= nullfunc_loop_profiled_time (just measured)
* a >= 0, b >= 0, c >= 0, d >= 0.
* Assume that ab and cd are equal to the minimums.
*/
p->cputime_overhead = CALIB_DOSCALE(cputime_overhead);
p->mcount_overhead = CALIB_DOSCALE(mcount_overhead - cputime_overhead);
p->mexitcount_overhead = CALIB_DOSCALE(mexitcount_overhead
- cputime_overhead);
nullfunc_loop_overhead = nullfunc_loop_profiled_time - empty_loop_time;
p->mexitcount_post_overhead = CALIB_DOSCALE((mcount_overhead
- nullfunc_loop_overhead)
/ 4);
p->mexitcount_pre_overhead = p->mexitcount_overhead
+ p->cputime_overhead
- p->mexitcount_post_overhead;
p->mcount_pre_overhead = CALIB_DOSCALE(nullfunc_loop_overhead)
- p->mexitcount_post_overhead;
p->mcount_post_overhead = p->mcount_overhead
+ p->cputime_overhead
- p->mcount_pre_overhead;
printf(
"Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d nsec\n",
c2n(p->cputime_overhead, p->profrate),
c2n(p->mcount_overhead, p->profrate),
c2n(p->mcount_pre_overhead, p->profrate),
c2n(p->mcount_post_overhead, p->profrate),
c2n(p->cputime_overhead, p->profrate),
c2n(p->mexitcount_overhead, p->profrate),
c2n(p->mexitcount_pre_overhead, p->profrate),
c2n(p->mexitcount_post_overhead, p->profrate));
printf(
"Profiling overheads: mcount: %d+%d, %d+%d; mexitcount: %d+%d, %d+%d cycles\n",
p->cputime_overhead, p->mcount_overhead,
p->mcount_pre_overhead, p->mcount_post_overhead,
p->cputime_overhead, p->mexitcount_overhead,
p->mexitcount_pre_overhead, p->mexitcount_post_overhead);
#endif /* GUPROF */
}
/*
* Return kernel profiling information.
*/
static int
sysctl_kern_prof(SYSCTL_HANDLER_ARGS)
{
int *name = (int *) arg1;
u_int namelen = arg2;
struct gmonparam *gp = &_gmonparam;
int error;
int state;
/* all sysctl names at this level are terminal */
if (namelen != 1)
return (ENOTDIR); /* overloaded */
switch (name[0]) {
case GPROF_STATE:
state = gp->state;
error = sysctl_handle_int(oidp, &state, 0, req);
if (error)
return (error);
if (!req->newptr)
return (0);
if (state == GMON_PROF_OFF) {
gp->state = state;
PROC_LOCK(&proc0);
stopprofclock(&proc0);
PROC_UNLOCK(&proc0);
stopguprof(gp);
} else if (state == GMON_PROF_ON) {
gp->state = GMON_PROF_OFF;
stopguprof(gp);
gp->profrate = profhz;
PROC_LOCK(&proc0);
startprofclock(&proc0);
PROC_UNLOCK(&proc0);
gp->state = state;
#ifdef GUPROF
} else if (state == GMON_PROF_HIRES) {
gp->state = GMON_PROF_OFF;
PROC_LOCK(&proc0);
stopprofclock(&proc0);
PROC_UNLOCK(&proc0);
startguprof(gp);
gp->state = state;
#endif
} else if (state != gp->state)
return (EINVAL);
return (0);
case GPROF_COUNT:
return (sysctl_handle_opaque(oidp,
gp->kcount, gp->kcountsize, req));
case GPROF_FROMS:
return (sysctl_handle_opaque(oidp,
gp->froms, gp->fromssize, req));
case GPROF_TOS:
return (sysctl_handle_opaque(oidp,
gp->tos, gp->tossize, req));
case GPROF_GMONPARAM:
return (sysctl_handle_opaque(oidp, gp, sizeof *gp, req));
default:
return (EOPNOTSUPP);
}
/* NOTREACHED */
}
static SYSCTL_NODE(_kern, KERN_PROF, prof, CTLFLAG_RW, sysctl_kern_prof, "");
#endif /* GPROF */
/*
* Profiling system call.
*
* The scale factor is a fixed point number with 16 bits of fraction, so that
* 1.0 is represented as 0x10000. A scale factor of 0 turns off profiling.
*/
#ifndef _SYS_SYSPROTO_H_
struct profil_args {
caddr_t samples;
size_t size;
size_t offset;
u_int scale;
};
#endif
/* ARGSUSED */
int
sys_profil(struct thread *td, struct profil_args *uap)
{
struct uprof *upp;
struct proc *p;
if (uap->scale > (1 << 16))
return (EINVAL);
p = td->td_proc;
if (uap->scale == 0) {
PROC_LOCK(p);
stopprofclock(p);
PROC_UNLOCK(p);
return (0);
}
PROC_LOCK(p);
upp = &td->td_proc->p_stats->p_prof;
PROC_PROFLOCK(p);
upp->pr_off = uap->offset;
upp->pr_scale = uap->scale;
upp->pr_base = uap->samples;
upp->pr_size = uap->size;
PROC_PROFUNLOCK(p);
startprofclock(p);
PROC_UNLOCK(p);
return (0);
}
/*
* Scale is a fixed-point number with the binary point 16 bits
* into the value, and is <= 1.0. pc is at most 32 bits, so the
* intermediate result is at most 48 bits.
*/
#define PC_TO_INDEX(pc, prof) \
((int)(((u_quad_t)((pc) - (prof)->pr_off) * \
(u_quad_t)((prof)->pr_scale)) >> 16) & ~1)
/*
* Collect user-level profiling statistics; called on a profiling tick,
* when a process is running in user-mode. This routine may be called
* from an interrupt context. We try to update the user profiling buffers
* cheaply with fuswintr() and suswintr(). If that fails, we revert to
* an AST that will vector us to trap() with a context in which copyin
* and copyout will work. Trap will then call addupc_task().
*
* Note that we may (rarely) not get around to the AST soon enough, and
* lose profile ticks when the next tick overwrites this one, but in this
* case the system is overloaded and the profile is probably already
* inaccurate.
*/
void
addupc_intr(struct thread *td, uintfptr_t pc, u_int ticks)
{
struct uprof *prof;
caddr_t addr;
u_int i;
int v;
if (ticks == 0)
return;
prof = &td->td_proc->p_stats->p_prof;
PROC_PROFLOCK(td->td_proc);
if (pc < prof->pr_off ||
(i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) {
PROC_PROFUNLOCK(td->td_proc);
return; /* out of range; ignore */
}
addr = prof->pr_base + i;
PROC_PROFUNLOCK(td->td_proc);
if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) {
td->td_profil_addr = pc;
td->td_profil_ticks = ticks;
td->td_pflags |= TDP_OWEUPC;
thread_lock(td);
td->td_flags |= TDF_ASTPENDING;
thread_unlock(td);
}
}
/*
* Much like before, but we can afford to take faults here. If the
* update fails, we simply turn off profiling.
*/
void
addupc_task(struct thread *td, uintfptr_t pc, u_int ticks)
{
struct proc *p = td->td_proc;
struct uprof *prof;
caddr_t addr;
u_int i;
u_short v;
int stop = 0;
if (ticks == 0)
return;
PROC_LOCK(p);
if (!(p->p_flag & P_PROFIL)) {
PROC_UNLOCK(p);
return;
}
p->p_profthreads++;
prof = &p->p_stats->p_prof;
PROC_PROFLOCK(p);
if (pc < prof->pr_off ||
(i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) {
PROC_PROFUNLOCK(p);
goto out;
}
addr = prof->pr_base + i;
PROC_PROFUNLOCK(p);
PROC_UNLOCK(p);
if (copyin(addr, &v, sizeof(v)) == 0) {
v += ticks;
if (copyout(&v, addr, sizeof(v)) == 0) {
PROC_LOCK(p);
goto out;
}
}
stop = 1;
PROC_LOCK(p);
out:
if (--p->p_profthreads == 0) {
if (p->p_flag & P_STOPPROF) {
wakeup(&p->p_profthreads);
p->p_flag &= ~P_STOPPROF;
stop = 0;
}
}
if (stop)
stopprofclock(p);
PROC_UNLOCK(p);
}