Index: head/sys/kern/kern_clock.c =================================================================== --- head/sys/kern/kern_clock.c (revision 333760) +++ head/sys/kern/kern_clock.c (revision 333761) @@ -1,895 +1,898 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1991, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, 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. * 3. 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. * * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_kdb.h" #include "opt_device_polling.h" #include "opt_hwpmc_hooks.h" #include "opt_ntp.h" #include "opt_watchdog.h" #include #include #include +#include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef GPROF #include #endif #ifdef HWPMC_HOOKS #include PMC_SOFT_DEFINE( , , clock, hard); PMC_SOFT_DEFINE( , , clock, stat); PMC_SOFT_DEFINE_EX( , , clock, prof, \ cpu_startprofclock, cpu_stopprofclock); #endif #ifdef DEVICE_POLLING extern void hardclock_device_poll(void); #endif /* DEVICE_POLLING */ static void initclocks(void *dummy); SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL); /* Spin-lock protecting profiling statistics. */ static struct mtx time_lock; SDT_PROVIDER_DECLARE(sched); SDT_PROBE_DEFINE2(sched, , , tick, "struct thread *", "struct proc *"); static int sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS) { int error; long cp_time[CPUSTATES]; #ifdef SCTL_MASK32 int i; unsigned int cp_time32[CPUSTATES]; #endif read_cpu_time(cp_time); #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { if (!req->oldptr) return SYSCTL_OUT(req, 0, sizeof(cp_time32)); for (i = 0; i < CPUSTATES; i++) cp_time32[i] = (unsigned int)cp_time[i]; error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); } else #endif { if (!req->oldptr) return SYSCTL_OUT(req, 0, sizeof(cp_time)); error = SYSCTL_OUT(req, cp_time, sizeof(cp_time)); } return error; } SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE, 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics"); static long empty[CPUSTATES]; static int sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS) { struct pcpu *pcpu; int error; int c; long *cp_time; #ifdef SCTL_MASK32 unsigned int cp_time32[CPUSTATES]; int i; #endif if (!req->oldptr) { #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1)); else #endif return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1)); } for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) { if (!CPU_ABSENT(c)) { pcpu = pcpu_find(c); cp_time = pcpu->pc_cp_time; } else { cp_time = empty; } #ifdef SCTL_MASK32 if (req->flags & SCTL_MASK32) { for (i = 0; i < CPUSTATES; i++) cp_time32[i] = (unsigned int)cp_time[i]; error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); } else #endif error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES); } return error; } SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE, 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics"); #ifdef DEADLKRES static const char *blessed[] = { "getblk", "so_snd_sx", "so_rcv_sx", NULL }; static int slptime_threshold = 1800; static int blktime_threshold = 900; static int sleepfreq = 3; static void deadlkres(void) { struct proc *p; struct thread *td; void *wchan; int blkticks, i, slpticks, slptype, tryl, tticks; tryl = 0; for (;;) { blkticks = blktime_threshold * hz; slpticks = slptime_threshold * hz; /* * Avoid to sleep on the sx_lock in order to avoid a possible * priority inversion problem leading to starvation. * If the lock can't be held after 100 tries, panic. */ if (!sx_try_slock(&allproc_lock)) { if (tryl > 100) panic("%s: possible deadlock detected on allproc_lock\n", __func__); tryl++; pause("allproc", sleepfreq * hz); continue; } tryl = 0; FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW) { PROC_UNLOCK(p); continue; } FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (TD_ON_LOCK(td)) { /* * The thread should be blocked on a * turnstile, simply check if the * turnstile channel is in good state. */ MPASS(td->td_blocked != NULL); tticks = ticks - td->td_blktick; thread_unlock(td); if (tticks > blkticks) { /* * Accordingly with provided * thresholds, this thread is * stuck for too long on a * turnstile. */ PROC_UNLOCK(p); sx_sunlock(&allproc_lock); panic("%s: possible deadlock detected for %p, blocked for %d ticks\n", __func__, td, tticks); } } else if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) { /* * Check if the thread is sleeping on a * lock, otherwise skip the check. * Drop the thread lock in order to * avoid a LOR with the sleepqueue * spinlock. */ wchan = td->td_wchan; tticks = ticks - td->td_slptick; thread_unlock(td); slptype = sleepq_type(wchan); if ((slptype == SLEEPQ_SX || slptype == SLEEPQ_LK) && tticks > slpticks) { /* * Accordingly with provided * thresholds, this thread is * stuck for too long on a * sleepqueue. * However, being on a * sleepqueue, we might still * check for the blessed * list. */ tryl = 0; for (i = 0; blessed[i] != NULL; i++) { if (!strcmp(blessed[i], td->td_wmesg)) { tryl = 1; break; } } if (tryl != 0) { tryl = 0; continue; } PROC_UNLOCK(p); sx_sunlock(&allproc_lock); panic("%s: possible deadlock detected for %p, blocked for %d ticks\n", __func__, td, tticks); } } else thread_unlock(td); } PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); /* Sleep for sleepfreq seconds. */ pause("-", sleepfreq * hz); } } static struct kthread_desc deadlkres_kd = { "deadlkres", deadlkres, (struct thread **)NULL }; SYSINIT(deadlkres, SI_SUB_CLOCKS, SI_ORDER_ANY, kthread_start, &deadlkres_kd); static SYSCTL_NODE(_debug, OID_AUTO, deadlkres, CTLFLAG_RW, 0, "Deadlock resolver"); SYSCTL_INT(_debug_deadlkres, OID_AUTO, slptime_threshold, CTLFLAG_RW, &slptime_threshold, 0, "Number of seconds within is valid to sleep on a sleepqueue"); SYSCTL_INT(_debug_deadlkres, OID_AUTO, blktime_threshold, CTLFLAG_RW, &blktime_threshold, 0, "Number of seconds within is valid to block on a turnstile"); SYSCTL_INT(_debug_deadlkres, OID_AUTO, sleepfreq, CTLFLAG_RW, &sleepfreq, 0, "Number of seconds between any deadlock resolver thread run"); #endif /* DEADLKRES */ void read_cpu_time(long *cp_time) { struct pcpu *pc; int i, j; /* Sum up global cp_time[]. */ bzero(cp_time, sizeof(long) * CPUSTATES); CPU_FOREACH(i) { pc = pcpu_find(i); for (j = 0; j < CPUSTATES; j++) cp_time[j] += pc->pc_cp_time[j]; } } #include static int watchdog_ticks; static int watchdog_enabled; static void watchdog_fire(void); static void watchdog_config(void *, u_int, int *); static void watchdog_attach(void) { EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0); } /* * Clock handling routines. * * This code is written to operate with two timers that run independently of * each other. * * The main timer, running hz times per second, is used to trigger interval * timers, timeouts and rescheduling as needed. * * The second timer handles kernel and user profiling, * and does resource use estimation. If the second timer is programmable, * it is randomized to avoid aliasing between the two clocks. For example, * the randomization prevents an adversary from always giving up the cpu * just before its quantum expires. Otherwise, it would never accumulate * cpu ticks. The mean frequency of the second timer is stathz. * * If no second timer exists, stathz will be zero; in this case we drive * profiling and statistics off the main clock. This WILL NOT be accurate; * do not do it unless absolutely necessary. * * The statistics clock may (or may not) be run at a higher rate while * profiling. This profile clock runs at profhz. We require that profhz * be an integral multiple of stathz. * * If the statistics clock is running fast, it must be divided by the ratio * profhz/stathz for statistics. (For profiling, every tick counts.) * * Time-of-day is maintained using a "timecounter", which may or may * not be related to the hardware generating the above mentioned * interrupts. */ int stathz; int profhz; int profprocs; volatile int ticks; int psratio; static DPCPU_DEFINE(int, pcputicks); /* Per-CPU version of ticks. */ #ifdef DEVICE_POLLING static int devpoll_run = 0; #endif /* * Initialize clock frequencies and start both clocks running. */ /* ARGSUSED*/ static void initclocks(void *dummy) { int i; /* * Set divisors to 1 (normal case) and let the machine-specific * code do its bit. */ mtx_init(&time_lock, "time lock", NULL, MTX_DEF); cpu_initclocks(); /* * Compute profhz/stathz, and fix profhz if needed. */ i = stathz ? stathz : hz; if (profhz == 0) profhz = i; psratio = profhz / i; #ifdef SW_WATCHDOG /* Enable hardclock watchdog now, even if a hardware watchdog exists. */ watchdog_attach(); #else /* Volunteer to run a software watchdog. */ if (wdog_software_attach == NULL) wdog_software_attach = watchdog_attach; #endif } /* * Each time the real-time timer fires, this function is called on all CPUs. * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only * the other CPUs in the system need to call this function. */ void hardclock_cpu(int usermode) { struct pstats *pstats; struct thread *td = curthread; struct proc *p = td->td_proc; int flags; /* * Run current process's virtual and profile time, as needed. */ pstats = p->p_stats; flags = 0; if (usermode && timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) { PROC_ITIMLOCK(p); if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) flags |= TDF_ALRMPEND | TDF_ASTPENDING; PROC_ITIMUNLOCK(p); } if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) { PROC_ITIMLOCK(p); if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) flags |= TDF_PROFPEND | TDF_ASTPENDING; PROC_ITIMUNLOCK(p); } thread_lock(td); td->td_flags |= flags; thread_unlock(td); #ifdef HWPMC_HOOKS if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); if (td->td_intr_frame != NULL) PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame); #endif callout_process(sbinuptime()); + epoch_pcpu_poll(); } /* * The real-time timer, interrupting hz times per second. */ void hardclock(int usermode, uintfptr_t pc) { atomic_add_int(&ticks, 1); hardclock_cpu(usermode); tc_ticktock(1); cpu_tick_calibration(); /* * If no separate statistics clock is available, run it from here. * * XXX: this only works for UP */ if (stathz == 0) { profclock(usermode, pc); statclock(usermode); } #ifdef DEVICE_POLLING hardclock_device_poll(); /* this is very short and quick */ #endif /* DEVICE_POLLING */ if (watchdog_enabled > 0 && --watchdog_ticks <= 0) watchdog_fire(); } void hardclock_cnt(int cnt, int usermode) { struct pstats *pstats; struct thread *td = curthread; struct proc *p = td->td_proc; int *t = DPCPU_PTR(pcputicks); int flags, global, newticks; int i; /* * Update per-CPU and possibly global ticks values. */ *t += cnt; do { global = ticks; newticks = *t - global; if (newticks <= 0) { if (newticks < -1) *t = global - 1; newticks = 0; break; } } while (!atomic_cmpset_int(&ticks, global, *t)); /* * Run current process's virtual and profile time, as needed. */ pstats = p->p_stats; flags = 0; if (usermode && timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) { PROC_ITIMLOCK(p); if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick * cnt) == 0) flags |= TDF_ALRMPEND | TDF_ASTPENDING; PROC_ITIMUNLOCK(p); } if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) { PROC_ITIMLOCK(p); if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick * cnt) == 0) flags |= TDF_PROFPEND | TDF_ASTPENDING; PROC_ITIMUNLOCK(p); } if (flags != 0) { thread_lock(td); td->td_flags |= flags; thread_unlock(td); } #ifdef HWPMC_HOOKS if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); if (td->td_intr_frame != NULL) PMC_SOFT_CALL_TF( , , clock, hard, td->td_intr_frame); #endif /* We are in charge to handle this tick duty. */ if (newticks > 0) { tc_ticktock(newticks); #ifdef DEVICE_POLLING /* Dangerous and no need to call these things concurrently. */ if (atomic_cmpset_acq_int(&devpoll_run, 0, 1)) { /* This is very short and quick. */ hardclock_device_poll(); atomic_store_rel_int(&devpoll_run, 0); } #endif /* DEVICE_POLLING */ if (watchdog_enabled > 0) { i = atomic_fetchadd_int(&watchdog_ticks, -newticks); if (i > 0 && i <= newticks) watchdog_fire(); } } if (curcpu == CPU_FIRST()) cpu_tick_calibration(); + epoch_pcpu_poll(); } void hardclock_sync(int cpu) { int *t; KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu)); t = DPCPU_ID_PTR(cpu, pcputicks); *t = ticks; } /* * Compute number of ticks in the specified amount of time. */ int tvtohz(struct timeval *tv) { unsigned long ticks; long sec, usec; /* * If the number of usecs in the whole seconds part of the time * difference fits in a long, then the total number of usecs will * fit in an unsigned long. Compute the total and convert it to * ticks, rounding up and adding 1 to allow for the current tick * to expire. Rounding also depends on unsigned long arithmetic * to avoid overflow. * * Otherwise, if the number of ticks in the whole seconds part of * the time difference fits in a long, then convert the parts to * ticks separately and add, using similar rounding methods and * overflow avoidance. This method would work in the previous * case but it is slightly slower and assumes that hz is integral. * * Otherwise, round the time difference down to the maximum * representable value. * * If ints have 32 bits, then the maximum value for any timeout in * 10ms ticks is 248 days. */ sec = tv->tv_sec; usec = tv->tv_usec; if (usec < 0) { sec--; usec += 1000000; } if (sec < 0) { #ifdef DIAGNOSTIC if (usec > 0) { sec++; usec -= 1000000; } printf("tvotohz: negative time difference %ld sec %ld usec\n", sec, usec); #endif ticks = 1; } else if (sec <= LONG_MAX / 1000000) ticks = howmany(sec * 1000000 + (unsigned long)usec, tick) + 1; else if (sec <= LONG_MAX / hz) ticks = sec * hz + howmany((unsigned long)usec, tick) + 1; else ticks = LONG_MAX; if (ticks > INT_MAX) ticks = INT_MAX; return ((int)ticks); } /* * Start profiling on a process. * * Kernel profiling passes proc0 which never exits and hence * keeps the profile clock running constantly. */ void startprofclock(struct proc *p) { PROC_LOCK_ASSERT(p, MA_OWNED); if (p->p_flag & P_STOPPROF) return; if ((p->p_flag & P_PROFIL) == 0) { p->p_flag |= P_PROFIL; mtx_lock(&time_lock); if (++profprocs == 1) cpu_startprofclock(); mtx_unlock(&time_lock); } } /* * Stop profiling on a process. */ void stopprofclock(struct proc *p) { PROC_LOCK_ASSERT(p, MA_OWNED); if (p->p_flag & P_PROFIL) { if (p->p_profthreads != 0) { while (p->p_profthreads != 0) { p->p_flag |= P_STOPPROF; msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, "stopprof", 0); } } if ((p->p_flag & P_PROFIL) == 0) return; p->p_flag &= ~P_PROFIL; mtx_lock(&time_lock); if (--profprocs == 0) cpu_stopprofclock(); mtx_unlock(&time_lock); } } /* * Statistics clock. Updates rusage information and calls the scheduler * to adjust priorities of the active thread. * * This should be called by all active processors. */ void statclock(int usermode) { statclock_cnt(1, usermode); } void statclock_cnt(int cnt, int usermode) { struct rusage *ru; struct vmspace *vm; struct thread *td; struct proc *p; long rss; long *cp_time; td = curthread; p = td->td_proc; cp_time = (long *)PCPU_PTR(cp_time); if (usermode) { /* * Charge the time as appropriate. */ td->td_uticks += cnt; if (p->p_nice > NZERO) cp_time[CP_NICE] += cnt; else cp_time[CP_USER] += cnt; } else { /* * Came from kernel mode, so we were: * - handling an interrupt, * - doing syscall or trap work on behalf of the current * user process, or * - spinning in the idle loop. * Whichever it is, charge the time as appropriate. * Note that we charge interrupts to the current process, * regardless of whether they are ``for'' that process, * so that we know how much of its real time was spent * in ``non-process'' (i.e., interrupt) work. */ if ((td->td_pflags & TDP_ITHREAD) || td->td_intr_nesting_level >= 2) { td->td_iticks += cnt; cp_time[CP_INTR] += cnt; } else { td->td_pticks += cnt; td->td_sticks += cnt; if (!TD_IS_IDLETHREAD(td)) cp_time[CP_SYS] += cnt; else cp_time[CP_IDLE] += cnt; } } /* Update resource usage integrals and maximums. */ MPASS(p->p_vmspace != NULL); vm = p->p_vmspace; ru = &td->td_ru; ru->ru_ixrss += pgtok(vm->vm_tsize) * cnt; ru->ru_idrss += pgtok(vm->vm_dsize) * cnt; ru->ru_isrss += pgtok(vm->vm_ssize) * cnt; rss = pgtok(vmspace_resident_count(vm)); if (ru->ru_maxrss < rss) ru->ru_maxrss = rss; KTR_POINT2(KTR_SCHED, "thread", sched_tdname(td), "statclock", "prio:%d", td->td_priority, "stathz:%d", (stathz)?stathz:hz); SDT_PROBE2(sched, , , tick, td, td->td_proc); thread_lock_flags(td, MTX_QUIET); for ( ; cnt > 0; cnt--) sched_clock(td); thread_unlock(td); #ifdef HWPMC_HOOKS if (td->td_intr_frame != NULL) PMC_SOFT_CALL_TF( , , clock, stat, td->td_intr_frame); #endif } void profclock(int usermode, uintfptr_t pc) { profclock_cnt(1, usermode, pc); } void profclock_cnt(int cnt, int usermode, uintfptr_t pc) { struct thread *td; #ifdef GPROF struct gmonparam *g; uintfptr_t i; #endif td = curthread; if (usermode) { /* * Came from user mode; CPU was in user state. * If this process is being profiled, record the tick. * if there is no related user location yet, don't * bother trying to count it. */ if (td->td_proc->p_flag & P_PROFIL) addupc_intr(td, pc, cnt); } #ifdef GPROF else { /* * Kernel statistics are just like addupc_intr, only easier. */ g = &_gmonparam; if (g->state == GMON_PROF_ON && pc >= g->lowpc) { i = PC_TO_I(g, pc); if (i < g->textsize) { KCOUNT(g, i) += cnt; } } } #endif #ifdef HWPMC_HOOKS if (td->td_intr_frame != NULL) PMC_SOFT_CALL_TF( , , clock, prof, td->td_intr_frame); #endif } /* * Return information about system clocks. */ static int sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) { struct clockinfo clkinfo; /* * Construct clockinfo structure. */ bzero(&clkinfo, sizeof(clkinfo)); clkinfo.hz = hz; clkinfo.tick = tick; clkinfo.profhz = profhz; clkinfo.stathz = stathz ? stathz : hz; return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); } SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD|CTLFLAG_MPSAFE, 0, 0, sysctl_kern_clockrate, "S,clockinfo", "Rate and period of various kernel clocks"); static void watchdog_config(void *unused __unused, u_int cmd, int *error) { u_int u; u = cmd & WD_INTERVAL; if (u >= WD_TO_1SEC) { watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz; watchdog_enabled = 1; *error = 0; } else { watchdog_enabled = 0; } } /* * Handle a watchdog timeout by dumping interrupt information and * then either dropping to DDB or panicking. */ static void watchdog_fire(void) { int nintr; uint64_t inttotal; u_long *curintr; char *curname; curintr = intrcnt; curname = intrnames; inttotal = 0; nintr = sintrcnt / sizeof(u_long); printf("interrupt total\n"); while (--nintr >= 0) { if (*curintr) printf("%-12s %20lu\n", curname, *curintr); curname += strlen(curname) + 1; inttotal += *curintr++; } printf("Total %20ju\n", (uintmax_t)inttotal); #if defined(KDB) && !defined(KDB_UNATTENDED) kdb_backtrace(); kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout"); #else panic("watchdog timeout"); #endif } Index: head/sys/kern/subr_epoch.c =================================================================== --- head/sys/kern/subr_epoch.c (revision 333760) +++ head/sys/kern/subr_epoch.c (revision 333761) @@ -1,576 +1,593 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2018, Matthew Macy * * 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. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include +#include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_EPOCH, "epoch", "epoch based reclamation"); /* arbitrary --- needs benchmarking */ #define MAX_ADAPTIVE_SPIN 1000 +#define MAX_EPOCHS 64 -#define EPOCH_EXITING 0x1 #ifdef __amd64__ #define EPOCH_ALIGN CACHE_LINE_SIZE*2 #else #define EPOCH_ALIGN CACHE_LINE_SIZE #endif CTASSERT(sizeof(epoch_section_t) == sizeof(ck_epoch_section_t)); CTASSERT(sizeof(ck_epoch_entry_t) == sizeof(struct epoch_context)); SYSCTL_NODE(_kern, OID_AUTO, epoch, CTLFLAG_RW, 0, "epoch information"); SYSCTL_NODE(_kern_epoch, OID_AUTO, stats, CTLFLAG_RW, 0, "epoch stats"); -static int poll_intvl; -SYSCTL_INT(_kern_epoch, OID_AUTO, poll_intvl, CTLFLAG_RWTUN, - &poll_intvl, 0, "# of ticks to wait between garbage collecting deferred frees"); + /* Stats. */ static counter_u64_t block_count; SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, nblocked, CTLFLAG_RW, &block_count, "# of times a thread was in an epoch when epoch_wait was called"); static counter_u64_t migrate_count; SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, migrations, CTLFLAG_RW, &migrate_count, "# of times thread was migrated to another CPU in epoch_wait"); static counter_u64_t turnstile_count; SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, ncontended, CTLFLAG_RW, &turnstile_count, "# of times a thread was blocked on a lock in an epoch during an epoch_wait"); static counter_u64_t switch_count; SYSCTL_COUNTER_U64(_kern_epoch_stats, OID_AUTO, switches, CTLFLAG_RW, &switch_count, "# of times a thread voluntarily context switched in epoch_wait"); TAILQ_HEAD(threadlist, thread); CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry, ck_epoch_entry_container) typedef struct epoch_record { ck_epoch_record_t er_record; volatile struct threadlist er_tdlist; volatile uint32_t er_gen; uint32_t er_cpuid; } *epoch_record_t; struct epoch_pcpu_state { struct epoch_record eps_record; } __aligned(EPOCH_ALIGN); struct epoch { struct ck_epoch e_epoch __aligned(EPOCH_ALIGN); - struct grouptask e_gtask; - struct callout e_timer; - struct mtx e_lock; - int e_flags; - /* make sure that immutable data doesn't overlap with the gtask, callout, and mutex*/ struct epoch_pcpu_state *e_pcpu_dom[MAXMEMDOM] __aligned(EPOCH_ALIGN); counter_u64_t e_frees; uint64_t e_free_last; + int e_idx; struct epoch_pcpu_state *e_pcpu[0]; }; +epoch_t allepochs[MAX_EPOCHS]; + +static DPCPU_DEFINE(struct grouptask, cb_task); +static DPCPU_DEFINE(int, cb_count); + static __read_mostly int domcount[MAXMEMDOM]; static __read_mostly int domoffsets[MAXMEMDOM]; static __read_mostly int inited; +static __read_mostly int epoch_count; __read_mostly epoch_t global_epoch; +static __read_mostly epoch_t private_epoch; -static void epoch_call_task(void *context); +static void epoch_call_task(void *context __unused); #if defined(__powerpc64__) || defined(__powerpc__) || !defined(NUMA) static bool usedomains = false; #else static bool usedomains = true; #endif static void epoch_init(void *arg __unused) { - int domain, count; + int domain, count, cpu; - if (poll_intvl == 0) - poll_intvl = hz; - block_count = counter_u64_alloc(M_WAITOK); migrate_count = counter_u64_alloc(M_WAITOK); turnstile_count = counter_u64_alloc(M_WAITOK); switch_count = counter_u64_alloc(M_WAITOK); if (usedomains == false) goto done; count = domain = 0; domoffsets[0] = 0; for (domain = 0; domain < vm_ndomains; domain++) { domcount[domain] = CPU_COUNT(&cpuset_domain[domain]); if (bootverbose) printf("domcount[%d] %d\n", domain, domcount[domain]); } for (domain = 1; domain < vm_ndomains; domain++) domoffsets[domain] = domoffsets[domain-1] + domcount[domain-1]; for (domain = 0; domain < vm_ndomains; domain++) { if (domcount[domain] == 0) { usedomains = false; break; } } done: + CPU_FOREACH(cpu) { + GROUPTASK_INIT(DPCPU_ID_PTR(cpu, cb_task), 0, epoch_call_task, NULL); + taskqgroup_attach_cpu(qgroup_softirq, DPCPU_ID_PTR(cpu, cb_task), NULL, cpu, -1, "epoch call task"); + } inited = 1; global_epoch = epoch_alloc(); + private_epoch = epoch_alloc(); } SYSINIT(epoch, SI_SUB_TASKQ + 1, SI_ORDER_FIRST, epoch_init, NULL); static void epoch_init_numa(epoch_t epoch) { int domain, cpu_offset; struct epoch_pcpu_state *eps; epoch_record_t er; for (domain = 0; domain < vm_ndomains; domain++) { eps = malloc_domain(sizeof(*eps)*domcount[domain], M_EPOCH, domain, M_ZERO|M_WAITOK); epoch->e_pcpu_dom[domain] = eps; cpu_offset = domoffsets[domain]; for (int i = 0; i < domcount[domain]; i++, eps++) { epoch->e_pcpu[cpu_offset + i] = eps; er = &eps->eps_record; ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL); TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist); er->er_cpuid = cpu_offset + i; } } } static void epoch_init_legacy(epoch_t epoch) { struct epoch_pcpu_state *eps; epoch_record_t er; eps = malloc(sizeof(*eps)*mp_ncpus, M_EPOCH, M_ZERO|M_WAITOK); epoch->e_pcpu_dom[0] = eps; for (int i = 0; i < mp_ncpus; i++, eps++) { epoch->e_pcpu[i] = eps; er = &eps->eps_record; ck_epoch_register(&epoch->e_epoch, &er->er_record, NULL); TAILQ_INIT((struct threadlist *)(uintptr_t)&er->er_tdlist); er->er_cpuid = i; } } -static void -epoch_callout(void *arg) -{ - epoch_t epoch; - uint64_t frees; - - epoch = arg; - frees = counter_u64_fetch(epoch->e_frees); - /* pick some better value */ - if (frees - epoch->e_free_last > 10) { - GROUPTASK_ENQUEUE(&epoch->e_gtask); - epoch->e_free_last = frees; - } - if ((epoch->e_flags & EPOCH_EXITING) == 0) - callout_reset(&epoch->e_timer, poll_intvl, epoch_callout, epoch); -} - epoch_t epoch_alloc(void) { epoch_t epoch; if (__predict_false(!inited)) panic("%s called too early in boot", __func__); epoch = malloc(sizeof(struct epoch) + mp_ncpus*sizeof(void*), M_EPOCH, M_ZERO|M_WAITOK); ck_epoch_init(&epoch->e_epoch); epoch->e_frees = counter_u64_alloc(M_WAITOK); - mtx_init(&epoch->e_lock, "epoch callout", NULL, MTX_DEF); - callout_init_mtx(&epoch->e_timer, &epoch->e_lock, 0); - taskqgroup_config_gtask_init(epoch, &epoch->e_gtask, epoch_call_task, "epoch call task"); if (usedomains) epoch_init_numa(epoch); else epoch_init_legacy(epoch); - callout_reset(&epoch->e_timer, poll_intvl, epoch_callout, epoch); + MPASS(epoch_count < MAX_EPOCHS-2); + epoch->e_idx = epoch_count; + allepochs[epoch_count++] = epoch; return (epoch); } void epoch_free(epoch_t epoch) { int domain; #ifdef INVARIANTS struct epoch_pcpu_state *eps; int cpu; CPU_FOREACH(cpu) { eps = epoch->e_pcpu[cpu]; MPASS(TAILQ_EMPTY(&eps->eps_record.er_tdlist)); } #endif - mtx_lock(&epoch->e_lock); - epoch->e_flags |= EPOCH_EXITING; - mtx_unlock(&epoch->e_lock); + allepochs[epoch->e_idx] = NULL; + epoch_wait(private_epoch); /* * Execute any lingering callbacks */ - GROUPTASK_ENQUEUE(&epoch->e_gtask); - gtaskqueue_drain(epoch->e_gtask.gt_taskqueue, &epoch->e_gtask.gt_task); - callout_drain(&epoch->e_timer); - mtx_destroy(&epoch->e_lock); counter_u64_free(epoch->e_frees); - taskqgroup_config_gtask_deinit(&epoch->e_gtask); if (usedomains) for (domain = 0; domain < vm_ndomains; domain++) free_domain(epoch->e_pcpu_dom[domain], M_EPOCH); else free(epoch->e_pcpu_dom[0], M_EPOCH); free(epoch, M_EPOCH); } #define INIT_CHECK(epoch) \ do { \ if (__predict_false((epoch) == NULL)) \ return; \ } while (0) void epoch_enter_internal(epoch_t epoch, struct thread *td) { struct epoch_pcpu_state *eps; INIT_CHECK(epoch); critical_enter(); td->td_pre_epoch_prio = td->td_priority; eps = epoch->e_pcpu[curcpu]; #ifdef INVARIANTS MPASS(td->td_epochnest < UCHAR_MAX - 2); if (td->td_epochnest > 1) { struct thread *curtd; int found = 0; TAILQ_FOREACH(curtd, &eps->eps_record.er_tdlist, td_epochq) if (curtd == td) found = 1; KASSERT(found, ("recursing on a second epoch")); critical_exit(); return; } #endif TAILQ_INSERT_TAIL(&eps->eps_record.er_tdlist, td, td_epochq); sched_pin(); ck_epoch_begin(&eps->eps_record.er_record, (ck_epoch_section_t*)&td->td_epoch_section); critical_exit(); } + +static void +epoch_enter_private(ck_epoch_section_t *section) +{ + struct epoch_pcpu_state *eps; + + MPASS(curthread->td_critnest); + eps = private_epoch->e_pcpu[curcpu]; + ck_epoch_begin(&eps->eps_record.er_record, section); +} + void epoch_exit_internal(epoch_t epoch, struct thread *td) { struct epoch_pcpu_state *eps; - td = curthread; MPASS(td->td_epochnest == 0); INIT_CHECK(epoch); critical_enter(); eps = epoch->e_pcpu[curcpu]; ck_epoch_end(&eps->eps_record.er_record, (ck_epoch_section_t*)&td->td_epoch_section); TAILQ_REMOVE(&eps->eps_record.er_tdlist, td, td_epochq); eps->eps_record.er_gen++; sched_unpin(); if (__predict_false(td->td_pre_epoch_prio != td->td_priority)) { thread_lock(td); sched_prio(td, td->td_pre_epoch_prio); thread_unlock(td); } critical_exit(); } +static void +epoch_exit_private(ck_epoch_section_t *section) +{ + struct epoch_pcpu_state *eps; + + MPASS(curthread->td_critnest); + eps = private_epoch->e_pcpu[curcpu]; + ck_epoch_end(&eps->eps_record.er_record, section); +} + /* * epoch_block_handler is a callback from the ck code when another thread is * currently in an epoch section. */ static void epoch_block_handler(struct ck_epoch *global __unused, ck_epoch_record_t *cr, void *arg __unused) { epoch_record_t record; struct epoch_pcpu_state *eps; struct thread *td, *tdwait, *owner; struct turnstile *ts; struct lock_object *lock; int spincount, gen; eps = arg; record = __containerof(cr, struct epoch_record, er_record); td = curthread; spincount = 0; counter_u64_add(block_count, 1); if (record->er_cpuid != curcpu) { /* * If the head of the list is running, we can wait for it * to remove itself from the list and thus save us the * overhead of a migration */ if ((tdwait = TAILQ_FIRST(&record->er_tdlist)) != NULL && TD_IS_RUNNING(tdwait)) { gen = record->er_gen; thread_unlock(td); do { cpu_spinwait(); } while (tdwait == TAILQ_FIRST(&record->er_tdlist) && gen == record->er_gen && TD_IS_RUNNING(tdwait) && spincount++ < MAX_ADAPTIVE_SPIN); thread_lock(td); return; } /* * Being on the same CPU as that of the record on which * we need to wait allows us access to the thread * list associated with that CPU. We can then examine the * oldest thread in the queue and wait on its turnstile * until it resumes and so on until a grace period * elapses. * */ counter_u64_add(migrate_count, 1); sched_bind(td, record->er_cpuid); /* * At this point we need to return to the ck code * to scan to see if a grace period has elapsed. * We can't move on to check the thread list, because * in the meantime new threads may have arrived that * in fact belong to a different epoch. */ return; } /* * Try to find a thread in an epoch section on this CPU * waiting on a turnstile. Otherwise find the lowest * priority thread (highest prio value) and drop our priority * to match to allow it to run. */ TAILQ_FOREACH(tdwait, &record->er_tdlist, td_epochq) { /* * Propagate our priority to any other waiters to prevent us * from starving them. They will have their original priority * restore on exit from epoch_wait(). */ if (!TD_IS_INHIBITED(tdwait) && tdwait->td_priority > td->td_priority) { critical_enter(); thread_unlock(td); thread_lock(tdwait); sched_prio(tdwait, td->td_priority); thread_unlock(tdwait); thread_lock(td); critical_exit(); } if (TD_IS_INHIBITED(tdwait) && TD_ON_LOCK(tdwait) && ((ts = tdwait->td_blocked) != NULL)) { /* * We unlock td to allow turnstile_wait to reacquire the * the thread lock. Before unlocking it we enter a critical * section to prevent preemption after we reenable interrupts * by dropping the thread lock in order to prevent tdwait * from getting to run. */ critical_enter(); thread_unlock(td); owner = turnstile_lock(ts, &lock); /* * The owner pointer indicates that the lock succeeded. Only * in case we hold the lock and the turnstile we locked is still * the one that tdwait is blocked on can we continue. Otherwise * The turnstile pointer has been changed out from underneath * us, as in the case where the lock holder has signalled tdwait, * and we need to continue. */ if (owner != NULL && ts == tdwait->td_blocked) { MPASS(TD_IS_INHIBITED(tdwait) && TD_ON_LOCK(tdwait)); critical_exit(); turnstile_wait(ts, owner, tdwait->td_tsqueue); counter_u64_add(turnstile_count, 1); thread_lock(td); return; } else if (owner != NULL) turnstile_unlock(ts, lock); thread_lock(td); critical_exit(); KASSERT(td->td_locks == 0, ("%d locks held", td->td_locks)); } } /* * We didn't find any threads actually blocked on a lock * so we have nothing to do except context switch away. */ counter_u64_add(switch_count, 1); mi_switch(SW_VOL | SWT_RELINQUISH, NULL); /* * Release the thread lock while yielding to * allow other threads to acquire the lock * pointed to by TDQ_LOCKPTR(td). Else a * deadlock like situation might happen. (HPS) */ thread_unlock(td); thread_lock(td); } void epoch_wait(epoch_t epoch) { struct thread *td; int was_bound; int old_cpu; int old_pinned; u_char old_prio; #ifdef INVARIANTS int locks; locks = curthread->td_locks; #endif INIT_CHECK(epoch); WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "epoch_wait() can sleep"); td = curthread; KASSERT(td->td_epochnest == 0, ("epoch_wait() in the middle of an epoch section")); thread_lock(td); DROP_GIANT(); old_cpu = PCPU_GET(cpuid); old_pinned = td->td_pinned; old_prio = td->td_priority; was_bound = sched_is_bound(td); sched_unbind(td); td->td_pinned = 0; sched_bind(td, old_cpu); ck_epoch_synchronize_wait(&epoch->e_epoch, epoch_block_handler, NULL); /* restore CPU binding, if any */ if (was_bound != 0) { sched_bind(td, old_cpu); } else { /* get thread back to initial CPU, if any */ if (old_pinned != 0) sched_bind(td, old_cpu); sched_unbind(td); } /* restore pinned after bind */ td->td_pinned = old_pinned; /* restore thread priority */ sched_prio(td, old_prio); thread_unlock(td); PICKUP_GIANT(); KASSERT(td->td_locks == locks, ("%d residual locks held", td->td_locks - locks)); } void epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t)) { struct epoch_pcpu_state *eps; ck_epoch_entry_t *cb; cb = (void *)ctx; MPASS(callback); /* too early in boot to have epoch set up */ if (__predict_false(epoch == NULL)) goto boottime; counter_u64_add(epoch->e_frees, 1); critical_enter(); + *DPCPU_PTR(cb_count) += 1; eps = epoch->e_pcpu[curcpu]; ck_epoch_call(&eps->eps_record.er_record, cb, (ck_epoch_cb_t*)callback); critical_exit(); return; boottime: callback(ctx); } + static void -epoch_call_task(void *context) +epoch_call_task(void *arg __unused) { - struct epoch_pcpu_state *eps; + ck_stack_entry_t *cursor, *head, *next; + ck_epoch_record_t *record; + ck_epoch_section_t section; epoch_t epoch; - struct thread *td; - ck_stack_entry_t *cursor; - ck_stack_t deferred; - int cpu; + ck_stack_t cb_stack; + int i, npending, total; - epoch = context; - td = curthread; - ck_stack_init(&deferred); - thread_lock(td); - CPU_FOREACH(cpu) { - sched_bind(td, cpu); - eps = epoch->e_pcpu[cpu]; - ck_epoch_poll_deferred(&eps->eps_record.er_record, &deferred); + ck_stack_init(&cb_stack); + critical_enter(); + epoch_enter_private(§ion); + for (total = i = 0; i < epoch_count; i++) { + if (__predict_false((epoch = allepochs[i]) == NULL)) + continue; + record = &epoch->e_pcpu[curcpu]->eps_record.er_record; + if ((npending = record->n_pending) == 0) + continue; + ck_epoch_poll_deferred(record, &cb_stack); + total += npending - record->n_pending; } - sched_unbind(td); - thread_unlock(td); - while((cursor = ck_stack_pop_npsc(&deferred)) != NULL) { + epoch_exit_private(§ion); + *DPCPU_PTR(cb_count) -= total; + critical_exit(); + + head = ck_stack_batch_pop_npsc(&cb_stack); + for (cursor = head; cursor != NULL; cursor = next) { struct ck_epoch_entry *entry = ck_epoch_entry_container(cursor); + next = CK_STACK_NEXT(cursor); entry->function(entry); } +} + +void +epoch_pcpu_poll(void) +{ + + if (DPCPU_GET(cb_count)) + GROUPTASK_ENQUEUE(DPCPU_PTR(cb_task)); } int in_epoch(void) { return (curthread->td_epochnest != 0); } Index: head/sys/sys/epoch.h =================================================================== --- head/sys/sys/epoch.h (revision 333760) +++ head/sys/sys/epoch.h (revision 333761) @@ -1,79 +1,80 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2018, Matthew Macy * * 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. * * $FreeBSD$ */ #ifndef _SYS_EPOCH_H_ #define _SYS_EPOCH_H_ #include #include struct epoch; typedef struct epoch *epoch_t; extern epoch_t global_epoch; struct epoch_context { void *data[2]; } __aligned(sizeof(void *)); typedef struct epoch_context *epoch_context_t; epoch_t epoch_alloc(void); void epoch_free(epoch_t epoch); void epoch_enter_internal(epoch_t epoch, struct thread *td); void epoch_exit_internal(epoch_t epoch, struct thread *td); void epoch_wait(epoch_t epoch); void epoch_call(epoch_t epoch, epoch_context_t ctx, void (*callback) (epoch_context_t)); +void epoch_pcpu_poll(void); int in_epoch(void); static __inline void epoch_enter(epoch_t epoch) { struct thread *td; int nesting; td = curthread; nesting = td->td_epochnest++; #ifndef INVARIANTS if (nesting == 0) #endif epoch_enter_internal(epoch, td); } static __inline void epoch_exit(epoch_t epoch) { struct thread *td; td = curthread; MPASS(td->td_epochnest); if (td->td_epochnest-- == 1) epoch_exit_internal(epoch, td); } #endif