diff options
Diffstat (limited to 'sys/kern')
| -rw-r--r-- | sys/kern/init_main.c | 1 | ||||
| -rw-r--r-- | sys/kern/kern_clock.c | 55 | ||||
| -rw-r--r-- | sys/kern/kern_exec.c | 3 | ||||
| -rw-r--r-- | sys/kern/kern_exit.c | 18 | ||||
| -rw-r--r-- | sys/kern/kern_fork.c | 2 | ||||
| -rw-r--r-- | sys/kern/kern_kse.c | 1159 | ||||
| -rw-r--r-- | sys/kern/kern_lock.c | 40 | ||||
| -rw-r--r-- | sys/kern/kern_resource.c | 134 | ||||
| -rw-r--r-- | sys/kern/kern_sig.c | 6 | ||||
| -rw-r--r-- | sys/kern/kern_switch.c | 234 | ||||
| -rw-r--r-- | sys/kern/kern_thread.c | 1159 | ||||
| -rw-r--r-- | sys/kern/subr_prof.c | 69 | ||||
| -rw-r--r-- | sys/kern/subr_trap.c | 55 | ||||
| -rw-r--r-- | sys/kern/subr_witness.c | 2 |
14 files changed, 1548 insertions, 1389 deletions
diff --git a/sys/kern/init_main.c b/sys/kern/init_main.c index 74424cc..4fdf604 100644 --- a/sys/kern/init_main.c +++ b/sys/kern/init_main.c @@ -379,7 +379,6 @@ proc0_init(void *dummy __unused) ke->ke_oncpu = 0; ke->ke_state = KES_THREAD; ke->ke_thread = td; - ke->ke_owner = td; p->p_peers = 0; p->p_leader = p; diff --git a/sys/kern/kern_clock.c b/sys/kern/kern_clock.c index c5e1b4a..ab2c9ee 100644 --- a/sys/kern/kern_clock.c +++ b/sys/kern/kern_clock.c @@ -320,6 +320,10 @@ startprofclock(p) * cover psdiv, etc. as well. */ mtx_lock_spin(&sched_lock); + if (p->p_sflag & PS_STOPPROF) { + mtx_unlock_spin(&sched_lock); + return; + } if ((p->p_sflag & PS_PROFIL) == 0) { p->p_sflag |= PS_PROFIL; if (++profprocs == 1 && stathz != 0) { @@ -341,9 +345,19 @@ stopprofclock(p) { int s; + PROC_LOCK_ASSERT(p, MA_OWNED); + +retry: mtx_lock_spin(&sched_lock); if (p->p_sflag & PS_PROFIL) { - p->p_sflag &= ~PS_PROFIL; + if (p->p_profthreads) { + p->p_sflag |= PS_STOPPROF; + mtx_unlock_spin(&sched_lock); + msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, + "stopprof", NULL); + goto retry; + } + p->p_sflag &= ~(PS_PROFIL|PS_STOPPROF); if (--profprocs == 0 && stathz != 0) { s = splstatclock(); psdiv = pscnt = 1; @@ -363,10 +377,7 @@ stopprofclock(p) * this function's relationship to statclock. */ void -statclock_process(ke, pc, user) - struct kse *ke; - register_t pc; - int user; +statclock_process(struct thread *td, register_t pc, int user) { #ifdef GPROF struct gmonparam *g; @@ -376,27 +387,31 @@ statclock_process(ke, pc, user) long rss; struct rusage *ru; struct vmspace *vm; - struct proc *p = ke->ke_proc; - struct thread *td = ke->ke_thread; /* current thread */ + struct proc *p = td->td_proc; - KASSERT(ke == curthread->td_kse, ("statclock_process: td != curthread")); mtx_assert(&sched_lock, MA_OWNED); if (user) { /* * Came from user mode; CPU was in user state. * If this process is being profiled, record the tick. */ - if (p->p_sflag & PS_PROFIL) - addupc_intr(ke, pc, 1); + if (p->p_sflag & PS_PROFIL) { + /* Only when thread is not in transition */ + if (!(td->td_flags & TDF_UPCALLING)) + addupc_intr(td, pc, 1); + } if (pscnt < psdiv) return; /* * Charge the time as appropriate. */ if (p->p_flag & P_KSES) - thread_add_ticks_intr(1, 1); - ke->ke_uticks++; - if (ke->ke_ksegrp->kg_nice > NZERO) + thread_statclock(1); + /* + td->td_uticks++; + */ + p->p_uticks++; + if (td->td_ksegrp->kg_nice > NZERO) cp_time[CP_NICE]++; else cp_time[CP_USER]++; @@ -429,12 +444,16 @@ statclock_process(ke, pc, user) * in ``non-process'' (i.e., interrupt) work. */ if ((td->td_ithd != NULL) || td->td_intr_nesting_level >= 2) { - ke->ke_iticks++; + p->p_iticks++; + /* + td->td_iticks++; + */ cp_time[CP_INTR]++; } else { if (p->p_flag & P_KSES) - thread_add_ticks_intr(0, 1); - ke->ke_sticks++; + thread_statclock(0); + td->td_sticks++; + p->p_sticks++; if (p != PCPU_GET(idlethread)->td_proc) cp_time[CP_SYS]++; else @@ -442,7 +461,7 @@ statclock_process(ke, pc, user) } } - sched_clock(ke->ke_thread); + sched_clock(td); /* Update resource usage integrals and maximums. */ if ((pstats = p->p_stats) != NULL && @@ -472,7 +491,7 @@ statclock(frame) mtx_lock_spin_flags(&sched_lock, MTX_QUIET); if (--pscnt == 0) pscnt = psdiv; - statclock_process(curthread->td_kse, CLKF_PC(frame), CLKF_USERMODE(frame)); + statclock_process(curthread, CLKF_PC(frame), CLKF_USERMODE(frame)); mtx_unlock_spin_flags(&sched_lock, MTX_QUIET); } diff --git a/sys/kern/kern_exec.c b/sys/kern/kern_exec.c index 33a0764..b6d77d2 100644 --- a/sys/kern/kern_exec.c +++ b/sys/kern/kern_exec.c @@ -210,10 +210,7 @@ kern_execve(td, fname, argv, envv, mac_p) * so unset the associated flags and lose KSE mode. */ p->p_flag &= ~P_KSES; - td->td_flags &= ~TDF_UNBOUND; td->td_mailbox = NULL; - td->td_kse->ke_mailbox = NULL; - td->td_kse->ke_flags &= ~KEF_DOUPCALL; thread_single_end(); } p->p_flag |= P_INEXEC; diff --git a/sys/kern/kern_exit.c b/sys/kern/kern_exit.c index 0b2c2e8..ce9a18c 100644 --- a/sys/kern/kern_exit.c +++ b/sys/kern/kern_exit.c @@ -147,7 +147,7 @@ exit1(td, rv) } /* - * XXXXKSE: MUST abort all other threads before proceeding past here. + * XXXKSE: MUST abort all other threads before proceeding past here. */ PROC_LOCK(p); if (p->p_flag & P_KSES) { @@ -156,17 +156,6 @@ exit1(td, rv) * if so, act apropriatly, (exit or suspend); */ thread_suspend_check(0); - /* - * Here is a trick.. - * We need to free up our KSE to process other threads - * so that we can safely set the UNBOUND flag - * (whether or not we have a mailbox) as we are NEVER - * going to return to the user. - * The flag will not be set yet if we are exiting - * because of a signal, pagefault, or similar - * (or even an exit(2) from the UTS). - */ - td->td_flags |= TDF_UNBOUND; /* * Kill off the other threads. This requires @@ -192,7 +181,6 @@ exit1(td, rv) * Turn off threading support. */ p->p_flag &= ~P_KSES; - td->td_flags &= ~TDF_UNBOUND; thread_single_end(); /* Don't need this any more. */ } /* @@ -237,8 +225,10 @@ exit1(td, rv) */ TAILQ_FOREACH(ep, &exit_list, next) (*ep->function)(p); - + + PROC_LOCK(p); stopprofclock(p); + PROC_UNLOCK(p); MALLOC(p->p_ru, struct rusage *, sizeof(struct rusage), M_ZOMBIE, 0); diff --git a/sys/kern/kern_fork.c b/sys/kern/kern_fork.c index f84afa8..6c896eb 100644 --- a/sys/kern/kern_fork.c +++ b/sys/kern/kern_fork.c @@ -492,9 +492,7 @@ again: /* Set up the thread as an active thread (as if runnable). */ ke2->ke_state = KES_THREAD; ke2->ke_thread = td2; - ke2->ke_owner = td2; td2->td_kse = ke2; - td2->td_flags &= ~TDF_UNBOUND; /* For the rest of this syscall. */ /* * Duplicate sub-structures as needed. diff --git a/sys/kern/kern_kse.c b/sys/kern/kern_kse.c index 78bce30..883f4d1 100644 --- a/sys/kern/kern_kse.c +++ b/sys/kern/kern_kse.c @@ -63,6 +63,7 @@ static uma_zone_t ksegrp_zone; static uma_zone_t kse_zone; static uma_zone_t thread_zone; +static uma_zone_t upcall_zone; /* DEBUG ONLY */ SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation"); @@ -78,16 +79,52 @@ static int max_groups_per_proc = 5; SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW, &max_groups_per_proc, 0, "Limit on thread groups per proc"); +static int virtual_cpu; + #define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start)) -struct threadqueue zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); +TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); TAILQ_HEAD(, kse) zombie_kses = TAILQ_HEAD_INITIALIZER(zombie_kses); TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps); -struct mtx zombie_thread_lock; -MTX_SYSINIT(zombie_thread_lock, &zombie_thread_lock, - "zombie_thread_lock", MTX_SPIN); +TAILQ_HEAD(, kse_upcall) zombie_upcalls = + TAILQ_HEAD_INITIALIZER(zombie_upcalls); +struct mtx kse_zombie_lock; +MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN); static void kse_purge(struct proc *p, struct thread *td); +static void kse_purge_group(struct thread *td); +static int thread_update_usr_ticks(struct thread *td); +static int thread_update_sys_ticks(struct thread *td); +static void thread_alloc_spare(struct thread *td, struct thread *spare); + +static int +sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS) +{ + int error, new_val; + int def_val; + +#ifdef SMP + def_val = mp_ncpus; +#else + def_val = 1; +#endif + if (virtual_cpu == 0) + new_val = def_val; + else + new_val = virtual_cpu; + error = sysctl_handle_int(oidp, &new_val, 0, req); + if (error != 0 || req->newptr == NULL) + return (error); + if (new_val < 0) + return (EINVAL); + virtual_cpu = new_val; + return (0); +} + +/* DEBUG ONLY */ +SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW, + 0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I", + "debug virtual cpus"); /* * Prepare a thread for use. @@ -99,7 +136,6 @@ thread_ctor(void *mem, int size, void *arg) td = (struct thread *)mem; td->td_state = TDS_INACTIVE; - td->td_flags |= TDF_UNBOUND; } /* @@ -161,6 +197,7 @@ thread_fini(void *mem, int size) td = (struct thread *)mem; pmap_dispose_thread(td); } + /* * Initialize type-stable parts of a kse (when newly created). */ @@ -172,6 +209,7 @@ kse_init(void *mem, int size) ke = (struct kse *)mem; ke->ke_sched = (struct ke_sched *)&ke[1]; } + /* * Initialize type-stable parts of a ksegrp (when newly created). */ @@ -185,7 +223,7 @@ ksegrp_init(void *mem, int size) } /* - * KSE is linked onto the idle queue. + * KSE is linked into kse group. */ void kse_link(struct kse *ke, struct ksegrp *kg) @@ -194,12 +232,12 @@ kse_link(struct kse *ke, struct ksegrp *kg) TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist); kg->kg_kses++; - ke->ke_state = KES_UNQUEUED; + ke->ke_state = KES_UNQUEUED; ke->ke_proc = p; ke->ke_ksegrp = kg; - ke->ke_owner = NULL; ke->ke_thread = NULL; - ke->ke_oncpu = NOCPU; + ke->ke_oncpu = NOCPU; + ke->ke_flags = 0; } void @@ -209,11 +247,13 @@ kse_unlink(struct kse *ke) mtx_assert(&sched_lock, MA_OWNED); kg = ke->ke_ksegrp; - TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist); - if (--kg->kg_kses == 0) { - ksegrp_unlink(kg); + if (ke->ke_state == KES_IDLE) { + TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist); + kg->kg_idle_kses--; } + if (--kg->kg_kses == 0) + ksegrp_unlink(kg); /* * Aggregate stats from the KSE */ @@ -228,15 +268,20 @@ ksegrp_link(struct ksegrp *kg, struct proc *p) TAILQ_INIT(&kg->kg_runq); /* links with td_runq */ TAILQ_INIT(&kg->kg_slpq); /* links with td_runq */ TAILQ_INIT(&kg->kg_kseq); /* all kses in ksegrp */ - TAILQ_INIT(&kg->kg_lq); /* loan kses in ksegrp */ - kg->kg_proc = p; -/* the following counters are in the -zero- section and may not need clearing */ + TAILQ_INIT(&kg->kg_iq); /* all idle kses in ksegrp */ + TAILQ_INIT(&kg->kg_upcalls); /* all upcall structure in ksegrp */ + kg->kg_proc = p; + /* + * the following counters are in the -zero- section + * and may not need clearing + */ kg->kg_numthreads = 0; - kg->kg_runnable = 0; - kg->kg_kses = 0; - kg->kg_loan_kses = 0; - kg->kg_runq_kses = 0; /* XXXKSE change name */ -/* link it in now that it's consistent */ + kg->kg_runnable = 0; + kg->kg_kses = 0; + kg->kg_runq_kses = 0; /* XXXKSE change name */ + kg->kg_idle_kses = 0; + kg->kg_numupcalls = 0; + /* link it in now that it's consistent */ p->p_numksegrps++; TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp); } @@ -247,9 +292,11 @@ ksegrp_unlink(struct ksegrp *kg) struct proc *p; mtx_assert(&sched_lock, MA_OWNED); + KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads")); + KASSERT((kg->kg_kses == 0), ("ksegrp_unlink: residual kses")); + KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls")); + p = kg->kg_proc; - KASSERT(((kg->kg_numthreads == 0) && (kg->kg_kses == 0)), - ("kseg_unlink: residual threads or KSEs")); TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp); p->p_numksegrps--; /* @@ -258,13 +305,63 @@ ksegrp_unlink(struct ksegrp *kg) ksegrp_stash(kg); } +struct kse_upcall * +upcall_alloc(void) +{ + struct kse_upcall *ku; + + ku = uma_zalloc(upcall_zone, 0); + bzero(ku, sizeof(*ku)); + return (ku); +} + +void +upcall_free(struct kse_upcall *ku) +{ + + uma_zfree(upcall_zone, ku); +} + +void +upcall_link(struct kse_upcall *ku, struct ksegrp *kg) +{ + + mtx_assert(&sched_lock, MA_OWNED); + TAILQ_INSERT_TAIL(&kg->kg_upcalls, ku, ku_link); + ku->ku_ksegrp = kg; + kg->kg_numupcalls++; +} + +void +upcall_unlink(struct kse_upcall *ku) +{ + struct ksegrp *kg = ku->ku_ksegrp; + + mtx_assert(&sched_lock, MA_OWNED); + KASSERT(ku->ku_owner == NULL, ("%s: have owner", __func__)); + TAILQ_REMOVE(&kg->kg_upcalls, ku, ku_link); + kg->kg_numupcalls--; + upcall_stash(ku); +} + +void +upcall_remove(struct thread *td) +{ + + if (td->td_upcall) { + td->td_upcall->ku_owner = NULL; + upcall_unlink(td->td_upcall); + td->td_upcall = 0; + } +} + /* - * for a newly created process, - * link up a the structure and its initial threads etc. + * For a newly created process, + * link up all the structures and its initial threads etc. */ void proc_linkup(struct proc *p, struct ksegrp *kg, - struct kse *ke, struct thread *td) + struct kse *ke, struct thread *td) { TAILQ_INIT(&p->p_ksegrps); /* all ksegrps in proc */ @@ -278,6 +375,11 @@ proc_linkup(struct proc *p, struct ksegrp *kg, thread_link(td, kg); } +/* +struct kse_thr_interrupt_args { + struct kse_thr_mailbox * tmbx; +}; +*/ int kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) { @@ -285,10 +387,7 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) struct thread *td2; p = td->td_proc; - /* KSE-enabled processes only, please. */ - if (!(p->p_flag & P_KSES)) - return (EINVAL); - if (uap->tmbx == NULL) + if (!(p->p_flag & P_KSES) || (uap->tmbx == NULL)) return (EINVAL); mtx_lock_spin(&sched_lock); FOREACH_THREAD_IN_PROC(p, td2) { @@ -299,7 +398,7 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) cv_abort(td2); else abortsleep(td2); - } + } mtx_unlock_spin(&sched_lock); return (0); } @@ -308,6 +407,11 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) return (ESRCH); } +/* +struct kse_exit_args { + register_t dummy; +}; +*/ int kse_exit(struct thread *td, struct kse_exit_args *uap) { @@ -316,11 +420,16 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) struct kse *ke; p = td->td_proc; - /* Only UTS can do the syscall */ - if (!(p->p_flag & P_KSES) || (td->td_mailbox != NULL)) + /* + * Only UTS can call the syscall and current group + * should be a threaded group. + */ + if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0)) return (EINVAL); + KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__)); + kg = td->td_ksegrp; - /* serialize killing kse */ + /* Serialize killing KSE */ PROC_LOCK(p); mtx_lock_spin(&sched_lock); if ((kg->kg_kses == 1) && (kg->kg_numthreads > 1)) { @@ -329,14 +438,17 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) return (EDEADLK); } ke = td->td_kse; + upcall_remove(td); if (p->p_numthreads == 1) { - ke->ke_flags &= ~KEF_DOUPCALL; - ke->ke_mailbox = NULL; + kse_purge(p, td); p->p_flag &= ~P_KSES; mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); } else { - ke->ke_flags |= KEF_EXIT; + if (kg->kg_numthreads == 1) { /* Shutdown a group */ + kse_purge_group(td); + ke->ke_flags |= KEF_EXIT; + } thread_exit(); /* NOTREACHED */ } @@ -345,10 +457,15 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) /* * Either becomes an upcall or waits for an awakening event and - * THEN becomes an upcall. Only error cases return. + * then becomes an upcall. Only error cases return. */ +/* +struct kse_release_args { + register_t dummy; +}; +*/ int -kse_release(struct thread * td, struct kse_release_args * uap) +kse_release(struct thread *td, struct kse_release_args *uap) { struct proc *p; struct ksegrp *kg; @@ -356,28 +473,25 @@ kse_release(struct thread * td, struct kse_release_args * uap) p = td->td_proc; kg = td->td_ksegrp; /* - * kse must have a mailbox ready for upcall, and only UTS can - * do the syscall. - */ - if (!(p->p_flag & P_KSES) || - (td->td_mailbox != NULL) || - (td->td_kse->ke_mailbox == NULL)) + * Only UTS can call the syscall and current group + * should be a threaded group. + */ + if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0)) return (EINVAL); + KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__)); PROC_LOCK(p); mtx_lock_spin(&sched_lock); /* Change OURSELF to become an upcall. */ - td->td_flags = TDF_UPCALLING; /* BOUND */ - if (!(td->td_kse->ke_flags & (KEF_DOUPCALL|KEF_ASTPENDING)) && + td->td_flags = TDF_UPCALLING; + if ((td->td_upcall->ku_flags & KUF_DOUPCALL) == 0 && (kg->kg_completed == NULL)) { - /* - * The KSE will however be lendable. - */ - TD_SET_IDLE(td); - PROC_UNLOCK(p); - p->p_stats->p_ru.ru_nvcsw++; - mi_switch(); + kg->kg_upsleeps++; mtx_unlock_spin(&sched_lock); + msleep(&kg->kg_completed, &p->p_mtx, PPAUSE|PCATCH, "ksepause", + NULL); + kg->kg_upsleeps--; + PROC_UNLOCK(p); } else { mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); @@ -392,61 +506,59 @@ int kse_wakeup(struct thread *td, struct kse_wakeup_args *uap) { struct proc *p; - struct kse *ke; struct ksegrp *kg; + struct kse_upcall *ku; struct thread *td2; p = td->td_proc; td2 = NULL; + ku = NULL; /* KSE-enabled processes only, please. */ if (!(p->p_flag & P_KSES)) - return EINVAL; + return (EINVAL); + PROC_LOCK(p); mtx_lock_spin(&sched_lock); if (uap->mbx) { FOREACH_KSEGRP_IN_PROC(p, kg) { - FOREACH_KSE_IN_GROUP(kg, ke) { - if (ke->ke_mailbox != uap->mbx) - continue; - td2 = ke->ke_owner; - KASSERT((td2 != NULL),("KSE with no owner")); - break; + FOREACH_UPCALL_IN_GROUP(kg, ku) { + if (ku->ku_mailbox == uap->mbx) + break; } - if (td2) { + if (ku) break; - } } } else { - /* - * look for any idle KSE to resurrect. - */ kg = td->td_ksegrp; - FOREACH_KSE_IN_GROUP(kg, ke) { - td2 = ke->ke_owner; - KASSERT((td2 != NULL),("KSE with no owner2")); - if (TD_IS_IDLE(td2)) - break; + if (kg->kg_upsleeps) { + wakeup_one(&kg->kg_completed); + mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); + return (0); } - KASSERT((td2 != NULL), ("no thread(s)")); + ku = TAILQ_FIRST(&kg->kg_upcalls); } - if (td2) { - if (TD_IS_IDLE(td2)) { - TD_CLR_IDLE(td2); - setrunnable(td2); - } else if (td != td2) { - /* guarantee do an upcall ASAP */ - td2->td_kse->ke_flags |= KEF_DOUPCALL; + if (ku) { + if ((td2 = ku->ku_owner) == NULL) { + panic("%s: no owner", __func__); + } else if (TD_ON_SLEEPQ(td2) && + (td2->td_wchan == &kg->kg_completed)) { + abortsleep(td2); + } else { + ku->ku_flags |= KUF_DOUPCALL; } mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); return (0); } mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); return (ESRCH); } /* * No new KSEG: first call: use current KSE, don't schedule an upcall - * All other situations, do allocate a new KSE and schedule an upcall on it. + * All other situations, do allocate max new KSEs and schedule an upcall. */ /* struct kse_create_args { struct kse_mailbox *mbx; @@ -456,112 +568,140 @@ int kse_create(struct thread *td, struct kse_create_args *uap) { struct kse *newke; - struct kse *ke; struct ksegrp *newkg; struct ksegrp *kg; struct proc *p; struct kse_mailbox mbx; - int err; + struct kse_upcall *newku; + int err, ncpus; p = td->td_proc; if ((err = copyin(uap->mbx, &mbx, sizeof(mbx)))) return (err); - p->p_flag |= P_KSES; /* easier to just set it than to test and set */ + /* Too bad, why hasn't kernel always a cpu counter !? */ +#ifdef SMP + ncpus = mp_ncpus; +#else + ncpus = 1; +#endif + if (thread_debug && virtual_cpu != 0) + ncpus = virtual_cpu; + + /* Easier to just set it than to test and set */ + p->p_flag |= P_KSES; kg = td->td_ksegrp; if (uap->newgroup) { + /* Have race condition but it is cheap */ if (p->p_numksegrps >= max_groups_per_proc) return (EPROCLIM); /* * If we want a new KSEGRP it doesn't matter whether * we have already fired up KSE mode before or not. - * We put the process in KSE mode and create a new KSEGRP - * and KSE. If our KSE has not got a mailbox yet then - * that doesn't matter, just leave it that way. It will - * ensure that this thread stay BOUND. It's possible - * that the call came form a threaded library and the main - * program knows nothing of threads. + * We put the process in KSE mode and create a new KSEGRP. */ newkg = ksegrp_alloc(); bzero(&newkg->kg_startzero, RANGEOF(struct ksegrp, - kg_startzero, kg_endzero)); + kg_startzero, kg_endzero)); bcopy(&kg->kg_startcopy, &newkg->kg_startcopy, RANGEOF(struct ksegrp, kg_startcopy, kg_endcopy)); - newke = kse_alloc(); + mtx_lock_spin(&sched_lock); + ksegrp_link(newkg, p); + if (p->p_numksegrps >= max_groups_per_proc) { + ksegrp_unlink(newkg); + mtx_unlock_spin(&sched_lock); + return (EPROCLIM); + } + mtx_unlock_spin(&sched_lock); } else { - /* - * Otherwise, if we have already set this KSE - * to have a mailbox, we want to make another KSE here, - * but only if there are not already the limit, which - * is 1 per CPU max. - * - * If the current KSE doesn't have a mailbox we just use it - * and give it one. - * - * Because we don't like to access - * the KSE outside of schedlock if we are UNBOUND, - * (because it can change if we are preempted by an interrupt) - * we can deduce it as having a mailbox if we are UNBOUND, - * and only need to actually look at it if we are BOUND, - * which is safe. + newkg = kg; + } + + /* + * Creating upcalls more than number of physical cpu does + * not help performance. + */ + if (newkg->kg_numupcalls >= ncpus) + return (EPROCLIM); + + if (newkg->kg_numupcalls == 0) { + /* + * Initialize KSE group, optimized for MP. + * Create KSEs as many as physical cpus, this increases + * concurrent even if userland is not MP safe and can only run + * on single CPU (for early version of libpthread, it is true). + * In ideal world, every physical cpu should execute a thread. + * If there is enough KSEs, threads in kernel can be + * executed parallel on different cpus with full speed, + * Concurrent in kernel shouldn't be restricted by number of + * upcalls userland provides. + * Adding more upcall structures only increases concurrent + * in userland. + * Highest performance configuration is: + * N kses = N upcalls = N phyiscal cpus */ - if ((td->td_flags & TDF_UNBOUND) || td->td_kse->ke_mailbox) { - if (thread_debug == 0) { /* if debugging, allow more */ -#ifdef SMP - if (kg->kg_kses > mp_ncpus) -#endif - return (EPROCLIM); - } + while (newkg->kg_kses < ncpus) { newke = kse_alloc(); - } else { - newke = NULL; - } - newkg = NULL; - } - if (newke) { - bzero(&newke->ke_startzero, RANGEOF(struct kse, - ke_startzero, ke_endzero)); + bzero(&newke->ke_startzero, RANGEOF(struct kse, + ke_startzero, ke_endzero)); #if 0 - bcopy(&ke->ke_startcopy, &newke->ke_startcopy, - RANGEOF(struct kse, ke_startcopy, ke_endcopy)); + mtx_lock_spin(&sched_lock); + bcopy(&ke->ke_startcopy, &newke->ke_startcopy, + RANGEOF(struct kse, ke_startcopy, ke_endcopy)); + mtx_unlock_spin(&sched_lock); #endif - /* For the first call this may not have been set */ - if (td->td_standin == NULL) { - td->td_standin = thread_alloc(); - } - mtx_lock_spin(&sched_lock); - if (newkg) { - if (p->p_numksegrps >= max_groups_per_proc) { - mtx_unlock_spin(&sched_lock); - ksegrp_free(newkg); - kse_free(newke); - return (EPROCLIM); - } - ksegrp_link(newkg, p); + mtx_lock_spin(&sched_lock); + kse_link(newke, newkg); + if (p->p_sflag & PS_NEEDSIGCHK) + newke->ke_flags |= KEF_ASTPENDING; + /* Add engine */ + kse_reassign(newke); + mtx_unlock_spin(&sched_lock); } - else - newkg = kg; - kse_link(newke, newkg); - if (p->p_sflag & PS_NEEDSIGCHK) - newke->ke_flags |= KEF_ASTPENDING; - newke->ke_mailbox = uap->mbx; - newke->ke_upcall = mbx.km_func; - bcopy(&mbx.km_stack, &newke->ke_stack, sizeof(stack_t)); - thread_schedule_upcall(td, newke); + } + newku = upcall_alloc(); + newku->ku_mailbox = uap->mbx; + newku->ku_func = mbx.km_func; + bcopy(&mbx.km_stack, &newku->ku_stack, sizeof(stack_t)); + + /* For the first call this may not have been set */ + if (td->td_standin == NULL) + thread_alloc_spare(td, NULL); + + mtx_lock_spin(&sched_lock); + if (newkg->kg_numupcalls >= ncpus) { + upcall_free(newku); mtx_unlock_spin(&sched_lock); + return (EPROCLIM); + } + upcall_link(newku, newkg); + + /* + * Each upcall structure has an owner thread, find which + * one owns it. + */ + if (uap->newgroup) { + /* + * Because new ksegrp hasn't thread, + * create an initial upcall thread to own it. + */ + thread_schedule_upcall(td, newku); } else { /* - * If we didn't allocate a new KSE then the we are using - * the exisiting (BOUND) kse. + * If current thread hasn't an upcall structure, + * just assign the upcall to it. */ - ke = td->td_kse; - ke->ke_mailbox = uap->mbx; - ke->ke_upcall = mbx.km_func; - bcopy(&mbx.km_stack, &ke->ke_stack, sizeof(stack_t)); + if (td->td_upcall == NULL) { + newku->ku_owner = td; + td->td_upcall = newku; + } else { + /* + * Create a new upcall thread to own it. + */ + thread_schedule_upcall(td, newku); + } } - /* - * Fill out the KSE-mode specific fields of the new kse. - */ + mtx_unlock_spin(&sched_lock); return (0); } @@ -642,6 +782,8 @@ threadinit(void) kse_zone = uma_zcreate("KSE", sched_sizeof_kse(), NULL, NULL, kse_init, NULL, UMA_ALIGN_CACHE, 0); + upcall_zone = uma_zcreate("UPCALL", sizeof(struct kse_upcall), + NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); } /* @@ -650,9 +792,9 @@ threadinit(void) void thread_stash(struct thread *td) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); } /* @@ -661,9 +803,21 @@ thread_stash(struct thread *td) void kse_stash(struct kse *ke) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_kses, ke, ke_procq); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); +} + +/* + * Stash an embarasingly extra upcall into the zombie upcall queue. + */ + +void +upcall_stash(struct kse_upcall *ku) +{ + mtx_lock_spin(&kse_zombie_lock); + TAILQ_INSERT_HEAD(&zombie_upcalls, ku, ku_link); + mtx_unlock_spin(&kse_zombie_lock); } /* @@ -672,13 +826,13 @@ kse_stash(struct kse *ke) void ksegrp_stash(struct ksegrp *kg) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); } /* - * Reap zombie threads. + * Reap zombie kse resource. */ void thread_reap(void) @@ -686,27 +840,34 @@ thread_reap(void) struct thread *td_first, *td_next; struct kse *ke_first, *ke_next; struct ksegrp *kg_first, * kg_next; + struct kse_upcall *ku_first, *ku_next; /* - * don't even bother to lock if none at this instant - * We really don't care about the next instant.. + * Don't even bother to lock if none at this instant, + * we really don't care about the next instant.. */ if ((!TAILQ_EMPTY(&zombie_threads)) || (!TAILQ_EMPTY(&zombie_kses)) - || (!TAILQ_EMPTY(&zombie_ksegrps))) { - mtx_lock_spin(&zombie_thread_lock); + || (!TAILQ_EMPTY(&zombie_ksegrps)) + || (!TAILQ_EMPTY(&zombie_upcalls))) { + mtx_lock_spin(&kse_zombie_lock); td_first = TAILQ_FIRST(&zombie_threads); ke_first = TAILQ_FIRST(&zombie_kses); kg_first = TAILQ_FIRST(&zombie_ksegrps); + ku_first = TAILQ_FIRST(&zombie_upcalls); if (td_first) TAILQ_INIT(&zombie_threads); if (ke_first) TAILQ_INIT(&zombie_kses); if (kg_first) TAILQ_INIT(&zombie_ksegrps); - mtx_unlock_spin(&zombie_thread_lock); + if (ku_first) + TAILQ_INIT(&zombie_upcalls); + mtx_unlock_spin(&kse_zombie_lock); while (td_first) { td_next = TAILQ_NEXT(td_first, td_runq); + if (td_first->td_ucred) + crfree(td_first->td_ucred); thread_free(td_first); td_first = td_next; } @@ -720,6 +881,11 @@ thread_reap(void) ksegrp_free(kg_first); kg_first = kg_next; } + while (ku_first) { + ku_next = TAILQ_NEXT(ku_first, ku_link); + upcall_free(ku_first); + ku_first = ku_next; + } } } @@ -792,20 +958,14 @@ thread_export_context(struct thread *td) struct ksegrp *kg; uintptr_t mbx; void *addr; - int error; + int error,temp; ucontext_t uc; - uint temp; p = td->td_proc; kg = td->td_ksegrp; /* Export the user/machine context. */ -#if 0 - addr = (caddr_t)td->td_mailbox + - offsetof(struct kse_thr_mailbox, tm_context); -#else /* if user pointer arithmetic is valid in the kernel */ - addr = (void *)(&td->td_mailbox->tm_context); -#endif + addr = (void *)(&td->td_mailbox->tm_context); error = copyin(addr, &uc, sizeof(ucontext_t)); if (error) goto bad; @@ -815,13 +975,14 @@ thread_export_context(struct thread *td) if (error) goto bad; - /* get address in latest mbox of list pointer */ -#if 0 - addr = (caddr_t)td->td_mailbox - + offsetof(struct kse_thr_mailbox , tm_next); -#else /* if user pointer arithmetic is valid in the kernel */ + /* Exports clock ticks in kernel mode */ + addr = (caddr_t)(&td->td_mailbox->tm_sticks); + temp = fuword(addr) + td->td_usticks; + if (suword(addr, temp)) + goto bad; + + /* Get address in latest mbox of list pointer */ addr = (void *)(&td->td_mailbox->tm_next); -#endif /* * Put the saved address of the previous first * entry into this one @@ -835,42 +996,43 @@ thread_export_context(struct thread *td) PROC_LOCK(p); if (mbx == (uintptr_t)kg->kg_completed) { kg->kg_completed = td->td_mailbox; + /* + * The thread context may be taken away by + * other upcall threads when we unlock + * process lock. it's no longer valid to + * use it again in any other places. + */ + td->td_mailbox = NULL; PROC_UNLOCK(p); break; } PROC_UNLOCK(p); } - addr = (caddr_t)td->td_mailbox - + offsetof(struct kse_thr_mailbox, tm_sticks); - temp = fuword(addr) + td->td_usticks; - if (suword(addr, temp)) - goto bad; + td->td_usticks = 0; return (0); bad: PROC_LOCK(p); psignal(p, SIGSEGV); PROC_UNLOCK(p); + /* The mailbox is bad, don't use it */ + td->td_mailbox = NULL; + td->td_usticks = 0; return (error); } /* * Take the list of completed mailboxes for this KSEGRP and put them on this - * KSE's mailbox as it's the next one going up. + * upcall's mailbox as it's the next one going up. */ static int -thread_link_mboxes(struct ksegrp *kg, struct kse *ke) +thread_link_mboxes(struct ksegrp *kg, struct kse_upcall *ku) { struct proc *p = kg->kg_proc; void *addr; uintptr_t mbx; -#if 0 - addr = (caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_completed); -#else /* if user pointer arithmetic is valid in the kernel */ - addr = (void *)(&ke->ke_mailbox->km_completed); -#endif + addr = (void *)(&ku->ku_mailbox->km_completed); for (;;) { mbx = (uintptr_t)kg->kg_completed; if (suword(addr, mbx)) { @@ -895,69 +1057,91 @@ thread_link_mboxes(struct ksegrp *kg, struct kse *ke) * This function should be called at statclock interrupt time */ int -thread_add_ticks_intr(int user, uint ticks) +thread_statclock(int user) { struct thread *td = curthread; - struct kse *ke = td->td_kse; - if (ke->ke_mailbox == NULL) - return -1; + if (td->td_ksegrp->kg_numupcalls == 0) + return (-1); if (user) { /* Current always do via ast() */ - ke->ke_flags |= KEF_ASTPENDING; - ke->ke_uuticks += ticks; + td->td_flags |= (TDF_ASTPENDING|TDF_USTATCLOCK); + td->td_uuticks += ticks; } else { if (td->td_mailbox != NULL) td->td_usticks += ticks; - else - ke->ke_usticks += ticks; + else { + /* XXXKSE + * We will call thread_user_enter() for every + * kernel entry in future, so if the thread mailbox + * is NULL, it must be a UTS kernel, don't account + * clock ticks for it. + */ + } } - return 0; + return (0); } +/* + * Export user mode state clock ticks + */ static int -thread_update_uticks(void) +thread_update_usr_ticks(struct thread *td) { - struct thread *td = curthread; struct proc *p = td->td_proc; - struct kse *ke = td->td_kse; struct kse_thr_mailbox *tmbx; + struct kse_upcall *ku; caddr_t addr; - uint uticks, sticks; - - if (ke->ke_mailbox == NULL) - return 0; + uint uticks; - uticks = ke->ke_uuticks; - ke->ke_uuticks = 0; - sticks = ke->ke_usticks; - ke->ke_usticks = 0; -#if 0 - tmbx = (void *)fuword((caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_curthread)); -#else /* if user pointer arithmetic is ok in the kernel */ - tmbx = (void *)fuword( (void *)&ke->ke_mailbox->km_curthread); -#endif + if ((ku = td->td_upcall) == NULL) + return (-1); + + tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread); if ((tmbx == NULL) || (tmbx == (void *)-1)) - return 0; + return (-1); + uticks = td->td_uuticks; + td->td_uuticks = 0; if (uticks) { - addr = (caddr_t)tmbx + offsetof(struct kse_thr_mailbox, tm_uticks); + addr = (caddr_t)&tmbx->tm_uticks; uticks += fuword(addr); - if (suword(addr, uticks)) - goto bad; + if (suword(addr, uticks)) { + PROC_LOCK(p); + psignal(p, SIGSEGV); + PROC_UNLOCK(p); + return (-2); + } } - if (sticks) { - addr = (caddr_t)tmbx + offsetof(struct kse_thr_mailbox, tm_sticks); - sticks += fuword(addr); - if (suword(addr, sticks)) - goto bad; + return (0); +} + +/* + * Export kernel mode state clock ticks + */ + +static int +thread_update_sys_ticks(struct thread *td) +{ + struct proc *p = td->td_proc; + caddr_t addr; + int sticks; + + if (td->td_mailbox == NULL) + return (-1); + if (td->td_usticks == 0) + return (0); + addr = (caddr_t)&td->td_mailbox->tm_sticks; + sticks = fuword(addr); + /* XXXKSE use XCHG instead */ + sticks += td->td_usticks; + td->td_usticks = 0; + if (suword(addr, sticks)) { + PROC_LOCK(p); + psignal(p, SIGSEGV); + PROC_UNLOCK(p); + return (-2); } - return 0; -bad: - PROC_LOCK(p); - psignal(p, SIGSEGV); - PROC_UNLOCK(p); - return -1; + return (0); } /* @@ -1013,6 +1197,7 @@ thread_exit(void) p->p_numthreads--; TAILQ_REMOVE(&kg->kg_threads, td, td_kglist); kg->kg_numthreads--; + /* * The test below is NOT true if we are the * sole exiting thread. P_STOPPED_SNGL is unset @@ -1024,25 +1209,28 @@ thread_exit(void) } } - /* Reassign this thread's KSE. */ + /* + * Because each upcall structure has an owner thread, + * owner thread exits only when process is in exiting + * state, so upcall to userland is no longer needed, + * deleting upcall structure is safe here. + * So when all threads in a group is exited, all upcalls + * in the group should be automatically freed. + */ + if (td->td_upcall) + upcall_remove(td); + ke->ke_state = KES_UNQUEUED; - + ke->ke_thread = NULL; /* * Decide what to do with the KSE attached to this thread. - * XXX Possibly kse_reassign should do both cases as it already - * does some of this. */ - if (ke->ke_flags & KEF_EXIT) { - KASSERT((ke->ke_owner == td), - ("thread_exit: KSE exiting with non-owner thread")); - ke->ke_thread = NULL; - td->td_kse = NULL; + if (ke->ke_flags & KEF_EXIT) kse_unlink(ke); - } else { - TD_SET_EXITING(td); /* definitly not runnable */ + else kse_reassign(ke); - } PROC_UNLOCK(p); + td->td_kse = NULL; td->td_state = TDS_INACTIVE; td->td_proc = NULL; td->td_ksegrp = NULL; @@ -1090,10 +1278,12 @@ thread_link(struct thread *td, struct ksegrp *kg) struct proc *p; p = kg->kg_proc; - td->td_state = TDS_INACTIVE; - td->td_proc = p; - td->td_ksegrp = kg; - td->td_last_kse = NULL; + td->td_state = TDS_INACTIVE; + td->td_proc = p; + td->td_ksegrp = kg; + td->td_last_kse = NULL; + td->td_flags = 0; + td->td_kse = NULL; LIST_INIT(&td->td_contested); callout_init(&td->td_slpcallout, 1); @@ -1101,116 +1291,139 @@ thread_link(struct thread *td, struct ksegrp *kg) TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist); p->p_numthreads++; kg->kg_numthreads++; - td->td_kse = NULL; } +/* + * Purge a ksegrp resource. When a ksegrp is preparing to + * exit, it calls this function. + */ +void +kse_purge_group(struct thread *td) +{ + struct ksegrp *kg; + struct kse *ke; + + kg = td->td_ksegrp; + KASSERT(kg->kg_numthreads == 1, ("%s: bad thread number", __func__)); + while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) { + KASSERT(ke->ke_state == KES_IDLE, + ("%s: wrong idle KSE state", __func__)); + kse_unlink(ke); + } + KASSERT((kg->kg_kses == 1), + ("%s: ksegrp still has %d KSEs", __func__, kg->kg_kses)); + KASSERT((kg->kg_numupcalls == 0), + ("%s: ksegrp still has %d upcall datas", + __func__, kg->kg_numupcalls)); +} + +/* + * Purge a process's KSE resource. When a process is preparing to + * exit, it calls kse_purge to release any extra KSE resources in + * the process. + */ void kse_purge(struct proc *p, struct thread *td) { - /* XXXKSE think about this.. - may need to wake up threads on loan queue. */ struct ksegrp *kg; + struct kse *ke; KASSERT(p->p_numthreads == 1, ("bad thread number")); mtx_lock_spin(&sched_lock); while ((kg = TAILQ_FIRST(&p->p_ksegrps)) != NULL) { TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp); p->p_numksegrps--; + /* + * There is no ownership for KSE, after all threads + * in the group exited, it is possible that some KSEs + * were left in idle queue, gc them now. + */ + while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) { + KASSERT(ke->ke_state == KES_IDLE, + ("%s: wrong idle KSE state", __func__)); + TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist); + kg->kg_idle_kses--; + TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist); + kg->kg_kses--; + kse_stash(ke); + } KASSERT(((kg->kg_kses == 0) && (kg != td->td_ksegrp)) || - ((kg->kg_kses == 1) && (kg == td->td_ksegrp)), - ("wrong kg_kses")); - if (kg != td->td_ksegrp) { + ((kg->kg_kses == 1) && (kg == td->td_ksegrp)), + ("ksegrp has wrong kg_kses: %d", kg->kg_kses)); + KASSERT((kg->kg_numupcalls == 0), + ("%s: ksegrp still has %d upcall datas", + __func__, kg->kg_numupcalls)); + + if (kg != td->td_ksegrp) ksegrp_stash(kg); - } } TAILQ_INSERT_HEAD(&p->p_ksegrps, td->td_ksegrp, kg_ksegrp); p->p_numksegrps++; mtx_unlock_spin(&sched_lock); } +/* + * This function is intended to be used to initialize a spare thread + * for upcall. Initialize thread's large data area outside sched_lock + * for thread_schedule_upcall(). + */ +void +thread_alloc_spare(struct thread *td, struct thread *spare) +{ + if (td->td_standin) + return; + if (spare == NULL) + spare = thread_alloc(); + td->td_standin = spare; + bzero(&spare->td_startzero, + (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); + spare->td_proc = td->td_proc; + /* Setup PCB and fork address */ + cpu_set_upcall(spare, td->td_pcb); + /* + * XXXKSE do we really need this? (default values for the + * frame). + */ + bcopy(td->td_frame, spare->td_frame, sizeof(struct trapframe)); + spare->td_ucred = crhold(td->td_ucred); +} /* * Create a thread and schedule it for upcall on the KSE given. * Use our thread's standin so that we don't have to allocate one. */ struct thread * -thread_schedule_upcall(struct thread *td, struct kse *ke) +thread_schedule_upcall(struct thread *td, struct kse_upcall *ku) { struct thread *td2; - int newkse; mtx_assert(&sched_lock, MA_OWNED); - newkse = (ke != td->td_kse); /* - * If the owner and kse are BOUND then that thread is planning to - * go to userland and upcalls are not expected. So don't make one. - * If it is not bound then make it so with the spare thread - * anf then borrw back the KSE to allow us to complete some in-kernel - * work. When we complete, the Bound thread will have the chance to - * complete. This thread will sleep as planned. Hopefully there will - * eventually be un unbound thread that can be converted to an - * upcall to report the completion of this thread. + * Schedule an upcall thread on specified kse_upcall, + * the kse_upcall must be free. + * td must have a spare thread. */ - + KASSERT(ku->ku_owner == NULL, ("%s: upcall has owner", __func__)); if ((td2 = td->td_standin) != NULL) { td->td_standin = NULL; } else { - if (newkse) - panic("no reserve thread when called with a new kse"); - /* - * If called from (e.g.) sleep and we do not have - * a reserve thread, then we've used it, so do not - * create an upcall. - */ + panic("no reserve thread when scheduling an upcall"); return (NULL); } CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)", td2, td->td_proc->p_pid, td->td_proc->p_comm); - bzero(&td2->td_startzero, - (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); bcopy(&td->td_startcopy, &td2->td_startcopy, (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy)); - thread_link(td2, ke->ke_ksegrp); - cpu_set_upcall(td2, td->td_pcb); - - /* - * XXXKSE do we really need this? (default values for the - * frame). - */ - bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe)); - - /* - * Bind the new thread to the KSE, - * and if it's our KSE, lend it back to ourself - * so we can continue running. - */ - td2->td_ucred = crhold(td->td_ucred); - td2->td_flags = TDF_UPCALLING; /* note: BOUND */ - td2->td_kse = ke; - td2->td_state = TDS_CAN_RUN; + thread_link(td2, ku->ku_ksegrp); + /* Let the new thread become owner of the upcall */ + ku->ku_owner = td2; + td2->td_upcall = ku; + td2->td_flags = TDF_UPCALLING; + td2->td_kse = NULL; + td2->td_state = TDS_CAN_RUN; td2->td_inhibitors = 0; - ke->ke_owner = td2; - /* - * If called from kse_reassign(), we are working on the current - * KSE so fake that we borrowed it. If called from - * kse_create(), don't, as we have a new kse too. - */ - if (!newkse) { - /* - * This thread will be scheduled when the current thread - * blocks, exits or tries to enter userspace, (which ever - * happens first). When that happens the KSe will "revert" - * to this thread in a BOUND manner. Since we are called - * from msleep() this is going to be "very soon" in nearly - * all cases. - */ - TD_SET_LOAN(td2); - } else { - ke->ke_thread = td2; - ke->ke_state = KES_THREAD; - setrunqueue(td2); - } + setrunqueue(td2); return (td2); /* bogus.. should be a void function */ } @@ -1222,14 +1435,16 @@ thread_schedule_upcall(struct thread *td, struct kse *ke) struct thread * signal_upcall(struct proc *p, int sig) { +#if 0 struct thread *td, *td2; struct kse *ke; sigset_t ss; int error; +#endif PROC_LOCK_ASSERT(p, MA_OWNED); return (NULL); - +#if 0 td = FIRST_THREAD_IN_PROC(p); ke = td->td_kse; PROC_UNLOCK(p); @@ -1244,28 +1459,31 @@ return (NULL); if (error) return (NULL); if (td->td_standin == NULL) - td->td_standin = thread_alloc(); + thread_alloc_spare(td, NULL); mtx_lock_spin(&sched_lock); td2 = thread_schedule_upcall(td, ke); /* Bogus JRE */ mtx_unlock_spin(&sched_lock); return (td2); +#endif } /* - * setup done on the thread when it enters the kernel. + * Setup done on the thread when it enters the kernel. * XXXKSE Presently only for syscalls but eventually all kernel entries. */ void thread_user_enter(struct proc *p, struct thread *td) { - struct kse *ke; + struct ksegrp *kg; + struct kse_upcall *ku; + kg = td->td_ksegrp; /* * First check that we shouldn't just abort. * But check if we are the single thread first! * XXX p_singlethread not locked, but should be safe. */ - if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) { + if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { PROC_LOCK(p); mtx_lock_spin(&sched_lock); thread_exit(); @@ -1278,43 +1496,37 @@ thread_user_enter(struct proc *p, struct thread *td) * possibility that we could do this lazily (in kse_reassign()), * but for now do it every time. */ - ke = td->td_kse; - td->td_flags &= ~TDF_UNBOUND; - if (ke->ke_mailbox != NULL) { -#if 0 - td->td_mailbox = (void *)fuword((caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_curthread)); -#else /* if user pointer arithmetic is ok in the kernel */ + kg = td->td_ksegrp; + if (kg->kg_numupcalls) { + ku = td->td_upcall; + KASSERT(ku, ("%s: no upcall owned", __func__)); + KASSERT((ku->ku_owner == td), ("%s: wrong owner", __func__)); td->td_mailbox = - (void *)fuword( (void *)&ke->ke_mailbox->km_curthread); -#endif + (void *)fuword((void *)&ku->ku_mailbox->km_curthread); if ((td->td_mailbox == NULL) || (td->td_mailbox == (void *)-1)) { - td->td_mailbox = NULL; /* single thread it.. */ + /* Don't schedule upcall when blocked */ + td->td_mailbox = NULL; mtx_lock_spin(&sched_lock); - td->td_flags &= ~(TDF_UNBOUND|TDF_CAN_UNBIND); + td->td_flags &= ~TDF_CAN_UNBIND; mtx_unlock_spin(&sched_lock); } else { - /* - * when thread limit reached, act like that the thread - * has already done an upcall. - */ if (p->p_numthreads > max_threads_per_proc) { - if (td->td_standin != NULL) { - thread_stash(td->td_standin); - td->td_standin = NULL; - } + /* + * Since kernel thread limit reached, + * don't schedule upcall anymore. + * XXXKSE These code in fact needn't. + */ + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); } else { if (td->td_standin == NULL) - td->td_standin = thread_alloc(); + thread_alloc_spare(td, NULL); + mtx_lock_spin(&sched_lock); + td->td_flags |= TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); } - mtx_lock_spin(&sched_lock); - td->td_flags |= TDF_CAN_UNBIND; - mtx_unlock_spin(&sched_lock); - KASSERT((ke->ke_owner == td), - ("thread_user_enter: No starting owner ")); - ke->ke_owner = td; - td->td_usticks = 0; } } } @@ -1335,165 +1547,90 @@ int thread_userret(struct thread *td, struct trapframe *frame) { int error; - int unbound; - struct kse *ke; + struct kse_upcall *ku; struct ksegrp *kg; - struct thread *worktodo; struct proc *p; struct timespec ts; - KASSERT((td->td_kse && td->td_kse->ke_thread && td->td_kse->ke_owner), - ("thread_userret: bad thread/kse pointers")); - KASSERT((td == curthread), - ("thread_userret: bad thread argument")); - - - kg = td->td_ksegrp; p = td->td_proc; - error = 0; - unbound = TD_IS_UNBOUND(td); + kg = td->td_ksegrp; - mtx_lock_spin(&sched_lock); - if ((worktodo = kg->kg_last_assigned)) - worktodo = TAILQ_NEXT(worktodo, td_runq); - else - worktodo = TAILQ_FIRST(&kg->kg_runq); + /* Nothing to do with non-threaded group/process */ + if (td->td_ksegrp->kg_numupcalls == 0) + return (0); /* - * Permanently bound threads never upcall but they may - * loan out their KSE at this point. - * Upcalls imply bound.. They also may want to do some Philantropy. - * Temporarily bound threads on the other hand either yield - * to other work and transform into an upcall, or proceed back to - * userland. + * State clock interrupt hit in userland, it + * is returning from interrupt, charge thread's + * userland time for UTS. */ + if (td->td_flags & TDF_USTATCLOCK) { + thread_update_usr_ticks(td); + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_USTATCLOCK; + mtx_unlock_spin(&sched_lock); + } + /* + * Optimisation: + * This thread has not started any upcall. + * If there is no work to report other than ourself, + * then it can return direct to userland. + */ if (TD_CAN_UNBIND(td)) { - td->td_flags &= ~(TDF_UNBOUND|TDF_CAN_UNBIND); - if (!worktodo && (kg->kg_completed == NULL) && - !(td->td_kse->ke_flags & KEF_DOUPCALL)) { - /* - * This thread has not started any upcall. - * If there is no work to report other than - * ourself, then it can return direct to userland. - */ -justreturn: - mtx_unlock_spin(&sched_lock); - thread_update_uticks(); + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); + if ((kg->kg_completed == NULL) && + (td->td_upcall->ku_flags & KUF_DOUPCALL) == 0) { + thread_update_sys_ticks(td); td->td_mailbox = NULL; return (0); } - mtx_unlock_spin(&sched_lock); error = thread_export_context(td); - td->td_usticks = 0; if (error) { /* - * As we are not running on a borrowed KSE, - * failing to do the KSE operation just defaults + * Failing to do the KSE operation just defaults * back to synchonous operation, so just return from * the syscall. */ - goto justreturn; + return (0); } - mtx_lock_spin(&sched_lock); /* - * Turn ourself into a bound upcall. - * We will rely on kse_reassign() - * to make us run at a later time. + * There is something to report, and we own an upcall + * strucuture, we can go to userland. + * Turn ourself into an upcall thread. */ + mtx_lock_spin(&sched_lock); td->td_flags |= TDF_UPCALLING; - - /* there may be more work since we re-locked schedlock */ - if ((worktodo = kg->kg_last_assigned)) - worktodo = TAILQ_NEXT(worktodo, td_runq); - else - worktodo = TAILQ_FIRST(&kg->kg_runq); - } else if (unbound) { - /* - * We are an unbound thread, looking to - * return to user space. There must be another owner - * of this KSE. - * We are using a borrowed KSE. save state and exit. - * kse_reassign() will recycle the kse as needed, - */ mtx_unlock_spin(&sched_lock); + } else if (td->td_mailbox) { error = thread_export_context(td); - td->td_usticks = 0; if (error) { - /* - * There is nothing we can do. - * We just lose that context. We - * probably should note this somewhere and send - * the process a signal. - */ PROC_LOCK(td->td_proc); - psignal(td->td_proc, SIGSEGV); mtx_lock_spin(&sched_lock); - ke = td->td_kse; /* possibly upcall with error? */ } else { + PROC_LOCK(td->td_proc); + mtx_lock_spin(&sched_lock); /* - * Don't make an upcall, just exit so that the owner - * can get its KSE if it wants it. - * Our context is already safely stored for later - * use by the UTS. + * There are upcall threads waiting for + * work to do, wake one of them up. + * XXXKSE Maybe wake all of them up. */ - PROC_LOCK(p); - mtx_lock_spin(&sched_lock); - ke = td->td_kse; - } - /* - * If the owner is idling, we now have something for it - * to report, so make it runnable. - * If the owner is not an upcall, make an attempt to - * ensure that at least one of any IDLED upcalls can - * wake up. - */ - if (ke->ke_owner->td_flags & TDF_UPCALLING) { - TD_CLR_IDLE(ke->ke_owner); - } else { - FOREACH_KSE_IN_GROUP(kg, ke) { - if (TD_IS_IDLE(ke->ke_owner)) { - TD_CLR_IDLE(ke->ke_owner); - setrunnable(ke->ke_owner); - break; - } - } + if (kg->kg_upsleeps) + wakeup_one(&kg->kg_completed); } thread_exit(); + /* NOTREACHED */ } - /* - * We ARE going back to userland with this KSE. - * We are permanently bound. We may be an upcall. - * If an upcall, check for threads that need to borrow the KSE. - * Any other thread that comes ready after this missed the boat. - */ - ke = td->td_kse; - /* - * If not upcalling, go back to userspace. - * If we are, get the upcall set up. - */ if (td->td_flags & TDF_UPCALLING) { - if (worktodo) { - /* - * force a switch to more urgent 'in kernel' - * work. Control will return to this thread - * when there is no more work to do. - * kse_reassign() will do that for us. - */ - TD_SET_LOAN(td); - p->p_stats->p_ru.ru_nvcsw++; - mi_switch(); /* kse_reassign() will (re)find worktodo */ - } - td->td_flags &= ~TDF_UPCALLING; - if (ke->ke_flags & KEF_DOUPCALL) - ke->ke_flags &= ~KEF_DOUPCALL; - mtx_unlock_spin(&sched_lock); - + KASSERT(TD_CAN_UNBIND(td) == 0, ("upcall thread can unbind")); + ku = td->td_upcall; /* * There is no more work to do and we are going to ride - * this thread/KSE up to userland as an upcall. + * this thread up to userland as an upcall. * Do the last parts of the setup needed for the upcall. */ CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)", @@ -1504,16 +1641,27 @@ justreturn: * Will use Giant in cpu_thread_clean() because it uses * kmem_free(kernel_map, ...) */ - cpu_set_upcall_kse(td, ke); + cpu_set_upcall_kse(td, ku); - /* + /* + * Clear TDF_UPCALLING after set upcall context, + * profiling code looks TDF_UPCALLING to avoid account + * a wrong user %EIP + */ + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_UPCALLING; + if (ku->ku_flags & KUF_DOUPCALL) + ku->ku_flags &= ~KUF_DOUPCALL; + mtx_unlock_spin(&sched_lock); + + /* * Unhook the list of completed threads. * anything that completes after this gets to * come in next time. * Put the list of completed thread mailboxes on * this KSE's mailbox. */ - error = thread_link_mboxes(kg, ke); + error = thread_link_mboxes(kg, ku); if (error) goto bad; @@ -1524,34 +1672,33 @@ justreturn: * it would be nice if this all happenned only on the first * time through. (the scan for extra work etc.) */ -#if 0 - error = suword((caddr_t)ke->ke_mailbox + - offsetof(struct kse_mailbox, km_curthread), 0); -#else /* if user pointer arithmetic is ok in the kernel */ - error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0); -#endif - ke->ke_uuticks = ke->ke_usticks = 0; + error = suword((caddr_t)&ku->ku_mailbox->km_curthread, 0); if (error) goto bad; + + /* Export current system time */ nanotime(&ts); if (copyout(&ts, - (caddr_t)&ke->ke_mailbox->km_timeofday, sizeof(ts))) { + (caddr_t)&ku->ku_mailbox->km_timeofday, sizeof(ts))) { goto bad; } - } else { - mtx_unlock_spin(&sched_lock); } /* * Optimisation: * Ensure that we have a spare thread available, * for when we re-enter the kernel. */ - if (td->td_standin == NULL) { - td->td_standin = thread_alloc(); - } + if (td->td_standin == NULL) + thread_alloc_spare(td, NULL); - thread_update_uticks(); + /* + * Clear thread mailbox first, then clear system tick count. + * The order is important because thread_statclock() use + * mailbox pointer to see if it is an userland thread or + * an UTS kernel thread. + */ td->td_mailbox = NULL; + td->td_usticks = 0; return (0); bad: @@ -1563,6 +1710,7 @@ bad: psignal(td->td_proc, SIGSEGV); PROC_UNLOCK(td->td_proc); td->td_mailbox = NULL; + td->td_usticks = 0; return (error); /* go sync */ } @@ -1601,7 +1749,6 @@ thread_single(int force_exit) if (force_exit == SINGLE_EXIT) { p->p_flag |= P_SINGLE_EXIT; - td->td_flags &= ~TDF_UNBOUND; } else p->p_flag &= ~P_SINGLE_EXIT; p->p_flag |= P_STOPPED_SINGLE; @@ -1624,17 +1771,16 @@ thread_single(int force_exit) else abortsleep(td2); } - if (TD_IS_IDLE(td2)) { - TD_CLR_IDLE(td2); - } } else { if (TD_IS_SUSPENDED(td2)) continue; - /* maybe other inhibitted states too? */ + /* + * maybe other inhibitted states too? + * XXXKSE Is it totally safe to + * suspend a non-interruptable thread? + */ if (td2->td_inhibitors & - (TDI_SLEEPING | TDI_SWAPPED | - TDI_LOAN | TDI_IDLE | - TDI_EXITING)) + (TDI_SLEEPING | TDI_SWAPPED)) thread_suspend_one(td2); } } @@ -1660,8 +1806,14 @@ thread_single(int force_exit) mtx_lock(&Giant); PROC_LOCK(p); } - if (force_exit == SINGLE_EXIT) + if (force_exit == SINGLE_EXIT) { + if (td->td_upcall) { + mtx_lock_spin(&sched_lock); + upcall_remove(td); + mtx_unlock_spin(&sched_lock); + } kse_purge(p, td); + } return (0); } @@ -1703,7 +1855,6 @@ thread_suspend_check(int return_instead) { struct thread *td; struct proc *p; - struct kse *ke; struct ksegrp *kg; td = curthread; @@ -1735,16 +1886,6 @@ thread_suspend_check(int return_instead) mtx_lock_spin(&sched_lock); while (mtx_owned(&Giant)) mtx_unlock(&Giant); - /* - * All threads should be exiting - * Unless they are the active "singlethread". - * destroy un-needed KSEs as we go.. - * KSEGRPS may implode too as #kses -> 0. - */ - ke = td->td_kse; - if (ke->ke_owner == td && - (kg->kg_kses >= kg->kg_numthreads )) - ke->ke_flags |= KEF_EXIT; thread_exit(); } @@ -1752,14 +1893,6 @@ thread_suspend_check(int return_instead) * When a thread suspends, it just * moves to the processes's suspend queue * and stays there. - * - * XXXKSE if TDF_BOUND is true - * it will not release it's KSE which might - * lead to deadlock if there are not enough KSEs - * to complete all waiting threads. - * Maybe be able to 'lend' it out again. - * (lent kse's can not go back to userland?) - * and can only be lent in STOPPED state. */ mtx_lock_spin(&sched_lock); if ((p->p_flag & P_STOPPED_SIG) && diff --git a/sys/kern/kern_lock.c b/sys/kern/kern_lock.c index d96547b..9000bc9 100644 --- a/sys/kern/kern_lock.c +++ b/sys/kern/kern_lock.c @@ -219,7 +219,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) #endif { int error; - pid_t pid; + struct thread *thr; int extflags, lockflags; CTR5(KTR_LOCKMGR, @@ -228,9 +228,9 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) error = 0; if (td == NULL) - pid = LK_KERNPROC; + thr = LK_KERNPROC; else - pid = td->td_proc->p_pid; + thr = td; mtx_lock(lkp->lk_interlock); if (flags & LK_INTERLOCK) { @@ -257,7 +257,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) * lock requests or upgrade requests ( but not the exclusive * lock itself ). */ - if (lkp->lk_lockholder != pid) { + if (lkp->lk_lockholder != thr) { lockflags = LK_HAVE_EXCL; mtx_lock_spin(&sched_lock); if (td != NULL && !(td->td_flags & TDF_DEADLKTREAT)) @@ -268,7 +268,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) break; sharelock(lkp, 1); #if defined(DEBUG_LOCKS) - lkp->lk_slockholder = pid; + lkp->lk_slockholder = thr; lkp->lk_sfilename = file; lkp->lk_slineno = line; lkp->lk_slockername = name; @@ -283,14 +283,14 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) /* FALLTHROUGH downgrade */ case LK_DOWNGRADE: - KASSERT(lkp->lk_lockholder == pid && lkp->lk_exclusivecount != 0, + KASSERT(lkp->lk_lockholder == thr && lkp->lk_exclusivecount != 0, ("lockmgr: not holding exclusive lock " - "(owner pid (%d) != pid (%d), exlcnt (%d) != 0", - lkp->lk_lockholder, pid, lkp->lk_exclusivecount)); + "(owner thread (%p) != thread (%p), exlcnt (%d) != 0", + lkp->lk_lockholder, thr, lkp->lk_exclusivecount)); sharelock(lkp, lkp->lk_exclusivecount); lkp->lk_exclusivecount = 0; lkp->lk_flags &= ~LK_HAVE_EXCL; - lkp->lk_lockholder = LK_NOPROC; + lkp->lk_lockholder = (struct thread *)LK_NOPROC; if (lkp->lk_waitcount) wakeup((void *)lkp); break; @@ -317,7 +317,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) * after the upgrade). If we return an error, the file * will always be unlocked. */ - if ((lkp->lk_lockholder == pid) || (lkp->lk_sharecount <= 0)) + if ((lkp->lk_lockholder == thr) || (lkp->lk_sharecount <= 0)) panic("lockmgr: upgrade exclusive lock"); shareunlock(lkp, 1); /* @@ -342,7 +342,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) if (error) break; lkp->lk_flags |= LK_HAVE_EXCL; - lkp->lk_lockholder = pid; + lkp->lk_lockholder = thr; if (lkp->lk_exclusivecount != 0) panic("lockmgr: non-zero exclusive count"); lkp->lk_exclusivecount = 1; @@ -364,7 +364,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) /* FALLTHROUGH exclusive request */ case LK_EXCLUSIVE: - if (lkp->lk_lockholder == pid && pid != LK_KERNPROC) { + if (lkp->lk_lockholder == thr && thr != LK_KERNPROC) { /* * Recursive lock. */ @@ -398,7 +398,7 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) if (error) break; lkp->lk_flags |= LK_HAVE_EXCL; - lkp->lk_lockholder = pid; + lkp->lk_lockholder = thr; if (lkp->lk_exclusivecount != 0) panic("lockmgr: non-zero exclusive count"); lkp->lk_exclusivecount = 1; @@ -411,10 +411,10 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) case LK_RELEASE: if (lkp->lk_exclusivecount != 0) { - if (lkp->lk_lockholder != pid && + if (lkp->lk_lockholder != thr && lkp->lk_lockholder != LK_KERNPROC) { - panic("lockmgr: pid %d, not %s %d unlocking", - pid, "exclusive lock holder", + panic("lockmgr: thread %p, not %s %p unlocking", + thr, "exclusive lock holder", lkp->lk_lockholder); } if (lkp->lk_exclusivecount == 1) { @@ -437,14 +437,14 @@ debuglockmgr(lkp, flags, interlkp, td, name, file, line) * check for holding a shared lock, but at least we can * check for an exclusive one. */ - if (lkp->lk_lockholder == pid) + if (lkp->lk_lockholder == thr) panic("lockmgr: draining against myself"); error = acquiredrain(lkp, extflags); if (error) break; lkp->lk_flags |= LK_DRAINING | LK_HAVE_EXCL; - lkp->lk_lockholder = pid; + lkp->lk_lockholder = thr; lkp->lk_exclusivecount = 1; #if defined(DEBUG_LOCKS) lkp->lk_filename = file; @@ -589,7 +589,7 @@ lockstatus(lkp, td) mtx_lock(lkp->lk_interlock); if (lkp->lk_exclusivecount != 0) { - if (td == NULL || lkp->lk_lockholder == td->td_proc->p_pid) + if (td == NULL || lkp->lk_lockholder == td) lock_type = LK_EXCLUSIVE; else lock_type = LK_EXCLOTHER; @@ -627,7 +627,7 @@ lockmgr_printinfo(lkp) printf(" lock type %s: SHARED (count %d)", lkp->lk_wmesg, lkp->lk_sharecount); else if (lkp->lk_flags & LK_HAVE_EXCL) - printf(" lock type %s: EXCL (count %d) by pid %d", + printf(" lock type %s: EXCL (count %d) by thread %p", lkp->lk_wmesg, lkp->lk_exclusivecount, lkp->lk_lockholder); if (lkp->lk_waitcount > 0) printf(" with %d pending", lkp->lk_waitcount); diff --git a/sys/kern/kern_resource.c b/sys/kern/kern_resource.c index 4ade890..cc7c493 100644 --- a/sys/kern/kern_resource.c +++ b/sys/kern/kern_resource.c @@ -671,32 +671,23 @@ calcru(p, up, sp, ip) { /* {user, system, interrupt, total} {ticks, usec}; previous tu: */ u_int64_t ut, uu, st, su, it, iu, tt, tu, ptu; - u_int64_t uut = 0, sut = 0, iut = 0; - int s; struct timeval tv; struct bintime bt; - struct kse *ke; - struct ksegrp *kg; mtx_assert(&sched_lock, MA_OWNED); /* XXX: why spl-protect ? worst case is an off-by-one report */ - FOREACH_KSEGRP_IN_PROC(p, kg) { - /* we could accumulate per ksegrp and per process here*/ - FOREACH_KSE_IN_GROUP(kg, ke) { - s = splstatclock(); - ut = ke->ke_uticks; - st = ke->ke_sticks; - it = ke->ke_iticks; - splx(s); - - tt = ut + st + it; - if (tt == 0) { - st = 1; - tt = 1; - } + ut = p->p_uticks; + st = p->p_sticks; + it = p->p_iticks; + + tt = ut + st + it; + if (tt == 0) { + st = 1; + tt = 1; + } - if (ke == curthread->td_kse) { + if (curthread->td_proc == p) { /* * Adjust for the current time slice. This is actually fairly * important since the error here is on the order of a time @@ -705,64 +696,59 @@ calcru(p, up, sp, ip) * processors also being 'current'. */ - binuptime(&bt); - bintime_sub(&bt, PCPU_PTR(switchtime)); - bintime_add(&bt, &p->p_runtime); - } else { - bt = p->p_runtime; - } - bintime2timeval(&bt, &tv); - tu = (u_int64_t)tv.tv_sec * 1000000 + tv.tv_usec; - ptu = ke->ke_uu + ke->ke_su + ke->ke_iu; - if (tu < ptu || (int64_t)tu < 0) { - /* XXX no %qd in kernel. Truncate. */ - printf("calcru: negative time of %ld usec for pid %d (%s)\n", - (long)tu, p->p_pid, p->p_comm); - tu = ptu; - } + binuptime(&bt); + bintime_sub(&bt, PCPU_PTR(switchtime)); + bintime_add(&bt, &p->p_runtime); + } else { + bt = p->p_runtime; + } + bintime2timeval(&bt, &tv); + tu = (u_int64_t)tv.tv_sec * 1000000 + tv.tv_usec; + ptu = p->p_uu + p->p_su + p->p_iu; + if (tu < ptu || (int64_t)tu < 0) { + /* XXX no %qd in kernel. Truncate. */ + printf("calcru: negative time of %ld usec for pid %d (%s)\n", + (long)tu, p->p_pid, p->p_comm); + tu = ptu; + } - /* Subdivide tu. */ - uu = (tu * ut) / tt; - su = (tu * st) / tt; - iu = tu - uu - su; + /* Subdivide tu. */ + uu = (tu * ut) / tt; + su = (tu * st) / tt; + iu = tu - uu - su; - /* Enforce monotonicity. */ - if (uu < ke->ke_uu || su < ke->ke_su || iu < ke->ke_iu) { - if (uu < ke->ke_uu) - uu = ke->ke_uu; - else if (uu + ke->ke_su + ke->ke_iu > tu) - uu = tu - ke->ke_su - ke->ke_iu; - if (st == 0) - su = ke->ke_su; - else { - su = ((tu - uu) * st) / (st + it); - if (su < ke->ke_su) - su = ke->ke_su; - else if (uu + su + ke->ke_iu > tu) - su = tu - uu - ke->ke_iu; - } - KASSERT(uu + su + ke->ke_iu <= tu, - ("calcru: monotonisation botch 1")); - iu = tu - uu - su; - KASSERT(iu >= ke->ke_iu, - ("calcru: monotonisation botch 2")); - } - ke->ke_uu = uu; - ke->ke_su = su; - ke->ke_iu = iu; - uut += uu; - sut += su; - iut += iu; - - } /* end kse loop */ - } /* end kseg loop */ - up->tv_sec = uut / 1000000; - up->tv_usec = uut % 1000000; - sp->tv_sec = sut / 1000000; - sp->tv_usec = sut % 1000000; + /* Enforce monotonicity. */ + if (uu < p->p_uu || su < p->p_su || iu < p->p_iu) { + if (uu < p->p_uu) + uu = p->p_uu; + else if (uu + p->p_su + p->p_iu > tu) + uu = tu - p->p_su - p->p_iu; + if (st == 0) + su = p->p_su; + else { + su = ((tu - uu) * st) / (st + it); + if (su < p->p_su) + su = p->p_su; + else if (uu + su + p->p_iu > tu) + su = tu - uu - p->p_iu; + } + KASSERT(uu + su + p->p_iu <= tu, + ("calcru: monotonisation botch 1")); + iu = tu - uu - su; + KASSERT(iu >= p->p_iu, + ("calcru: monotonisation botch 2")); + } + p->p_uu = uu; + p->p_su = su; + p->p_iu = iu; + + up->tv_sec = uu / 1000000; + up->tv_usec = uu % 1000000; + sp->tv_sec = su / 1000000; + sp->tv_usec = su % 1000000; if (ip != NULL) { - ip->tv_sec = iut / 1000000; - ip->tv_usec = iut % 1000000; + ip->tv_sec = iu / 1000000; + ip->tv_usec = iu % 1000000; } } diff --git a/sys/kern/kern_sig.c b/sys/kern/kern_sig.c index c84cbd1..a6ecbf8 100644 --- a/sys/kern/kern_sig.c +++ b/sys/kern/kern_sig.c @@ -1525,9 +1525,6 @@ psignal(p, sig) if (TD_IS_SLEEPING(td) && (td->td_flags & TDF_SINTR)) thread_suspend_one(td); - else if (TD_IS_IDLE(td)) { - thread_suspend_one(td); - } } if (p->p_suspcount == p->p_numthreads) { mtx_unlock_spin(&sched_lock); @@ -1640,9 +1637,6 @@ tdsignal(struct thread *td, int sig, sig_t action) cv_abort(td); else abortsleep(td); - } else if (TD_IS_IDLE(td)) { - TD_CLR_IDLE(td); - setrunnable(td); } #ifdef SMP else { diff --git a/sys/kern/kern_switch.c b/sys/kern/kern_switch.c index 6651f70..5cefb1c 100644 --- a/sys/kern/kern_switch.c +++ b/sys/kern/kern_switch.c @@ -111,7 +111,7 @@ static void runq_readjust(struct runq *rq, struct kse *ke); * Functions that manipulate runnability from a thread perspective. * ************************************************************************/ /* - * Select the KSE that will be run next. From that find the thread, and x + * Select the KSE that will be run next. From that find the thread, and * remove it from the KSEGRP's run queue. If there is thread clustering, * this will be what does it. */ @@ -127,7 +127,7 @@ retry: td = ke->ke_thread; KASSERT((td->td_kse == ke), ("kse/thread mismatch")); kg = ke->ke_ksegrp; - if (TD_IS_UNBOUND(td)) { + if (td->td_proc->p_flag & P_KSES) { TAILQ_REMOVE(&kg->kg_runq, td, td_runq); if (kg->kg_last_assigned == td) { kg->kg_last_assigned = TAILQ_PREV(td, @@ -158,9 +158,8 @@ retry: } /* - * Given a KSE (now surplus or at least loanable), either assign a new - * runable thread to it (and put it in the run queue) or put it in - * the ksegrp's idle KSE list. + * Given a surplus KSE, either assign a new runable thread to it + * (and put it in the run queue) or put it in the ksegrp's idle KSE list. * Or maybe give it back to its owner if it's been loaned. * Assumes that the original thread is either not runnable or * already on the run queue @@ -170,108 +169,54 @@ kse_reassign(struct kse *ke) { struct ksegrp *kg; struct thread *td; - struct thread *owner; struct thread *original; - int loaned; + struct kse_upcall *ku; - KASSERT((ke->ke_owner), ("reassigning KSE with no owner")); - KASSERT((ke->ke_thread && TD_IS_INHIBITED(ke->ke_thread)), - ("reassigning KSE with no or runnable thread")); mtx_assert(&sched_lock, MA_OWNED); - kg = ke->ke_ksegrp; - owner = ke->ke_owner; - loaned = TD_LENDER(owner); original = ke->ke_thread; - - if (TD_CAN_UNBIND(original) && (original->td_standin)) { - KASSERT((owner == original), - ("Early thread borrowing?")); + KASSERT(original == NULL || TD_IS_INHIBITED(original), + ("reassigning KSE with runnable thread")); + kg = ke->ke_ksegrp; + if (original) { /* - * The outgoing thread is "threaded" and has never - * scheduled an upcall. - * decide whether this is a short or long term event - * and thus whether or not to schedule an upcall. - * if it is a short term event, just suspend it in + * If the outgoing thread is in threaded group and has never + * scheduled an upcall, decide whether this is a short + * or long term event and thus whether or not to schedule + * an upcall. + * If it is a short term event, just suspend it in * a way that takes its KSE with it. * Select the events for which we want to schedule upcalls. * For now it's just sleep. - * Other threads that still have not fired an upcall - * are held to their KSE using the temorary Binding. + * XXXKSE eventually almost any inhibition could do. */ - if (TD_ON_SLEEPQ(original)) { - /* - * An bound thread that can still unbind itself - * has been scheduled out. - * If it is sleeping, then we need to schedule an - * upcall. - * XXXKSE eventually almost any inhibition could do. + if (TD_CAN_UNBIND(original) && (original->td_standin) && + TD_ON_SLEEPQ(original)) { + /* + * Release ownership of upcall, and schedule an upcall + * thread, this new upcall thread becomes the owner of + * the upcall structure. */ + ku = original->td_upcall; + ku->ku_owner = NULL; + original->td_upcall = NULL; original->td_flags &= ~TDF_CAN_UNBIND; - original->td_flags |= TDF_UNBOUND; - thread_schedule_upcall(original, ke); - owner = ke->ke_owner; - loaned = 1; + thread_schedule_upcall(original, ku); } + original->td_kse = NULL; } - /* - * If the current thread was borrowing, then make things consistent - * by giving it back to the owner for the moment. The original thread - * must be unbound and have already used its chance for - * firing off an upcall. Threads that have not yet made an upcall - * can not borrow KSEs. - */ - if (loaned) { - TD_CLR_LOAN(owner); - ke->ke_thread = owner; - original->td_kse = NULL; /* give it amnesia */ - /* - * Upcalling threads have lower priority than all - * in-kernel threads, However threads that have loaned out - * their KSE and are NOT upcalling have the priority that - * they have. In other words, only look for other work if - * the owner is not runnable, OR is upcalling. - */ - if (TD_CAN_RUN(owner) && - ((owner->td_flags & TDF_UPCALLING) == 0)) { - setrunnable(owner); - CTR2(KTR_RUNQ, "kse_reassign: ke%p -> td%p (give back)", - ke, owner); - return; - } - } - /* - * Either the owner is not runnable, or is an upcall. * Find the first unassigned thread - * If there is a 'last assigned' then see what's next. - * otherwise look at what is first. */ - if ((td = kg->kg_last_assigned)) { + if ((td = kg->kg_last_assigned) != NULL) td = TAILQ_NEXT(td, td_runq); - } else { + else td = TAILQ_FIRST(&kg->kg_runq); - } /* - * If we found one assign it the kse, otherwise idle the kse. + * If we found one, assign it the kse, otherwise idle the kse. */ if (td) { - /* - * Assign the new thread to the KSE. - * and make the KSE runnable again, - */ - if (TD_IS_BOUND(owner)) { - /* - * If there is a reason to keep the previous - * owner, do so. - */ - TD_SET_LOAN(owner); - } else { - /* otherwise, cut it free */ - ke->ke_owner = td; - owner->td_kse = NULL; - } kg->kg_last_assigned = td; td->td_kse = ke; ke->ke_thread = td; @@ -280,43 +225,11 @@ kse_reassign(struct kse *ke) return; } - /* - * Now handle any waiting upcall. - * Since we didn't make them runnable before. - */ - if (TD_CAN_RUN(owner)) { - setrunnable(owner); - CTR2(KTR_RUNQ, "kse_reassign: ke%p -> td%p (give back)", - ke, owner); - return; - } - - /* - * It is possible that this is the last thread in the group - * because the KSE is being shut down or the process - * is exiting. - */ - if (TD_IS_EXITING(owner) || (ke->ke_flags & KEF_EXIT)) { - ke->ke_thread = NULL; - owner->td_kse = NULL; - kse_unlink(ke); - return; - } - - /* - * At this stage all we know is that the owner - * is the same as the 'active' thread in the KSE - * and that it is - * Presently NOT loaned out. - * Put it on the loanable queue. Make it fifo - * so that long term sleepers donate their KSE's first. - */ - KASSERT((TD_IS_BOUND(owner)), ("kse_reassign: UNBOUND lender")); - ke->ke_state = KES_THREAD; - ke->ke_flags |= KEF_ONLOANQ; - TAILQ_INSERT_TAIL(&kg->kg_lq, ke, ke_kgrlist); - kg->kg_loan_kses++; - CTR1(KTR_RUNQ, "kse_reassign: ke%p on loan queue", ke); + ke->ke_state = KES_IDLE; + ke->ke_thread = NULL; + TAILQ_INSERT_TAIL(&kg->kg_iq, ke, ke_kgrlist); + kg->kg_idle_kses++; + CTR1(KTR_RUNQ, "kse_reassign: ke%p on idle queue", ke); return; } @@ -325,7 +238,7 @@ kse_reassign(struct kse *ke) * Remove a thread from its KSEGRP's run queue. * This in turn may remove it from a KSE if it was already assigned * to one, possibly causing a new thread to be assigned to the KSE - * and the KSE getting a new priority (unless it's a BOUND thread/KSE pair). + * and the KSE getting a new priority. */ static void remrunqueue(struct thread *td) @@ -335,17 +248,16 @@ remrunqueue(struct thread *td) struct kse *ke; mtx_assert(&sched_lock, MA_OWNED); - KASSERT ((TD_ON_RUNQ(td)), ("remrunqueue: Bad state on run queue")); + KASSERT((TD_ON_RUNQ(td)), ("remrunqueue: Bad state on run queue")); kg = td->td_ksegrp; ke = td->td_kse; - /* - * If it's a bound thread/KSE pair, take the shortcut. All non-KSE - * threads are BOUND. - */ CTR1(KTR_RUNQ, "remrunqueue: td%p", td); kg->kg_runnable--; TD_SET_CAN_RUN(td); - if (TD_IS_BOUND(td)) { + /* + * If it is not a threaded process, take the shortcut. + */ + if ((td->td_proc->p_flag & P_KSES) == 0) { /* Bring its kse with it, leave the thread attached */ sched_rem(ke); ke->ke_state = KES_THREAD; @@ -363,7 +275,7 @@ remrunqueue(struct thread *td) sched_rem(ke); ke->ke_state = KES_THREAD; td2 = kg->kg_last_assigned; - KASSERT((td2 != NULL), ("last assigned has wrong value ")); + KASSERT((td2 != NULL), ("last assigned has wrong value")); if (td2 == td) kg->kg_last_assigned = td3; kse_reassign(ke); @@ -381,14 +293,14 @@ adjustrunqueue( struct thread *td, int newpri) struct kse *ke; mtx_assert(&sched_lock, MA_OWNED); - KASSERT ((TD_ON_RUNQ(td)), ("adjustrunqueue: Bad state on run queue")); - /* - * If it's a bound thread/KSE pair, take the shortcut. All non-KSE - * threads are BOUND. - */ + KASSERT((TD_ON_RUNQ(td)), ("adjustrunqueue: Bad state on run queue")); + ke = td->td_kse; CTR1(KTR_RUNQ, "adjustrunqueue: td%p", td); - if (TD_IS_BOUND(td)) { + /* + * If it is not a threaded process, take the shortcut. + */ + if ((td->td_proc->p_flag & P_KSES) == 0) { /* We only care about the kse in the run queue. */ td->td_priority = newpri; if (ke->ke_rqindex != (newpri / RQ_PPQ)) { @@ -397,9 +309,8 @@ adjustrunqueue( struct thread *td, int newpri) } return; } - /* - * An unbound thread. This is not optimised yet. - */ + + /* It is a threaded process */ kg = td->td_ksegrp; kg->kg_runnable--; TD_SET_CAN_RUN(td); @@ -439,48 +350,17 @@ setrunqueue(struct thread *td) sched_add(td->td_kse); return; } - /* - * If the process is threaded but the thread is bound then - * there is still a little extra to do re. KSE loaning. - */ - if (TD_IS_BOUND(td)) { - KASSERT((td->td_kse != NULL), - ("queueing BAD thread to run queue")); - ke = td->td_kse; - KASSERT((ke->ke_owner == ke->ke_thread), - ("setrunqueue: Hey KSE loaned out")); - if (ke->ke_flags & KEF_ONLOANQ) { - ke->ke_flags &= ~KEF_ONLOANQ; - TAILQ_REMOVE(&kg->kg_lq, ke, ke_kgrlist); - kg->kg_loan_kses--; - } - sched_add(td->td_kse); - return; - } - /* - * Ok, so we are threading with this thread. - * We don't have a KSE, see if we can get one.. - */ tda = kg->kg_last_assigned; if ((ke = td->td_kse) == NULL) { - /* - * We will need a KSE, see if there is one.. - * First look for a free one, before getting desperate. - * If we can't get one, our priority is not high enough.. - * that's ok.. - */ - if (kg->kg_loan_kses) { + if (kg->kg_idle_kses) { /* - * Failing that see if we can borrow one. + * There is a free one so it's ours for the asking.. */ - ke = TAILQ_FIRST(&kg->kg_lq); - TAILQ_REMOVE(&kg->kg_lq, ke, ke_kgrlist); - ke->ke_flags &= ~KEF_ONLOANQ; + ke = TAILQ_FIRST(&kg->kg_iq); + TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist); ke->ke_state = KES_THREAD; - TD_SET_LOAN(ke->ke_owner); - ke->ke_thread = NULL; - kg->kg_loan_kses--; + kg->kg_idle_kses--; } else if (tda && (tda->td_priority > td->td_priority)) { /* * None free, but there is one we can commandeer. @@ -495,11 +375,7 @@ setrunqueue(struct thread *td) } else { /* * Temporarily disassociate so it looks like the other cases. - * If the owner wasn't lending before, then it is now.. */ - if (!TD_LENDER(ke->ke_owner)) { - TD_SET_LOAN(ke->ke_owner); - } ke->ke_thread = NULL; td->td_kse = NULL; } @@ -831,6 +707,7 @@ thread_sanity_check(struct thread *td, char *string) if (kg->kg_last_assigned && (saw_lastassigned == 0)) { panc(string, "where on earth does lastassigned point?"); } +#if 0 FOREACH_THREAD_IN_GROUP(kg, td2) { if (((td2->td_flags & TDF_UNBOUND) == 0) && (TD_ON_RUNQ(td2))) { @@ -840,6 +717,7 @@ thread_sanity_check(struct thread *td, char *string) } } } +#endif #if 0 if ((unassigned + assigned) != kg->kg_runnable) { panc(string, "wrong number in runnable"); diff --git a/sys/kern/kern_thread.c b/sys/kern/kern_thread.c index 78bce30..883f4d1 100644 --- a/sys/kern/kern_thread.c +++ b/sys/kern/kern_thread.c @@ -63,6 +63,7 @@ static uma_zone_t ksegrp_zone; static uma_zone_t kse_zone; static uma_zone_t thread_zone; +static uma_zone_t upcall_zone; /* DEBUG ONLY */ SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation"); @@ -78,16 +79,52 @@ static int max_groups_per_proc = 5; SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW, &max_groups_per_proc, 0, "Limit on thread groups per proc"); +static int virtual_cpu; + #define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start)) -struct threadqueue zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); +TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); TAILQ_HEAD(, kse) zombie_kses = TAILQ_HEAD_INITIALIZER(zombie_kses); TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps); -struct mtx zombie_thread_lock; -MTX_SYSINIT(zombie_thread_lock, &zombie_thread_lock, - "zombie_thread_lock", MTX_SPIN); +TAILQ_HEAD(, kse_upcall) zombie_upcalls = + TAILQ_HEAD_INITIALIZER(zombie_upcalls); +struct mtx kse_zombie_lock; +MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN); static void kse_purge(struct proc *p, struct thread *td); +static void kse_purge_group(struct thread *td); +static int thread_update_usr_ticks(struct thread *td); +static int thread_update_sys_ticks(struct thread *td); +static void thread_alloc_spare(struct thread *td, struct thread *spare); + +static int +sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS) +{ + int error, new_val; + int def_val; + +#ifdef SMP + def_val = mp_ncpus; +#else + def_val = 1; +#endif + if (virtual_cpu == 0) + new_val = def_val; + else + new_val = virtual_cpu; + error = sysctl_handle_int(oidp, &new_val, 0, req); + if (error != 0 || req->newptr == NULL) + return (error); + if (new_val < 0) + return (EINVAL); + virtual_cpu = new_val; + return (0); +} + +/* DEBUG ONLY */ +SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW, + 0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I", + "debug virtual cpus"); /* * Prepare a thread for use. @@ -99,7 +136,6 @@ thread_ctor(void *mem, int size, void *arg) td = (struct thread *)mem; td->td_state = TDS_INACTIVE; - td->td_flags |= TDF_UNBOUND; } /* @@ -161,6 +197,7 @@ thread_fini(void *mem, int size) td = (struct thread *)mem; pmap_dispose_thread(td); } + /* * Initialize type-stable parts of a kse (when newly created). */ @@ -172,6 +209,7 @@ kse_init(void *mem, int size) ke = (struct kse *)mem; ke->ke_sched = (struct ke_sched *)&ke[1]; } + /* * Initialize type-stable parts of a ksegrp (when newly created). */ @@ -185,7 +223,7 @@ ksegrp_init(void *mem, int size) } /* - * KSE is linked onto the idle queue. + * KSE is linked into kse group. */ void kse_link(struct kse *ke, struct ksegrp *kg) @@ -194,12 +232,12 @@ kse_link(struct kse *ke, struct ksegrp *kg) TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist); kg->kg_kses++; - ke->ke_state = KES_UNQUEUED; + ke->ke_state = KES_UNQUEUED; ke->ke_proc = p; ke->ke_ksegrp = kg; - ke->ke_owner = NULL; ke->ke_thread = NULL; - ke->ke_oncpu = NOCPU; + ke->ke_oncpu = NOCPU; + ke->ke_flags = 0; } void @@ -209,11 +247,13 @@ kse_unlink(struct kse *ke) mtx_assert(&sched_lock, MA_OWNED); kg = ke->ke_ksegrp; - TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist); - if (--kg->kg_kses == 0) { - ksegrp_unlink(kg); + if (ke->ke_state == KES_IDLE) { + TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist); + kg->kg_idle_kses--; } + if (--kg->kg_kses == 0) + ksegrp_unlink(kg); /* * Aggregate stats from the KSE */ @@ -228,15 +268,20 @@ ksegrp_link(struct ksegrp *kg, struct proc *p) TAILQ_INIT(&kg->kg_runq); /* links with td_runq */ TAILQ_INIT(&kg->kg_slpq); /* links with td_runq */ TAILQ_INIT(&kg->kg_kseq); /* all kses in ksegrp */ - TAILQ_INIT(&kg->kg_lq); /* loan kses in ksegrp */ - kg->kg_proc = p; -/* the following counters are in the -zero- section and may not need clearing */ + TAILQ_INIT(&kg->kg_iq); /* all idle kses in ksegrp */ + TAILQ_INIT(&kg->kg_upcalls); /* all upcall structure in ksegrp */ + kg->kg_proc = p; + /* + * the following counters are in the -zero- section + * and may not need clearing + */ kg->kg_numthreads = 0; - kg->kg_runnable = 0; - kg->kg_kses = 0; - kg->kg_loan_kses = 0; - kg->kg_runq_kses = 0; /* XXXKSE change name */ -/* link it in now that it's consistent */ + kg->kg_runnable = 0; + kg->kg_kses = 0; + kg->kg_runq_kses = 0; /* XXXKSE change name */ + kg->kg_idle_kses = 0; + kg->kg_numupcalls = 0; + /* link it in now that it's consistent */ p->p_numksegrps++; TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp); } @@ -247,9 +292,11 @@ ksegrp_unlink(struct ksegrp *kg) struct proc *p; mtx_assert(&sched_lock, MA_OWNED); + KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads")); + KASSERT((kg->kg_kses == 0), ("ksegrp_unlink: residual kses")); + KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls")); + p = kg->kg_proc; - KASSERT(((kg->kg_numthreads == 0) && (kg->kg_kses == 0)), - ("kseg_unlink: residual threads or KSEs")); TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp); p->p_numksegrps--; /* @@ -258,13 +305,63 @@ ksegrp_unlink(struct ksegrp *kg) ksegrp_stash(kg); } +struct kse_upcall * +upcall_alloc(void) +{ + struct kse_upcall *ku; + + ku = uma_zalloc(upcall_zone, 0); + bzero(ku, sizeof(*ku)); + return (ku); +} + +void +upcall_free(struct kse_upcall *ku) +{ + + uma_zfree(upcall_zone, ku); +} + +void +upcall_link(struct kse_upcall *ku, struct ksegrp *kg) +{ + + mtx_assert(&sched_lock, MA_OWNED); + TAILQ_INSERT_TAIL(&kg->kg_upcalls, ku, ku_link); + ku->ku_ksegrp = kg; + kg->kg_numupcalls++; +} + +void +upcall_unlink(struct kse_upcall *ku) +{ + struct ksegrp *kg = ku->ku_ksegrp; + + mtx_assert(&sched_lock, MA_OWNED); + KASSERT(ku->ku_owner == NULL, ("%s: have owner", __func__)); + TAILQ_REMOVE(&kg->kg_upcalls, ku, ku_link); + kg->kg_numupcalls--; + upcall_stash(ku); +} + +void +upcall_remove(struct thread *td) +{ + + if (td->td_upcall) { + td->td_upcall->ku_owner = NULL; + upcall_unlink(td->td_upcall); + td->td_upcall = 0; + } +} + /* - * for a newly created process, - * link up a the structure and its initial threads etc. + * For a newly created process, + * link up all the structures and its initial threads etc. */ void proc_linkup(struct proc *p, struct ksegrp *kg, - struct kse *ke, struct thread *td) + struct kse *ke, struct thread *td) { TAILQ_INIT(&p->p_ksegrps); /* all ksegrps in proc */ @@ -278,6 +375,11 @@ proc_linkup(struct proc *p, struct ksegrp *kg, thread_link(td, kg); } +/* +struct kse_thr_interrupt_args { + struct kse_thr_mailbox * tmbx; +}; +*/ int kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) { @@ -285,10 +387,7 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) struct thread *td2; p = td->td_proc; - /* KSE-enabled processes only, please. */ - if (!(p->p_flag & P_KSES)) - return (EINVAL); - if (uap->tmbx == NULL) + if (!(p->p_flag & P_KSES) || (uap->tmbx == NULL)) return (EINVAL); mtx_lock_spin(&sched_lock); FOREACH_THREAD_IN_PROC(p, td2) { @@ -299,7 +398,7 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) cv_abort(td2); else abortsleep(td2); - } + } mtx_unlock_spin(&sched_lock); return (0); } @@ -308,6 +407,11 @@ kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap) return (ESRCH); } +/* +struct kse_exit_args { + register_t dummy; +}; +*/ int kse_exit(struct thread *td, struct kse_exit_args *uap) { @@ -316,11 +420,16 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) struct kse *ke; p = td->td_proc; - /* Only UTS can do the syscall */ - if (!(p->p_flag & P_KSES) || (td->td_mailbox != NULL)) + /* + * Only UTS can call the syscall and current group + * should be a threaded group. + */ + if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0)) return (EINVAL); + KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__)); + kg = td->td_ksegrp; - /* serialize killing kse */ + /* Serialize killing KSE */ PROC_LOCK(p); mtx_lock_spin(&sched_lock); if ((kg->kg_kses == 1) && (kg->kg_numthreads > 1)) { @@ -329,14 +438,17 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) return (EDEADLK); } ke = td->td_kse; + upcall_remove(td); if (p->p_numthreads == 1) { - ke->ke_flags &= ~KEF_DOUPCALL; - ke->ke_mailbox = NULL; + kse_purge(p, td); p->p_flag &= ~P_KSES; mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); } else { - ke->ke_flags |= KEF_EXIT; + if (kg->kg_numthreads == 1) { /* Shutdown a group */ + kse_purge_group(td); + ke->ke_flags |= KEF_EXIT; + } thread_exit(); /* NOTREACHED */ } @@ -345,10 +457,15 @@ kse_exit(struct thread *td, struct kse_exit_args *uap) /* * Either becomes an upcall or waits for an awakening event and - * THEN becomes an upcall. Only error cases return. + * then becomes an upcall. Only error cases return. */ +/* +struct kse_release_args { + register_t dummy; +}; +*/ int -kse_release(struct thread * td, struct kse_release_args * uap) +kse_release(struct thread *td, struct kse_release_args *uap) { struct proc *p; struct ksegrp *kg; @@ -356,28 +473,25 @@ kse_release(struct thread * td, struct kse_release_args * uap) p = td->td_proc; kg = td->td_ksegrp; /* - * kse must have a mailbox ready for upcall, and only UTS can - * do the syscall. - */ - if (!(p->p_flag & P_KSES) || - (td->td_mailbox != NULL) || - (td->td_kse->ke_mailbox == NULL)) + * Only UTS can call the syscall and current group + * should be a threaded group. + */ + if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0)) return (EINVAL); + KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__)); PROC_LOCK(p); mtx_lock_spin(&sched_lock); /* Change OURSELF to become an upcall. */ - td->td_flags = TDF_UPCALLING; /* BOUND */ - if (!(td->td_kse->ke_flags & (KEF_DOUPCALL|KEF_ASTPENDING)) && + td->td_flags = TDF_UPCALLING; + if ((td->td_upcall->ku_flags & KUF_DOUPCALL) == 0 && (kg->kg_completed == NULL)) { - /* - * The KSE will however be lendable. - */ - TD_SET_IDLE(td); - PROC_UNLOCK(p); - p->p_stats->p_ru.ru_nvcsw++; - mi_switch(); + kg->kg_upsleeps++; mtx_unlock_spin(&sched_lock); + msleep(&kg->kg_completed, &p->p_mtx, PPAUSE|PCATCH, "ksepause", + NULL); + kg->kg_upsleeps--; + PROC_UNLOCK(p); } else { mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); @@ -392,61 +506,59 @@ int kse_wakeup(struct thread *td, struct kse_wakeup_args *uap) { struct proc *p; - struct kse *ke; struct ksegrp *kg; + struct kse_upcall *ku; struct thread *td2; p = td->td_proc; td2 = NULL; + ku = NULL; /* KSE-enabled processes only, please. */ if (!(p->p_flag & P_KSES)) - return EINVAL; + return (EINVAL); + PROC_LOCK(p); mtx_lock_spin(&sched_lock); if (uap->mbx) { FOREACH_KSEGRP_IN_PROC(p, kg) { - FOREACH_KSE_IN_GROUP(kg, ke) { - if (ke->ke_mailbox != uap->mbx) - continue; - td2 = ke->ke_owner; - KASSERT((td2 != NULL),("KSE with no owner")); - break; + FOREACH_UPCALL_IN_GROUP(kg, ku) { + if (ku->ku_mailbox == uap->mbx) + break; } - if (td2) { + if (ku) break; - } } } else { - /* - * look for any idle KSE to resurrect. - */ kg = td->td_ksegrp; - FOREACH_KSE_IN_GROUP(kg, ke) { - td2 = ke->ke_owner; - KASSERT((td2 != NULL),("KSE with no owner2")); - if (TD_IS_IDLE(td2)) - break; + if (kg->kg_upsleeps) { + wakeup_one(&kg->kg_completed); + mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); + return (0); } - KASSERT((td2 != NULL), ("no thread(s)")); + ku = TAILQ_FIRST(&kg->kg_upcalls); } - if (td2) { - if (TD_IS_IDLE(td2)) { - TD_CLR_IDLE(td2); - setrunnable(td2); - } else if (td != td2) { - /* guarantee do an upcall ASAP */ - td2->td_kse->ke_flags |= KEF_DOUPCALL; + if (ku) { + if ((td2 = ku->ku_owner) == NULL) { + panic("%s: no owner", __func__); + } else if (TD_ON_SLEEPQ(td2) && + (td2->td_wchan == &kg->kg_completed)) { + abortsleep(td2); + } else { + ku->ku_flags |= KUF_DOUPCALL; } mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); return (0); } mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); return (ESRCH); } /* * No new KSEG: first call: use current KSE, don't schedule an upcall - * All other situations, do allocate a new KSE and schedule an upcall on it. + * All other situations, do allocate max new KSEs and schedule an upcall. */ /* struct kse_create_args { struct kse_mailbox *mbx; @@ -456,112 +568,140 @@ int kse_create(struct thread *td, struct kse_create_args *uap) { struct kse *newke; - struct kse *ke; struct ksegrp *newkg; struct ksegrp *kg; struct proc *p; struct kse_mailbox mbx; - int err; + struct kse_upcall *newku; + int err, ncpus; p = td->td_proc; if ((err = copyin(uap->mbx, &mbx, sizeof(mbx)))) return (err); - p->p_flag |= P_KSES; /* easier to just set it than to test and set */ + /* Too bad, why hasn't kernel always a cpu counter !? */ +#ifdef SMP + ncpus = mp_ncpus; +#else + ncpus = 1; +#endif + if (thread_debug && virtual_cpu != 0) + ncpus = virtual_cpu; + + /* Easier to just set it than to test and set */ + p->p_flag |= P_KSES; kg = td->td_ksegrp; if (uap->newgroup) { + /* Have race condition but it is cheap */ if (p->p_numksegrps >= max_groups_per_proc) return (EPROCLIM); /* * If we want a new KSEGRP it doesn't matter whether * we have already fired up KSE mode before or not. - * We put the process in KSE mode and create a new KSEGRP - * and KSE. If our KSE has not got a mailbox yet then - * that doesn't matter, just leave it that way. It will - * ensure that this thread stay BOUND. It's possible - * that the call came form a threaded library and the main - * program knows nothing of threads. + * We put the process in KSE mode and create a new KSEGRP. */ newkg = ksegrp_alloc(); bzero(&newkg->kg_startzero, RANGEOF(struct ksegrp, - kg_startzero, kg_endzero)); + kg_startzero, kg_endzero)); bcopy(&kg->kg_startcopy, &newkg->kg_startcopy, RANGEOF(struct ksegrp, kg_startcopy, kg_endcopy)); - newke = kse_alloc(); + mtx_lock_spin(&sched_lock); + ksegrp_link(newkg, p); + if (p->p_numksegrps >= max_groups_per_proc) { + ksegrp_unlink(newkg); + mtx_unlock_spin(&sched_lock); + return (EPROCLIM); + } + mtx_unlock_spin(&sched_lock); } else { - /* - * Otherwise, if we have already set this KSE - * to have a mailbox, we want to make another KSE here, - * but only if there are not already the limit, which - * is 1 per CPU max. - * - * If the current KSE doesn't have a mailbox we just use it - * and give it one. - * - * Because we don't like to access - * the KSE outside of schedlock if we are UNBOUND, - * (because it can change if we are preempted by an interrupt) - * we can deduce it as having a mailbox if we are UNBOUND, - * and only need to actually look at it if we are BOUND, - * which is safe. + newkg = kg; + } + + /* + * Creating upcalls more than number of physical cpu does + * not help performance. + */ + if (newkg->kg_numupcalls >= ncpus) + return (EPROCLIM); + + if (newkg->kg_numupcalls == 0) { + /* + * Initialize KSE group, optimized for MP. + * Create KSEs as many as physical cpus, this increases + * concurrent even if userland is not MP safe and can only run + * on single CPU (for early version of libpthread, it is true). + * In ideal world, every physical cpu should execute a thread. + * If there is enough KSEs, threads in kernel can be + * executed parallel on different cpus with full speed, + * Concurrent in kernel shouldn't be restricted by number of + * upcalls userland provides. + * Adding more upcall structures only increases concurrent + * in userland. + * Highest performance configuration is: + * N kses = N upcalls = N phyiscal cpus */ - if ((td->td_flags & TDF_UNBOUND) || td->td_kse->ke_mailbox) { - if (thread_debug == 0) { /* if debugging, allow more */ -#ifdef SMP - if (kg->kg_kses > mp_ncpus) -#endif - return (EPROCLIM); - } + while (newkg->kg_kses < ncpus) { newke = kse_alloc(); - } else { - newke = NULL; - } - newkg = NULL; - } - if (newke) { - bzero(&newke->ke_startzero, RANGEOF(struct kse, - ke_startzero, ke_endzero)); + bzero(&newke->ke_startzero, RANGEOF(struct kse, + ke_startzero, ke_endzero)); #if 0 - bcopy(&ke->ke_startcopy, &newke->ke_startcopy, - RANGEOF(struct kse, ke_startcopy, ke_endcopy)); + mtx_lock_spin(&sched_lock); + bcopy(&ke->ke_startcopy, &newke->ke_startcopy, + RANGEOF(struct kse, ke_startcopy, ke_endcopy)); + mtx_unlock_spin(&sched_lock); #endif - /* For the first call this may not have been set */ - if (td->td_standin == NULL) { - td->td_standin = thread_alloc(); - } - mtx_lock_spin(&sched_lock); - if (newkg) { - if (p->p_numksegrps >= max_groups_per_proc) { - mtx_unlock_spin(&sched_lock); - ksegrp_free(newkg); - kse_free(newke); - return (EPROCLIM); - } - ksegrp_link(newkg, p); + mtx_lock_spin(&sched_lock); + kse_link(newke, newkg); + if (p->p_sflag & PS_NEEDSIGCHK) + newke->ke_flags |= KEF_ASTPENDING; + /* Add engine */ + kse_reassign(newke); + mtx_unlock_spin(&sched_lock); } - else - newkg = kg; - kse_link(newke, newkg); - if (p->p_sflag & PS_NEEDSIGCHK) - newke->ke_flags |= KEF_ASTPENDING; - newke->ke_mailbox = uap->mbx; - newke->ke_upcall = mbx.km_func; - bcopy(&mbx.km_stack, &newke->ke_stack, sizeof(stack_t)); - thread_schedule_upcall(td, newke); + } + newku = upcall_alloc(); + newku->ku_mailbox = uap->mbx; + newku->ku_func = mbx.km_func; + bcopy(&mbx.km_stack, &newku->ku_stack, sizeof(stack_t)); + + /* For the first call this may not have been set */ + if (td->td_standin == NULL) + thread_alloc_spare(td, NULL); + + mtx_lock_spin(&sched_lock); + if (newkg->kg_numupcalls >= ncpus) { + upcall_free(newku); mtx_unlock_spin(&sched_lock); + return (EPROCLIM); + } + upcall_link(newku, newkg); + + /* + * Each upcall structure has an owner thread, find which + * one owns it. + */ + if (uap->newgroup) { + /* + * Because new ksegrp hasn't thread, + * create an initial upcall thread to own it. + */ + thread_schedule_upcall(td, newku); } else { /* - * If we didn't allocate a new KSE then the we are using - * the exisiting (BOUND) kse. + * If current thread hasn't an upcall structure, + * just assign the upcall to it. */ - ke = td->td_kse; - ke->ke_mailbox = uap->mbx; - ke->ke_upcall = mbx.km_func; - bcopy(&mbx.km_stack, &ke->ke_stack, sizeof(stack_t)); + if (td->td_upcall == NULL) { + newku->ku_owner = td; + td->td_upcall = newku; + } else { + /* + * Create a new upcall thread to own it. + */ + thread_schedule_upcall(td, newku); + } } - /* - * Fill out the KSE-mode specific fields of the new kse. - */ + mtx_unlock_spin(&sched_lock); return (0); } @@ -642,6 +782,8 @@ threadinit(void) kse_zone = uma_zcreate("KSE", sched_sizeof_kse(), NULL, NULL, kse_init, NULL, UMA_ALIGN_CACHE, 0); + upcall_zone = uma_zcreate("UPCALL", sizeof(struct kse_upcall), + NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); } /* @@ -650,9 +792,9 @@ threadinit(void) void thread_stash(struct thread *td) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); } /* @@ -661,9 +803,21 @@ thread_stash(struct thread *td) void kse_stash(struct kse *ke) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_kses, ke, ke_procq); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); +} + +/* + * Stash an embarasingly extra upcall into the zombie upcall queue. + */ + +void +upcall_stash(struct kse_upcall *ku) +{ + mtx_lock_spin(&kse_zombie_lock); + TAILQ_INSERT_HEAD(&zombie_upcalls, ku, ku_link); + mtx_unlock_spin(&kse_zombie_lock); } /* @@ -672,13 +826,13 @@ kse_stash(struct kse *ke) void ksegrp_stash(struct ksegrp *kg) { - mtx_lock_spin(&zombie_thread_lock); + mtx_lock_spin(&kse_zombie_lock); TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp); - mtx_unlock_spin(&zombie_thread_lock); + mtx_unlock_spin(&kse_zombie_lock); } /* - * Reap zombie threads. + * Reap zombie kse resource. */ void thread_reap(void) @@ -686,27 +840,34 @@ thread_reap(void) struct thread *td_first, *td_next; struct kse *ke_first, *ke_next; struct ksegrp *kg_first, * kg_next; + struct kse_upcall *ku_first, *ku_next; /* - * don't even bother to lock if none at this instant - * We really don't care about the next instant.. + * Don't even bother to lock if none at this instant, + * we really don't care about the next instant.. */ if ((!TAILQ_EMPTY(&zombie_threads)) || (!TAILQ_EMPTY(&zombie_kses)) - || (!TAILQ_EMPTY(&zombie_ksegrps))) { - mtx_lock_spin(&zombie_thread_lock); + || (!TAILQ_EMPTY(&zombie_ksegrps)) + || (!TAILQ_EMPTY(&zombie_upcalls))) { + mtx_lock_spin(&kse_zombie_lock); td_first = TAILQ_FIRST(&zombie_threads); ke_first = TAILQ_FIRST(&zombie_kses); kg_first = TAILQ_FIRST(&zombie_ksegrps); + ku_first = TAILQ_FIRST(&zombie_upcalls); if (td_first) TAILQ_INIT(&zombie_threads); if (ke_first) TAILQ_INIT(&zombie_kses); if (kg_first) TAILQ_INIT(&zombie_ksegrps); - mtx_unlock_spin(&zombie_thread_lock); + if (ku_first) + TAILQ_INIT(&zombie_upcalls); + mtx_unlock_spin(&kse_zombie_lock); while (td_first) { td_next = TAILQ_NEXT(td_first, td_runq); + if (td_first->td_ucred) + crfree(td_first->td_ucred); thread_free(td_first); td_first = td_next; } @@ -720,6 +881,11 @@ thread_reap(void) ksegrp_free(kg_first); kg_first = kg_next; } + while (ku_first) { + ku_next = TAILQ_NEXT(ku_first, ku_link); + upcall_free(ku_first); + ku_first = ku_next; + } } } @@ -792,20 +958,14 @@ thread_export_context(struct thread *td) struct ksegrp *kg; uintptr_t mbx; void *addr; - int error; + int error,temp; ucontext_t uc; - uint temp; p = td->td_proc; kg = td->td_ksegrp; /* Export the user/machine context. */ -#if 0 - addr = (caddr_t)td->td_mailbox + - offsetof(struct kse_thr_mailbox, tm_context); -#else /* if user pointer arithmetic is valid in the kernel */ - addr = (void *)(&td->td_mailbox->tm_context); -#endif + addr = (void *)(&td->td_mailbox->tm_context); error = copyin(addr, &uc, sizeof(ucontext_t)); if (error) goto bad; @@ -815,13 +975,14 @@ thread_export_context(struct thread *td) if (error) goto bad; - /* get address in latest mbox of list pointer */ -#if 0 - addr = (caddr_t)td->td_mailbox - + offsetof(struct kse_thr_mailbox , tm_next); -#else /* if user pointer arithmetic is valid in the kernel */ + /* Exports clock ticks in kernel mode */ + addr = (caddr_t)(&td->td_mailbox->tm_sticks); + temp = fuword(addr) + td->td_usticks; + if (suword(addr, temp)) + goto bad; + + /* Get address in latest mbox of list pointer */ addr = (void *)(&td->td_mailbox->tm_next); -#endif /* * Put the saved address of the previous first * entry into this one @@ -835,42 +996,43 @@ thread_export_context(struct thread *td) PROC_LOCK(p); if (mbx == (uintptr_t)kg->kg_completed) { kg->kg_completed = td->td_mailbox; + /* + * The thread context may be taken away by + * other upcall threads when we unlock + * process lock. it's no longer valid to + * use it again in any other places. + */ + td->td_mailbox = NULL; PROC_UNLOCK(p); break; } PROC_UNLOCK(p); } - addr = (caddr_t)td->td_mailbox - + offsetof(struct kse_thr_mailbox, tm_sticks); - temp = fuword(addr) + td->td_usticks; - if (suword(addr, temp)) - goto bad; + td->td_usticks = 0; return (0); bad: PROC_LOCK(p); psignal(p, SIGSEGV); PROC_UNLOCK(p); + /* The mailbox is bad, don't use it */ + td->td_mailbox = NULL; + td->td_usticks = 0; return (error); } /* * Take the list of completed mailboxes for this KSEGRP and put them on this - * KSE's mailbox as it's the next one going up. + * upcall's mailbox as it's the next one going up. */ static int -thread_link_mboxes(struct ksegrp *kg, struct kse *ke) +thread_link_mboxes(struct ksegrp *kg, struct kse_upcall *ku) { struct proc *p = kg->kg_proc; void *addr; uintptr_t mbx; -#if 0 - addr = (caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_completed); -#else /* if user pointer arithmetic is valid in the kernel */ - addr = (void *)(&ke->ke_mailbox->km_completed); -#endif + addr = (void *)(&ku->ku_mailbox->km_completed); for (;;) { mbx = (uintptr_t)kg->kg_completed; if (suword(addr, mbx)) { @@ -895,69 +1057,91 @@ thread_link_mboxes(struct ksegrp *kg, struct kse *ke) * This function should be called at statclock interrupt time */ int -thread_add_ticks_intr(int user, uint ticks) +thread_statclock(int user) { struct thread *td = curthread; - struct kse *ke = td->td_kse; - if (ke->ke_mailbox == NULL) - return -1; + if (td->td_ksegrp->kg_numupcalls == 0) + return (-1); if (user) { /* Current always do via ast() */ - ke->ke_flags |= KEF_ASTPENDING; - ke->ke_uuticks += ticks; + td->td_flags |= (TDF_ASTPENDING|TDF_USTATCLOCK); + td->td_uuticks += ticks; } else { if (td->td_mailbox != NULL) td->td_usticks += ticks; - else - ke->ke_usticks += ticks; + else { + /* XXXKSE + * We will call thread_user_enter() for every + * kernel entry in future, so if the thread mailbox + * is NULL, it must be a UTS kernel, don't account + * clock ticks for it. + */ + } } - return 0; + return (0); } +/* + * Export user mode state clock ticks + */ static int -thread_update_uticks(void) +thread_update_usr_ticks(struct thread *td) { - struct thread *td = curthread; struct proc *p = td->td_proc; - struct kse *ke = td->td_kse; struct kse_thr_mailbox *tmbx; + struct kse_upcall *ku; caddr_t addr; - uint uticks, sticks; - - if (ke->ke_mailbox == NULL) - return 0; + uint uticks; - uticks = ke->ke_uuticks; - ke->ke_uuticks = 0; - sticks = ke->ke_usticks; - ke->ke_usticks = 0; -#if 0 - tmbx = (void *)fuword((caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_curthread)); -#else /* if user pointer arithmetic is ok in the kernel */ - tmbx = (void *)fuword( (void *)&ke->ke_mailbox->km_curthread); -#endif + if ((ku = td->td_upcall) == NULL) + return (-1); + + tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread); if ((tmbx == NULL) || (tmbx == (void *)-1)) - return 0; + return (-1); + uticks = td->td_uuticks; + td->td_uuticks = 0; if (uticks) { - addr = (caddr_t)tmbx + offsetof(struct kse_thr_mailbox, tm_uticks); + addr = (caddr_t)&tmbx->tm_uticks; uticks += fuword(addr); - if (suword(addr, uticks)) - goto bad; + if (suword(addr, uticks)) { + PROC_LOCK(p); + psignal(p, SIGSEGV); + PROC_UNLOCK(p); + return (-2); + } } - if (sticks) { - addr = (caddr_t)tmbx + offsetof(struct kse_thr_mailbox, tm_sticks); - sticks += fuword(addr); - if (suword(addr, sticks)) - goto bad; + return (0); +} + +/* + * Export kernel mode state clock ticks + */ + +static int +thread_update_sys_ticks(struct thread *td) +{ + struct proc *p = td->td_proc; + caddr_t addr; + int sticks; + + if (td->td_mailbox == NULL) + return (-1); + if (td->td_usticks == 0) + return (0); + addr = (caddr_t)&td->td_mailbox->tm_sticks; + sticks = fuword(addr); + /* XXXKSE use XCHG instead */ + sticks += td->td_usticks; + td->td_usticks = 0; + if (suword(addr, sticks)) { + PROC_LOCK(p); + psignal(p, SIGSEGV); + PROC_UNLOCK(p); + return (-2); } - return 0; -bad: - PROC_LOCK(p); - psignal(p, SIGSEGV); - PROC_UNLOCK(p); - return -1; + return (0); } /* @@ -1013,6 +1197,7 @@ thread_exit(void) p->p_numthreads--; TAILQ_REMOVE(&kg->kg_threads, td, td_kglist); kg->kg_numthreads--; + /* * The test below is NOT true if we are the * sole exiting thread. P_STOPPED_SNGL is unset @@ -1024,25 +1209,28 @@ thread_exit(void) } } - /* Reassign this thread's KSE. */ + /* + * Because each upcall structure has an owner thread, + * owner thread exits only when process is in exiting + * state, so upcall to userland is no longer needed, + * deleting upcall structure is safe here. + * So when all threads in a group is exited, all upcalls + * in the group should be automatically freed. + */ + if (td->td_upcall) + upcall_remove(td); + ke->ke_state = KES_UNQUEUED; - + ke->ke_thread = NULL; /* * Decide what to do with the KSE attached to this thread. - * XXX Possibly kse_reassign should do both cases as it already - * does some of this. */ - if (ke->ke_flags & KEF_EXIT) { - KASSERT((ke->ke_owner == td), - ("thread_exit: KSE exiting with non-owner thread")); - ke->ke_thread = NULL; - td->td_kse = NULL; + if (ke->ke_flags & KEF_EXIT) kse_unlink(ke); - } else { - TD_SET_EXITING(td); /* definitly not runnable */ + else kse_reassign(ke); - } PROC_UNLOCK(p); + td->td_kse = NULL; td->td_state = TDS_INACTIVE; td->td_proc = NULL; td->td_ksegrp = NULL; @@ -1090,10 +1278,12 @@ thread_link(struct thread *td, struct ksegrp *kg) struct proc *p; p = kg->kg_proc; - td->td_state = TDS_INACTIVE; - td->td_proc = p; - td->td_ksegrp = kg; - td->td_last_kse = NULL; + td->td_state = TDS_INACTIVE; + td->td_proc = p; + td->td_ksegrp = kg; + td->td_last_kse = NULL; + td->td_flags = 0; + td->td_kse = NULL; LIST_INIT(&td->td_contested); callout_init(&td->td_slpcallout, 1); @@ -1101,116 +1291,139 @@ thread_link(struct thread *td, struct ksegrp *kg) TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist); p->p_numthreads++; kg->kg_numthreads++; - td->td_kse = NULL; } +/* + * Purge a ksegrp resource. When a ksegrp is preparing to + * exit, it calls this function. + */ +void +kse_purge_group(struct thread *td) +{ + struct ksegrp *kg; + struct kse *ke; + + kg = td->td_ksegrp; + KASSERT(kg->kg_numthreads == 1, ("%s: bad thread number", __func__)); + while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) { + KASSERT(ke->ke_state == KES_IDLE, + ("%s: wrong idle KSE state", __func__)); + kse_unlink(ke); + } + KASSERT((kg->kg_kses == 1), + ("%s: ksegrp still has %d KSEs", __func__, kg->kg_kses)); + KASSERT((kg->kg_numupcalls == 0), + ("%s: ksegrp still has %d upcall datas", + __func__, kg->kg_numupcalls)); +} + +/* + * Purge a process's KSE resource. When a process is preparing to + * exit, it calls kse_purge to release any extra KSE resources in + * the process. + */ void kse_purge(struct proc *p, struct thread *td) { - /* XXXKSE think about this.. - may need to wake up threads on loan queue. */ struct ksegrp *kg; + struct kse *ke; KASSERT(p->p_numthreads == 1, ("bad thread number")); mtx_lock_spin(&sched_lock); while ((kg = TAILQ_FIRST(&p->p_ksegrps)) != NULL) { TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp); p->p_numksegrps--; + /* + * There is no ownership for KSE, after all threads + * in the group exited, it is possible that some KSEs + * were left in idle queue, gc them now. + */ + while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) { + KASSERT(ke->ke_state == KES_IDLE, + ("%s: wrong idle KSE state", __func__)); + TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist); + kg->kg_idle_kses--; + TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist); + kg->kg_kses--; + kse_stash(ke); + } KASSERT(((kg->kg_kses == 0) && (kg != td->td_ksegrp)) || - ((kg->kg_kses == 1) && (kg == td->td_ksegrp)), - ("wrong kg_kses")); - if (kg != td->td_ksegrp) { + ((kg->kg_kses == 1) && (kg == td->td_ksegrp)), + ("ksegrp has wrong kg_kses: %d", kg->kg_kses)); + KASSERT((kg->kg_numupcalls == 0), + ("%s: ksegrp still has %d upcall datas", + __func__, kg->kg_numupcalls)); + + if (kg != td->td_ksegrp) ksegrp_stash(kg); - } } TAILQ_INSERT_HEAD(&p->p_ksegrps, td->td_ksegrp, kg_ksegrp); p->p_numksegrps++; mtx_unlock_spin(&sched_lock); } +/* + * This function is intended to be used to initialize a spare thread + * for upcall. Initialize thread's large data area outside sched_lock + * for thread_schedule_upcall(). + */ +void +thread_alloc_spare(struct thread *td, struct thread *spare) +{ + if (td->td_standin) + return; + if (spare == NULL) + spare = thread_alloc(); + td->td_standin = spare; + bzero(&spare->td_startzero, + (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); + spare->td_proc = td->td_proc; + /* Setup PCB and fork address */ + cpu_set_upcall(spare, td->td_pcb); + /* + * XXXKSE do we really need this? (default values for the + * frame). + */ + bcopy(td->td_frame, spare->td_frame, sizeof(struct trapframe)); + spare->td_ucred = crhold(td->td_ucred); +} /* * Create a thread and schedule it for upcall on the KSE given. * Use our thread's standin so that we don't have to allocate one. */ struct thread * -thread_schedule_upcall(struct thread *td, struct kse *ke) +thread_schedule_upcall(struct thread *td, struct kse_upcall *ku) { struct thread *td2; - int newkse; mtx_assert(&sched_lock, MA_OWNED); - newkse = (ke != td->td_kse); /* - * If the owner and kse are BOUND then that thread is planning to - * go to userland and upcalls are not expected. So don't make one. - * If it is not bound then make it so with the spare thread - * anf then borrw back the KSE to allow us to complete some in-kernel - * work. When we complete, the Bound thread will have the chance to - * complete. This thread will sleep as planned. Hopefully there will - * eventually be un unbound thread that can be converted to an - * upcall to report the completion of this thread. + * Schedule an upcall thread on specified kse_upcall, + * the kse_upcall must be free. + * td must have a spare thread. */ - + KASSERT(ku->ku_owner == NULL, ("%s: upcall has owner", __func__)); if ((td2 = td->td_standin) != NULL) { td->td_standin = NULL; } else { - if (newkse) - panic("no reserve thread when called with a new kse"); - /* - * If called from (e.g.) sleep and we do not have - * a reserve thread, then we've used it, so do not - * create an upcall. - */ + panic("no reserve thread when scheduling an upcall"); return (NULL); } CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)", td2, td->td_proc->p_pid, td->td_proc->p_comm); - bzero(&td2->td_startzero, - (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); bcopy(&td->td_startcopy, &td2->td_startcopy, (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy)); - thread_link(td2, ke->ke_ksegrp); - cpu_set_upcall(td2, td->td_pcb); - - /* - * XXXKSE do we really need this? (default values for the - * frame). - */ - bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe)); - - /* - * Bind the new thread to the KSE, - * and if it's our KSE, lend it back to ourself - * so we can continue running. - */ - td2->td_ucred = crhold(td->td_ucred); - td2->td_flags = TDF_UPCALLING; /* note: BOUND */ - td2->td_kse = ke; - td2->td_state = TDS_CAN_RUN; + thread_link(td2, ku->ku_ksegrp); + /* Let the new thread become owner of the upcall */ + ku->ku_owner = td2; + td2->td_upcall = ku; + td2->td_flags = TDF_UPCALLING; + td2->td_kse = NULL; + td2->td_state = TDS_CAN_RUN; td2->td_inhibitors = 0; - ke->ke_owner = td2; - /* - * If called from kse_reassign(), we are working on the current - * KSE so fake that we borrowed it. If called from - * kse_create(), don't, as we have a new kse too. - */ - if (!newkse) { - /* - * This thread will be scheduled when the current thread - * blocks, exits or tries to enter userspace, (which ever - * happens first). When that happens the KSe will "revert" - * to this thread in a BOUND manner. Since we are called - * from msleep() this is going to be "very soon" in nearly - * all cases. - */ - TD_SET_LOAN(td2); - } else { - ke->ke_thread = td2; - ke->ke_state = KES_THREAD; - setrunqueue(td2); - } + setrunqueue(td2); return (td2); /* bogus.. should be a void function */ } @@ -1222,14 +1435,16 @@ thread_schedule_upcall(struct thread *td, struct kse *ke) struct thread * signal_upcall(struct proc *p, int sig) { +#if 0 struct thread *td, *td2; struct kse *ke; sigset_t ss; int error; +#endif PROC_LOCK_ASSERT(p, MA_OWNED); return (NULL); - +#if 0 td = FIRST_THREAD_IN_PROC(p); ke = td->td_kse; PROC_UNLOCK(p); @@ -1244,28 +1459,31 @@ return (NULL); if (error) return (NULL); if (td->td_standin == NULL) - td->td_standin = thread_alloc(); + thread_alloc_spare(td, NULL); mtx_lock_spin(&sched_lock); td2 = thread_schedule_upcall(td, ke); /* Bogus JRE */ mtx_unlock_spin(&sched_lock); return (td2); +#endif } /* - * setup done on the thread when it enters the kernel. + * Setup done on the thread when it enters the kernel. * XXXKSE Presently only for syscalls but eventually all kernel entries. */ void thread_user_enter(struct proc *p, struct thread *td) { - struct kse *ke; + struct ksegrp *kg; + struct kse_upcall *ku; + kg = td->td_ksegrp; /* * First check that we shouldn't just abort. * But check if we are the single thread first! * XXX p_singlethread not locked, but should be safe. */ - if ((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) { + if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { PROC_LOCK(p); mtx_lock_spin(&sched_lock); thread_exit(); @@ -1278,43 +1496,37 @@ thread_user_enter(struct proc *p, struct thread *td) * possibility that we could do this lazily (in kse_reassign()), * but for now do it every time. */ - ke = td->td_kse; - td->td_flags &= ~TDF_UNBOUND; - if (ke->ke_mailbox != NULL) { -#if 0 - td->td_mailbox = (void *)fuword((caddr_t)ke->ke_mailbox - + offsetof(struct kse_mailbox, km_curthread)); -#else /* if user pointer arithmetic is ok in the kernel */ + kg = td->td_ksegrp; + if (kg->kg_numupcalls) { + ku = td->td_upcall; + KASSERT(ku, ("%s: no upcall owned", __func__)); + KASSERT((ku->ku_owner == td), ("%s: wrong owner", __func__)); td->td_mailbox = - (void *)fuword( (void *)&ke->ke_mailbox->km_curthread); -#endif + (void *)fuword((void *)&ku->ku_mailbox->km_curthread); if ((td->td_mailbox == NULL) || (td->td_mailbox == (void *)-1)) { - td->td_mailbox = NULL; /* single thread it.. */ + /* Don't schedule upcall when blocked */ + td->td_mailbox = NULL; mtx_lock_spin(&sched_lock); - td->td_flags &= ~(TDF_UNBOUND|TDF_CAN_UNBIND); + td->td_flags &= ~TDF_CAN_UNBIND; mtx_unlock_spin(&sched_lock); } else { - /* - * when thread limit reached, act like that the thread - * has already done an upcall. - */ if (p->p_numthreads > max_threads_per_proc) { - if (td->td_standin != NULL) { - thread_stash(td->td_standin); - td->td_standin = NULL; - } + /* + * Since kernel thread limit reached, + * don't schedule upcall anymore. + * XXXKSE These code in fact needn't. + */ + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); } else { if (td->td_standin == NULL) - td->td_standin = thread_alloc(); + thread_alloc_spare(td, NULL); + mtx_lock_spin(&sched_lock); + td->td_flags |= TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); } - mtx_lock_spin(&sched_lock); - td->td_flags |= TDF_CAN_UNBIND; - mtx_unlock_spin(&sched_lock); - KASSERT((ke->ke_owner == td), - ("thread_user_enter: No starting owner ")); - ke->ke_owner = td; - td->td_usticks = 0; } } } @@ -1335,165 +1547,90 @@ int thread_userret(struct thread *td, struct trapframe *frame) { int error; - int unbound; - struct kse *ke; + struct kse_upcall *ku; struct ksegrp *kg; - struct thread *worktodo; struct proc *p; struct timespec ts; - KASSERT((td->td_kse && td->td_kse->ke_thread && td->td_kse->ke_owner), - ("thread_userret: bad thread/kse pointers")); - KASSERT((td == curthread), - ("thread_userret: bad thread argument")); - - - kg = td->td_ksegrp; p = td->td_proc; - error = 0; - unbound = TD_IS_UNBOUND(td); + kg = td->td_ksegrp; - mtx_lock_spin(&sched_lock); - if ((worktodo = kg->kg_last_assigned)) - worktodo = TAILQ_NEXT(worktodo, td_runq); - else - worktodo = TAILQ_FIRST(&kg->kg_runq); + /* Nothing to do with non-threaded group/process */ + if (td->td_ksegrp->kg_numupcalls == 0) + return (0); /* - * Permanently bound threads never upcall but they may - * loan out their KSE at this point. - * Upcalls imply bound.. They also may want to do some Philantropy. - * Temporarily bound threads on the other hand either yield - * to other work and transform into an upcall, or proceed back to - * userland. + * State clock interrupt hit in userland, it + * is returning from interrupt, charge thread's + * userland time for UTS. */ + if (td->td_flags & TDF_USTATCLOCK) { + thread_update_usr_ticks(td); + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_USTATCLOCK; + mtx_unlock_spin(&sched_lock); + } + /* + * Optimisation: + * This thread has not started any upcall. + * If there is no work to report other than ourself, + * then it can return direct to userland. + */ if (TD_CAN_UNBIND(td)) { - td->td_flags &= ~(TDF_UNBOUND|TDF_CAN_UNBIND); - if (!worktodo && (kg->kg_completed == NULL) && - !(td->td_kse->ke_flags & KEF_DOUPCALL)) { - /* - * This thread has not started any upcall. - * If there is no work to report other than - * ourself, then it can return direct to userland. - */ -justreturn: - mtx_unlock_spin(&sched_lock); - thread_update_uticks(); + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_CAN_UNBIND; + mtx_unlock_spin(&sched_lock); + if ((kg->kg_completed == NULL) && + (td->td_upcall->ku_flags & KUF_DOUPCALL) == 0) { + thread_update_sys_ticks(td); td->td_mailbox = NULL; return (0); } - mtx_unlock_spin(&sched_lock); error = thread_export_context(td); - td->td_usticks = 0; if (error) { /* - * As we are not running on a borrowed KSE, - * failing to do the KSE operation just defaults + * Failing to do the KSE operation just defaults * back to synchonous operation, so just return from * the syscall. */ - goto justreturn; + return (0); } - mtx_lock_spin(&sched_lock); /* - * Turn ourself into a bound upcall. - * We will rely on kse_reassign() - * to make us run at a later time. + * There is something to report, and we own an upcall + * strucuture, we can go to userland. + * Turn ourself into an upcall thread. */ + mtx_lock_spin(&sched_lock); td->td_flags |= TDF_UPCALLING; - - /* there may be more work since we re-locked schedlock */ - if ((worktodo = kg->kg_last_assigned)) - worktodo = TAILQ_NEXT(worktodo, td_runq); - else - worktodo = TAILQ_FIRST(&kg->kg_runq); - } else if (unbound) { - /* - * We are an unbound thread, looking to - * return to user space. There must be another owner - * of this KSE. - * We are using a borrowed KSE. save state and exit. - * kse_reassign() will recycle the kse as needed, - */ mtx_unlock_spin(&sched_lock); + } else if (td->td_mailbox) { error = thread_export_context(td); - td->td_usticks = 0; if (error) { - /* - * There is nothing we can do. - * We just lose that context. We - * probably should note this somewhere and send - * the process a signal. - */ PROC_LOCK(td->td_proc); - psignal(td->td_proc, SIGSEGV); mtx_lock_spin(&sched_lock); - ke = td->td_kse; /* possibly upcall with error? */ } else { + PROC_LOCK(td->td_proc); + mtx_lock_spin(&sched_lock); /* - * Don't make an upcall, just exit so that the owner - * can get its KSE if it wants it. - * Our context is already safely stored for later - * use by the UTS. + * There are upcall threads waiting for + * work to do, wake one of them up. + * XXXKSE Maybe wake all of them up. */ - PROC_LOCK(p); - mtx_lock_spin(&sched_lock); - ke = td->td_kse; - } - /* - * If the owner is idling, we now have something for it - * to report, so make it runnable. - * If the owner is not an upcall, make an attempt to - * ensure that at least one of any IDLED upcalls can - * wake up. - */ - if (ke->ke_owner->td_flags & TDF_UPCALLING) { - TD_CLR_IDLE(ke->ke_owner); - } else { - FOREACH_KSE_IN_GROUP(kg, ke) { - if (TD_IS_IDLE(ke->ke_owner)) { - TD_CLR_IDLE(ke->ke_owner); - setrunnable(ke->ke_owner); - break; - } - } + if (kg->kg_upsleeps) + wakeup_one(&kg->kg_completed); } thread_exit(); + /* NOTREACHED */ } - /* - * We ARE going back to userland with this KSE. - * We are permanently bound. We may be an upcall. - * If an upcall, check for threads that need to borrow the KSE. - * Any other thread that comes ready after this missed the boat. - */ - ke = td->td_kse; - /* - * If not upcalling, go back to userspace. - * If we are, get the upcall set up. - */ if (td->td_flags & TDF_UPCALLING) { - if (worktodo) { - /* - * force a switch to more urgent 'in kernel' - * work. Control will return to this thread - * when there is no more work to do. - * kse_reassign() will do that for us. - */ - TD_SET_LOAN(td); - p->p_stats->p_ru.ru_nvcsw++; - mi_switch(); /* kse_reassign() will (re)find worktodo */ - } - td->td_flags &= ~TDF_UPCALLING; - if (ke->ke_flags & KEF_DOUPCALL) - ke->ke_flags &= ~KEF_DOUPCALL; - mtx_unlock_spin(&sched_lock); - + KASSERT(TD_CAN_UNBIND(td) == 0, ("upcall thread can unbind")); + ku = td->td_upcall; /* * There is no more work to do and we are going to ride - * this thread/KSE up to userland as an upcall. + * this thread up to userland as an upcall. * Do the last parts of the setup needed for the upcall. */ CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)", @@ -1504,16 +1641,27 @@ justreturn: * Will use Giant in cpu_thread_clean() because it uses * kmem_free(kernel_map, ...) */ - cpu_set_upcall_kse(td, ke); + cpu_set_upcall_kse(td, ku); - /* + /* + * Clear TDF_UPCALLING after set upcall context, + * profiling code looks TDF_UPCALLING to avoid account + * a wrong user %EIP + */ + mtx_lock_spin(&sched_lock); + td->td_flags &= ~TDF_UPCALLING; + if (ku->ku_flags & KUF_DOUPCALL) + ku->ku_flags &= ~KUF_DOUPCALL; + mtx_unlock_spin(&sched_lock); + + /* * Unhook the list of completed threads. * anything that completes after this gets to * come in next time. * Put the list of completed thread mailboxes on * this KSE's mailbox. */ - error = thread_link_mboxes(kg, ke); + error = thread_link_mboxes(kg, ku); if (error) goto bad; @@ -1524,34 +1672,33 @@ justreturn: * it would be nice if this all happenned only on the first * time through. (the scan for extra work etc.) */ -#if 0 - error = suword((caddr_t)ke->ke_mailbox + - offsetof(struct kse_mailbox, km_curthread), 0); -#else /* if user pointer arithmetic is ok in the kernel */ - error = suword((caddr_t)&ke->ke_mailbox->km_curthread, 0); -#endif - ke->ke_uuticks = ke->ke_usticks = 0; + error = suword((caddr_t)&ku->ku_mailbox->km_curthread, 0); if (error) goto bad; + + /* Export current system time */ nanotime(&ts); if (copyout(&ts, - (caddr_t)&ke->ke_mailbox->km_timeofday, sizeof(ts))) { + (caddr_t)&ku->ku_mailbox->km_timeofday, sizeof(ts))) { goto bad; } - } else { - mtx_unlock_spin(&sched_lock); } /* * Optimisation: * Ensure that we have a spare thread available, * for when we re-enter the kernel. */ - if (td->td_standin == NULL) { - td->td_standin = thread_alloc(); - } + if (td->td_standin == NULL) + thread_alloc_spare(td, NULL); - thread_update_uticks(); + /* + * Clear thread mailbox first, then clear system tick count. + * The order is important because thread_statclock() use + * mailbox pointer to see if it is an userland thread or + * an UTS kernel thread. + */ td->td_mailbox = NULL; + td->td_usticks = 0; return (0); bad: @@ -1563,6 +1710,7 @@ bad: psignal(td->td_proc, SIGSEGV); PROC_UNLOCK(td->td_proc); td->td_mailbox = NULL; + td->td_usticks = 0; return (error); /* go sync */ } @@ -1601,7 +1749,6 @@ thread_single(int force_exit) if (force_exit == SINGLE_EXIT) { p->p_flag |= P_SINGLE_EXIT; - td->td_flags &= ~TDF_UNBOUND; } else p->p_flag &= ~P_SINGLE_EXIT; p->p_flag |= P_STOPPED_SINGLE; @@ -1624,17 +1771,16 @@ thread_single(int force_exit) else abortsleep(td2); } - if (TD_IS_IDLE(td2)) { - TD_CLR_IDLE(td2); - } } else { if (TD_IS_SUSPENDED(td2)) continue; - /* maybe other inhibitted states too? */ + /* + * maybe other inhibitted states too? + * XXXKSE Is it totally safe to + * suspend a non-interruptable thread? + */ if (td2->td_inhibitors & - (TDI_SLEEPING | TDI_SWAPPED | - TDI_LOAN | TDI_IDLE | - TDI_EXITING)) + (TDI_SLEEPING | TDI_SWAPPED)) thread_suspend_one(td2); } } @@ -1660,8 +1806,14 @@ thread_single(int force_exit) mtx_lock(&Giant); PROC_LOCK(p); } - if (force_exit == SINGLE_EXIT) + if (force_exit == SINGLE_EXIT) { + if (td->td_upcall) { + mtx_lock_spin(&sched_lock); + upcall_remove(td); + mtx_unlock_spin(&sched_lock); + } kse_purge(p, td); + } return (0); } @@ -1703,7 +1855,6 @@ thread_suspend_check(int return_instead) { struct thread *td; struct proc *p; - struct kse *ke; struct ksegrp *kg; td = curthread; @@ -1735,16 +1886,6 @@ thread_suspend_check(int return_instead) mtx_lock_spin(&sched_lock); while (mtx_owned(&Giant)) mtx_unlock(&Giant); - /* - * All threads should be exiting - * Unless they are the active "singlethread". - * destroy un-needed KSEs as we go.. - * KSEGRPS may implode too as #kses -> 0. - */ - ke = td->td_kse; - if (ke->ke_owner == td && - (kg->kg_kses >= kg->kg_numthreads )) - ke->ke_flags |= KEF_EXIT; thread_exit(); } @@ -1752,14 +1893,6 @@ thread_suspend_check(int return_instead) * When a thread suspends, it just * moves to the processes's suspend queue * and stays there. - * - * XXXKSE if TDF_BOUND is true - * it will not release it's KSE which might - * lead to deadlock if there are not enough KSEs - * to complete all waiting threads. - * Maybe be able to 'lend' it out again. - * (lent kse's can not go back to userland?) - * and can only be lent in STOPPED state. */ mtx_lock_spin(&sched_lock); if ((p->p_flag & P_STOPPED_SIG) && diff --git a/sys/kern/subr_prof.c b/sys/kern/subr_prof.c index 2c22a92..ecf309e 100644 --- a/sys/kern/subr_prof.c +++ b/sys/kern/subr_prof.c @@ -358,7 +358,9 @@ sysctl_kern_prof(SYSCTL_HANDLER_ARGS) 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; @@ -369,7 +371,9 @@ sysctl_kern_prof(SYSCTL_HANDLER_ARGS) #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 @@ -419,7 +423,7 @@ profil(td, uap) struct thread *td; register struct profil_args *uap; { - register struct uprof *upp; + struct uprof *upp; int s; int error = 0; @@ -430,7 +434,9 @@ profil(td, uap) goto done2; } if (uap->scale == 0) { + PROC_LOCK(td->td_proc); stopprofclock(td->td_proc); + PROC_UNLOCK(td->td_proc); goto done2; } upp = &td->td_proc->p_stats->p_prof; @@ -472,19 +478,16 @@ done2: * inaccurate. */ void -addupc_intr(ke, pc, ticks) - register struct kse *ke; - register uintptr_t pc; - u_int ticks; +addupc_intr(struct thread *td, uintptr_t pc, u_int ticks) { - register struct uprof *prof; - register caddr_t addr; - register u_int i; - register int v; + struct uprof *prof; + caddr_t addr; + u_int i; + int v; if (ticks == 0) return; - prof = &ke->ke_proc->p_stats->p_prof; + prof = &td->td_proc->p_stats->p_prof; if (pc < prof->pr_off || (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) return; /* out of range; ignore */ @@ -492,9 +495,9 @@ addupc_intr(ke, pc, ticks) addr = prof->pr_base + i; if ((v = fuswintr(addr)) == -1 || suswintr(addr, v + ticks) == -1) { mtx_lock_spin(&sched_lock); - prof->pr_addr = pc; - prof->pr_ticks = ticks; - ke->ke_flags |= KEF_OWEUPC | KEF_ASTPENDING ; + td->td_praddr = pc; + td->td_prticks = ticks; + td->td_flags |= (TDF_OWEUPC | TDF_ASTPENDING); mtx_unlock_spin(&sched_lock); } } @@ -502,34 +505,56 @@ addupc_intr(ke, pc, ticks) /* * Much like before, but we can afford to take faults here. If the * update fails, we simply turn off profiling. + * XXXKSE, don't use kse unless we got sched lock. */ void -addupc_task(ke, pc, ticks) - register struct kse *ke; - register uintptr_t pc; - u_int ticks; +addupc_task(struct thread *td, uintptr_t pc, u_int ticks) { - struct proc *p = ke->ke_proc; + struct proc *p = td->td_proc; register struct uprof *prof; register caddr_t addr; register u_int i; u_short v; + int stop = 0; if (ticks == 0) return; + PROC_LOCK(p); + mtx_lock_spin(&sched_lock); + if (!(p->p_sflag & PS_PROFIL)) { + mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); + return; + } + p->p_profthreads++; + mtx_unlock_spin(&sched_lock); + PROC_UNLOCK(p); prof = &p->p_stats->p_prof; if (pc < prof->pr_off || - (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) - return; + (i = PC_TO_INDEX(pc, prof)) >= prof->pr_size) { + goto out; + } addr = prof->pr_base + i; if (copyin(addr, &v, sizeof(v)) == 0) { v += ticks; if (copyout(&v, addr, sizeof(v)) == 0) - return; + goto out; + } + stop = 1; + +out: + PROC_LOCK(p); + if (--p->p_profthreads == 0) { + if (p->p_sflag & PS_STOPPROF) { + wakeup(&p->p_profthreads); + stop = 0; + } } - stopprofclock(p); + if (stop) + stopprofclock(p); + PROC_UNLOCK(p); } #if defined(__i386__) && __GNUC__ >= 2 diff --git a/sys/kern/subr_trap.c b/sys/kern/subr_trap.c index eec2ae6..ecd00c1 100644 --- a/sys/kern/subr_trap.c +++ b/sys/kern/subr_trap.c @@ -73,15 +73,21 @@ userret(td, frame, oticks) u_int oticks; { struct proc *p = td->td_proc; - struct kse *ke = td->td_kse; +#ifdef INVARIANTS + struct kse *ke; +#endif CTR3(KTR_SYSC, "userret: thread %p (pid %d, %s)", td, p->p_pid, p->p_comm); #ifdef INVARIANTS - /* Check that we called signotify() enough. */ + /* + * Check that we called signotify() enough. + * XXXKSE this checking is bogus for threaded program, + */ mtx_lock(&Giant); PROC_LOCK(p); mtx_lock_spin(&sched_lock); + ke = td->td_kse; if (SIGPENDING(p) && ((p->p_sflag & PS_NEEDSIGCHK) == 0 || (td->td_kse->ke_flags & KEF_ASTPENDING) == 0)) printf("failed to set signal flags properly for ast()\n"); @@ -96,6 +102,18 @@ userret(td, frame, oticks) sched_userret(td); /* + * Charge system time if profiling. + * + * XXX should move PS_PROFIL to a place that can obviously be + * accessed safely without sched_lock. + */ + + if (p->p_sflag & PS_PROFIL) { + ticks = td->td_sticks - oticks; + addupc_task(td, TRAPF_PC(frame), (u_int)ticks * psratio); + } + + /* * We need to check to see if we have to exit or wait due to a * single threading requirement or some other STOP condition. * Don't bother doing all the work if the stop bits are not set @@ -113,21 +131,6 @@ userret(td, frame, oticks) if (p->p_flag & P_KSES) { thread_userret(td, frame); } - - /* - * Charge system time if profiling. - * - * XXX should move PS_PROFIL to a place that can obviously be - * accessed safely without sched_lock. - */ - if (p->p_sflag & PS_PROFIL) { - quad_t ticks; - - mtx_lock_spin(&sched_lock); - ticks = ke->ke_sticks - oticks; - mtx_unlock_spin(&sched_lock); - addupc_task(ke, TRAPF_PC(frame), (u_int)ticks * psratio); - } } /* @@ -146,6 +149,7 @@ ast(struct trapframe *framep) u_int prticks, sticks; int sflag; int flags; + int tflags; int sig; #if defined(DEV_NPX) && !defined(SMP) int ucode; @@ -175,19 +179,21 @@ ast(struct trapframe *framep) */ mtx_lock_spin(&sched_lock); ke = td->td_kse; - sticks = ke->ke_sticks; + sticks = td->td_sticks; + tflags = td->td_flags; flags = ke->ke_flags; sflag = p->p_sflag; p->p_sflag &= ~(PS_ALRMPEND | PS_NEEDSIGCHK | PS_PROFPEND | PS_XCPU); #ifdef MAC p->p_sflag &= ~PS_MACPEND; #endif - ke->ke_flags &= ~(KEF_ASTPENDING | KEF_NEEDRESCHED | KEF_OWEUPC); + ke->ke_flags &= ~(KEF_ASTPENDING | KEF_NEEDRESCHED); + td->td_flags &= ~(TDF_ASTPENDING | TDF_OWEUPC); cnt.v_soft++; prticks = 0; - if (flags & KEF_OWEUPC && sflag & PS_PROFIL) { - prticks = p->p_stats->p_prof.pr_ticks; - p->p_stats->p_prof.pr_ticks = 0; + if (tflags & TDF_OWEUPC && sflag & PS_PROFIL) { + prticks = td->td_prticks; + td->td_prticks = 0; } mtx_unlock_spin(&sched_lock); /* @@ -200,8 +206,9 @@ ast(struct trapframe *framep) if (td->td_ucred != p->p_ucred) cred_update_thread(td); - if (flags & KEF_OWEUPC && sflag & PS_PROFIL) - addupc_task(ke, p->p_stats->p_prof.pr_addr, prticks); + if (tflags & TDF_OWEUPC && sflag & PS_PROFIL) { + addupc_task(td, td->td_praddr, prticks); + } if (sflag & PS_ALRMPEND) { PROC_LOCK(p); psignal(p, SIGVTALRM); diff --git a/sys/kern/subr_witness.c b/sys/kern/subr_witness.c index 3ab4a33..b958ceb 100644 --- a/sys/kern/subr_witness.c +++ b/sys/kern/subr_witness.c @@ -244,7 +244,7 @@ static struct witness_order_list_entry order_lists[] = { #endif { "clk", &lock_class_mtx_spin }, { "mutex profiling lock", &lock_class_mtx_spin }, - { "zombie_thread_lock", &lock_class_mtx_spin }, + { "kse zombie lock", &lock_class_mtx_spin }, { "ALD Queue", &lock_class_mtx_spin }, #ifdef __ia64__ { "MCA spin lock", &lock_class_mtx_spin }, |
