d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

1. get rid of the usage of next_kn in kqueue_scan. The latter is hell to maintain and lead to countless bugs.

Browse Source 
2. slightly reduce lock contention in the knote_drop path.
This commit is contained in:
Oscar Zhao 2020-05-06 03:42:29 -04:00
parent 61067bc214
commit 96330609b7
1 changed files with 28 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

60
sys/kern/kern_event.c
View File

@ -3050,7 +3050,7 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
{
struct kqueue *kq;
struct kevent *kevp;
struct knote *kn, *marker, *rtmarker, *nextkn;
struct knote *kn, *marker, *rtmarker;
struct knlist *knl;
sbintime_t asbt, rsbt, fsbt;
int count, error, haskqglobal, influx, nkev, touch, fevent;
@ -3077,7 +3077,6 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
/* activate kq if not already activated */
kevq_activate(kevq, td);
}
KEVQ_LOCK(kevq);
/* release processing knotes first */
kevq_rel_proc_kn(kevq);
@ -3280,7 +3279,6 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
influx = 0;
kn = NULL;
nextkn = NULL;
while (count < maxevents) {
KEVQ_OWNED(kevq);
@ -3289,20 +3287,16 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
if (curr < rtlimit) {
kntq = &kevq->kn_rt_head;
kncnt = &kevq->kn_rt_count;
kn = TAILQ_FIRST(kntq);
} else {
/* we've reached the limit, dequeue the realtime marker knote */
nextkn = rtmarker;
// otherwise just dequeue the rtmarker
kn = rtmarker;
}
} else {
kntq = &kevq->kn_head;
kncnt = &kevq->kn_count;
kn = TAILQ_FIRST(kntq);
}
if (nextkn == NULL) {
nextkn = TAILQ_FIRST(kntq);
}
kn = nextkn;
KASSERT(kn != NULL, ("kqueue_scan dequeued NULL"));
@ -3326,14 +3320,6 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
continue;
}
nextkn = TAILQ_NEXT(kn, kn_tqe);
if ((kn->kn_status & KN_PROCESSING) == KN_PROCESSING) {
// ignore knotes being processed
KN_FLUX_UNLOCK(kn);
continue;
}
// now this kn is going to be always dequeued from the kevq
TAILQ_REMOVE(kntq, kn, kn_tqe);
@ -3348,7 +3334,6 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
if (kn == rtmarker) {
rdrained = 1;
nextkn = NULL;
continue;
}
@ -3359,6 +3344,14 @@ kqueue_scan(struct kevq *kevq, int maxevents, struct kevent_copyops *k_ops,
goto done;
}
if ((kn->kn_status & KN_PROCESSING) == KN_PROCESSING) {
// reinsert at the end of queue
TAILQ_INSERT_TAIL(kntq, kn, kn_tqe);
KN_FLUX_UNLOCK(kn);
continue;
}
// now process the knote
kn->kn_status &= ~(KN_QUEUED | KN_WS);
(*kncnt)--;
@ -4078,8 +4071,8 @@ kevq_wakeup(struct kevq* kevq)
KEVQ_OWNED(kevq);
if ((kevq->kevq_state & KEVQ_SLEEP) == KEVQ_SLEEP) {
kevq->kevq_state &= ~KEVQ_SLEEP;
wakeup(kevq);
}
wakeup(kevq);
}
static void
@ -4591,19 +4584,10 @@ knote_drop_detached(struct knote *kn, struct thread *td)
("knote %p still attached", kn));
KQ_NOTOWNED(kq);
KQ_LOCK(kq);
KASSERT(kn->kn_influx == 1,
("knote_drop called on %p with influx %d", kn, kn->kn_influx));
if (kn->kn_fop->f_isfd)
list = &kq->kq_knlist[kn->kn_id];
else
list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
if (!SLIST_EMPTY(list))
SLIST_REMOVE(list, kn, knote, kn_link);
// drop from kevqs
if (kn->kn_status & KN_QUEUED) {
kevq = kn->kn_kevq;
KEVQ_LOCK(kevq);
@ -4617,9 +4601,21 @@ knote_drop_detached(struct knote *kn, struct thread *td)
knote_proc_dequeue(kn);
KEVQ_UNLOCK(kevq);
}
// drop from kq
KQ_LOCK(kq);
if (kn->kn_fop->f_isfd)
list = &kq->kq_knlist[kn->kn_id];
else
list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
if (!SLIST_EMPTY(list))
SLIST_REMOVE(list, kn, knote, kn_link);
knote_leave_flux_ul(kn);
KQ_UNLOCK(kq);
knote_leave_flux_ul(kn);
if (kn->kn_fop->f_isfd) {
fdrop(kn->kn_fp, td);