- remove blocked/unblocked upcall ordering.

- always wait for unblocked upcall if we have to continue a blocked
  thread.

=> removes wakeup from sys_sa_stacks when a stack is returned.
=> avoids extra sa_unblockyield syscall when unblocked upcall is
   delivered before blocked upcall is processed.
=> avoids double pagefault if we continued a thread before the
   pagefault was resolved.
=> avoids losing unblocked state if we continued a thread after
   skipping the unblocked upcall.
This commit is contained in:
cl 2003-11-17 22:57:52 +00:00
parent 2e11d201c2
commit ac4fd64ba5

View File

@ -1,4 +1,4 @@
/* $NetBSD: kern_sa.c,v 1.43 2003/11/17 22:52:09 cl Exp $ */
/* $NetBSD: kern_sa.c,v 1.44 2003/11/17 22:57:52 cl Exp $ */
/*-
* Copyright (c) 2001 The NetBSD Foundation, Inc.
@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: kern_sa.c,v 1.43 2003/11/17 22:52:09 cl Exp $");
__KERNEL_RCSID(0, "$NetBSD: kern_sa.c,v 1.44 2003/11/17 22:57:52 cl Exp $");
#include <sys/param.h>
#include <sys/systm.h>
@ -250,7 +250,6 @@ sys_sa_stacks(struct lwp *l, void *v, register_t *retval)
if ((l2 = sast->sast_blocker)) {
l2->l_upcallstack = NULL;
sast->sast_blocker = NULL;
wakeup(&l2->l_upcallstack);
}
if (SLIST_NEXT(sast, sast_list) != (void *)-1) {
count = i;
@ -440,20 +439,22 @@ sa_preempt(struct lwp *l)
/*
* Help userspace library resolve locks and critical sections
* Help userspace library resolve locks and critical sections:
* - return if the unblocked upcall has already been delivered.
* This case is usually already detected in userspace.
* - recycles the calling LWP and its stack if it was not preempted
* and idle the VP until the sa_id LWP unblocks
* - recycles the to be unblocked LWP if the calling LWP was preempted
* and returns control to the userspace library so it can switch to
* the blocked thread
* This is used if a thread blocks because of a pagefault and is in a
* critical section in the userspace library and the critical section
* resolving code cannot continue until the blocked thread is unblocked.
* If the userspace library switches to the blocked thread in the second
* case, it will either continue (because the pagefault has been handled)
* or it will pagefault again. The second pagefault will be detected by
* the double pagefault code and the VP will idle until the pagefault
* has been handled.
* and the unblocked upcall was not yet delivered. Put the sa_id
* LWP on the VP and wait until it unblocks or switch to it if it's
* ready. There will be no unblocked upcall.
* - recycles the blocked LWP if up_preempted == NULL. This is used
* if the blocked LWP is an idle thread and we don't care for the
* unblocked upcall.
* - otherwise, wait for the blocked LWP to get ready. The unblocked
* upcall is delivered when we return.
* This is used if a thread blocks (mostly because of a pagefault) and
* is in a critical section in the userspace library and the critical
* section resolving code cannot continue until the blocked thread is
* unblocked.
*/
int
sys_sa_unblockyield(struct lwp *l, void *v, register_t *retval)
@ -465,7 +466,7 @@ sys_sa_unblockyield(struct lwp *l, void *v, register_t *retval)
} */ *uap = v;
struct sadata *sa = l->l_proc->p_sa;
struct proc *p = l->l_proc;
struct lwp *l2;
struct lwp *l2, **hp;
struct sastack sast;
int error, f, s;
void *preempted;
@ -498,79 +499,100 @@ sys_sa_unblockyield(struct lwp *l, void *v, register_t *retval)
break;
}
}
if (l2 && (l2->l_upcallstack == NULL ||
(l2->l_upcallstack->sast_blocker != l2 &&
l2->l_upcallstack->sast_blocker != NULL)))
l2 = NULL;
if (l2 == NULL) {
SCHED_UNLOCK(s);
return (ESRCH);
}
if (l2->l_upcallstack->sast_blocker != l2 ||
sast.sast_stack.ss_sp != l2->l_upcallstack->sast_stack.ss_sp) {
SCHED_UNLOCK(s);
return (EINVAL);
}
/*
* upcall not interrupted: (*up_preempted == NULL)
* - lwp ready: (wchan == upcallstacks)
* ==> recycle stack, put lwp on vp,
* unsleep lwp, make runnable, recycle upcall lwp (=l)
* - lwp not ready:
* ==> recycle stack, put lwp on vp, recycle upcall lwp (=l)
*
* upcall interrupted: (*up_preempted != NULL || up_preempted == NULL)
* ==> recycle upcall lwp
*/
if (preempted != NULL) {
/* just return, prevented in userland most of the time */
DPRINTFN(11,("sys_sa_unblockyield(%d.%d) unblocked upcall done\n",
p->p_pid, l->l_lid));
KDASSERT(preempted != NULL);
} else if (preempted == (void *)-1) {
/* recycle blocked LWP */
DPRINTFN(11,("sys_sa_unblockyield(%d.%d) recycle %d "
"(was %sready) upcall stack %p\n",
p->p_pid, l->l_lid, l2->l_lid,
(l2->l_wchan == &l2->l_upcallstack) ? "" :
(l2->l_upcallstack->sast_blocker == NULL) ? "" :
"not ", l2->l_upcallstack->sast_stack.ss_sp));
l2->l_upcallstack->sast_blocker = NULL;
if (l2->l_wchan == &l2->l_upcallstack) {
unsleep(l2);
if (l2->l_stat == LSSLEEP) {
l2->l_slptime = 0;
l2->l_stat = LSRUN;
l2->l_proc->p_nrlwps++;
if (l2->l_flag & L_INMEM)
setrunqueue(l2);
else
sched_wakeup((caddr_t)&proc0);
}
if (l2->l_upcallstack->sast_blocker == NULL) {
/*
* l2 is on the wokenq, remove it and put l2
* in the cache
*/
hp = &sa->sa_wokenq_head;
while (*hp != l2)
hp = &(*hp)->l_forw;
*hp = l2->l_forw;
if (sa->sa_wokenq_tailp == &l2->l_forw)
sa->sa_wokenq_tailp = hp;
l2->l_flag &= ~L_SA_BLOCKING;
l2->l_upcallstack = NULL;
sa_putcachelwp(p, l2); /* PHOLD from sa_setwoken */
} else
/* let sa_setwoken put it in the cache */
l2->l_upcallstack->sast_blocker = NULL;
} else if (preempted != NULL) {
/* wait for the blocked LWP to get ready, then return */
DPRINTFN(11,("sys_sa_unblockyield(%d.%d) waiting for %d "
"(was %sready) upcall stack %p\n",
p->p_pid, l->l_lid, l2->l_lid,
(l2->l_upcallstack->sast_blocker == NULL) ? "" :
"not ", l2->l_upcallstack->sast_stack.ss_sp));
if (l2->l_upcallstack->sast_blocker != NULL) {
l2->l_upcallstack->sast_blocker = l;
SCHED_UNLOCK(s); /* XXXcl we're still holding
* the kernel lock, is that
* good enough? */
SA_LWP_STATE_LOCK(l, f);
tsleep((caddr_t) &l2->l_upcallstack, PWAIT,
"saunblock", 0);
SA_LWP_STATE_UNLOCK(l, f);
if (p->p_flag & P_WEXIT)
lwp_exit(l);
return(0);
}
} else {
/* recycle calling LWP and resume blocked LWP */
DPRINTFN(11,("sys_sa_unblockyield(%d.%d) resuming %d "
"(is %sready) upcall stack %p\n",
p->p_pid, l->l_lid, l2->l_lid,
(l2->l_wchan == &l2->l_upcallstack) ? "" :
(l2->l_upcallstack->sast_blocker == NULL) ? "" :
"not ", l2->l_upcallstack->sast_stack.ss_sp));
sa->sa_vp = l2;
l2->l_flag &= ~L_SA_BLOCKING;
l2->l_upcallstack->sast_blocker = NULL;
SLIST_INSERT_HEAD(&sa->sa_stackslist, l2->l_upcallstack,
sast_list);
l2->l_upcallstack = NULL;
if (l2->l_wchan == &l2->l_upcallstack) {
unsleep(l2);
if (l2->l_stat == LSSLEEP) {
l2->l_slptime = 0;
l2->l_stat = LSRUN;
l2->l_proc->p_nrlwps++;
if (l2->l_flag & L_INMEM)
setrunqueue(l2);
else
sched_wakeup((caddr_t)&proc0);
}
if (l2->l_upcallstack->sast_blocker == NULL) {
/*
* l2 is on the wokenq, remove it and
* continue l2
*/
hp = &sa->sa_wokenq_head;
while (*hp != l2)
hp = &(*hp)->l_forw;
*hp = l2->l_forw;
if (sa->sa_wokenq_tailp == &l2->l_forw)
sa->sa_wokenq_tailp = hp;
l2->l_upcallstack = NULL;
setrunnable(l2);
PRELE(l2); /* PHOLD from sa_setwoken */
} else {
/* cleanup l_upcallstack */
l2->l_upcallstack->sast_blocker = NULL;
l2->l_upcallstack = NULL;
l2 = NULL; /* don't continue l2 yet */
}
p->p_nrlwps--;
PHOLD(l);
sa_putcachelwp(p, l);
mi_switch(l, NULL);
mi_switch(l, l2);
/* mostly NOTREACHED */
SCHED_ASSERT_UNLOCKED();
splx(s);
@ -856,6 +878,8 @@ sa_switch(struct lwp *l, int type)
return;
}
DPRINTFN(8,("sa_switch(%d.%d) blocked upcall %d, stack %p\n",
p->p_pid, l->l_lid, l2->l_lid, sast->sast_stack.ss_sp));
SIMPLEQ_INSERT_TAIL(&sa->sa_upcalls, sau, sau_next);
l2->l_flag |= L_SA_UPCALL;
@ -1036,50 +1060,24 @@ sa_unblock_userret(struct lwp *l)
DPRINTFN(7,("sa_unblock_userret(%d.%d %x) \n", p->p_pid, l->l_lid,
l->l_flag));
while (l->l_upcallstack != NULL) {
if (l->l_upcallstack->sast_blocker == NULL) {
SCHED_LOCK(s);
l->l_flag &= ~(L_SA_UPCALL|L_SA_BLOCKING);
l->l_upcallstack = NULL;
p->p_nrlwps--;
PHOLD(l);
sa_putcachelwp(p, l);
SA_LWP_STATE_UNLOCK(l, f);
mi_switch(l, NULL);
/* mostly NOTREACHED */
SCHED_ASSERT_UNLOCKED();
splx(s);
KDASSERT(p->p_flag & P_WEXIT);
lwp_exit(l);
}
if ((l->l_flag & L_SA_BLOCKING) == 0) {
l->l_upcallstack->sast_blocker = NULL;
l->l_upcallstack = NULL;
break;
}
tsleep((caddr_t) &l->l_upcallstack, PWAIT,
"saunblock", 0);
if (p->p_flag & P_WEXIT)
lwp_exit(l);
}
sa_setwoken(l);
/* maybe NOTREACHED */
if (l->l_flag & L_SA_BLOCKING) {
SCHED_LOCK(s);
if (l != sa->sa_vp) {
/* Invoke an "unblocked" upcall */
sa_setwoken(l);
/* maybe NOTREACHED */
DPRINTFN(8,("sa_unblock_userret(%d.%d) unblocking\n",
p->p_pid, l->l_lid));
SCHED_ASSERT_UNLOCKED();
l2 = sa_vp_repossess(l);
SCHED_UNLOCK(s);
if (l2 == NULL)
lwp_exit(l);
SCHED_ASSERT_UNLOCKED();
PHOLD(l2);
sau = sadata_upcall_alloc(1);
sau->sau_arg = NULL;
@ -1113,13 +1111,14 @@ sa_unblock_userret(struct lwp *l)
/* NOTREACHED */
}
SCHED_LOCK(s);
SIMPLEQ_INSERT_TAIL(&sa->sa_upcalls, sau, sau_next);
l->l_flag |= L_SA_UPCALL;
l->l_flag &= ~L_SA_BLOCKING;
SCHED_LOCK(s);
sa_putcachelwp(p, l2); /* PHOLD from sa_vp_repossess */
SCHED_UNLOCK(s);
l->l_upcallstack = NULL;
sa_putcachelwp(p, l2);
}
SCHED_UNLOCK(s);
SA_LWP_STATE_UNLOCK(l, f);
KERNEL_PROC_UNLOCK(l);
@ -1170,7 +1169,6 @@ sa_upcall_userret(struct lwp *l)
SCHED_UNLOCK(s);
PHOLD(l2);
if (p->p_flag & P_WEXIT)
lwp_exit(l);
@ -1201,8 +1199,9 @@ sa_upcall_userret(struct lwp *l)
SIMPLEQ_INSERT_TAIL(&sa->sa_upcalls, sau, sau_next);
l2->l_flag &= ~L_SA_BLOCKING;
l2->l_upcallstack = NULL;
SCHED_LOCK(s);
sa_putcachelwp(p, l2);
sa_putcachelwp(p, l2); /* PHOLD from sa_setwoken */
}
SCHED_UNLOCK(s);
@ -1314,11 +1313,11 @@ sa_upcall_userret(struct lwp *l)
eventq = l2->l_forw;
DPRINTFN(8,("sa_upcall_userret(%d.%d) unblocking extra %d\n",
p->p_pid, l->l_lid, l2->l_lid));
PHOLD(l2);
sa_upcall_getstate(&e_ss, l2);
l2->l_flag &= ~L_SA_BLOCKING;
SCHED_LOCK(s);
sa_putcachelwp(p, l2);
l2->l_flag &= ~L_SA_BLOCKING;
l2->l_upcallstack = NULL;
sa_putcachelwp(p, l2); /* PHOLD from sa_setwoken */
SCHED_UNLOCK(s);
if (copyout(&e_ss.ss_captured.ss_ctx,
e_ss.ss_captured.ss_sa.sa_context,
@ -1399,6 +1398,11 @@ sa_setwoken(struct lwp *l)
SCHED_LOCK(s);
if ((l->l_flag & L_SA_BLOCKING) == 0) {
SCHED_UNLOCK(s);
return;
}
p = l->l_proc;
sa = p->p_sa;
vp_lwp = sa->sa_vp;
@ -1409,6 +1413,23 @@ sa_setwoken(struct lwp *l)
l->l_proc->p_pid, l->l_lid, l->l_flag,
vp_lwp->l_lid));
if (l->l_upcallstack && l->l_upcallstack->sast_blocker == NULL) {
DPRINTFN(11,("sa_setwoken(%d.%d) recycle\n",
l->l_proc->p_pid, l->l_lid));
l->l_flag &= ~(L_SA_UPCALL|L_SA_BLOCKING);
l->l_flag |= L_SA;
l->l_upcallstack = NULL;
p->p_nrlwps--;
PHOLD(l);
sa_putcachelwp(p, l);
mi_switch(l, NULL);
/* mostly NOTREACHED */
SCHED_ASSERT_UNLOCKED();
splx(s);
KDASSERT(p->p_flag & P_WEXIT);
lwp_exit(l);
}
#if notyet
if (vp_lwp->l_flag & L_SA_IDLE) {
KDASSERT((vp_lwp->l_flag & L_SA_UPCALL) == 0);
@ -1426,6 +1447,7 @@ sa_setwoken(struct lwp *l)
l->l_proc->p_pid, l->l_lid, vp_lwp->l_lid,
vp_lwp->l_stat));
PHOLD(l);
if (sa->sa_wokenq_head == NULL)
sa->sa_wokenq_head = l;
else
@ -1482,6 +1504,12 @@ sa_setwoken(struct lwp *l)
panic("sa_vp LWP not sleeping/onproc/runnable");
}
if (l->l_upcallstack) {
if (l->l_upcallstack->sast_blocker != l)
sched_wakeup(&l->l_upcallstack);
l->l_upcallstack->sast_blocker = NULL;
}
l->l_stat = LSSUSPENDED;
p->p_nrlwps--;
mi_switch(l, l2);
@ -1498,7 +1526,8 @@ sa_vp_repossess(struct lwp *l)
struct lwp *l2;
struct proc *p = l->l_proc;
struct sadata *sa = p->p_sa;
int s;
SCHED_ASSERT_LOCKED();
/*
* Put ourselves on the virtual processor and note that the
@ -1514,8 +1543,6 @@ sa_vp_repossess(struct lwp *l)
KDASSERT(l2 != l);
if (l2) {
PHOLD(l2);
SCHED_LOCK(s);
switch (l2->l_stat) {
case LSRUN:
remrunqueue(l2);
@ -1539,7 +1566,6 @@ sa_vp_repossess(struct lwp *l)
#endif
}
l2->l_stat = LSSUSPENDED;
SCHED_UNLOCK(s);
}
return l2;
}