Aren
/
NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Retry enabling spin mutexes. We should be able to avoid poking the 

scheduler by just making wakeup from cv_wait() honor the same locking
order as when a spin mutex is acquired though mutex_enter().
*fingers crossed*
This commit is contained in:
pooka 2013-05-02 20:33:54 +00:00
parent 939c69c331
commit 9a6702d07c
2 changed files with 48 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
lib/librumpuser/rumpuser_pth.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: rumpuser_pth.c,v 1.23 2013/05/02 19:14:59 pooka Exp $ */
/* $NetBSD: rumpuser_pth.c,v 1.24 2013/05/02 20:33:54 pooka Exp $ */
/*
* Copyright (c) 2007-2010 Antti Kantee. All Rights Reserved.
@ -28,7 +28,7 @@
#include "rumpuser_port.h"
#if !defined(lint)
__RCSID("$NetBSD: rumpuser_pth.c,v 1.23 2013/05/02 19:14:59 pooka Exp $");
__RCSID("$NetBSD: rumpuser_pth.c,v 1.24 2013/05/02 20:33:54 pooka Exp $");
#endif /* !lint */
#include <sys/queue.h>
@ -378,17 +378,53 @@ rumpuser_cv_destroy(struct rumpuser_cv *cv)
free(cv);
}
static void
cv_unschedule(struct rumpuser_mtx *mtx, int *nlocks)
{
rumpkern_unsched(nlocks, mtx);
mtxexit(mtx);
}
static void
cv_reschedule(struct rumpuser_mtx *mtx, int nlocks)
{
/*
* If the cv interlock is a spin mutex, we must first release
* the mutex that was reacquired by pthread_cond_wait(),
* acquire the CPU context and only then relock the mutex.
* This is to preserve resource allocation order so that
* we don't deadlock. Non-spinning mutexes don't have this
* problem since they don't use a hold-and-wait approach
* to acquiring the mutex wrt the rump kernel CPU context.
*
* The more optimal solution would be to rework rumpkern_sched()
* so that it's possible to tell the scheduler
* "if you need to block, drop this lock first", but I'm not
* going poking there without some numbers on how often this
* path is taken for spin mutexes.
*/
if ((mtx->flags & (RUMPUSER_MTX_SPIN | RUMPUSER_MTX_KMUTEX)) ==
(RUMPUSER_MTX_SPIN | RUMPUSER_MTX_KMUTEX)) {
NOFAIL_ERRNO(pthread_mutex_unlock(&mtx->pthmtx));
rumpkern_sched(nlocks, mtx);
rumpuser_mutex_enter_nowrap(mtx);
} else {
mtxenter(mtx);
rumpkern_sched(nlocks, mtx);
}
}
void
rumpuser_cv_wait(struct rumpuser_cv *cv, struct rumpuser_mtx *mtx)
{
int nlocks;
cv->nwaiters++;
rumpkern_unsched(&nlocks, mtx);
mtxexit(mtx);
cv_unschedule(mtx, &nlocks);
NOFAIL_ERRNO(pthread_cond_wait(&cv->pthcv, &mtx->pthmtx));
mtxenter(mtx);
rumpkern_sched(nlocks, mtx);
cv_reschedule(mtx, nlocks);
cv->nwaiters--;
}
@ -420,8 +456,7 @@ rumpuser_cv_timedwait(struct rumpuser_cv *cv, struct rumpuser_mtx *mtx,
clock_gettime(CLOCK_REALTIME, &ts);
cv->nwaiters++;
rumpkern_unsched(&nlocks, mtx);
mtxexit(mtx);
cv_unschedule(mtx, &nlocks);
ts.tv_sec += sec;
ts.tv_nsec += nsec;
@ -430,8 +465,8 @@ rumpuser_cv_timedwait(struct rumpuser_cv *cv, struct rumpuser_mtx *mtx,
ts.tv_nsec -= 1000*1000*1000;
}
rv = pthread_cond_timedwait(&cv->pthcv, &mtx->pthmtx, &ts);
mtxenter(mtx);
rumpkern_sched(nlocks, mtx);
cv_reschedule(mtx, nlocks);
cv->nwaiters--;
ET(rv);

8
sys/rump/librump/rumpkern/locks.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: locks.c,v 1.60 2013/04/30 00:03:53 pooka Exp $ */
/* $NetBSD: locks.c,v 1.61 2013/05/02 20:33:54 pooka Exp $ */
/*
* Copyright (c) 2007-2011 Antti Kantee. All Rights Reserved.
@ -26,7 +26,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: locks.c,v 1.60 2013/04/30 00:03:53 pooka Exp $");
__KERNEL_RCSID(0, "$NetBSD: locks.c,v 1.61 2013/05/02 20:33:54 pooka Exp $");
#include <sys/param.h>
#include <sys/kmem.h>
@ -113,11 +113,9 @@ mutex_init(kmutex_t *mtx, kmutex_type_t type, int ipl)
isspin = 1;
}
#if 0
/* spin mutex support needs some cpu scheduler rework */
if (isspin)
ruflags |= RUMPUSER_MTX_SPIN;
#endif
rumpuser_mutex_init((struct rumpuser_mtx **)mtx, ruflags);
ALLOCK(mtx, &mutex_lockops);
}
@ -144,7 +142,7 @@ mutex_spin_enter(kmutex_t *mtx)
{
WANTLOCK(mtx, false, false);
rumpuser_mutex_enter(RUMPMTX(mtx));
rumpuser_mutex_enter_nowrap(RUMPMTX(mtx));
LOCKED(mtx, false);
}