Home | History | Annotate | Line # | Download | only in rumpkern
scheduler.c revision 1.25
      1 /*      $NetBSD: scheduler.c,v 1.25 2011/01/28 16:58:28 pooka Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2010 Antti Kantee.  All Rights Reserved.
      5  *
      6  * Redistribution and use in source and binary forms, with or without
      7  * modification, are permitted provided that the following conditions
      8  * are met:
      9  * 1. Redistributions of source code must retain the above copyright
     10  *    notice, this list of conditions and the following disclaimer.
     11  * 2. Redistributions in binary form must reproduce the above copyright
     12  *    notice, this list of conditions and the following disclaimer in the
     13  *    documentation and/or other materials provided with the distribution.
     14  *
     15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     16  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     17  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     18  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     21  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     25  * SUCH DAMAGE.
     26  */
     27 
     28 #include <sys/cdefs.h>
     29 __KERNEL_RCSID(0, "$NetBSD: scheduler.c,v 1.25 2011/01/28 16:58:28 pooka Exp $");
     30 
     31 #include <sys/param.h>
     32 #include <sys/atomic.h>
     33 #include <sys/cpu.h>
     34 #include <sys/kmem.h>
     35 #include <sys/mutex.h>
     36 #include <sys/namei.h>
     37 #include <sys/queue.h>
     38 #include <sys/select.h>
     39 #include <sys/systm.h>
     40 
     41 #include <rump/rumpuser.h>
     42 
     43 #include "rump_private.h"
     44 
     45 static struct cpu_info rump_cpus[MAXCPUS];
     46 static struct rumpcpu {
     47 	/* needed in fastpath */
     48 	struct cpu_info *rcpu_ci;
     49 	void *rcpu_prevlwp;
     50 
     51 	/* needed in slowpath */
     52 	struct rumpuser_mtx *rcpu_mtx;
     53 	struct rumpuser_cv *rcpu_cv;
     54 	int rcpu_wanted;
     55 
     56 	/* offset 20 (P=4) or 36 (P=8) here */
     57 
     58 	/*
     59 	 * Some stats.  Not really that necessary, but we should
     60 	 * have room.  Note that these overflow quite fast, so need
     61 	 * to be collected often.
     62 	 */
     63 	unsigned int rcpu_fastpath;
     64 	unsigned int rcpu_slowpath;
     65 	unsigned int rcpu_migrated;
     66 
     67 	/* offset 32 (P=4) or 50 (P=8) */
     68 
     69 	int rcpu_align[0] __aligned(CACHE_LINE_SIZE);
     70 } rcpu_storage[MAXCPUS];
     71 struct cpu_info *rump_cpu = &rump_cpus[0];
     72 int ncpu;
     73 
     74 #define RCPULWP_BUSY	((void *)-1)
     75 #define RCPULWP_WANTED	((void *)-2)
     76 
     77 static struct rumpuser_mtx *lwp0mtx;
     78 static struct rumpuser_cv *lwp0cv;
     79 static unsigned nextcpu;
     80 
     81 kmutex_t unruntime_lock; /* unruntime lwp lock.  practically unused */
     82 
     83 static bool lwp0isbusy = false;
     84 
     85 /*
     86  * Keep some stats.
     87  *
     88  * Keeping track of there is not really critical for speed, unless
     89  * stats happen to be on a different cache line (CACHE_LINE_SIZE is
     90  * really just a coarse estimate), so default for the performant case
     91  * (i.e. no stats).
     92  */
     93 #ifdef RUMPSCHED_STATS
     94 #define SCHED_FASTPATH(rcpu) rcpu->rcpu_fastpath++;
     95 #define SCHED_SLOWPATH(rcpu) rcpu->rcpu_slowpath++;
     96 #define SCHED_MIGRATED(rcpu) rcpu->rcpu_migrated++;
     97 #else
     98 #define SCHED_FASTPATH(rcpu)
     99 #define SCHED_SLOWPATH(rcpu)
    100 #define SCHED_MIGRATED(rcpu)
    101 #endif
    102 
    103 struct cpu_info *
    104 cpu_lookup(u_int index)
    105 {
    106 
    107 	return &rump_cpus[index];
    108 }
    109 
    110 static inline struct rumpcpu *
    111 getnextcpu(void)
    112 {
    113 	unsigned newcpu;
    114 
    115 	newcpu = atomic_inc_uint_nv(&nextcpu);
    116 	if (__predict_false(ncpu > UINT_MAX/2))
    117 		atomic_and_uint(&nextcpu, 0);
    118 	newcpu = newcpu % ncpu;
    119 
    120 	return &rcpu_storage[newcpu];
    121 }
    122 
    123 /* this could/should be mi_attach_cpu? */
    124 void
    125 rump_cpus_bootstrap(int *nump)
    126 {
    127 	struct rumpcpu *rcpu;
    128 	struct cpu_info *ci;
    129 	int num = *nump;
    130 	int i;
    131 
    132 	if (num > MAXCPUS) {
    133 		aprint_verbose("CPU limit: %d wanted, %d (MAXCPUS) "
    134 		    "available (adjusted)\n", num, MAXCPUS);
    135 		num = MAXCPUS;
    136 	}
    137 
    138 	for (i = 0; i < num; i++) {
    139 		rcpu = &rcpu_storage[i];
    140 		ci = &rump_cpus[i];
    141 		ci->ci_index = i;
    142 	}
    143 
    144 	/* attach first cpu for bootstrap */
    145 	rump_cpu_attach(&rump_cpus[0]);
    146 	ncpu = 1;
    147 	*nump = num;
    148 }
    149 
    150 void
    151 rump_scheduler_init(int numcpu)
    152 {
    153 	struct rumpcpu *rcpu;
    154 	struct cpu_info *ci;
    155 	int i;
    156 
    157 	rumpuser_mutex_init(&lwp0mtx);
    158 	rumpuser_cv_init(&lwp0cv);
    159 	for (i = 0; i < numcpu; i++) {
    160 		rcpu = &rcpu_storage[i];
    161 		ci = &rump_cpus[i];
    162 		rcpu->rcpu_ci = ci;
    163 		ci->ci_schedstate.spc_mutex =
    164 		    mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
    165 		ci->ci_schedstate.spc_flags = SPCF_RUNNING;
    166 		rcpu->rcpu_wanted = 0;
    167 		rumpuser_cv_init(&rcpu->rcpu_cv);
    168 		rumpuser_mutex_init(&rcpu->rcpu_mtx);
    169 	}
    170 
    171 	mutex_init(&unruntime_lock, MUTEX_DEFAULT, IPL_NONE);
    172 }
    173 
    174 /*
    175  * condvar ops using scheduler lock as the rumpuser interlock.
    176  */
    177 void
    178 rump_schedlock_cv_wait(struct rumpuser_cv *cv)
    179 {
    180 	struct lwp *l = curlwp;
    181 	struct rumpcpu *rcpu = &rcpu_storage[l->l_cpu-&rump_cpus[0]];
    182 
    183 	/* mutex will be taken and released in cpu schedule/unschedule */
    184 	rumpuser_cv_wait(cv, rcpu->rcpu_mtx);
    185 }
    186 
    187 int
    188 rump_schedlock_cv_timedwait(struct rumpuser_cv *cv, const struct timespec *ts)
    189 {
    190 	struct lwp *l = curlwp;
    191 	struct rumpcpu *rcpu = &rcpu_storage[l->l_cpu-&rump_cpus[0]];
    192 
    193 	/* mutex will be taken and released in cpu schedule/unschedule */
    194 	return rumpuser_cv_timedwait(cv, rcpu->rcpu_mtx,
    195 	    ts->tv_sec, ts->tv_nsec);
    196 }
    197 
    198 static void
    199 lwp0busy(void)
    200 {
    201 
    202 	/* busy lwp0 */
    203 	KASSERT(curlwp == NULL || curlwp->l_stat != LSONPROC);
    204 	rumpuser_mutex_enter_nowrap(lwp0mtx);
    205 	while (lwp0isbusy)
    206 		rumpuser_cv_wait_nowrap(lwp0cv, lwp0mtx);
    207 	lwp0isbusy = true;
    208 	rumpuser_mutex_exit(lwp0mtx);
    209 }
    210 
    211 static void
    212 lwp0rele(void)
    213 {
    214 
    215 	rumpuser_mutex_enter_nowrap(lwp0mtx);
    216 	KASSERT(lwp0isbusy == true);
    217 	lwp0isbusy = false;
    218 	rumpuser_cv_signal(lwp0cv);
    219 	rumpuser_mutex_exit(lwp0mtx);
    220 }
    221 
    222 void
    223 rump_schedule()
    224 {
    225 	struct lwp *l;
    226 
    227 	/*
    228 	 * If there is no dedicated lwp, allocate a temp one and
    229 	 * set it to be free'd upon unschedule().  Use lwp0 context
    230 	 * for reserving the necessary resources.  Don't optimize
    231 	 * for this case -- anyone who cares about performance will
    232 	 * start a real thread.
    233 	 */
    234 	if (__predict_true((l = rumpuser_get_curlwp()) != NULL)) {
    235 		rump_schedule_cpu(l);
    236 		LWP_CACHE_CREDS(l, l->l_proc);
    237 	} else {
    238 		lwp0busy();
    239 
    240 		/* schedule cpu and use lwp0 */
    241 		rump_schedule_cpu(&lwp0);
    242 		rumpuser_set_curlwp(&lwp0);
    243 
    244 		/* allocate thread, switch to it, and release lwp0 */
    245 		l = rump__lwproc_alloclwp(initproc);
    246 		rump_lwproc_switch(l);
    247 		lwp0rele();
    248 
    249 		/*
    250 		 * mark new thread dead-on-unschedule.  this
    251 		 * means that we'll be running with l_refcnt == 0.
    252 		 * relax, it's fine.
    253 		 */
    254 		rump_lwproc_releaselwp();
    255 	}
    256 }
    257 
    258 void
    259 rump_schedule_cpu(struct lwp *l)
    260 {
    261 
    262 	rump_schedule_cpu_interlock(l, NULL);
    263 }
    264 
    265 /*
    266  * Schedule a CPU.  This optimizes for the case where we schedule
    267  * the same thread often, and we have nCPU >= nFrequently-Running-Thread
    268  * (where CPU is virtual rump cpu, not host CPU).
    269  */
    270 void
    271 rump_schedule_cpu_interlock(struct lwp *l, void *interlock)
    272 {
    273 	struct rumpcpu *rcpu;
    274 	void *old;
    275 	bool domigrate;
    276 	bool bound = l->l_pflag & LP_BOUND;
    277 
    278 	l->l_stat = LSRUN;
    279 
    280 	/*
    281 	 * First, try fastpath: if we were the previous user of the
    282 	 * CPU, everything is in order cachewise and we can just
    283 	 * proceed to use it.
    284 	 *
    285 	 * If we are a different thread (i.e. CAS fails), we must go
    286 	 * through a memory barrier to ensure we get a truthful
    287 	 * view of the world.
    288 	 */
    289 
    290 	KASSERT(l->l_target_cpu != NULL);
    291 	rcpu = &rcpu_storage[l->l_target_cpu-&rump_cpus[0]];
    292 	if (atomic_cas_ptr(&rcpu->rcpu_prevlwp, l, RCPULWP_BUSY) == l) {
    293 		if (__predict_true(interlock == rcpu->rcpu_mtx))
    294 			rumpuser_mutex_exit(rcpu->rcpu_mtx);
    295 		SCHED_FASTPATH(rcpu);
    296 		/* jones, you're the man */
    297 		goto fastlane;
    298 	}
    299 
    300 	/*
    301 	 * Else, it's the slowpath for us.  First, determine if we
    302 	 * can migrate.
    303 	 */
    304 	if (ncpu == 1)
    305 		domigrate = false;
    306 	else
    307 		domigrate = true;
    308 
    309 	/* Take lock.  This acts as a load barrier too. */
    310 	if (__predict_true(interlock != rcpu->rcpu_mtx))
    311 		rumpuser_mutex_enter_nowrap(rcpu->rcpu_mtx);
    312 
    313 	for (;;) {
    314 		SCHED_SLOWPATH(rcpu);
    315 		old = atomic_swap_ptr(&rcpu->rcpu_prevlwp, RCPULWP_WANTED);
    316 
    317 		/* CPU is free? */
    318 		if (old != RCPULWP_BUSY && old != RCPULWP_WANTED) {
    319 			if (atomic_cas_ptr(&rcpu->rcpu_prevlwp,
    320 			    RCPULWP_WANTED, RCPULWP_BUSY) == RCPULWP_WANTED) {
    321 				break;
    322 			}
    323 		}
    324 
    325 		/*
    326 		 * Do we want to migrate once?
    327 		 * This may need a slightly better algorithm, or we
    328 		 * might cache pingpong eternally for non-frequent
    329 		 * threads.
    330 		 */
    331 		if (domigrate && !bound) {
    332 			domigrate = false;
    333 			SCHED_MIGRATED(rcpu);
    334 			rumpuser_mutex_exit(rcpu->rcpu_mtx);
    335 			rcpu = getnextcpu();
    336 			rumpuser_mutex_enter_nowrap(rcpu->rcpu_mtx);
    337 			continue;
    338 		}
    339 
    340 		/* Want CPU, wait until it's released an retry */
    341 		rcpu->rcpu_wanted++;
    342 		rumpuser_cv_wait_nowrap(rcpu->rcpu_cv, rcpu->rcpu_mtx);
    343 		rcpu->rcpu_wanted--;
    344 	}
    345 	rumpuser_mutex_exit(rcpu->rcpu_mtx);
    346 
    347  fastlane:
    348 	l->l_cpu = l->l_target_cpu = rcpu->rcpu_ci;
    349 	l->l_mutex = rcpu->rcpu_ci->ci_schedstate.spc_mutex;
    350 	l->l_ncsw++;
    351 	l->l_stat = LSONPROC;
    352 
    353 	rcpu->rcpu_ci->ci_curlwp = l;
    354 }
    355 
    356 void
    357 rump_unschedule()
    358 {
    359 	struct lwp *l = rumpuser_get_curlwp();
    360 #ifdef DIAGNOSTIC
    361 	int nlock;
    362 
    363 	KERNEL_UNLOCK_ALL(l, &nlock);
    364 	KASSERT(nlock == 0);
    365 #endif
    366 
    367 	KASSERT(l->l_mutex == l->l_cpu->ci_schedstate.spc_mutex);
    368 	rump_unschedule_cpu(l);
    369 	l->l_mutex = &unruntime_lock;
    370 	l->l_stat = LSSTOP;
    371 
    372 	/*
    373 	 * Check special conditions:
    374 	 *  1) do we need to free the lwp which just unscheduled?
    375 	 *     (locking order: lwp0, cpu)
    376 	 *  2) do we want to clear curlwp for the current host thread
    377 	 */
    378 	if (__predict_false(l->l_flag & LW_WEXIT)) {
    379 		lwp0busy();
    380 
    381 		/* Now that we have lwp0, we can schedule a CPU again */
    382 		rump_schedule_cpu(l);
    383 
    384 		/* switch to lwp0.  this frees the old thread */
    385 		KASSERT(l->l_flag & LW_WEXIT);
    386 		rump_lwproc_switch(&lwp0);
    387 
    388 		/* release lwp0 */
    389 		rump_unschedule_cpu(&lwp0);
    390 		lwp0.l_mutex = &unruntime_lock;
    391 		lwp0.l_pflag &= ~LP_RUNNING;
    392 		lwp0rele();
    393 		rumpuser_set_curlwp(NULL);
    394 
    395 	} else if (__predict_false(l->l_flag & LW_RUMP_CLEAR)) {
    396 		rumpuser_set_curlwp(NULL);
    397 		l->l_flag &= ~LW_RUMP_CLEAR;
    398 	}
    399 }
    400 
    401 void
    402 rump_unschedule_cpu(struct lwp *l)
    403 {
    404 
    405 	rump_unschedule_cpu_interlock(l, NULL);
    406 }
    407 
    408 void
    409 rump_unschedule_cpu_interlock(struct lwp *l, void *interlock)
    410 {
    411 
    412 	if ((l->l_pflag & LP_INTR) == 0)
    413 		rump_softint_run(l->l_cpu);
    414 	rump_unschedule_cpu1(l, interlock);
    415 }
    416 
    417 void
    418 rump_unschedule_cpu1(struct lwp *l, void *interlock)
    419 {
    420 	struct rumpcpu *rcpu;
    421 	struct cpu_info *ci;
    422 	void *old;
    423 
    424 	ci = l->l_cpu;
    425 	ci->ci_curlwp = NULL;
    426 	rcpu = &rcpu_storage[ci-&rump_cpus[0]];
    427 
    428 	KASSERT(rcpu->rcpu_ci == ci);
    429 
    430 	/*
    431 	 * Make sure all stores are seen before the CPU release.  This
    432 	 * is relevant only in the non-fastpath scheduling case, but
    433 	 * we don't know here if that's going to happen, so need to
    434 	 * expect the worst.
    435 	 */
    436 	membar_exit();
    437 
    438 	/* Release the CPU. */
    439 	old = atomic_swap_ptr(&rcpu->rcpu_prevlwp, l);
    440 
    441 	/* No waiters?  No problems.  We're outta here. */
    442 	if (old == RCPULWP_BUSY) {
    443 		/* Was the scheduler interlock requested? */
    444 		if (__predict_false(interlock == rcpu->rcpu_mtx))
    445 			rumpuser_mutex_enter_nowrap(rcpu->rcpu_mtx);
    446 		return;
    447 	}
    448 
    449 	KASSERT(old == RCPULWP_WANTED);
    450 
    451 	/*
    452 	 * Ok, things weren't so snappy.
    453 	 *
    454 	 * Snailpath: take lock and signal anyone waiting for this CPU.
    455 	 */
    456 
    457 	rumpuser_mutex_enter_nowrap(rcpu->rcpu_mtx);
    458 	if (rcpu->rcpu_wanted)
    459 		rumpuser_cv_broadcast(rcpu->rcpu_cv);
    460 
    461 	if (__predict_true(interlock != rcpu->rcpu_mtx))
    462 		rumpuser_mutex_exit(rcpu->rcpu_mtx);
    463 }
    464 
    465 /* Give up and retake CPU (perhaps a different one) */
    466 void
    467 yield()
    468 {
    469 	struct lwp *l = curlwp;
    470 	int nlocks;
    471 
    472 	KERNEL_UNLOCK_ALL(l, &nlocks);
    473 	rump_unschedule_cpu(l);
    474 	rump_schedule_cpu(l);
    475 	KERNEL_LOCK(nlocks, l);
    476 }
    477 
    478 void
    479 preempt()
    480 {
    481 
    482 	yield();
    483 }
    484 
    485 bool
    486 kpreempt(uintptr_t where)
    487 {
    488 
    489 	return false;
    490 }
    491 
    492 /*
    493  * There is no kernel thread preemption in rump currently.  But call
    494  * the implementing macros anyway in case they grow some side-effects
    495  * down the road.
    496  */
    497 void
    498 kpreempt_disable(void)
    499 {
    500 
    501 	KPREEMPT_DISABLE(curlwp);
    502 }
    503 
    504 void
    505 kpreempt_enable(void)
    506 {
    507 
    508 	KPREEMPT_ENABLE(curlwp);
    509 }
    510 
    511 void
    512 suspendsched(void)
    513 {
    514 
    515 	/*
    516 	 * Could wait until everyone is out and block further entries,
    517 	 * but skip that for now.
    518 	 */
    519 }
    520 
    521 void
    522 sched_nice(struct proc *p, int level)
    523 {
    524 
    525 	/* nothing to do for now */
    526 }
    527