Home | History | Annotate | Line # | Download | only in kern
subr_xcall.c revision 1.13
      1 /*	$NetBSD: subr_xcall.c,v 1.13 2011/05/13 22:16:44 rmind Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2007-2010 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Andrew Doran and Mindaugas Rasiukevicius.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 /*
     33  * Cross call support
     34  *
     35  * Background
     36  *
     37  *	Sometimes it is necessary to modify hardware state that is tied
     38  *	directly to individual CPUs (such as a CPU's local timer), and
     39  *	these updates can not be done remotely by another CPU.  The LWP
     40  *	requesting the update may be unable to guarantee that it will be
     41  *	running on the CPU where the update must occur, when the update
     42  *	occurs.
     43  *
     44  *	Additionally, it's sometimes necessary to modify per-CPU software
     45  *	state from a remote CPU.  Where these update operations are so
     46  *	rare or the access to the per-CPU data so frequent that the cost
     47  *	of using locking or atomic operations to provide coherency is
     48  *	prohibitive, another way must be found.
     49  *
     50  *	Cross calls help to solve these types of problem by allowing
     51  *	any CPU in the system to request that an arbitrary function be
     52  *	executed on any other CPU.
     53  *
     54  * Implementation
     55  *
     56  *	A slow mechanism for making 'low priority' cross calls is
     57  *	provided.  The function to be executed runs on the remote CPU
     58  *	within a bound kthread.  No queueing is provided, and the
     59  *	implementation uses global state.  The function being called may
     60  *	block briefly on locks, but in doing so must be careful to not
     61  *	interfere with other cross calls in the system.  The function is
     62  *	called with thread context and not from a soft interrupt, so it
     63  *	can ensure that it is not interrupting other code running on the
     64  *	CPU, and so has exclusive access to the CPU.  Since this facility
     65  *	is heavyweight, it's expected that it will not be used often.
     66  *
     67  *	Cross calls must not allocate memory, as the pagedaemon uses
     68  *	them (and memory allocation may need to wait on the pagedaemon).
     69  *
     70  *	A low-overhead mechanism for high priority calls (XC_HIGHPRI) is
     71  *	also provided.  The function to be executed runs on a software
     72  *	interrupt context, at SOFTINT_CLOCK level, and is expected to be
     73  *	very lightweight, e.g. avoid blocking.
     74  */
     75 
     76 #include <sys/cdefs.h>
     77 __KERNEL_RCSID(0, "$NetBSD: subr_xcall.c,v 1.13 2011/05/13 22:16:44 rmind Exp $");
     78 
     79 #include <sys/types.h>
     80 #include <sys/param.h>
     81 #include <sys/xcall.h>
     82 #include <sys/mutex.h>
     83 #include <sys/condvar.h>
     84 #include <sys/evcnt.h>
     85 #include <sys/kthread.h>
     86 #include <sys/cpu.h>
     87 
     88 /* Cross-call state box. */
     89 typedef struct {
     90 	kmutex_t	xc_lock;
     91 	kcondvar_t	xc_busy;
     92 	xcfunc_t	xc_func;
     93 	void *		xc_arg1;
     94 	void *		xc_arg2;
     95 	uint64_t	xc_headp;
     96 	uint64_t	xc_donep;
     97 } xc_state_t;
     98 
     99 /* Bit indicating high (1) or low (0) priority. */
    100 #define	XC_PRI_BIT	(1ULL << 63)
    101 
    102 /* Low priority xcall structures. */
    103 static xc_state_t	xc_low_pri	__cacheline_aligned;
    104 static uint64_t		xc_tailp	__cacheline_aligned;
    105 
    106 /* High priority xcall structures. */
    107 static xc_state_t	xc_high_pri	__cacheline_aligned;
    108 static void *		xc_sih		__cacheline_aligned;
    109 
    110 /* Event counters. */
    111 static struct evcnt	xc_unicast_ev	__cacheline_aligned;
    112 static struct evcnt	xc_broadcast_ev	__cacheline_aligned;
    113 
    114 static void		xc_init(void);
    115 static void		xc_thread(void *);
    116 static void		xc_highpri_intr(void *);
    117 
    118 static inline uint64_t	xc_highpri(xcfunc_t, void *, void *, struct cpu_info *);
    119 static inline uint64_t	xc_lowpri(xcfunc_t, void *, void *, struct cpu_info *);
    120 
    121 /*
    122  * xc_init:
    123  *
    124  *	Initialize low and high priority cross-call structures.
    125  */
    126 static void
    127 xc_init(void)
    128 {
    129 	xc_state_t *xclo = &xc_low_pri, *xchi = &xc_high_pri;
    130 
    131 	memset(xclo, 0, sizeof(xc_state_t));
    132 	mutex_init(&xclo->xc_lock, MUTEX_DEFAULT, IPL_NONE);
    133 	cv_init(&xclo->xc_busy, "xclocv");
    134 	xc_tailp = 0;
    135 
    136 	memset(xchi, 0, sizeof(xc_state_t));
    137 	mutex_init(&xchi->xc_lock, MUTEX_DEFAULT, IPL_SOFTCLOCK);
    138 	cv_init(&xchi->xc_busy, "xchicv");
    139 	xc_sih = softint_establish(SOFTINT_CLOCK | SOFTINT_MPSAFE,
    140 	    xc_highpri_intr, NULL);
    141 	KASSERT(xc_sih != NULL);
    142 
    143 	evcnt_attach_dynamic(&xc_unicast_ev, EVCNT_TYPE_MISC, NULL,
    144 	   "crosscall", "unicast");
    145 	evcnt_attach_dynamic(&xc_broadcast_ev, EVCNT_TYPE_MISC, NULL,
    146 	   "crosscall", "broadcast");
    147 }
    148 
    149 /*
    150  * xc_init_cpu:
    151  *
    152  *	Initialize the cross-call subsystem.  Called once for each CPU
    153  *	in the system as they are attached.
    154  */
    155 void
    156 xc_init_cpu(struct cpu_info *ci)
    157 {
    158 	static bool again = false;
    159 	int error;
    160 
    161 	if (!again) {
    162 		/* Autoconfiguration will prevent re-entry. */
    163 		xc_init();
    164 		again = true;
    165 	}
    166 	cv_init(&ci->ci_data.cpu_xcall, "xcall");
    167 	error = kthread_create(PRI_XCALL, KTHREAD_MPSAFE, ci, xc_thread,
    168 	    NULL, NULL, "xcall/%u", ci->ci_index);
    169 	KASSERT(error == 0);
    170 }
    171 
    172 /*
    173  * xc_broadcast:
    174  *
    175  *	Trigger a call on all CPUs in the system.
    176  */
    177 uint64_t
    178 xc_broadcast(u_int flags, xcfunc_t func, void *arg1, void *arg2)
    179 {
    180 
    181 	KASSERT(!cpu_intr_p() && !cpu_softintr_p());
    182 
    183 	if ((flags & XC_HIGHPRI) != 0) {
    184 		return xc_highpri(func, arg1, arg2, NULL);
    185 	} else {
    186 		return xc_lowpri(func, arg1, arg2, NULL);
    187 	}
    188 }
    189 
    190 /*
    191  * xc_unicast:
    192  *
    193  *	Trigger a call on one CPU.
    194  */
    195 uint64_t
    196 xc_unicast(u_int flags, xcfunc_t func, void *arg1, void *arg2,
    197 	   struct cpu_info *ci)
    198 {
    199 
    200 	KASSERT(ci != NULL);
    201 	KASSERT(!cpu_intr_p() && !cpu_softintr_p());
    202 
    203 	if ((flags & XC_HIGHPRI) != 0) {
    204 		return xc_highpri(func, arg1, arg2, ci);
    205 	} else {
    206 		return xc_lowpri(func, arg1, arg2, ci);
    207 	}
    208 }
    209 
    210 /*
    211  * xc_wait:
    212  *
    213  *	Wait for a cross call to complete.
    214  */
    215 void
    216 xc_wait(uint64_t where)
    217 {
    218 	xc_state_t *xc;
    219 
    220 	KASSERT(!cpu_intr_p() && !cpu_softintr_p());
    221 
    222 	/* Determine whether it is high or low priority cross-call. */
    223 	if ((where & XC_PRI_BIT) != 0) {
    224 		xc = &xc_high_pri;
    225 		where &= ~XC_PRI_BIT;
    226 	} else {
    227 		xc = &xc_low_pri;
    228 	}
    229 
    230 	/* Fast path, if already done. */
    231 	if (xc->xc_donep >= where) {
    232 		return;
    233 	}
    234 
    235 	/* Slow path: block until awoken. */
    236 	mutex_enter(&xc->xc_lock);
    237 	while (xc->xc_donep < where) {
    238 		cv_wait(&xc->xc_busy, &xc->xc_lock);
    239 	}
    240 	mutex_exit(&xc->xc_lock);
    241 }
    242 
    243 /*
    244  * xc_lowpri:
    245  *
    246  *	Trigger a low priority call on one or more CPUs.
    247  */
    248 static inline uint64_t
    249 xc_lowpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci)
    250 {
    251 	xc_state_t *xc = &xc_low_pri;
    252 	CPU_INFO_ITERATOR cii;
    253 	uint64_t where;
    254 
    255 	mutex_enter(&xc->xc_lock);
    256 	while (xc->xc_headp != xc_tailp) {
    257 		cv_wait(&xc->xc_busy, &xc->xc_lock);
    258 	}
    259 	xc->xc_arg1 = arg1;
    260 	xc->xc_arg2 = arg2;
    261 	xc->xc_func = func;
    262 	if (ci == NULL) {
    263 		xc_broadcast_ev.ev_count++;
    264 		for (CPU_INFO_FOREACH(cii, ci)) {
    265 			if ((ci->ci_schedstate.spc_flags & SPCF_RUNNING) == 0)
    266 				continue;
    267 			xc->xc_headp += 1;
    268 			ci->ci_data.cpu_xcall_pending = true;
    269 			cv_signal(&ci->ci_data.cpu_xcall);
    270 		}
    271 	} else {
    272 		xc_unicast_ev.ev_count++;
    273 		xc->xc_headp += 1;
    274 		ci->ci_data.cpu_xcall_pending = true;
    275 		cv_signal(&ci->ci_data.cpu_xcall);
    276 	}
    277 	KASSERT(xc_tailp < xc->xc_headp);
    278 	where = xc->xc_headp;
    279 	mutex_exit(&xc->xc_lock);
    280 
    281 	/* Return a low priority ticket. */
    282 	KASSERT((where & XC_PRI_BIT) == 0);
    283 	return where;
    284 }
    285 
    286 /*
    287  * xc_thread:
    288  *
    289  *	One thread per-CPU to dispatch low priority calls.
    290  */
    291 static void
    292 xc_thread(void *cookie)
    293 {
    294 	struct cpu_info *ci = curcpu();
    295 	xc_state_t *xc = &xc_low_pri;
    296 	void *arg1, *arg2;
    297 	xcfunc_t func;
    298 
    299 	mutex_enter(&xc->xc_lock);
    300 	for (;;) {
    301 		while (!ci->ci_data.cpu_xcall_pending) {
    302 			if (xc->xc_headp == xc_tailp) {
    303 				cv_broadcast(&xc->xc_busy);
    304 			}
    305 			cv_wait(&ci->ci_data.cpu_xcall, &xc->xc_lock);
    306 			KASSERT(ci == curcpu());
    307 		}
    308 		ci->ci_data.cpu_xcall_pending = false;
    309 		func = xc->xc_func;
    310 		arg1 = xc->xc_arg1;
    311 		arg2 = xc->xc_arg2;
    312 		xc_tailp++;
    313 		mutex_exit(&xc->xc_lock);
    314 
    315 		KASSERT(func != NULL);
    316 		(*func)(arg1, arg2);
    317 
    318 		mutex_enter(&xc->xc_lock);
    319 		xc->xc_donep++;
    320 	}
    321 	/* NOTREACHED */
    322 }
    323 
    324 /*
    325  * xc_ipi_handler:
    326  *
    327  *	Handler of cross-call IPI.
    328  */
    329 void
    330 xc_ipi_handler(void)
    331 {
    332 	/* Executes xc_highpri_intr() via software interrupt. */
    333 	softint_schedule(xc_sih);
    334 }
    335 
    336 /*
    337  * xc_highpri_intr:
    338  *
    339  *	A software interrupt handler for high priority calls.
    340  */
    341 static void
    342 xc_highpri_intr(void *dummy)
    343 {
    344 	xc_state_t *xc = &xc_high_pri;
    345 	void *arg1, *arg2;
    346 	xcfunc_t func;
    347 
    348 	KASSERT(cpu_softintr_p());
    349 	/*
    350 	 * Lock-less fetch of function and its arguments.
    351 	 * Safe since it cannot change at this point.
    352 	 */
    353 	KASSERT(xc->xc_donep < xc->xc_headp);
    354 	func = xc->xc_func;
    355 	arg1 = xc->xc_arg1;
    356 	arg2 = xc->xc_arg2;
    357 
    358 	KASSERT(func != NULL);
    359 	(*func)(arg1, arg2);
    360 
    361 	/*
    362 	 * Note the request as done, and if we have reached the head,
    363 	 * cross-call has been processed - notify waiters, if any.
    364 	 */
    365 	mutex_enter(&xc->xc_lock);
    366 	if (++xc->xc_donep == xc->xc_headp) {
    367 		cv_broadcast(&xc->xc_busy);
    368 	}
    369 	mutex_exit(&xc->xc_lock);
    370 }
    371 
    372 /*
    373  * xc_highpri:
    374  *
    375  *	Trigger a high priority call on one or more CPUs.
    376  */
    377 static inline uint64_t
    378 xc_highpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci)
    379 {
    380 	xc_state_t *xc = &xc_high_pri;
    381 	uint64_t where;
    382 
    383 	mutex_enter(&xc->xc_lock);
    384 	while (xc->xc_headp != xc->xc_donep) {
    385 		cv_wait(&xc->xc_busy, &xc->xc_lock);
    386 	}
    387 	xc->xc_func = func;
    388 	xc->xc_arg1 = arg1;
    389 	xc->xc_arg2 = arg2;
    390 	xc->xc_headp += (ci ? 1 : ncpu);
    391 	where = xc->xc_headp;
    392 	mutex_exit(&xc->xc_lock);
    393 
    394 	/*
    395 	 * Send the IPI once lock is released.
    396 	 * Note: it will handle the local CPU case.
    397 	 */
    398 
    399 #ifdef MULTIPROCESSOR
    400 	kpreempt_disable();
    401 	if (curcpu() == ci) {
    402 		/* Unicast: local CPU. */
    403 		xc_ipi_handler();
    404 	} else if (ci) {
    405 		/* Unicast: remote CPU. */
    406 		xc_send_ipi(ci);
    407 	} else {
    408 		/* Broadcast: all, including local. */
    409 		xc_send_ipi(NULL);
    410 		xc_ipi_handler();
    411 	}
    412 	kpreempt_enable();
    413 #else
    414 	KASSERT(curcpu() == ci);
    415 	xc_ipi_handler();
    416 #endif
    417 
    418 	/* Indicate a high priority ticket. */
    419 	return (where | XC_PRI_BIT);
    420 }
    421