File:  [DragonFly] / src / sys / kern / kern_poll.c
Revision 1.10: download - view: text, annotated - select for diffs
Sun Apr 11 05:14:47 2004 UTC (10 years, 7 months ago) by hsu
Branches: MAIN
CVS tags: HEAD
Fix compilation errors with missing header files and misnamed formal parameter.

Reported by:	David Rhodus

    1: /*-
    2:  * Copyright (c) 2001-2002 Luigi Rizzo
    3:  *
    4:  * Supported by: the Xorp Project (www.xorp.org)
    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 AUTHORS AND CONTRIBUTORS ``AS IS'' AND
   16:  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   17:  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   18:  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS 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
   21:  * OR 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:  * $FreeBSD: src/sys/kern/kern_poll.c,v 1.2.2.4 2002/06/27 23:26:33 luigi Exp $
   28:  * $DragonFly: src/sys/kern/kern_poll.c,v 1.10 2004/04/11 05:14:47 hsu Exp $
   29:  */
   30: 
   31: #include <sys/param.h>
   32: #include <sys/systm.h>
   33: #include <sys/kernel.h>
   34: #include <sys/socket.h>			/* needed by net/if.h		*/
   35: #include <sys/sysctl.h>
   36: 
   37: #include <sys/thread2.h>
   38: #include <sys/msgport2.h>
   39: 
   40: #include <i386/include/md_var.h>	/* for vm_page_zero_idle()	*/
   41: #include <net/if.h>			/* for IFF_* flags		*/
   42: #include <net/netisr.h>			/* for NETISR_POLL		*/
   43: 
   44: #ifdef SMP
   45: #include "opt_lint.h"
   46: #ifndef COMPILING_LINT
   47: #error DEVICE_POLLING is not compatible with SMP
   48: #endif
   49: #endif
   50: 
   51: /* the two netisr handlers */
   52: static void netisr_poll(struct netmsg *);
   53: static void netisr_pollmore(struct netmsg *);
   54: 
   55: void init_device_poll(void);		/* init routine			*/
   56: void hardclock_device_poll(void);	/* hook from hardclock		*/
   57: void ether_poll(int);			/* polling while in trap	*/
   58: int idle_poll(void);			/* poll while in idle loop	*/
   59: 
   60: /*
   61:  * Polling support for [network] device drivers.
   62:  *
   63:  * Drivers which support this feature try to register with the
   64:  * polling code.
   65:  *
   66:  * If registration is successful, the driver must disable interrupts,
   67:  * and further I/O is performed through the handler, which is invoked
   68:  * (at least once per clock tick) with 3 arguments: the "arg" passed at
   69:  * register time (a struct ifnet pointer), a command, and a "count" limit.
   70:  *
   71:  * The command can be one of the following:
   72:  *  POLL_ONLY: quick move of "count" packets from input/output queues.
   73:  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
   74:  *	other more expensive operations. This command is issued periodically
   75:  *	but less frequently than POLL_ONLY.
   76:  *  POLL_DEREGISTER: deregister and return to interrupt mode.
   77:  *
   78:  * The first two commands are only issued if the interface is marked as
   79:  * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
   80:  *
   81:  * The count limit specifies how much work the handler can do during the
   82:  * call -- typically this is the number of packets to be received, or
   83:  * transmitted, etc. (drivers are free to interpret this number, as long
   84:  * as the max time spent in the function grows roughly linearly with the
   85:  * count).
   86:  *
   87:  * Deregistration can be requested by the driver itself (typically in the
   88:  * *_stop() routine), or by the polling code, by invoking the handler.
   89:  *
   90:  * Polling can be globally enabled or disabled with the sysctl variable
   91:  * kern.polling.enable (default is 0, disabled)
   92:  *
   93:  * A second variable controls the sharing of CPU between polling/kernel
   94:  * network processing, and other activities (typically userlevel tasks):
   95:  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
   96:  * of CPU allocated to user tasks. CPU is allocated proportionally to the
   97:  * shares, by dynamically adjusting the "count" (poll_burst).
   98:  *
   99:  * Other parameters can should be left to their default values.
  100:  * The following constraints hold
  101:  *
  102:  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
  103:  *	0 <= poll_in_trap <= poll_each_burst
  104:  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
  105:  */
  106: 
  107: #define MIN_POLL_BURST_MAX	10
  108: #define MAX_POLL_BURST_MAX	1000
  109: 
  110: SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
  111: 	"Device polling parameters");
  112: 
  113: static u_int32_t poll_burst = 5;
  114: SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
  115: 	&poll_burst, 0, "Current polling burst size");
  116: 
  117: static u_int32_t poll_each_burst = 5;
  118: SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
  119: 	&poll_each_burst, 0, "Max size of each burst");
  120: 
  121: static u_int32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
  122: SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
  123: 	&poll_burst_max, 0, "Max Polling burst size");
  124: 
  125: static u_int32_t poll_in_idle_loop=1;		/* do we poll in idle loop ? */
  126: SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
  127: 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
  128: 
  129: u_int32_t poll_in_trap;			/* used in trap.c */
  130: SYSCTL_UINT(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW,
  131: 	&poll_in_trap, 0, "Poll burst size during a trap");
  132: 
  133: static u_int32_t user_frac = 50;
  134: SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
  135: 	&user_frac, 0, "Desired user fraction of cpu time");
  136: 
  137: static u_int32_t reg_frac = 20 ;
  138: SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
  139: 	&reg_frac, 0, "Every this many cycles poll register");
  140: 
  141: static u_int32_t short_ticks;
  142: SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
  143: 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
  144: 
  145: static u_int32_t lost_polls;
  146: SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
  147: 	&lost_polls, 0, "How many times we would have lost a poll tick");
  148: 
  149: static u_int32_t pending_polls;
  150: SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
  151: 	&pending_polls, 0, "Do we need to poll again");
  152: 
  153: static int residual_burst = 0;
  154: SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
  155: 	&residual_burst, 0, "# of residual cycles in burst");
  156: 
  157: static u_int32_t poll_handlers; /* next free entry in pr[]. */
  158: SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
  159: 	&poll_handlers, 0, "Number of registered poll handlers");
  160: 
  161: static int polling = 0;		/* global polling enable */
  162: SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
  163: 	&polling, 0, "Polling enabled");
  164: 
  165: static u_int32_t phase;
  166: SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
  167: 	&phase, 0, "Polling phase");
  168: 
  169: static u_int32_t suspect;
  170: SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
  171: 	&suspect, 0, "suspect event");
  172: 
  173: static u_int32_t stalled;
  174: SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
  175: 	&stalled, 0, "potential stalls");
  176: 
  177: 
  178: #define POLL_LIST_LEN  128
  179: struct pollrec {
  180: 	poll_handler_t	*handler;
  181: 	struct ifnet	*ifp;
  182: };
  183: 
  184: static struct pollrec pr[POLL_LIST_LEN];
  185: 
  186: /*
  187:  * register relevant netisr. Called from kern_clock.c:
  188:  */
  189: void
  190: init_device_poll(void)
  191: {
  192: 	netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
  193: 	netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
  194: }
  195: 
  196: /*
  197:  * Hook from hardclock. Tries to schedule a netisr, but keeps track
  198:  * of lost ticks due to the previous handler taking too long.
  199:  * Normally, this should not happen, because polling handler should
  200:  * run for a short time. However, in some cases (e.g. when there are
  201:  * changes in link status etc.) the drivers take a very long time
  202:  * (even in the order of milliseconds) to reset and reconfigure the
  203:  * device, causing apparent lost polls.
  204:  *
  205:  * The first part of the code is just for debugging purposes, and tries
  206:  * to count how often hardclock ticks are shorter than they should,
  207:  * meaning either stray interrupts or delayed events.
  208:  *
  209:  * WARNING! called from fastint or IPI, the MP lock might not be held.
  210:  */
  211: void
  212: hardclock_device_poll(void)
  213: {
  214: 	static struct timeval prev_t, t;
  215: 	int delta;
  216: 
  217: 	if (poll_handlers == 0)
  218: 		return;
  219: 
  220: 	microuptime(&t);
  221: 	delta = (t.tv_usec - prev_t.tv_usec) +
  222: 		(t.tv_sec - prev_t.tv_sec)*1000000;
  223: 	if (delta * hz < 500000)
  224: 		short_ticks++;
  225: 	else
  226: 		prev_t = t;
  227: 
  228: 	if (pending_polls > 100) {
  229: 		/*
  230: 		 * Too much, assume it has stalled (not always true
  231: 		 * see comment above).
  232: 		 */
  233: 		stalled++;
  234: 		pending_polls = 0;
  235: 		phase = 0;
  236: 	}
  237: 
  238: 	if (phase <= 2) {
  239: 		if (phase != 0)
  240: 			suspect++;
  241: 		phase = 1;
  242: 		schednetisr(NETISR_POLL);
  243: 		phase = 2;
  244: 	}
  245: 	if (pending_polls++ > 0)
  246: 		lost_polls++;
  247: }
  248: 
  249: /*
  250:  * ether_poll is called from the idle loop or from the trap handler.
  251:  */
  252: void
  253: ether_poll(int count)
  254: {
  255: 	int i;
  256: 	int s = splimp();
  257: 
  258: 	if (count > poll_each_burst)
  259: 		count = poll_each_burst;
  260: 	for (i = 0 ; i < poll_handlers ; i++)
  261: 		if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
  262: 		    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
  263: 			pr[i].handler(pr[i].ifp, 0, count); /* quick check */
  264: 	splx(s);
  265: }
  266: 
  267: /*
  268:  * idle_poll is replaces the body of the idle loop when DEVICE_POLLING
  269:  * is used.  YYY not currently implemented.
  270:  */
  271: int
  272: idle_poll(void)
  273: {
  274: 	if (poll_in_idle_loop && poll_handlers > 0) {
  275: 		int s = splimp();
  276: 		cpu_enable_intr();
  277: 		ether_poll(poll_each_burst);
  278: 		cpu_disable_intr();
  279: 		splx(s);
  280: 		vm_page_zero_idle();
  281: 		return 1;
  282: 	} else
  283: 		return vm_page_zero_idle();
  284: }
  285: 
  286: /*
  287:  * netisr_pollmore is called after other netisr's, possibly scheduling
  288:  * another NETISR_POLL call, or adapting the burst size for the next cycle.
  289:  *
  290:  * It is very bad to fetch large bursts of packets from a single card at once,
  291:  * because the burst could take a long time to be completely processed, or
  292:  * could saturate the intermediate queue (ipintrq or similar) leading to
  293:  * losses or unfairness. To reduce the problem, and also to account better for
  294:  * time spent in network-related processing, we split the burst in smaller
  295:  * chunks of fixed size, giving control to the other netisr's between chunks.
  296:  * This helps in improving the fairness, reducing livelock (because we
  297:  * emulate more closely the "process to completion" that we have with
  298:  * fastforwarding) and accounting for the work performed in low level
  299:  * handling and forwarding.
  300:  */
  301: 
  302: static struct timeval poll_start_t;
  303: 
  304: /* ARGSUSED */
  305: static void
  306: netisr_pollmore(struct netmsg *msg)
  307: {
  308: 	struct timeval t;
  309: 	int kern_load;
  310: 	int s = splhigh();
  311: 
  312: 	phase = 5;
  313: 	if (residual_burst > 0) {
  314: 		schednetisr(NETISR_POLL);
  315: 		/* will run immediately on return, followed by netisrs */
  316: 		goto out;
  317: 	}
  318: 	/* here we can account time spent in netisr's in this tick */
  319: 	microuptime(&t);
  320: 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
  321: 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
  322: 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
  323: 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
  324: 		if (poll_burst > 1)
  325: 			poll_burst--;
  326: 	} else {
  327: 		if (poll_burst < poll_burst_max)
  328: 			poll_burst++;
  329: 	}
  330: 
  331: 	pending_polls--;
  332: 	if (pending_polls == 0)	/* we are done */
  333: 		phase = 0;
  334: 	else {
  335: 		/*
  336: 		 * Last cycle was long and caused us to miss one or more
  337: 		 * hardclock ticks. Restart processing again, but slightly
  338: 		 * reduce the burst size to prevent that this happens again.
  339: 		 */
  340: 		poll_burst -= (poll_burst / 8);
  341: 		if (poll_burst < 1)
  342: 			poll_burst = 1;
  343: 		schednetisr(NETISR_POLL);
  344: 		phase = 6;
  345: 	}
  346: out:
  347: 	splx(s);
  348: 	lwkt_replymsg(&msg->nm_lmsg, 0);
  349: }
  350: 
  351: /*
  352:  * netisr_poll is scheduled by schednetisr when appropriate, typically once
  353:  * per tick. It is called at splnet() so first thing to do is to upgrade to
  354:  * splimp(), and call all registered handlers.
  355:  */
  356: /* ARGSUSED */
  357: static void
  358: netisr_poll(struct netmsg *msg)
  359: {
  360: 	static int reg_frac_count;
  361: 	int i, cycles;
  362: 	enum poll_cmd arg = POLL_ONLY;
  363: 	int s=splimp();
  364: 
  365: 	phase = 3;
  366: 	if (residual_burst == 0) { /* first call in this tick */
  367: 		microuptime(&poll_start_t);
  368: 		/*
  369: 		 * Check that paremeters are consistent with runtime
  370: 		 * variables. Some of these tests could be done at sysctl
  371: 		 * time, but the savings would be very limited because we
  372: 		 * still have to check against reg_frac_count and
  373: 		 * poll_each_burst. So, instead of writing separate sysctl
  374: 		 * handlers, we do all here.
  375: 		 */
  376: 
  377: 		if (reg_frac > hz)
  378: 			reg_frac = hz;
  379: 		else if (reg_frac < 1)
  380: 			reg_frac = 1;
  381: 		if (reg_frac_count > reg_frac)
  382: 			reg_frac_count = reg_frac - 1;
  383: 		if (reg_frac_count-- == 0) {
  384: 			arg = POLL_AND_CHECK_STATUS;
  385: 			reg_frac_count = reg_frac - 1;
  386: 		}
  387: 		if (poll_burst_max < MIN_POLL_BURST_MAX)
  388: 			poll_burst_max = MIN_POLL_BURST_MAX;
  389: 		else if (poll_burst_max > MAX_POLL_BURST_MAX)
  390: 			poll_burst_max = MAX_POLL_BURST_MAX;
  391: 
  392: 		if (poll_each_burst < 1)
  393: 			poll_each_burst = 1;
  394: 		else if (poll_each_burst > poll_burst_max)
  395: 			poll_each_burst = poll_burst_max;
  396: 
  397: 		residual_burst = poll_burst;
  398: 	}
  399: 	cycles = (residual_burst < poll_each_burst) ?
  400: 		residual_burst : poll_each_burst;
  401: 	residual_burst -= cycles;
  402: 
  403: 	if (polling) {
  404: 		for (i = 0 ; i < poll_handlers ; i++)
  405: 			if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
  406: 			    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
  407: 				pr[i].handler(pr[i].ifp, arg, cycles);
  408: 	} else {	/* unregister */
  409: 		for (i = 0 ; i < poll_handlers ; i++) {
  410: 			if (pr[i].handler &&
  411: 			    pr[i].ifp->if_flags & IFF_RUNNING) {
  412: 				pr[i].ifp->if_ipending &= ~IFF_POLLING;
  413: 				pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1);
  414: 			}
  415: 			pr[i].handler=NULL;
  416: 		}
  417: 		residual_burst = 0;
  418: 		poll_handlers = 0;
  419: 	}
  420: 	schednetisr(NETISR_POLLMORE);
  421: 	phase = 4;
  422: 	splx(s);
  423: 	lwkt_replymsg(&msg->nm_lmsg, 0);
  424: }
  425: 
  426: /*
  427:  * Try to register routine for polling. Returns 1 if successful
  428:  * (and polling should be enabled), 0 otherwise.
  429:  * A device is not supposed to register itself multiple times.
  430:  *
  431:  * This is called from within the *_intr() functions, so we do not need
  432:  * further locking.
  433:  */
  434: int
  435: ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
  436: {
  437: 	int s;
  438: 
  439: 	if (polling == 0) /* polling disabled, cannot register */
  440: 		return 0;
  441: 	if (h == NULL || ifp == NULL)		/* bad arguments	*/
  442: 		return 0;
  443: 	if ( !(ifp->if_flags & IFF_UP) )	/* must be up		*/
  444: 		return 0;
  445: 	if (ifp->if_ipending & IFF_POLLING)	/* already polling	*/
  446: 		return 0;
  447: 
  448: 	s = splhigh();
  449: 	if (poll_handlers >= POLL_LIST_LEN) {
  450: 		/*
  451: 		 * List full, cannot register more entries.
  452: 		 * This should never happen; if it does, it is probably a
  453: 		 * broken driver trying to register multiple times. Checking
  454: 		 * this at runtime is expensive, and won't solve the problem
  455: 		 * anyways, so just report a few times and then give up.
  456: 		 */
  457: 		static int verbose = 10 ;
  458: 		splx(s);
  459: 		if (verbose >0) {
  460: 			printf("poll handlers list full, "
  461: 				"maybe a broken driver ?\n");
  462: 			verbose--;
  463: 		}
  464: 		return 0; /* no polling for you */
  465: 	}
  466: 
  467: 	pr[poll_handlers].handler = h;
  468: 	pr[poll_handlers].ifp = ifp;
  469: 	poll_handlers++;
  470: 	ifp->if_ipending |= IFF_POLLING;
  471: 	splx(s);
  472: 	return 1; /* polling enabled in next call */
  473: }
  474: 
  475: /*
  476:  * Remove interface from the polling list. Normally called by *_stop().
  477:  * It is not an error to call it with IFF_POLLING clear, the call is
  478:  * sufficiently rare to be preferable to save the space for the extra
  479:  * test in each driver in exchange of one additional function call.
  480:  */
  481: int
  482: ether_poll_deregister(struct ifnet *ifp)
  483: {
  484: 	int i;
  485: 	int s = splimp();
  486: 	
  487: 	if ( !ifp || !(ifp->if_ipending & IFF_POLLING) ) {
  488: 		splx(s);
  489: 		return 0;
  490: 	}
  491: 	for (i = 0 ; i < poll_handlers ; i++)
  492: 		if (pr[i].ifp == ifp) /* found it */
  493: 			break;
  494: 	ifp->if_ipending &= ~IFF_POLLING; /* found or not... */
  495: 	if (i == poll_handlers) {
  496: 		splx(s);
  497: 		printf("ether_poll_deregister: ifp not found!!!\n");
  498: 		return 0;
  499: 	}
  500: 	poll_handlers--;
  501: 	if (i < poll_handlers) { /* Last entry replaces this one. */
  502: 		pr[i].handler = pr[poll_handlers].handler;
  503: 		pr[i].ifp = pr[poll_handlers].ifp;
  504: 	}
  505: 	splx(s);
  506: 	return 1;
  507: }