File:  [DragonFly] / src / sys / kern / kern_poll.c
Revision 1.9: download - view: text, annotated - select for diffs
Fri Apr 9 22:34:09 2004 UTC (10 years, 3 months ago) by hsu
Branches: MAIN
CVS tags: HEAD
Push the lwkt_replymsg() up one level from netisr_service_loop() to
the message handler so we can explicitly reply or not reply as appropriate.

    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 1.9 2004/04/09 22:34:09 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 <i386/include/md_var.h>	/* for vm_page_zero_idle()	*/
   38: #include <net/if.h>			/* for IFF_* flags		*/
   39: #include <net/netisr.h>			/* for NETISR_POLL		*/
   40: 
   41: #ifdef SMP
   42: #include "opt_lint.h"
   43: #ifndef COMPILING_LINT
   44: #error DEVICE_POLLING is not compatible with SMP
   45: #endif
   46: #endif
   47: 
   48: /* the two netisr handlers */
   49: static void netisr_poll(struct netmsg *);
   50: static void netisr_pollmore(struct netmsg *);
   51: 
   52: void init_device_poll(void);		/* init routine			*/
   53: void hardclock_device_poll(void);	/* hook from hardclock		*/
   54: void ether_poll(int);			/* polling while in trap	*/
   55: int idle_poll(void);			/* poll while in idle loop	*/
   56: 
   57: /*
   58:  * Polling support for [network] device drivers.
   59:  *
   60:  * Drivers which support this feature try to register with the
   61:  * polling code.
   62:  *
   63:  * If registration is successful, the driver must disable interrupts,
   64:  * and further I/O is performed through the handler, which is invoked
   65:  * (at least once per clock tick) with 3 arguments: the "arg" passed at
   66:  * register time (a struct ifnet pointer), a command, and a "count" limit.
   67:  *
   68:  * The command can be one of the following:
   69:  *  POLL_ONLY: quick move of "count" packets from input/output queues.
   70:  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
   71:  *	other more expensive operations. This command is issued periodically
   72:  *	but less frequently than POLL_ONLY.
   73:  *  POLL_DEREGISTER: deregister and return to interrupt mode.
   74:  *
   75:  * The first two commands are only issued if the interface is marked as
   76:  * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set.
   77:  *
   78:  * The count limit specifies how much work the handler can do during the
   79:  * call -- typically this is the number of packets to be received, or
   80:  * transmitted, etc. (drivers are free to interpret this number, as long
   81:  * as the max time spent in the function grows roughly linearly with the
   82:  * count).
   83:  *
   84:  * Deregistration can be requested by the driver itself (typically in the
   85:  * *_stop() routine), or by the polling code, by invoking the handler.
   86:  *
   87:  * Polling can be globally enabled or disabled with the sysctl variable
   88:  * kern.polling.enable (default is 0, disabled)
   89:  *
   90:  * A second variable controls the sharing of CPU between polling/kernel
   91:  * network processing, and other activities (typically userlevel tasks):
   92:  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
   93:  * of CPU allocated to user tasks. CPU is allocated proportionally to the
   94:  * shares, by dynamically adjusting the "count" (poll_burst).
   95:  *
   96:  * Other parameters can should be left to their default values.
   97:  * The following constraints hold
   98:  *
   99:  *	1 <= poll_each_burst <= poll_burst <= poll_burst_max
  100:  *	0 <= poll_in_trap <= poll_each_burst
  101:  *	MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
  102:  */
  103: 
  104: #define MIN_POLL_BURST_MAX	10
  105: #define MAX_POLL_BURST_MAX	1000
  106: 
  107: SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
  108: 	"Device polling parameters");
  109: 
  110: static u_int32_t poll_burst = 5;
  111: SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RW,
  112: 	&poll_burst, 0, "Current polling burst size");
  113: 
  114: static u_int32_t poll_each_burst = 5;
  115: SYSCTL_UINT(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW,
  116: 	&poll_each_burst, 0, "Max size of each burst");
  117: 
  118: static u_int32_t poll_burst_max = 150;	/* good for 100Mbit net and HZ=1000 */
  119: SYSCTL_UINT(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW,
  120: 	&poll_burst_max, 0, "Max Polling burst size");
  121: 
  122: static u_int32_t poll_in_idle_loop=1;		/* do we poll in idle loop ? */
  123: SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW,
  124: 	&poll_in_idle_loop, 0, "Enable device polling in idle loop");
  125: 
  126: u_int32_t poll_in_trap;			/* used in trap.c */
  127: SYSCTL_UINT(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW,
  128: 	&poll_in_trap, 0, "Poll burst size during a trap");
  129: 
  130: static u_int32_t user_frac = 50;
  131: SYSCTL_UINT(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW,
  132: 	&user_frac, 0, "Desired user fraction of cpu time");
  133: 
  134: static u_int32_t reg_frac = 20 ;
  135: SYSCTL_UINT(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW,
  136: 	&reg_frac, 0, "Every this many cycles poll register");
  137: 
  138: static u_int32_t short_ticks;
  139: SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW,
  140: 	&short_ticks, 0, "Hardclock ticks shorter than they should be");
  141: 
  142: static u_int32_t lost_polls;
  143: SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW,
  144: 	&lost_polls, 0, "How many times we would have lost a poll tick");
  145: 
  146: static u_int32_t pending_polls;
  147: SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RW,
  148: 	&pending_polls, 0, "Do we need to poll again");
  149: 
  150: static int residual_burst = 0;
  151: SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RW,
  152: 	&residual_burst, 0, "# of residual cycles in burst");
  153: 
  154: static u_int32_t poll_handlers; /* next free entry in pr[]. */
  155: SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD,
  156: 	&poll_handlers, 0, "Number of registered poll handlers");
  157: 
  158: static int polling = 0;		/* global polling enable */
  159: SYSCTL_UINT(_kern_polling, OID_AUTO, enable, CTLFLAG_RW,
  160: 	&polling, 0, "Polling enabled");
  161: 
  162: static u_int32_t phase;
  163: SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RW,
  164: 	&phase, 0, "Polling phase");
  165: 
  166: static u_int32_t suspect;
  167: SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RW,
  168: 	&suspect, 0, "suspect event");
  169: 
  170: static u_int32_t stalled;
  171: SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RW,
  172: 	&stalled, 0, "potential stalls");
  173: 
  174: 
  175: #define POLL_LIST_LEN  128
  176: struct pollrec {
  177: 	poll_handler_t	*handler;
  178: 	struct ifnet	*ifp;
  179: };
  180: 
  181: static struct pollrec pr[POLL_LIST_LEN];
  182: 
  183: /*
  184:  * register relevant netisr. Called from kern_clock.c:
  185:  */
  186: void
  187: init_device_poll(void)
  188: {
  189: 	netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
  190: 	netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
  191: }
  192: 
  193: /*
  194:  * Hook from hardclock. Tries to schedule a netisr, but keeps track
  195:  * of lost ticks due to the previous handler taking too long.
  196:  * Normally, this should not happen, because polling handler should
  197:  * run for a short time. However, in some cases (e.g. when there are
  198:  * changes in link status etc.) the drivers take a very long time
  199:  * (even in the order of milliseconds) to reset and reconfigure the
  200:  * device, causing apparent lost polls.
  201:  *
  202:  * The first part of the code is just for debugging purposes, and tries
  203:  * to count how often hardclock ticks are shorter than they should,
  204:  * meaning either stray interrupts or delayed events.
  205:  *
  206:  * WARNING! called from fastint or IPI, the MP lock might not be held.
  207:  */
  208: void
  209: hardclock_device_poll(void)
  210: {
  211: 	static struct timeval prev_t, t;
  212: 	int delta;
  213: 
  214: 	if (poll_handlers == 0)
  215: 		return;
  216: 
  217: 	microuptime(&t);
  218: 	delta = (t.tv_usec - prev_t.tv_usec) +
  219: 		(t.tv_sec - prev_t.tv_sec)*1000000;
  220: 	if (delta * hz < 500000)
  221: 		short_ticks++;
  222: 	else
  223: 		prev_t = t;
  224: 
  225: 	if (pending_polls > 100) {
  226: 		/*
  227: 		 * Too much, assume it has stalled (not always true
  228: 		 * see comment above).
  229: 		 */
  230: 		stalled++;
  231: 		pending_polls = 0;
  232: 		phase = 0;
  233: 	}
  234: 
  235: 	if (phase <= 2) {
  236: 		if (phase != 0)
  237: 			suspect++;
  238: 		phase = 1;
  239: 		schednetisr(NETISR_POLL);
  240: 		phase = 2;
  241: 	}
  242: 	if (pending_polls++ > 0)
  243: 		lost_polls++;
  244: }
  245: 
  246: /*
  247:  * ether_poll is called from the idle loop or from the trap handler.
  248:  */
  249: void
  250: ether_poll(int count)
  251: {
  252: 	int i;
  253: 	int s = splimp();
  254: 
  255: 	if (count > poll_each_burst)
  256: 		count = poll_each_burst;
  257: 	for (i = 0 ; i < poll_handlers ; i++)
  258: 		if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
  259: 		    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
  260: 			pr[i].handler(pr[i].ifp, 0, count); /* quick check */
  261: 	splx(s);
  262: }
  263: 
  264: /*
  265:  * idle_poll is replaces the body of the idle loop when DEVICE_POLLING
  266:  * is used.  YYY not currently implemented.
  267:  */
  268: int
  269: idle_poll(void)
  270: {
  271: 	if (poll_in_idle_loop && poll_handlers > 0) {
  272: 		int s = splimp();
  273: 		cpu_enable_intr();
  274: 		ether_poll(poll_each_burst);
  275: 		cpu_disable_intr();
  276: 		splx(s);
  277: 		vm_page_zero_idle();
  278: 		return 1;
  279: 	} else
  280: 		return vm_page_zero_idle();
  281: }
  282: 
  283: /*
  284:  * netisr_pollmore is called after other netisr's, possibly scheduling
  285:  * another NETISR_POLL call, or adapting the burst size for the next cycle.
  286:  *
  287:  * It is very bad to fetch large bursts of packets from a single card at once,
  288:  * because the burst could take a long time to be completely processed, or
  289:  * could saturate the intermediate queue (ipintrq or similar) leading to
  290:  * losses or unfairness. To reduce the problem, and also to account better for
  291:  * time spent in network-related processing, we split the burst in smaller
  292:  * chunks of fixed size, giving control to the other netisr's between chunks.
  293:  * This helps in improving the fairness, reducing livelock (because we
  294:  * emulate more closely the "process to completion" that we have with
  295:  * fastforwarding) and accounting for the work performed in low level
  296:  * handling and forwarding.
  297:  */
  298: 
  299: static struct timeval poll_start_t;
  300: 
  301: /* ARGSUSED */
  302: static void
  303: netisr_pollmore(struct netmsg *dummy __unused)
  304: {
  305: 	struct timeval t;
  306: 	int kern_load;
  307: 	int s = splhigh();
  308: 
  309: 	phase = 5;
  310: 	if (residual_burst > 0) {
  311: 		schednetisr(NETISR_POLL);
  312: 		/* will run immediately on return, followed by netisrs */
  313: 		goto out;
  314: 	}
  315: 	/* here we can account time spent in netisr's in this tick */
  316: 	microuptime(&t);
  317: 	kern_load = (t.tv_usec - poll_start_t.tv_usec) +
  318: 		(t.tv_sec - poll_start_t.tv_sec)*1000000;	/* us */
  319: 	kern_load = (kern_load * hz) / 10000;			/* 0..100 */
  320: 	if (kern_load > (100 - user_frac)) { /* try decrease ticks */
  321: 		if (poll_burst > 1)
  322: 			poll_burst--;
  323: 	} else {
  324: 		if (poll_burst < poll_burst_max)
  325: 			poll_burst++;
  326: 	}
  327: 
  328: 	pending_polls--;
  329: 	if (pending_polls == 0)	/* we are done */
  330: 		phase = 0;
  331: 	else {
  332: 		/*
  333: 		 * Last cycle was long and caused us to miss one or more
  334: 		 * hardclock ticks. Restart processing again, but slightly
  335: 		 * reduce the burst size to prevent that this happens again.
  336: 		 */
  337: 		poll_burst -= (poll_burst / 8);
  338: 		if (poll_burst < 1)
  339: 			poll_burst = 1;
  340: 		schednetisr(NETISR_POLL);
  341: 		phase = 6;
  342: 	}
  343: out:
  344: 	splx(s);
  345: 	lwkt_replymsg(&msg->nm_lmsg, 0);
  346: }
  347: 
  348: /*
  349:  * netisr_poll is scheduled by schednetisr when appropriate, typically once
  350:  * per tick. It is called at splnet() so first thing to do is to upgrade to
  351:  * splimp(), and call all registered handlers.
  352:  */
  353: /* ARGSUSED */
  354: static void
  355: netisr_poll(struct netmsg *dummy __unused)
  356: {
  357: 	static int reg_frac_count;
  358: 	int i, cycles;
  359: 	enum poll_cmd arg = POLL_ONLY;
  360: 	int s=splimp();
  361: 
  362: 	phase = 3;
  363: 	if (residual_burst == 0) { /* first call in this tick */
  364: 		microuptime(&poll_start_t);
  365: 		/*
  366: 		 * Check that paremeters are consistent with runtime
  367: 		 * variables. Some of these tests could be done at sysctl
  368: 		 * time, but the savings would be very limited because we
  369: 		 * still have to check against reg_frac_count and
  370: 		 * poll_each_burst. So, instead of writing separate sysctl
  371: 		 * handlers, we do all here.
  372: 		 */
  373: 
  374: 		if (reg_frac > hz)
  375: 			reg_frac = hz;
  376: 		else if (reg_frac < 1)
  377: 			reg_frac = 1;
  378: 		if (reg_frac_count > reg_frac)
  379: 			reg_frac_count = reg_frac - 1;
  380: 		if (reg_frac_count-- == 0) {
  381: 			arg = POLL_AND_CHECK_STATUS;
  382: 			reg_frac_count = reg_frac - 1;
  383: 		}
  384: 		if (poll_burst_max < MIN_POLL_BURST_MAX)
  385: 			poll_burst_max = MIN_POLL_BURST_MAX;
  386: 		else if (poll_burst_max > MAX_POLL_BURST_MAX)
  387: 			poll_burst_max = MAX_POLL_BURST_MAX;
  388: 
  389: 		if (poll_each_burst < 1)
  390: 			poll_each_burst = 1;
  391: 		else if (poll_each_burst > poll_burst_max)
  392: 			poll_each_burst = poll_burst_max;
  393: 
  394: 		residual_burst = poll_burst;
  395: 	}
  396: 	cycles = (residual_burst < poll_each_burst) ?
  397: 		residual_burst : poll_each_burst;
  398: 	residual_burst -= cycles;
  399: 
  400: 	if (polling) {
  401: 		for (i = 0 ; i < poll_handlers ; i++)
  402: 			if (pr[i].handler && (IFF_UP|IFF_RUNNING) ==
  403: 			    (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) )
  404: 				pr[i].handler(pr[i].ifp, arg, cycles);
  405: 	} else {	/* unregister */
  406: 		for (i = 0 ; i < poll_handlers ; i++) {
  407: 			if (pr[i].handler &&
  408: 			    pr[i].ifp->if_flags & IFF_RUNNING) {
  409: 				pr[i].ifp->if_ipending &= ~IFF_POLLING;
  410: 				pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1);
  411: 			}
  412: 			pr[i].handler=NULL;
  413: 		}
  414: 		residual_burst = 0;
  415: 		poll_handlers = 0;
  416: 	}
  417: 	schednetisr(NETISR_POLLMORE);
  418: 	phase = 4;
  419: 	splx(s);
  420: 	lwkt_replymsg(&msg->nm_lmsg, 0);
  421: }
  422: 
  423: /*
  424:  * Try to register routine for polling. Returns 1 if successful
  425:  * (and polling should be enabled), 0 otherwise.
  426:  * A device is not supposed to register itself multiple times.
  427:  *
  428:  * This is called from within the *_intr() functions, so we do not need
  429:  * further locking.
  430:  */
  431: int
  432: ether_poll_register(poll_handler_t *h, struct ifnet *ifp)
  433: {
  434: 	int s;
  435: 
  436: 	if (polling == 0) /* polling disabled, cannot register */
  437: 		return 0;
  438: 	if (h == NULL || ifp == NULL)		/* bad arguments	*/
  439: 		return 0;
  440: 	if ( !(ifp->if_flags & IFF_UP) )	/* must be up		*/
  441: 		return 0;
  442: 	if (ifp->if_ipending & IFF_POLLING)	/* already polling	*/
  443: 		return 0;
  444: 
  445: 	s = splhigh();
  446: 	if (poll_handlers >= POLL_LIST_LEN) {
  447: 		/*
  448: 		 * List full, cannot register more entries.
  449: 		 * This should never happen; if it does, it is probably a
  450: 		 * broken driver trying to register multiple times. Checking
  451: 		 * this at runtime is expensive, and won't solve the problem
  452: 		 * anyways, so just report a few times and then give up.
  453: 		 */
  454: 		static int verbose = 10 ;
  455: 		splx(s);
  456: 		if (verbose >0) {
  457: 			printf("poll handlers list full, "
  458: 				"maybe a broken driver ?\n");
  459: 			verbose--;
  460: 		}
  461: 		return 0; /* no polling for you */
  462: 	}
  463: 
  464: 	pr[poll_handlers].handler = h;
  465: 	pr[poll_handlers].ifp = ifp;
  466: 	poll_handlers++;
  467: 	ifp->if_ipending |= IFF_POLLING;
  468: 	splx(s);
  469: 	return 1; /* polling enabled in next call */
  470: }
  471: 
  472: /*
  473:  * Remove interface from the polling list. Normally called by *_stop().
  474:  * It is not an error to call it with IFF_POLLING clear, the call is
  475:  * sufficiently rare to be preferable to save the space for the extra
  476:  * test in each driver in exchange of one additional function call.
  477:  */
  478: int
  479: ether_poll_deregister(struct ifnet *ifp)
  480: {
  481: 	int i;
  482: 	int s = splimp();
  483: 	
  484: 	if ( !ifp || !(ifp->if_ipending & IFF_POLLING) ) {
  485: 		splx(s);
  486: 		return 0;
  487: 	}
  488: 	for (i = 0 ; i < poll_handlers ; i++)
  489: 		if (pr[i].ifp == ifp) /* found it */
  490: 			break;
  491: 	ifp->if_ipending &= ~IFF_POLLING; /* found or not... */
  492: 	if (i == poll_handlers) {
  493: 		splx(s);
  494: 		printf("ether_poll_deregister: ifp not found!!!\n");
  495: 		return 0;
  496: 	}
  497: 	poll_handlers--;
  498: 	if (i < poll_handlers) { /* Last entry replaces this one. */
  499: 		pr[i].handler = pr[poll_handlers].handler;
  500: 		pr[i].ifp = pr[poll_handlers].ifp;
  501: 	}
  502: 	splx(s);
  503: 	return 1;
  504: }