From f8ad22919e217e5aa0f3f7a246fc37aaee182364 Mon Sep 17 00:00:00 2001 From: luigi Date: Fri, 14 Dec 2001 17:56:12 +0000 Subject: Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications --- sys/kern/kern_clock.c | 12 ++ sys/kern/kern_intr.c | 14 ++ sys/kern/kern_poll.c | 422 ++++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 448 insertions(+) create mode 100644 sys/kern/kern_poll.c (limited to 'sys/kern') diff --git a/sys/kern/kern_clock.c b/sys/kern/kern_clock.c index 4c2ccc0..a338dca 100644 --- a/sys/kern/kern_clock.c +++ b/sys/kern/kern_clock.c @@ -69,6 +69,12 @@ #include #endif +#ifdef DEVICE_POLLING +#include /* for NETISR_POLL */ + +extern void ether_poll1(void); +extern void hardclock_device_poll(void); +#endif /* DEVICE_POLLING */ static void initclocks __P((void *dummy)); SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) @@ -140,6 +146,9 @@ initclocks(dummy) psdiv = pscnt = 1; cpu_initclocks(); +#ifdef DEVICE_POLLING + register_netisr(NETISR_POLL, ether_poll1); +#endif /* * Compute profhz/stathz, and fix profhz if needed. */ @@ -212,6 +221,9 @@ hardclock(frame) statclock(frame); tc_windup(); +#ifdef DEVICE_POLLING + hardclock_device_poll(); +#endif /* DEVICE_POLLING */ /* * Process callouts at a very low cpu priority, so we don't keep the diff --git a/sys/kern/kern_intr.c b/sys/kern/kern_intr.c index e873c6e..88868dc 100644 --- a/sys/kern/kern_intr.c +++ b/sys/kern/kern_intr.c @@ -623,7 +623,16 @@ swi_net(void *dummy) u_int bits; int i; +#ifdef DEVICE_POLLING + for (;;) { + int pollmore; +#endif bits = atomic_readandclear_int(&netisr); +#ifdef DEVICE_POLLING + if (bits == 0) + return; + pollmore = bits & (1 << NETISR_POLL); +#endif while ((i = ffs(bits)) != 0) { i--; if (netisrs[i] != NULL) @@ -632,6 +641,11 @@ swi_net(void *dummy) printf("swi_net: unregistered isr number: %d.\n", i); bits &= ~(1 << i); } +#ifdef DEVICE_POLLING + if (pollmore) + ether_pollmore(); + } +#endif } /* diff --git a/sys/kern/kern_poll.c b/sys/kern/kern_poll.c new file mode 100644 index 0000000..763e5cb --- /dev/null +++ b/sys/kern/kern_poll.c @@ -0,0 +1,422 @@ +/*- + * Copyright (c) 2001 Luigi Rizzo + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + * + * $FreeBSD$ + */ + +#include +#include +#include +#include /* needed by net/if.h */ +#include + +#include /* for IFF_* flags */ +#include /* for NETISR_POLL */ + +#ifdef SMP +#error DEVICE_POLLING is not compatible with SMP +#endif + +void ether_poll1(void); +void ether_poll(int); /* polling while in trap */ +void ether_pollmore(void); +void hardclock_device_poll(void); + +/* + * Polling support for [network] device drivers. + * + * Drivers which support this feature try to register with the + * polling code. + * + * If registration is successful, the driver must disable interrupts, + * and further I/O is performed through the handler, which is invoked + * (at least once per clock tick) with 3 arguments: the "arg" passed at + * register time (a struct ifnet pointer), a command, and a "count" limit. + * + * The command can be one of the following: + * POLL_ONLY: quick move of "count" packets from input/output queues. + * POLL_AND_CHECK_STATUS: as above, plus check status registers or do + * other more expensive operations. This command is issued periodically + * but less frequently than POLL_ONLY. + * POLL_DEREGISTER: deregister and return to interrupt mode. + * + * The first two commands are only issued if the interface is marked as + * 'IFF_UP and IFF_RUNNING', the last one only if IFF_RUNNING is set. + * + * The count limit specifies how much work the handler can do during the + * call -- typically this is the number of packets to be received, or + * transmitted, etc. (drivers are free to interpret this number, as long + * as the max time spent in the function grows roughly linearly with the + * count). + * + * Deregistration can be requested by the driver itself (typically in the + * *_stop() routine), or by the polling code, by invoking the handler. + * + * Polling can be globally enabled or disabled with the sysctl variable + * kern.polling.enable (default is 0, disabled) + * + * A second variable controls the sharing of CPU between polling/kernel + * network processing, and other activities (typically userlevel tasks): + * kern.polling.user_frac (between 0 and 100, default 50) sets the share + * of CPU allocated to user tasks. CPU is allocated proportionally to the + * shares, by dynamically adjusting the "count" (poll_burst). + * + * Other parameters can should be left to their default values. + * The following constraints hold + * + * 1 <= poll_each_burst <= poll_burst <= poll_burst_max + * 0 <= poll_in_trap <= poll_each_burst + * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX + */ + +#define MIN_POLL_BURST_MAX 10 +#define MAX_POLL_BURST_MAX 1000 + +SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0, + "Device polling parameters"); + +static u_int32_t poll_burst = 5; +SYSCTL_ULONG(_kern_polling, OID_AUTO, burst, CTLFLAG_RW, + &poll_burst, 0, "Current polling burst size"); + +static u_int32_t poll_each_burst = 5; +SYSCTL_ULONG(_kern_polling, OID_AUTO, each_burst, CTLFLAG_RW, + &poll_each_burst, 0, "Max size of each burst"); + +static u_int32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */ +SYSCTL_ULONG(_kern_polling, OID_AUTO, burst_max, CTLFLAG_RW, + &poll_burst_max, 0, "Max Polling burst size"); + +u_int32_t poll_in_trap; /* used in trap.c */ +SYSCTL_ULONG(_kern_polling, OID_AUTO, poll_in_trap, CTLFLAG_RW, + &poll_in_trap, 0, "Poll burst size during a trap"); + +static u_int32_t user_frac = 50; +SYSCTL_ULONG(_kern_polling, OID_AUTO, user_frac, CTLFLAG_RW, + &user_frac, 0, "Desired user fraction of cpu time"); + +static u_int32_t reg_frac = 20 ; +SYSCTL_ULONG(_kern_polling, OID_AUTO, reg_frac, CTLFLAG_RW, + ®_frac, 0, "Every this many cycles poll register"); + +static u_int32_t short_ticks; +SYSCTL_ULONG(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RW, + &short_ticks, 0, "Hardclock ticks shorter than they should be"); + +static u_int32_t lost_polls; +SYSCTL_ULONG(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RW, + &lost_polls, 0, "How many times we would have lost a poll tick"); + +static u_int32_t poll_handlers; /* next free entry in pr[]. */ +SYSCTL_ULONG(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD, + &poll_handlers, 0, "Number of registered poll handlers"); + +static int polling = 0; /* global polling enable */ +SYSCTL_ULONG(_kern_polling, OID_AUTO, enable, CTLFLAG_RW, + &polling, 0, "Polling enabled"); + + +static u_int32_t poll1_active; +static u_int32_t need_poll_again; + +#define POLL_LIST_LEN 128 +struct pollrec { + poll_handler_t *handler; + struct ifnet *ifp; +}; + +static struct pollrec pr[POLL_LIST_LEN]; + +/* + * Hook from hardclock. Tries to schedule a netisr, but keeps track + * of lost ticks due to the previous handler taking too long. + * The first part of the code is just for debugging purposes, and tries + * to count how often hardclock ticks are shorter than they should, + * meaning either stray interrupts or delayed events. + */ +void +hardclock_device_poll(void) +{ + static struct timeval prev_t, t; + int delta; + + microuptime(&t); + delta = (t.tv_usec - prev_t.tv_usec) + + (t.tv_sec - prev_t.tv_sec)*1000000; + if (delta * hz < 500000) + short_ticks++; + else + prev_t = t; + + if (poll_handlers > 0) { + if (poll1_active) { + lost_polls++; + need_poll_again++; + } else { + poll1_active = 1; + schednetisr(NETISR_POLL); + } + } +} + +/* + * ether_poll is called from the idle loop or from the trap handler. + */ +void +ether_poll(int count) +{ + int i; + int s = splimp(); + + mtx_lock(&Giant); + + if (count > poll_each_burst) + count = poll_each_burst; + for (i = 0 ; i < poll_handlers ; i++) + if (pr[i].handler && (IFF_UP|IFF_RUNNING) == + (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) ) + pr[i].handler(pr[i].ifp, 0, count); /* quick check */ + mtx_unlock(&Giant); + splx(s); +} + +/* + * ether_pollmore is called after other netisr's, possibly scheduling + * another NETISR_POLL call, or adapting the burst size for the next cycle. + * + * It is very bad to fetch large bursts of packets from a single card at once, + * because the burst could take a long time to be completely processed, or + * could saturate the intermediate queue (ipintrq or similar) leading to + * losses or unfairness. To reduce the problem, and also to account better for + * time spent in network-related processnig, we split the burst in smaller + * chunks of fixed size, giving control to the other netisr's between chunks. + * This helps in improving the fairness, reducing livelock (because we + * emulate more closely the "process to completion" that we have with + * fastforwarding) and accounting for the work performed in low level + * handling and forwarding. + */ + +static int residual_burst = 0; + +static struct timeval poll_start_t; + +void +ether_pollmore() +{ + struct timeval t; + int kern_load; + int s = splhigh(); + + if (residual_burst > 0) { + schednetisr(NETISR_POLL); + /* will run immediately on return, followed by netisrs */ + splx(s); + return ; + } + /* here we can account time spent in netisr's in this tick */ + microuptime(&t); + kern_load = (t.tv_usec - poll_start_t.tv_usec) + + (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */ + kern_load = (kern_load * hz) / 10000; /* 0..100 */ + if (kern_load > (100 - user_frac)) { /* try decrease ticks */ + if (poll_burst > 1) + poll_burst--; + } else { + if (poll_burst < poll_burst_max) + poll_burst++; + } + + if (need_poll_again) { + /* + * Last cycle was long and caused us to miss one or more + * hardclock ticks. Restart processnig again, but slightly + * reduce the burst size to prevent that this happens again. + */ + need_poll_again--; + poll_burst -= (poll_burst / 8); + if (poll_burst < 1) + poll_burst = 1; + schednetisr(NETISR_POLL); + } else + poll1_active = 0; + splx(s); +} + +/* + * ether_poll1 is called by schednetisr when appropriate, typically once + * per tick. It is called at splnet() so first thing to do is to upgrade to + * splimp(), and call all registered handlers. + */ +void +ether_poll1(void) +{ + static int reg_frac_count; + int i, cycles; + enum poll_cmd arg = POLL_ONLY; + int s=splimp(); + mtx_lock(&Giant); + + if (residual_burst == 0) { /* first call in this tick */ + microuptime(&poll_start_t); + /* + * Check that paremeters are consistent with runtime + * variables. Some of these tests could be done at sysctl + * time, but the savings would be very limited because we + * still have to check against reg_frac_count and + * poll_each_burst. So, instead of writing separate sysctl + * handlers, we do all here. + */ + + if (reg_frac > hz) + reg_frac = hz; + else if (reg_frac < 1) + reg_frac = 1; + if (reg_frac_count > reg_frac) + reg_frac_count = reg_frac - 1; + if (reg_frac_count-- == 0) { + arg = POLL_AND_CHECK_STATUS; + reg_frac_count = reg_frac - 1; + } + if (poll_burst_max < MIN_POLL_BURST_MAX) + poll_burst_max = MIN_POLL_BURST_MAX; + else if (poll_burst_max > MAX_POLL_BURST_MAX) + poll_burst_max = MAX_POLL_BURST_MAX; + + if (poll_each_burst < 1) + poll_each_burst = 1; + else if (poll_each_burst > poll_burst_max) + poll_each_burst = poll_burst_max; + + residual_burst = poll_burst; + } + cycles = (residual_burst < poll_each_burst) ? + residual_burst : poll_each_burst; + residual_burst -= cycles; + + if (polling) { + for (i = 0 ; i < poll_handlers ; i++) + if (pr[i].handler && (IFF_UP|IFF_RUNNING) == + (pr[i].ifp->if_flags & (IFF_UP|IFF_RUNNING)) ) + pr[i].handler(pr[i].ifp, arg, cycles); + } else { /* unregister */ + for (i = 0 ; i < poll_handlers ; i++) { + if (pr[i].handler && + pr[i].ifp->if_flags & IFF_RUNNING) { + pr[i].ifp->if_ipending &= ~IFF_POLLING; + pr[i].handler(pr[i].ifp, POLL_DEREGISTER, 1); + } + pr[i].handler=NULL; + } + residual_burst = 0; + poll_handlers = 0; + } + /* on -stable, schednetisr(NETISR_POLLMORE); */ + mtx_unlock(&Giant); + splx(s); +} + +/* + * Try to register routine for polling. Returns 1 if successful + * (and polling should be enabled), 0 otherwise. + * A device is not supposed to register itself multiple times. + * + * This is called from within the *_intr() function, so we should + * probably not need further locking. XXX + */ +int +ether_poll_register(poll_handler_t *h, struct ifnet *ifp) +{ + int s; + + if (polling == 0) /* polling disabled, cannot register */ + return 0; + if (h == NULL || ifp == NULL) /* bad arguments */ + return 0; + if ( !(ifp->if_flags & IFF_UP) ) /* must be up */ + return 0; + if (ifp->if_ipending & IFF_POLLING) /* already polling */ + return 0; + + s = splhigh(); + if (poll_handlers >= POLL_LIST_LEN) { + /* + * List full, cannot register more entries. + * This should never happen; if it does, it is probably a + * broken driver trying to register multiple times. Checking + * this at runtime is expensive, and won't solve the problem + * anyways, so just report a few times and then give up. + */ + static int verbose = 10 ; + splx(s); + if (verbose >0) { + printf("poll handlers list full, " + "maybe a broken driver ?\n"); + verbose--; + } + return 0; /* no polling for you */ + } + + pr[poll_handlers].handler = h; + pr[poll_handlers].ifp = ifp; + poll_handlers++; + ifp->if_ipending |= IFF_POLLING; + splx(s); + return 1; /* polling enabled in next call */ +} + +/* + * Remove the interface from the list of polling ones. + * Normally run by *_stop(). + * We allow it being called with IFF_POLLING clear, the + * call is sufficiently rare so it is preferable to save the + * space for the extra test in each device in exchange of one + * additional function call. + */ +int +ether_poll_deregister(struct ifnet *ifp) +{ + int i; + + mtx_lock(&Giant); + if ( !ifp || !(ifp->if_ipending & IFF_POLLING) ) { + mtx_unlock(&Giant); + return 0; + } + for (i = 0 ; i < poll_handlers ; i++) + if (pr[i].ifp == ifp) /* found it */ + break; + ifp->if_ipending &= ~IFF_POLLING; /* found or not... */ + if (i == poll_handlers) { + mtx_unlock(&Giant); + printf("ether_poll_deregister: ifp not found!!!\n"); + return 0; + } + poll_handlers--; + if (i < poll_handlers) { /* Last entry replaces this one. */ + pr[i].handler = pr[poll_handlers].handler; + pr[i].ifp = pr[poll_handlers].ifp; + } + mtx_unlock(&Giant); + return 1; +} -- cgit v1.1