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 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_device_polling.h" 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/kernel.h> 36 #include <sys/kthread.h> 37 #include <sys/proc.h> 38 #include <sys/eventhandler.h> 39 #include <sys/resourcevar.h> 40 #include <sys/socket.h> /* needed by net/if.h */ 41 #include <sys/sockio.h> 42 #include <sys/sysctl.h> 43 #include <sys/syslog.h> 44 #include <sys/vimage.h> 45 46 #include <net/if.h> /* for IFF_* flags */ 47 #include <net/netisr.h> /* for NETISR_POLL */ 48 #include <net/vnet.h> 49 50 static int poll_switch(SYSCTL_HANDLER_ARGS); 51 52 void hardclock_device_poll(void); /* hook from hardclock */ 53 54 static struct mtx poll_mtx; 55 56 /* 57 * Polling support for [network] device drivers. 58 * 59 * Drivers which support this feature can register with the 60 * polling code. 61 * 62 * If registration is successful, the driver must disable interrupts, 63 * and further I/O is performed through the handler, which is invoked 64 * (at least once per clock tick) with 3 arguments: the "arg" passed at 65 * register time (a struct ifnet pointer), a command, and a "count" limit. 66 * 67 * The command can be one of the following: 68 * POLL_ONLY: quick move of "count" packets from input/output queues. 69 * POLL_AND_CHECK_STATUS: as above, plus check status registers or do 70 * other more expensive operations. This command is issued periodically 71 * but less frequently than POLL_ONLY. 72 * 73 * The count limit specifies how much work the handler can do during the 74 * call -- typically this is the number of packets to be received, or 75 * transmitted, etc. (drivers are free to interpret this number, as long 76 * as the max time spent in the function grows roughly linearly with the 77 * count). 78 * 79 * Polling is enabled and disabled via setting IFCAP_POLLING flag on 80 * the interface. The driver ioctl handler should register interface 81 * with polling and disable interrupts, if registration was successful. 82 * 83 * A second variable controls the sharing of CPU between polling/kernel 84 * network processing, and other activities (typically userlevel tasks): 85 * kern.polling.user_frac (between 0 and 100, default 50) sets the share 86 * of CPU allocated to user tasks. CPU is allocated proportionally to the 87 * shares, by dynamically adjusting the "count" (poll_burst). 88 * 89 * Other parameters can should be left to their default values. 90 * The following constraints hold 91 * 92 * 1 <= poll_each_burst <= poll_burst <= poll_burst_max 93 * 0 <= poll_each_burst 94 * MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX 95 */ 96 97 #define MIN_POLL_BURST_MAX 10 98 #define MAX_POLL_BURST_MAX 1000 99 100 static uint32_t poll_burst = 5; 101 static uint32_t poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */ 102 static uint32_t poll_each_burst = 5; 103 104 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0, 105 "Device polling parameters"); 106 107 SYSCTL_UINT(_kern_polling, OID_AUTO, burst, CTLFLAG_RD, 108 &poll_burst, 0, "Current polling burst size"); 109 110 static int netisr_poll_scheduled; 111 static int netisr_pollmore_scheduled; 112 static int poll_shutting_down; 113 114 static int poll_burst_max_sysctl(SYSCTL_HANDLER_ARGS) 115 { 116 uint32_t val = poll_burst_max; 117 int error; 118 119 error = sysctl_handle_int(oidp, &val, 0, req); 120 if (error || !req->newptr ) 121 return (error); 122 if (val < MIN_POLL_BURST_MAX || val > MAX_POLL_BURST_MAX) 123 return (EINVAL); 124 125 mtx_lock(&poll_mtx); 126 poll_burst_max = val; 127 if (poll_burst > poll_burst_max) 128 poll_burst = poll_burst_max; 129 if (poll_each_burst > poll_burst_max) 130 poll_each_burst = MIN_POLL_BURST_MAX; 131 mtx_unlock(&poll_mtx); 132 133 return (0); 134 } 135 SYSCTL_PROC(_kern_polling, OID_AUTO, burst_max, CTLTYPE_UINT | CTLFLAG_RW, 136 0, sizeof(uint32_t), poll_burst_max_sysctl, "I", "Max Polling burst size"); 137 138 static int poll_each_burst_sysctl(SYSCTL_HANDLER_ARGS) 139 { 140 uint32_t val = poll_each_burst; 141 int error; 142 143 error = sysctl_handle_int(oidp, &val, 0, req); 144 if (error || !req->newptr ) 145 return (error); 146 if (val < 1) 147 return (EINVAL); 148 149 mtx_lock(&poll_mtx); 150 if (val > poll_burst_max) { 151 mtx_unlock(&poll_mtx); 152 return (EINVAL); 153 } 154 poll_each_burst = val; 155 mtx_unlock(&poll_mtx); 156 157 return (0); 158 } 159 SYSCTL_PROC(_kern_polling, OID_AUTO, each_burst, CTLTYPE_UINT | CTLFLAG_RW, 160 0, sizeof(uint32_t), poll_each_burst_sysctl, "I", 161 "Max size of each burst"); 162 163 static uint32_t poll_in_idle_loop=0; /* do we poll in idle loop ? */ 164 SYSCTL_UINT(_kern_polling, OID_AUTO, idle_poll, CTLFLAG_RW, 165 &poll_in_idle_loop, 0, "Enable device polling in idle loop"); 166 167 static uint32_t user_frac = 50; 168 static int user_frac_sysctl(SYSCTL_HANDLER_ARGS) 169 { 170 uint32_t val = user_frac; 171 int error; 172 173 error = sysctl_handle_int(oidp, &val, 0, req); 174 if (error || !req->newptr ) 175 return (error); 176 if (val < 0 || val > 99) 177 return (EINVAL); 178 179 mtx_lock(&poll_mtx); 180 user_frac = val; 181 mtx_unlock(&poll_mtx); 182 183 return (0); 184 } 185 SYSCTL_PROC(_kern_polling, OID_AUTO, user_frac, CTLTYPE_UINT | CTLFLAG_RW, 186 0, sizeof(uint32_t), user_frac_sysctl, "I", 187 "Desired user fraction of cpu time"); 188 189 static uint32_t reg_frac_count = 0; 190 static uint32_t reg_frac = 20 ; 191 static int reg_frac_sysctl(SYSCTL_HANDLER_ARGS) 192 { 193 uint32_t val = reg_frac; 194 int error; 195 196 error = sysctl_handle_int(oidp, &val, 0, req); 197 if (error || !req->newptr ) 198 return (error); 199 if (val < 1 || val > hz) 200 return (EINVAL); 201 202 mtx_lock(&poll_mtx); 203 reg_frac = val; 204 if (reg_frac_count >= reg_frac) 205 reg_frac_count = 0; 206 mtx_unlock(&poll_mtx); 207 208 return (0); 209 } 210 SYSCTL_PROC(_kern_polling, OID_AUTO, reg_frac, CTLTYPE_UINT | CTLFLAG_RW, 211 0, sizeof(uint32_t), reg_frac_sysctl, "I", 212 "Every this many cycles check registers"); 213 214 static uint32_t short_ticks; 215 SYSCTL_UINT(_kern_polling, OID_AUTO, short_ticks, CTLFLAG_RD, 216 &short_ticks, 0, "Hardclock ticks shorter than they should be"); 217 218 static uint32_t lost_polls; 219 SYSCTL_UINT(_kern_polling, OID_AUTO, lost_polls, CTLFLAG_RD, 220 &lost_polls, 0, "How many times we would have lost a poll tick"); 221 222 static uint32_t pending_polls; 223 SYSCTL_UINT(_kern_polling, OID_AUTO, pending_polls, CTLFLAG_RD, 224 &pending_polls, 0, "Do we need to poll again"); 225 226 static int residual_burst = 0; 227 SYSCTL_INT(_kern_polling, OID_AUTO, residual_burst, CTLFLAG_RD, 228 &residual_burst, 0, "# of residual cycles in burst"); 229 230 static uint32_t poll_handlers; /* next free entry in pr[]. */ 231 SYSCTL_UINT(_kern_polling, OID_AUTO, handlers, CTLFLAG_RD, 232 &poll_handlers, 0, "Number of registered poll handlers"); 233 234 static int polling = 0; 235 SYSCTL_PROC(_kern_polling, OID_AUTO, enable, CTLTYPE_UINT | CTLFLAG_RW, 236 0, sizeof(int), poll_switch, "I", "Switch polling for all interfaces"); 237 238 static uint32_t phase; 239 SYSCTL_UINT(_kern_polling, OID_AUTO, phase, CTLFLAG_RD, 240 &phase, 0, "Polling phase"); 241 242 static uint32_t suspect; 243 SYSCTL_UINT(_kern_polling, OID_AUTO, suspect, CTLFLAG_RD, 244 &suspect, 0, "suspect event"); 245 246 static uint32_t stalled; 247 SYSCTL_UINT(_kern_polling, OID_AUTO, stalled, CTLFLAG_RD, 248 &stalled, 0, "potential stalls"); 249 250 static uint32_t idlepoll_sleeping; /* idlepoll is sleeping */ 251 SYSCTL_UINT(_kern_polling, OID_AUTO, idlepoll_sleeping, CTLFLAG_RD, 252 &idlepoll_sleeping, 0, "idlepoll is sleeping"); 253 254 255 #define POLL_LIST_LEN 128 256 struct pollrec { 257 poll_handler_t *handler; 258 struct ifnet *ifp; 259 }; 260 261 static struct pollrec pr[POLL_LIST_LEN]; 262 263 static void 264 poll_shutdown(void *arg, int howto) 265 { 266 267 poll_shutting_down = 1; 268 } 269 270 static void 271 init_device_poll(void) 272 { 273 274 mtx_init(&poll_mtx, "polling", NULL, MTX_DEF); 275 EVENTHANDLER_REGISTER(shutdown_post_sync, poll_shutdown, NULL, 276 SHUTDOWN_PRI_LAST); 277 } 278 SYSINIT(device_poll, SI_SUB_CLOCKS, SI_ORDER_MIDDLE, init_device_poll, NULL); 279 280 281 /* 282 * Hook from hardclock. Tries to schedule a netisr, but keeps track 283 * of lost ticks due to the previous handler taking too long. 284 * Normally, this should not happen, because polling handler should 285 * run for a short time. However, in some cases (e.g. when there are 286 * changes in link status etc.) the drivers take a very long time 287 * (even in the order of milliseconds) to reset and reconfigure the 288 * device, causing apparent lost polls. 289 * 290 * The first part of the code is just for debugging purposes, and tries 291 * to count how often hardclock ticks are shorter than they should, 292 * meaning either stray interrupts or delayed events. 293 */ 294 void 295 hardclock_device_poll(void) 296 { 297 static struct timeval prev_t, t; 298 int delta; 299 300 if (poll_handlers == 0 || poll_shutting_down) 301 return; 302 303 microuptime(&t); 304 delta = (t.tv_usec - prev_t.tv_usec) + 305 (t.tv_sec - prev_t.tv_sec)*1000000; 306 if (delta * hz < 500000) 307 short_ticks++; 308 else 309 prev_t = t; 310 311 if (pending_polls > 100) { 312 /* 313 * Too much, assume it has stalled (not always true 314 * see comment above). 315 */ 316 stalled++; 317 pending_polls = 0; 318 phase = 0; 319 } 320 321 if (phase <= 2) { 322 if (phase != 0) 323 suspect++; 324 phase = 1; 325 netisr_poll_scheduled = 1; 326 netisr_pollmore_scheduled = 1; 327 netisr_sched_poll(); 328 phase = 2; 329 } 330 if (pending_polls++ > 0) 331 lost_polls++; 332 } 333 334 /* 335 * ether_poll is called from the idle loop. 336 */ 337 static void 338 ether_poll(int count) 339 { 340 int i; 341 342 mtx_lock(&poll_mtx); 343 344 if (count > poll_each_burst) 345 count = poll_each_burst; 346 347 for (i = 0 ; i < poll_handlers ; i++) 348 pr[i].handler(pr[i].ifp, POLL_ONLY, count); 349 350 mtx_unlock(&poll_mtx); 351 } 352 353 /* 354 * netisr_pollmore is called after other netisr's, possibly scheduling 355 * another NETISR_POLL call, or adapting the burst size for the next cycle. 356 * 357 * It is very bad to fetch large bursts of packets from a single card at once, 358 * because the burst could take a long time to be completely processed, or 359 * could saturate the intermediate queue (ipintrq or similar) leading to 360 * losses or unfairness. To reduce the problem, and also to account better for 361 * time spent in network-related processing, we split the burst in smaller 362 * chunks of fixed size, giving control to the other netisr's between chunks. 363 * This helps in improving the fairness, reducing livelock (because we 364 * emulate more closely the "process to completion" that we have with 365 * fastforwarding) and accounting for the work performed in low level 366 * handling and forwarding. 367 */ 368 369 static struct timeval poll_start_t; 370 371 void 372 netisr_pollmore() 373 { 374 struct timeval t; 375 int kern_load; 376 377 mtx_lock(&poll_mtx); 378 if (!netisr_pollmore_scheduled) { 379 mtx_unlock(&poll_mtx); 380 return; 381 } 382 netisr_pollmore_scheduled = 0; 383 phase = 5; 384 if (residual_burst > 0) { 385 netisr_poll_scheduled = 1; 386 netisr_pollmore_scheduled = 1; 387 netisr_sched_poll(); 388 mtx_unlock(&poll_mtx); 389 /* will run immediately on return, followed by netisrs */ 390 return; 391 } 392 /* here we can account time spent in netisr's in this tick */ 393 microuptime(&t); 394 kern_load = (t.tv_usec - poll_start_t.tv_usec) + 395 (t.tv_sec - poll_start_t.tv_sec)*1000000; /* us */ 396 kern_load = (kern_load * hz) / 10000; /* 0..100 */ 397 if (kern_load > (100 - user_frac)) { /* try decrease ticks */ 398 if (poll_burst > 1) 399 poll_burst--; 400 } else { 401 if (poll_burst < poll_burst_max) 402 poll_burst++; 403 } 404 405 pending_polls--; 406 if (pending_polls == 0) /* we are done */ 407 phase = 0; 408 else { 409 /* 410 * Last cycle was long and caused us to miss one or more 411 * hardclock ticks. Restart processing again, but slightly 412 * reduce the burst size to prevent that this happens again. 413 */ 414 poll_burst -= (poll_burst / 8); 415 if (poll_burst < 1) 416 poll_burst = 1; 417 netisr_poll_scheduled = 1; 418 netisr_pollmore_scheduled = 1; 419 netisr_sched_poll(); 420 phase = 6; 421 } 422 mtx_unlock(&poll_mtx); 423 } 424 425 /* 426 * netisr_poll is typically scheduled once per tick. 427 */ 428 void 429 netisr_poll(void) 430 { 431 int i, cycles; 432 enum poll_cmd arg = POLL_ONLY; 433 434 mtx_lock(&poll_mtx); 435 if (!netisr_poll_scheduled) { 436 mtx_unlock(&poll_mtx); 437 return; 438 } 439 netisr_poll_scheduled = 0; 440 phase = 3; 441 if (residual_burst == 0) { /* first call in this tick */ 442 microuptime(&poll_start_t); 443 if (++reg_frac_count == reg_frac) { 444 arg = POLL_AND_CHECK_STATUS; 445 reg_frac_count = 0; 446 } 447 448 residual_burst = poll_burst; 449 } 450 cycles = (residual_burst < poll_each_burst) ? 451 residual_burst : poll_each_burst; 452 residual_burst -= cycles; 453 454 for (i = 0 ; i < poll_handlers ; i++) 455 pr[i].handler(pr[i].ifp, arg, cycles); 456 457 phase = 4; 458 mtx_unlock(&poll_mtx); 459 } 460 461 /* 462 * Try to register routine for polling. Returns 0 if successful 463 * (and polling should be enabled), error code otherwise. 464 * A device is not supposed to register itself multiple times. 465 * 466 * This is called from within the *_ioctl() functions. 467 */ 468 int 469 ether_poll_register(poll_handler_t *h, struct ifnet *ifp) 470 { 471 int i; 472 473 KASSERT(h != NULL, ("%s: handler is NULL", __func__)); 474 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__)); 475 476 mtx_lock(&poll_mtx); 477 if (poll_handlers >= POLL_LIST_LEN) { 478 /* 479 * List full, cannot register more entries. 480 * This should never happen; if it does, it is probably a 481 * broken driver trying to register multiple times. Checking 482 * this at runtime is expensive, and won't solve the problem 483 * anyways, so just report a few times and then give up. 484 */ 485 static int verbose = 10 ; 486 if (verbose >0) { 487 log(LOG_ERR, "poll handlers list full, " 488 "maybe a broken driver ?\n"); 489 verbose--; 490 } 491 mtx_unlock(&poll_mtx); 492 return (ENOMEM); /* no polling for you */ 493 } 494 495 for (i = 0 ; i < poll_handlers ; i++) 496 if (pr[i].ifp == ifp && pr[i].handler != NULL) { 497 mtx_unlock(&poll_mtx); 498 log(LOG_DEBUG, "ether_poll_register: %s: handler" 499 " already registered\n", ifp->if_xname); 500 return (EEXIST); 501 } 502 503 pr[poll_handlers].handler = h; 504 pr[poll_handlers].ifp = ifp; 505 poll_handlers++; 506 mtx_unlock(&poll_mtx); 507 if (idlepoll_sleeping) 508 wakeup(&idlepoll_sleeping); 509 return (0); 510 } 511 512 /* 513 * Remove interface from the polling list. Called from *_ioctl(), too. 514 */ 515 int 516 ether_poll_deregister(struct ifnet *ifp) 517 { 518 int i; 519 520 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__)); 521 522 mtx_lock(&poll_mtx); 523 524 for (i = 0 ; i < poll_handlers ; i++) 525 if (pr[i].ifp == ifp) /* found it */ 526 break; 527 if (i == poll_handlers) { 528 log(LOG_DEBUG, "ether_poll_deregister: %s: not found!\n", 529 ifp->if_xname); 530 mtx_unlock(&poll_mtx); 531 return (ENOENT); 532 } 533 poll_handlers--; 534 if (i < poll_handlers) { /* Last entry replaces this one. */ 535 pr[i].handler = pr[poll_handlers].handler; 536 pr[i].ifp = pr[poll_handlers].ifp; 537 } 538 mtx_unlock(&poll_mtx); 539 return (0); 540 } 541 542 /* 543 * Legacy interface for turning polling on all interfaces at one time. 544 */ 545 static int 546 poll_switch(SYSCTL_HANDLER_ARGS) 547 { 548 INIT_VNET_NET(curvnet); 549 struct ifnet *ifp; 550 int error; 551 int val = polling; 552 553 error = sysctl_handle_int(oidp, &val, 0, req); 554 if (error || !req->newptr ) 555 return (error); 556 557 if (val == polling) 558 return (0); 559 560 if (val < 0 || val > 1) 561 return (EINVAL); 562 563 polling = val; 564 565 IFNET_RLOCK(); 566 TAILQ_FOREACH(ifp, &V_ifnet, if_link) { 567 if (ifp->if_capabilities & IFCAP_POLLING) { 568 struct ifreq ifr; 569 570 if (val == 1) 571 ifr.ifr_reqcap = 572 ifp->if_capenable | IFCAP_POLLING; 573 else 574 ifr.ifr_reqcap = 575 ifp->if_capenable & ~IFCAP_POLLING; 576 (void) (*ifp->if_ioctl)(ifp, SIOCSIFCAP, (caddr_t)&ifr); 577 } 578 } 579 IFNET_RUNLOCK(); 580 581 log(LOG_ERR, "kern.polling.enable is deprecated. Use ifconfig(8)"); 582 583 return (0); 584 } 585 586 static void 587 poll_idle(void) 588 { 589 struct thread *td = curthread; 590 struct rtprio rtp; 591 592 rtp.prio = RTP_PRIO_MAX; /* lowest priority */ 593 rtp.type = RTP_PRIO_IDLE; 594 PROC_SLOCK(td->td_proc); 595 rtp_to_pri(&rtp, td); 596 PROC_SUNLOCK(td->td_proc); 597 598 for (;;) { 599 if (poll_in_idle_loop && poll_handlers > 0) { 600 idlepoll_sleeping = 0; 601 ether_poll(poll_each_burst); 602 thread_lock(td); 603 mi_switch(SW_VOL, NULL); 604 thread_unlock(td); 605 } else { 606 idlepoll_sleeping = 1; 607 tsleep(&idlepoll_sleeping, 0, "pollid", hz * 3); 608 } 609 } 610 } 611 612 static struct proc *idlepoll; 613 static struct kproc_desc idlepoll_kp = { 614 "idlepoll", 615 poll_idle, 616 &idlepoll 617 }; 618 SYSINIT(idlepoll, SI_SUB_KTHREAD_VM, SI_ORDER_ANY, kproc_start, 619 &idlepoll_kp); 620