kern_cpuset.c revision 192895
1/*- 2 * Copyright (c) 2008, Jeffrey Roberson <jeff@freebsd.org> 3 * All rights reserved. 4 * 5 * Copyright (c) 2008 Nokia Corporation 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice unmodified, this list of conditions, and the following 13 * disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 * 29 */ 30 31#include <sys/cdefs.h> 32__FBSDID("$FreeBSD: head/sys/kern/kern_cpuset.c 192895 2009-05-27 14:11:23Z jamie $"); 33 34#include "opt_ddb.h" 35 36#include <sys/param.h> 37#include <sys/systm.h> 38#include <sys/sysproto.h> 39#include <sys/jail.h> 40#include <sys/kernel.h> 41#include <sys/lock.h> 42#include <sys/malloc.h> 43#include <sys/mutex.h> 44#include <sys/priv.h> 45#include <sys/proc.h> 46#include <sys/refcount.h> 47#include <sys/sched.h> 48#include <sys/smp.h> 49#include <sys/syscallsubr.h> 50#include <sys/cpuset.h> 51#include <sys/sx.h> 52#include <sys/refcount.h> 53#include <sys/queue.h> 54#include <sys/limits.h> 55#include <sys/bus.h> 56#include <sys/interrupt.h> 57 58#include <vm/uma.h> 59 60#ifdef DDB 61#include <ddb/ddb.h> 62#endif /* DDB */ 63 64/* 65 * cpusets provide a mechanism for creating and manipulating sets of 66 * processors for the purpose of constraining the scheduling of threads to 67 * specific processors. 68 * 69 * Each process belongs to an identified set, by default this is set 1. Each 70 * thread may further restrict the cpus it may run on to a subset of this 71 * named set. This creates an anonymous set which other threads and processes 72 * may not join by number. 73 * 74 * The named set is referred to herein as the 'base' set to avoid ambiguity. 75 * This set is usually a child of a 'root' set while the anonymous set may 76 * simply be referred to as a mask. In the syscall api these are referred to 77 * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here. 78 * 79 * Threads inherit their set from their creator whether it be anonymous or 80 * not. This means that anonymous sets are immutable because they may be 81 * shared. To modify an anonymous set a new set is created with the desired 82 * mask and the same parent as the existing anonymous set. This gives the 83 * illusion of each thread having a private mask.A 84 * 85 * Via the syscall apis a user may ask to retrieve or modify the root, base, 86 * or mask that is discovered via a pid, tid, or setid. Modifying a set 87 * modifies all numbered and anonymous child sets to comply with the new mask. 88 * Modifying a pid or tid's mask applies only to that tid but must still 89 * exist within the assigned parent set. 90 * 91 * A thread may not be assigned to a a group seperate from other threads in 92 * the process. This is to remove ambiguity when the setid is queried with 93 * a pid argument. There is no other technical limitation. 94 * 95 * This somewhat complex arrangement is intended to make it easy for 96 * applications to query available processors and bind their threads to 97 * specific processors while also allowing administrators to dynamically 98 * reprovision by changing sets which apply to groups of processes. 99 * 100 * A simple application should not concern itself with sets at all and 101 * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id 102 * meaning 'curthread'. It may query availble cpus for that tid with a 103 * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...). 104 */ 105static uma_zone_t cpuset_zone; 106static struct mtx cpuset_lock; 107static struct setlist cpuset_ids; 108static struct unrhdr *cpuset_unr; 109static struct cpuset *cpuset_zero; 110 111cpuset_t *cpuset_root; 112 113/* 114 * Acquire a reference to a cpuset, all pointers must be tracked with refs. 115 */ 116struct cpuset * 117cpuset_ref(struct cpuset *set) 118{ 119 120 refcount_acquire(&set->cs_ref); 121 return (set); 122} 123 124/* 125 * Walks up the tree from 'set' to find the root. Returns the root 126 * referenced. 127 */ 128static struct cpuset * 129cpuset_refroot(struct cpuset *set) 130{ 131 132 for (; set->cs_parent != NULL; set = set->cs_parent) 133 if (set->cs_flags & CPU_SET_ROOT) 134 break; 135 cpuset_ref(set); 136 137 return (set); 138} 139 140/* 141 * Find the first non-anonymous set starting from 'set'. Returns this set 142 * referenced. May return the passed in set with an extra ref if it is 143 * not anonymous. 144 */ 145static struct cpuset * 146cpuset_refbase(struct cpuset *set) 147{ 148 149 if (set->cs_id == CPUSET_INVALID) 150 set = set->cs_parent; 151 cpuset_ref(set); 152 153 return (set); 154} 155 156/* 157 * Release a reference in a context where it is safe to allocte. 158 */ 159void 160cpuset_rel(struct cpuset *set) 161{ 162 cpusetid_t id; 163 164 if (refcount_release(&set->cs_ref) == 0) 165 return; 166 mtx_lock_spin(&cpuset_lock); 167 LIST_REMOVE(set, cs_siblings); 168 id = set->cs_id; 169 if (id != CPUSET_INVALID) 170 LIST_REMOVE(set, cs_link); 171 mtx_unlock_spin(&cpuset_lock); 172 cpuset_rel(set->cs_parent); 173 uma_zfree(cpuset_zone, set); 174 if (id != CPUSET_INVALID) 175 free_unr(cpuset_unr, id); 176} 177 178/* 179 * Deferred release must be used when in a context that is not safe to 180 * allocate/free. This places any unreferenced sets on the list 'head'. 181 */ 182static void 183cpuset_rel_defer(struct setlist *head, struct cpuset *set) 184{ 185 186 if (refcount_release(&set->cs_ref) == 0) 187 return; 188 mtx_lock_spin(&cpuset_lock); 189 LIST_REMOVE(set, cs_siblings); 190 if (set->cs_id != CPUSET_INVALID) 191 LIST_REMOVE(set, cs_link); 192 LIST_INSERT_HEAD(head, set, cs_link); 193 mtx_unlock_spin(&cpuset_lock); 194} 195 196/* 197 * Complete a deferred release. Removes the set from the list provided to 198 * cpuset_rel_defer. 199 */ 200static void 201cpuset_rel_complete(struct cpuset *set) 202{ 203 LIST_REMOVE(set, cs_link); 204 cpuset_rel(set->cs_parent); 205 uma_zfree(cpuset_zone, set); 206} 207 208/* 209 * Find a set based on an id. Returns it with a ref. 210 */ 211static struct cpuset * 212cpuset_lookup(cpusetid_t setid, struct thread *td) 213{ 214 struct cpuset *set; 215 216 if (setid == CPUSET_INVALID) 217 return (NULL); 218 mtx_lock_spin(&cpuset_lock); 219 LIST_FOREACH(set, &cpuset_ids, cs_link) 220 if (set->cs_id == setid) 221 break; 222 if (set) 223 cpuset_ref(set); 224 mtx_unlock_spin(&cpuset_lock); 225 226 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__)); 227 if (set != NULL && jailed(td->td_ucred)) { 228 struct cpuset *jset, *tset; 229 230 jset = td->td_ucred->cr_prison->pr_cpuset; 231 for (tset = set; tset != NULL; tset = tset->cs_parent) 232 if (tset == jset) 233 break; 234 if (tset == NULL) { 235 cpuset_rel(set); 236 set = NULL; 237 } 238 } 239 240 return (set); 241} 242 243/* 244 * Create a set in the space provided in 'set' with the provided parameters. 245 * The set is returned with a single ref. May return EDEADLK if the set 246 * will have no valid cpu based on restrictions from the parent. 247 */ 248static int 249_cpuset_create(struct cpuset *set, struct cpuset *parent, cpuset_t *mask, 250 cpusetid_t id) 251{ 252 253 if (!CPU_OVERLAP(&parent->cs_mask, mask)) 254 return (EDEADLK); 255 CPU_COPY(mask, &set->cs_mask); 256 LIST_INIT(&set->cs_children); 257 refcount_init(&set->cs_ref, 1); 258 set->cs_flags = 0; 259 mtx_lock_spin(&cpuset_lock); 260 CPU_AND(mask, &parent->cs_mask); 261 set->cs_id = id; 262 set->cs_parent = cpuset_ref(parent); 263 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings); 264 if (set->cs_id != CPUSET_INVALID) 265 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 266 mtx_unlock_spin(&cpuset_lock); 267 268 return (0); 269} 270 271/* 272 * Create a new non-anonymous set with the requested parent and mask. May 273 * return failures if the mask is invalid or a new number can not be 274 * allocated. 275 */ 276static int 277cpuset_create(struct cpuset **setp, struct cpuset *parent, cpuset_t *mask) 278{ 279 struct cpuset *set; 280 cpusetid_t id; 281 int error; 282 283 id = alloc_unr(cpuset_unr); 284 if (id == -1) 285 return (ENFILE); 286 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK); 287 error = _cpuset_create(set, parent, mask, id); 288 if (error == 0) 289 return (0); 290 free_unr(cpuset_unr, id); 291 uma_zfree(cpuset_zone, set); 292 293 return (error); 294} 295 296/* 297 * Recursively check for errors that would occur from applying mask to 298 * the tree of sets starting at 'set'. Checks for sets that would become 299 * empty as well as RDONLY flags. 300 */ 301static int 302cpuset_testupdate(struct cpuset *set, cpuset_t *mask) 303{ 304 struct cpuset *nset; 305 cpuset_t newmask; 306 int error; 307 308 mtx_assert(&cpuset_lock, MA_OWNED); 309 if (set->cs_flags & CPU_SET_RDONLY) 310 return (EPERM); 311 if (!CPU_OVERLAP(&set->cs_mask, mask)) 312 return (EDEADLK); 313 CPU_COPY(&set->cs_mask, &newmask); 314 CPU_AND(&newmask, mask); 315 error = 0; 316 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 317 if ((error = cpuset_testupdate(nset, &newmask)) != 0) 318 break; 319 return (error); 320} 321 322/* 323 * Applies the mask 'mask' without checking for empty sets or permissions. 324 */ 325static void 326cpuset_update(struct cpuset *set, cpuset_t *mask) 327{ 328 struct cpuset *nset; 329 330 mtx_assert(&cpuset_lock, MA_OWNED); 331 CPU_AND(&set->cs_mask, mask); 332 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 333 cpuset_update(nset, &set->cs_mask); 334 335 return; 336} 337 338/* 339 * Modify the set 'set' to use a copy of the mask provided. Apply this new 340 * mask to restrict all children in the tree. Checks for validity before 341 * applying the changes. 342 */ 343static int 344cpuset_modify(struct cpuset *set, cpuset_t *mask) 345{ 346 struct cpuset *root; 347 int error; 348 349 error = priv_check(curthread, PRIV_SCHED_CPUSET); 350 if (error) 351 return (error); 352 /* 353 * In case we are called from within the jail 354 * we do not allow modifying the dedicated root 355 * cpuset of the jail but may still allow to 356 * change child sets. 357 */ 358 if (jailed(curthread->td_ucred) && 359 set->cs_flags & CPU_SET_ROOT) 360 return (EPERM); 361 /* 362 * Verify that we have access to this set of 363 * cpus. 364 */ 365 root = set->cs_parent; 366 if (root && !CPU_SUBSET(&root->cs_mask, mask)) 367 return (EINVAL); 368 mtx_lock_spin(&cpuset_lock); 369 error = cpuset_testupdate(set, mask); 370 if (error) 371 goto out; 372 cpuset_update(set, mask); 373 CPU_COPY(mask, &set->cs_mask); 374out: 375 mtx_unlock_spin(&cpuset_lock); 376 377 return (error); 378} 379 380/* 381 * Resolve the 'which' parameter of several cpuset apis. 382 * 383 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also 384 * checks for permission via p_cansched(). 385 * 386 * For WHICH_SET returns a valid set with a new reference. 387 * 388 * -1 may be supplied for any argument to mean the current proc/thread or 389 * the base set of the current thread. May fail with ESRCH/EPERM. 390 */ 391static int 392cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp, 393 struct cpuset **setp) 394{ 395 struct cpuset *set; 396 struct thread *td; 397 struct proc *p; 398 int error; 399 400 *pp = p = NULL; 401 *tdp = td = NULL; 402 *setp = set = NULL; 403 switch (which) { 404 case CPU_WHICH_PID: 405 if (id == -1) { 406 PROC_LOCK(curproc); 407 p = curproc; 408 break; 409 } 410 if ((p = pfind(id)) == NULL) 411 return (ESRCH); 412 break; 413 case CPU_WHICH_TID: 414 if (id == -1) { 415 PROC_LOCK(curproc); 416 p = curproc; 417 td = curthread; 418 break; 419 } 420 sx_slock(&allproc_lock); 421 FOREACH_PROC_IN_SYSTEM(p) { 422 PROC_LOCK(p); 423 FOREACH_THREAD_IN_PROC(p, td) 424 if (td->td_tid == id) 425 break; 426 if (td != NULL) 427 break; 428 PROC_UNLOCK(p); 429 } 430 sx_sunlock(&allproc_lock); 431 if (td == NULL) 432 return (ESRCH); 433 break; 434 case CPU_WHICH_CPUSET: 435 if (id == -1) { 436 thread_lock(curthread); 437 set = cpuset_refbase(curthread->td_cpuset); 438 thread_unlock(curthread); 439 } else 440 set = cpuset_lookup(id, curthread); 441 if (set) { 442 *setp = set; 443 return (0); 444 } 445 return (ESRCH); 446 case CPU_WHICH_JAIL: 447 { 448 /* Find `set' for prison with given id. */ 449 struct prison *pr; 450 451 sx_slock(&allprison_lock); 452 pr = prison_find_child(curthread->td_ucred->cr_prison, id); 453 sx_sunlock(&allprison_lock); 454 if (pr == NULL) 455 return (ESRCH); 456 cpuset_ref(pr->pr_cpuset); 457 *setp = pr->pr_cpuset; 458 mtx_unlock(&pr->pr_mtx); 459 return (0); 460 } 461 case CPU_WHICH_IRQ: 462 return (0); 463 default: 464 return (EINVAL); 465 } 466 error = p_cansched(curthread, p); 467 if (error) { 468 PROC_UNLOCK(p); 469 return (error); 470 } 471 if (td == NULL) 472 td = FIRST_THREAD_IN_PROC(p); 473 *pp = p; 474 *tdp = td; 475 return (0); 476} 477 478/* 479 * Create an anonymous set with the provided mask in the space provided by 480 * 'fset'. If the passed in set is anonymous we use its parent otherwise 481 * the new set is a child of 'set'. 482 */ 483static int 484cpuset_shadow(struct cpuset *set, struct cpuset *fset, cpuset_t *mask) 485{ 486 struct cpuset *parent; 487 488 if (set->cs_id == CPUSET_INVALID) 489 parent = set->cs_parent; 490 else 491 parent = set; 492 if (!CPU_SUBSET(&parent->cs_mask, mask)) 493 return (EDEADLK); 494 return (_cpuset_create(fset, parent, mask, CPUSET_INVALID)); 495} 496 497/* 498 * Handle two cases for replacing the base set or mask of an entire process. 499 * 500 * 1) Set is non-null and mask is null. This reparents all anonymous sets 501 * to the provided set and replaces all non-anonymous td_cpusets with the 502 * provided set. 503 * 2) Mask is non-null and set is null. This replaces or creates anonymous 504 * sets for every thread with the existing base as a parent. 505 * 506 * This is overly complicated because we can't allocate while holding a 507 * spinlock and spinlocks must be held while changing and examining thread 508 * state. 509 */ 510static int 511cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask) 512{ 513 struct setlist freelist; 514 struct setlist droplist; 515 struct cpuset *tdset; 516 struct cpuset *nset; 517 struct thread *td; 518 struct proc *p; 519 int threads; 520 int nfree; 521 int error; 522 /* 523 * The algorithm requires two passes due to locking considerations. 524 * 525 * 1) Lookup the process and acquire the locks in the required order. 526 * 2) If enough cpusets have not been allocated release the locks and 527 * allocate them. Loop. 528 */ 529 LIST_INIT(&freelist); 530 LIST_INIT(&droplist); 531 nfree = 0; 532 for (;;) { 533 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset); 534 if (error) 535 goto out; 536 if (nfree >= p->p_numthreads) 537 break; 538 threads = p->p_numthreads; 539 PROC_UNLOCK(p); 540 for (; nfree < threads; nfree++) { 541 nset = uma_zalloc(cpuset_zone, M_WAITOK); 542 LIST_INSERT_HEAD(&freelist, nset, cs_link); 543 } 544 } 545 PROC_LOCK_ASSERT(p, MA_OWNED); 546 /* 547 * Now that the appropriate locks are held and we have enough cpusets, 548 * make sure the operation will succeed before applying changes. The 549 * proc lock prevents td_cpuset from changing between calls. 550 */ 551 error = 0; 552 FOREACH_THREAD_IN_PROC(p, td) { 553 thread_lock(td); 554 tdset = td->td_cpuset; 555 /* 556 * Verify that a new mask doesn't specify cpus outside of 557 * the set the thread is a member of. 558 */ 559 if (mask) { 560 if (tdset->cs_id == CPUSET_INVALID) 561 tdset = tdset->cs_parent; 562 if (!CPU_SUBSET(&tdset->cs_mask, mask)) 563 error = EDEADLK; 564 /* 565 * Verify that a new set won't leave an existing thread 566 * mask without a cpu to run on. It can, however, restrict 567 * the set. 568 */ 569 } else if (tdset->cs_id == CPUSET_INVALID) { 570 if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask)) 571 error = EDEADLK; 572 } 573 thread_unlock(td); 574 if (error) 575 goto unlock_out; 576 } 577 /* 578 * Replace each thread's cpuset while using deferred release. We 579 * must do this because the thread lock must be held while operating 580 * on the thread and this limits the type of operations allowed. 581 */ 582 FOREACH_THREAD_IN_PROC(p, td) { 583 thread_lock(td); 584 /* 585 * If we presently have an anonymous set or are applying a 586 * mask we must create an anonymous shadow set. That is 587 * either parented to our existing base or the supplied set. 588 * 589 * If we have a base set with no anonymous shadow we simply 590 * replace it outright. 591 */ 592 tdset = td->td_cpuset; 593 if (tdset->cs_id == CPUSET_INVALID || mask) { 594 nset = LIST_FIRST(&freelist); 595 LIST_REMOVE(nset, cs_link); 596 if (mask) 597 error = cpuset_shadow(tdset, nset, mask); 598 else 599 error = _cpuset_create(nset, set, 600 &tdset->cs_mask, CPUSET_INVALID); 601 if (error) { 602 LIST_INSERT_HEAD(&freelist, nset, cs_link); 603 thread_unlock(td); 604 break; 605 } 606 } else 607 nset = cpuset_ref(set); 608 cpuset_rel_defer(&droplist, tdset); 609 td->td_cpuset = nset; 610 sched_affinity(td); 611 thread_unlock(td); 612 } 613unlock_out: 614 PROC_UNLOCK(p); 615out: 616 while ((nset = LIST_FIRST(&droplist)) != NULL) 617 cpuset_rel_complete(nset); 618 while ((nset = LIST_FIRST(&freelist)) != NULL) { 619 LIST_REMOVE(nset, cs_link); 620 uma_zfree(cpuset_zone, nset); 621 } 622 return (error); 623} 624 625/* 626 * Apply an anonymous mask to a single thread. 627 */ 628int 629cpuset_setthread(lwpid_t id, cpuset_t *mask) 630{ 631 struct cpuset *nset; 632 struct cpuset *set; 633 struct thread *td; 634 struct proc *p; 635 int error; 636 637 nset = uma_zalloc(cpuset_zone, M_WAITOK); 638 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set); 639 if (error) 640 goto out; 641 set = NULL; 642 thread_lock(td); 643 error = cpuset_shadow(td->td_cpuset, nset, mask); 644 if (error == 0) { 645 set = td->td_cpuset; 646 td->td_cpuset = nset; 647 sched_affinity(td); 648 nset = NULL; 649 } 650 thread_unlock(td); 651 PROC_UNLOCK(p); 652 if (set) 653 cpuset_rel(set); 654out: 655 if (nset) 656 uma_zfree(cpuset_zone, nset); 657 return (error); 658} 659 660/* 661 * Creates the cpuset for thread0. We make two sets: 662 * 663 * 0 - The root set which should represent all valid processors in the 664 * system. It is initially created with a mask of all processors 665 * because we don't know what processors are valid until cpuset_init() 666 * runs. This set is immutable. 667 * 1 - The default set which all processes are a member of until changed. 668 * This allows an administrator to move all threads off of given cpus to 669 * dedicate them to high priority tasks or save power etc. 670 */ 671struct cpuset * 672cpuset_thread0(void) 673{ 674 struct cpuset *set; 675 int error; 676 677 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL, 678 NULL, NULL, UMA_ALIGN_PTR, 0); 679 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE); 680 /* 681 * Create the root system set for the whole machine. Doesn't use 682 * cpuset_create() due to NULL parent. 683 */ 684 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); 685 set->cs_mask.__bits[0] = -1; 686 LIST_INIT(&set->cs_children); 687 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 688 set->cs_ref = 1; 689 set->cs_flags = CPU_SET_ROOT; 690 cpuset_zero = set; 691 cpuset_root = &set->cs_mask; 692 /* 693 * Now derive a default, modifiable set from that to give out. 694 */ 695 set = uma_zalloc(cpuset_zone, M_WAITOK); 696 error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1); 697 KASSERT(error == 0, ("Error creating default set: %d\n", error)); 698 /* 699 * Initialize the unit allocator. 0 and 1 are allocated above. 700 */ 701 cpuset_unr = new_unrhdr(2, INT_MAX, NULL); 702 703 return (set); 704} 705 706/* 707 * Create a cpuset, which would be cpuset_create() but 708 * mark the new 'set' as root. 709 * 710 * We are not going to reparent the td to it. Use cpuset_setproc_update_set() 711 * for that. 712 * 713 * In case of no error, returns the set in *setp locked with a reference. 714 */ 715int 716cpuset_create_root(struct prison *pr, struct cpuset **setp) 717{ 718 struct cpuset *set; 719 int error; 720 721 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__)); 722 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__)); 723 724 error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask); 725 if (error) 726 return (error); 727 728 KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data", 729 __func__, __LINE__)); 730 731 /* Mark the set as root. */ 732 set = *setp; 733 set->cs_flags |= CPU_SET_ROOT; 734 735 return (0); 736} 737 738int 739cpuset_setproc_update_set(struct proc *p, struct cpuset *set) 740{ 741 int error; 742 743 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__)); 744 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__)); 745 746 cpuset_ref(set); 747 error = cpuset_setproc(p->p_pid, set, NULL); 748 if (error) 749 return (error); 750 cpuset_rel(set); 751 return (0); 752} 753 754/* 755 * This is called once the final set of system cpus is known. Modifies 756 * the root set and all children and mark the root readonly. 757 */ 758static void 759cpuset_init(void *arg) 760{ 761 cpuset_t mask; 762 763 CPU_ZERO(&mask); 764#ifdef SMP 765 mask.__bits[0] = all_cpus; 766#else 767 mask.__bits[0] = 1; 768#endif 769 if (cpuset_modify(cpuset_zero, &mask)) 770 panic("Can't set initial cpuset mask.\n"); 771 cpuset_zero->cs_flags |= CPU_SET_RDONLY; 772} 773SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL); 774 775#ifndef _SYS_SYSPROTO_H_ 776struct cpuset_args { 777 cpusetid_t *setid; 778}; 779#endif 780int 781cpuset(struct thread *td, struct cpuset_args *uap) 782{ 783 struct cpuset *root; 784 struct cpuset *set; 785 int error; 786 787 thread_lock(td); 788 root = cpuset_refroot(td->td_cpuset); 789 thread_unlock(td); 790 error = cpuset_create(&set, root, &root->cs_mask); 791 cpuset_rel(root); 792 if (error) 793 return (error); 794 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id)); 795 if (error == 0) 796 error = cpuset_setproc(-1, set, NULL); 797 cpuset_rel(set); 798 return (error); 799} 800 801#ifndef _SYS_SYSPROTO_H_ 802struct cpuset_setid_args { 803 cpuwhich_t which; 804 id_t id; 805 cpusetid_t setid; 806}; 807#endif 808int 809cpuset_setid(struct thread *td, struct cpuset_setid_args *uap) 810{ 811 struct cpuset *set; 812 int error; 813 814 /* 815 * Presently we only support per-process sets. 816 */ 817 if (uap->which != CPU_WHICH_PID) 818 return (EINVAL); 819 set = cpuset_lookup(uap->setid, td); 820 if (set == NULL) 821 return (ESRCH); 822 error = cpuset_setproc(uap->id, set, NULL); 823 cpuset_rel(set); 824 return (error); 825} 826 827#ifndef _SYS_SYSPROTO_H_ 828struct cpuset_getid_args { 829 cpulevel_t level; 830 cpuwhich_t which; 831 id_t id; 832 cpusetid_t *setid; 833#endif 834int 835cpuset_getid(struct thread *td, struct cpuset_getid_args *uap) 836{ 837 struct cpuset *nset; 838 struct cpuset *set; 839 struct thread *ttd; 840 struct proc *p; 841 cpusetid_t id; 842 int error; 843 844 if (uap->level == CPU_LEVEL_WHICH && uap->which != CPU_WHICH_CPUSET) 845 return (EINVAL); 846 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 847 if (error) 848 return (error); 849 switch (uap->which) { 850 case CPU_WHICH_TID: 851 case CPU_WHICH_PID: 852 thread_lock(ttd); 853 set = cpuset_refbase(ttd->td_cpuset); 854 thread_unlock(ttd); 855 PROC_UNLOCK(p); 856 break; 857 case CPU_WHICH_CPUSET: 858 case CPU_WHICH_JAIL: 859 break; 860 case CPU_WHICH_IRQ: 861 return (EINVAL); 862 } 863 switch (uap->level) { 864 case CPU_LEVEL_ROOT: 865 nset = cpuset_refroot(set); 866 cpuset_rel(set); 867 set = nset; 868 break; 869 case CPU_LEVEL_CPUSET: 870 break; 871 case CPU_LEVEL_WHICH: 872 break; 873 } 874 id = set->cs_id; 875 cpuset_rel(set); 876 if (error == 0) 877 error = copyout(&id, uap->setid, sizeof(id)); 878 879 return (error); 880} 881 882#ifndef _SYS_SYSPROTO_H_ 883struct cpuset_getaffinity_args { 884 cpulevel_t level; 885 cpuwhich_t which; 886 id_t id; 887 size_t cpusetsize; 888 cpuset_t *mask; 889}; 890#endif 891int 892cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap) 893{ 894 struct thread *ttd; 895 struct cpuset *nset; 896 struct cpuset *set; 897 struct proc *p; 898 cpuset_t *mask; 899 int error; 900 size_t size; 901 902 if (uap->cpusetsize < sizeof(cpuset_t) || 903 uap->cpusetsize > CPU_MAXSIZE / NBBY) 904 return (ERANGE); 905 size = uap->cpusetsize; 906 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO); 907 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 908 if (error) 909 goto out; 910 switch (uap->level) { 911 case CPU_LEVEL_ROOT: 912 case CPU_LEVEL_CPUSET: 913 switch (uap->which) { 914 case CPU_WHICH_TID: 915 case CPU_WHICH_PID: 916 thread_lock(ttd); 917 set = cpuset_ref(ttd->td_cpuset); 918 thread_unlock(ttd); 919 break; 920 case CPU_WHICH_CPUSET: 921 case CPU_WHICH_JAIL: 922 break; 923 case CPU_WHICH_IRQ: 924 error = EINVAL; 925 goto out; 926 } 927 if (uap->level == CPU_LEVEL_ROOT) 928 nset = cpuset_refroot(set); 929 else 930 nset = cpuset_refbase(set); 931 CPU_COPY(&nset->cs_mask, mask); 932 cpuset_rel(nset); 933 break; 934 case CPU_LEVEL_WHICH: 935 switch (uap->which) { 936 case CPU_WHICH_TID: 937 thread_lock(ttd); 938 CPU_COPY(&ttd->td_cpuset->cs_mask, mask); 939 thread_unlock(ttd); 940 break; 941 case CPU_WHICH_PID: 942 FOREACH_THREAD_IN_PROC(p, ttd) { 943 thread_lock(ttd); 944 CPU_OR(mask, &ttd->td_cpuset->cs_mask); 945 thread_unlock(ttd); 946 } 947 break; 948 case CPU_WHICH_CPUSET: 949 case CPU_WHICH_JAIL: 950 CPU_COPY(&set->cs_mask, mask); 951 break; 952 case CPU_WHICH_IRQ: 953 error = intr_getaffinity(uap->id, mask); 954 break; 955 } 956 break; 957 default: 958 error = EINVAL; 959 break; 960 } 961 if (set) 962 cpuset_rel(set); 963 if (p) 964 PROC_UNLOCK(p); 965 if (error == 0) 966 error = copyout(mask, uap->mask, size); 967out: 968 free(mask, M_TEMP); 969 return (error); 970} 971 972#ifndef _SYS_SYSPROTO_H_ 973struct cpuset_setaffinity_args { 974 cpulevel_t level; 975 cpuwhich_t which; 976 id_t id; 977 size_t cpusetsize; 978 const cpuset_t *mask; 979}; 980#endif 981int 982cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap) 983{ 984 struct cpuset *nset; 985 struct cpuset *set; 986 struct thread *ttd; 987 struct proc *p; 988 cpuset_t *mask; 989 int error; 990 991 if (uap->cpusetsize < sizeof(cpuset_t) || 992 uap->cpusetsize > CPU_MAXSIZE / NBBY) 993 return (ERANGE); 994 mask = malloc(uap->cpusetsize, M_TEMP, M_WAITOK | M_ZERO); 995 error = copyin(uap->mask, mask, uap->cpusetsize); 996 if (error) 997 goto out; 998 /* 999 * Verify that no high bits are set. 1000 */ 1001 if (uap->cpusetsize > sizeof(cpuset_t)) { 1002 char *end; 1003 char *cp; 1004 1005 end = cp = (char *)&mask->__bits; 1006 end += uap->cpusetsize; 1007 cp += sizeof(cpuset_t); 1008 while (cp != end) 1009 if (*cp++ != 0) { 1010 error = EINVAL; 1011 goto out; 1012 } 1013 1014 } 1015 switch (uap->level) { 1016 case CPU_LEVEL_ROOT: 1017 case CPU_LEVEL_CPUSET: 1018 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 1019 if (error) 1020 break; 1021 switch (uap->which) { 1022 case CPU_WHICH_TID: 1023 case CPU_WHICH_PID: 1024 thread_lock(ttd); 1025 set = cpuset_ref(ttd->td_cpuset); 1026 thread_unlock(ttd); 1027 PROC_UNLOCK(p); 1028 break; 1029 case CPU_WHICH_CPUSET: 1030 case CPU_WHICH_JAIL: 1031 break; 1032 case CPU_WHICH_IRQ: 1033 error = EINVAL; 1034 goto out; 1035 } 1036 if (uap->level == CPU_LEVEL_ROOT) 1037 nset = cpuset_refroot(set); 1038 else 1039 nset = cpuset_refbase(set); 1040 error = cpuset_modify(nset, mask); 1041 cpuset_rel(nset); 1042 cpuset_rel(set); 1043 break; 1044 case CPU_LEVEL_WHICH: 1045 switch (uap->which) { 1046 case CPU_WHICH_TID: 1047 error = cpuset_setthread(uap->id, mask); 1048 break; 1049 case CPU_WHICH_PID: 1050 error = cpuset_setproc(uap->id, NULL, mask); 1051 break; 1052 case CPU_WHICH_CPUSET: 1053 case CPU_WHICH_JAIL: 1054 error = cpuset_which(uap->which, uap->id, &p, 1055 &ttd, &set); 1056 if (error == 0) { 1057 error = cpuset_modify(set, mask); 1058 cpuset_rel(set); 1059 } 1060 break; 1061 case CPU_WHICH_IRQ: 1062 error = intr_setaffinity(uap->id, mask); 1063 break; 1064 default: 1065 error = EINVAL; 1066 break; 1067 } 1068 break; 1069 default: 1070 error = EINVAL; 1071 break; 1072 } 1073out: 1074 free(mask, M_TEMP); 1075 return (error); 1076} 1077 1078#ifdef DDB 1079DB_SHOW_COMMAND(cpusets, db_show_cpusets) 1080{ 1081 struct cpuset *set; 1082 int cpu, once; 1083 1084 LIST_FOREACH(set, &cpuset_ids, cs_link) { 1085 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n", 1086 set, set->cs_id, set->cs_ref, set->cs_flags, 1087 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0); 1088 db_printf(" mask="); 1089 for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) { 1090 if (CPU_ISSET(cpu, &set->cs_mask)) { 1091 if (once == 0) { 1092 db_printf("%d", cpu); 1093 once = 1; 1094 } else 1095 db_printf(",%d", cpu); 1096 } 1097 } 1098 db_printf("\n"); 1099 if (db_pager_quit) 1100 break; 1101 } 1102} 1103#endif /* DDB */ 1104