kern_cpuset.c revision 251470
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 251470 2013-06-06 14:43:19Z jhb $"); 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/queue.h> 53#include <sys/libkern.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. 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 group separate 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 available 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; 110 111/* Return the size of cpuset_t at the kernel level */ 112SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD, 113 0, sizeof(cpuset_t), "sizeof(cpuset_t)"); 114 115cpuset_t *cpuset_root; 116 117/* 118 * Acquire a reference to a cpuset, all pointers must be tracked with refs. 119 */ 120struct cpuset * 121cpuset_ref(struct cpuset *set) 122{ 123 124 refcount_acquire(&set->cs_ref); 125 return (set); 126} 127 128/* 129 * Walks up the tree from 'set' to find the root. Returns the root 130 * referenced. 131 */ 132static struct cpuset * 133cpuset_refroot(struct cpuset *set) 134{ 135 136 for (; set->cs_parent != NULL; set = set->cs_parent) 137 if (set->cs_flags & CPU_SET_ROOT) 138 break; 139 cpuset_ref(set); 140 141 return (set); 142} 143 144/* 145 * Find the first non-anonymous set starting from 'set'. Returns this set 146 * referenced. May return the passed in set with an extra ref if it is 147 * not anonymous. 148 */ 149static struct cpuset * 150cpuset_refbase(struct cpuset *set) 151{ 152 153 if (set->cs_id == CPUSET_INVALID) 154 set = set->cs_parent; 155 cpuset_ref(set); 156 157 return (set); 158} 159 160/* 161 * Release a reference in a context where it is safe to allocate. 162 */ 163void 164cpuset_rel(struct cpuset *set) 165{ 166 cpusetid_t id; 167 168 if (refcount_release(&set->cs_ref) == 0) 169 return; 170 mtx_lock_spin(&cpuset_lock); 171 LIST_REMOVE(set, cs_siblings); 172 id = set->cs_id; 173 if (id != CPUSET_INVALID) 174 LIST_REMOVE(set, cs_link); 175 mtx_unlock_spin(&cpuset_lock); 176 cpuset_rel(set->cs_parent); 177 uma_zfree(cpuset_zone, set); 178 if (id != CPUSET_INVALID) 179 free_unr(cpuset_unr, id); 180} 181 182/* 183 * Deferred release must be used when in a context that is not safe to 184 * allocate/free. This places any unreferenced sets on the list 'head'. 185 */ 186static void 187cpuset_rel_defer(struct setlist *head, struct cpuset *set) 188{ 189 190 if (refcount_release(&set->cs_ref) == 0) 191 return; 192 mtx_lock_spin(&cpuset_lock); 193 LIST_REMOVE(set, cs_siblings); 194 if (set->cs_id != CPUSET_INVALID) 195 LIST_REMOVE(set, cs_link); 196 LIST_INSERT_HEAD(head, set, cs_link); 197 mtx_unlock_spin(&cpuset_lock); 198} 199 200/* 201 * Complete a deferred release. Removes the set from the list provided to 202 * cpuset_rel_defer. 203 */ 204static void 205cpuset_rel_complete(struct cpuset *set) 206{ 207 LIST_REMOVE(set, cs_link); 208 cpuset_rel(set->cs_parent); 209 uma_zfree(cpuset_zone, set); 210} 211 212/* 213 * Find a set based on an id. Returns it with a ref. 214 */ 215static struct cpuset * 216cpuset_lookup(cpusetid_t setid, struct thread *td) 217{ 218 struct cpuset *set; 219 220 if (setid == CPUSET_INVALID) 221 return (NULL); 222 mtx_lock_spin(&cpuset_lock); 223 LIST_FOREACH(set, &cpuset_ids, cs_link) 224 if (set->cs_id == setid) 225 break; 226 if (set) 227 cpuset_ref(set); 228 mtx_unlock_spin(&cpuset_lock); 229 230 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__)); 231 if (set != NULL && jailed(td->td_ucred)) { 232 struct cpuset *jset, *tset; 233 234 jset = td->td_ucred->cr_prison->pr_cpuset; 235 for (tset = set; tset != NULL; tset = tset->cs_parent) 236 if (tset == jset) 237 break; 238 if (tset == NULL) { 239 cpuset_rel(set); 240 set = NULL; 241 } 242 } 243 244 return (set); 245} 246 247/* 248 * Create a set in the space provided in 'set' with the provided parameters. 249 * The set is returned with a single ref. May return EDEADLK if the set 250 * will have no valid cpu based on restrictions from the parent. 251 */ 252static int 253_cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask, 254 cpusetid_t id) 255{ 256 257 if (!CPU_OVERLAP(&parent->cs_mask, mask)) 258 return (EDEADLK); 259 CPU_COPY(mask, &set->cs_mask); 260 LIST_INIT(&set->cs_children); 261 refcount_init(&set->cs_ref, 1); 262 set->cs_flags = 0; 263 mtx_lock_spin(&cpuset_lock); 264 CPU_AND(&set->cs_mask, &parent->cs_mask); 265 set->cs_id = id; 266 set->cs_parent = cpuset_ref(parent); 267 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings); 268 if (set->cs_id != CPUSET_INVALID) 269 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 270 mtx_unlock_spin(&cpuset_lock); 271 272 return (0); 273} 274 275/* 276 * Create a new non-anonymous set with the requested parent and mask. May 277 * return failures if the mask is invalid or a new number can not be 278 * allocated. 279 */ 280static int 281cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask) 282{ 283 struct cpuset *set; 284 cpusetid_t id; 285 int error; 286 287 id = alloc_unr(cpuset_unr); 288 if (id == -1) 289 return (ENFILE); 290 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK); 291 error = _cpuset_create(set, parent, mask, id); 292 if (error == 0) 293 return (0); 294 free_unr(cpuset_unr, id); 295 uma_zfree(cpuset_zone, set); 296 297 return (error); 298} 299 300/* 301 * Recursively check for errors that would occur from applying mask to 302 * the tree of sets starting at 'set'. Checks for sets that would become 303 * empty as well as RDONLY flags. 304 */ 305static int 306cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask) 307{ 308 struct cpuset *nset; 309 cpuset_t newmask; 310 int error; 311 312 mtx_assert(&cpuset_lock, MA_OWNED); 313 if (set->cs_flags & CPU_SET_RDONLY) 314 return (EPERM); 315 if (check_mask) { 316 if (!CPU_OVERLAP(&set->cs_mask, mask)) 317 return (EDEADLK); 318 CPU_COPY(&set->cs_mask, &newmask); 319 CPU_AND(&newmask, mask); 320 } else 321 CPU_COPY(mask, &newmask); 322 error = 0; 323 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 324 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0) 325 break; 326 return (error); 327} 328 329/* 330 * Applies the mask 'mask' without checking for empty sets or permissions. 331 */ 332static void 333cpuset_update(struct cpuset *set, cpuset_t *mask) 334{ 335 struct cpuset *nset; 336 337 mtx_assert(&cpuset_lock, MA_OWNED); 338 CPU_AND(&set->cs_mask, mask); 339 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 340 cpuset_update(nset, &set->cs_mask); 341 342 return; 343} 344 345/* 346 * Modify the set 'set' to use a copy of the mask provided. Apply this new 347 * mask to restrict all children in the tree. Checks for validity before 348 * applying the changes. 349 */ 350static int 351cpuset_modify(struct cpuset *set, cpuset_t *mask) 352{ 353 struct cpuset *root; 354 int error; 355 356 error = priv_check(curthread, PRIV_SCHED_CPUSET); 357 if (error) 358 return (error); 359 /* 360 * In case we are called from within the jail 361 * we do not allow modifying the dedicated root 362 * cpuset of the jail but may still allow to 363 * change child sets. 364 */ 365 if (jailed(curthread->td_ucred) && 366 set->cs_flags & CPU_SET_ROOT) 367 return (EPERM); 368 /* 369 * Verify that we have access to this set of 370 * cpus. 371 */ 372 root = set->cs_parent; 373 if (root && !CPU_SUBSET(&root->cs_mask, mask)) 374 return (EINVAL); 375 mtx_lock_spin(&cpuset_lock); 376 error = cpuset_testupdate(set, mask, 0); 377 if (error) 378 goto out; 379 CPU_COPY(mask, &set->cs_mask); 380 cpuset_update(set, mask); 381out: 382 mtx_unlock_spin(&cpuset_lock); 383 384 return (error); 385} 386 387/* 388 * Resolve the 'which' parameter of several cpuset apis. 389 * 390 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also 391 * checks for permission via p_cansched(). 392 * 393 * For WHICH_SET returns a valid set with a new reference. 394 * 395 * -1 may be supplied for any argument to mean the current proc/thread or 396 * the base set of the current thread. May fail with ESRCH/EPERM. 397 */ 398static int 399cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp, 400 struct cpuset **setp) 401{ 402 struct cpuset *set; 403 struct thread *td; 404 struct proc *p; 405 int error; 406 407 *pp = p = NULL; 408 *tdp = td = NULL; 409 *setp = set = NULL; 410 switch (which) { 411 case CPU_WHICH_PID: 412 if (id == -1) { 413 PROC_LOCK(curproc); 414 p = curproc; 415 break; 416 } 417 if ((p = pfind(id)) == NULL) 418 return (ESRCH); 419 break; 420 case CPU_WHICH_TID: 421 if (id == -1) { 422 PROC_LOCK(curproc); 423 p = curproc; 424 td = curthread; 425 break; 426 } 427 td = tdfind(id, -1); 428 if (td == NULL) 429 return (ESRCH); 430 p = td->td_proc; 431 break; 432 case CPU_WHICH_CPUSET: 433 if (id == -1) { 434 thread_lock(curthread); 435 set = cpuset_refbase(curthread->td_cpuset); 436 thread_unlock(curthread); 437 } else 438 set = cpuset_lookup(id, curthread); 439 if (set) { 440 *setp = set; 441 return (0); 442 } 443 return (ESRCH); 444 case CPU_WHICH_JAIL: 445 { 446 /* Find `set' for prison with given id. */ 447 struct prison *pr; 448 449 sx_slock(&allprison_lock); 450 pr = prison_find_child(curthread->td_ucred->cr_prison, id); 451 sx_sunlock(&allprison_lock); 452 if (pr == NULL) 453 return (ESRCH); 454 cpuset_ref(pr->pr_cpuset); 455 *setp = pr->pr_cpuset; 456 mtx_unlock(&pr->pr_mtx); 457 return (0); 458 } 459 case CPU_WHICH_IRQ: 460 return (0); 461 default: 462 return (EINVAL); 463 } 464 error = p_cansched(curthread, p); 465 if (error) { 466 PROC_UNLOCK(p); 467 return (error); 468 } 469 if (td == NULL) 470 td = FIRST_THREAD_IN_PROC(p); 471 *pp = p; 472 *tdp = td; 473 return (0); 474} 475 476/* 477 * Create an anonymous set with the provided mask in the space provided by 478 * 'fset'. If the passed in set is anonymous we use its parent otherwise 479 * the new set is a child of 'set'. 480 */ 481static int 482cpuset_shadow(struct cpuset *set, struct cpuset *fset, const cpuset_t *mask) 483{ 484 struct cpuset *parent; 485 486 if (set->cs_id == CPUSET_INVALID) 487 parent = set->cs_parent; 488 else 489 parent = set; 490 if (!CPU_SUBSET(&parent->cs_mask, mask)) 491 return (EDEADLK); 492 return (_cpuset_create(fset, parent, mask, CPUSET_INVALID)); 493} 494 495/* 496 * Handle two cases for replacing the base set or mask of an entire process. 497 * 498 * 1) Set is non-null and mask is null. This reparents all anonymous sets 499 * to the provided set and replaces all non-anonymous td_cpusets with the 500 * provided set. 501 * 2) Mask is non-null and set is null. This replaces or creates anonymous 502 * sets for every thread with the existing base as a parent. 503 * 504 * This is overly complicated because we can't allocate while holding a 505 * spinlock and spinlocks must be held while changing and examining thread 506 * state. 507 */ 508static int 509cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask) 510{ 511 struct setlist freelist; 512 struct setlist droplist; 513 struct cpuset *tdset; 514 struct cpuset *nset; 515 struct thread *td; 516 struct proc *p; 517 int threads; 518 int nfree; 519 int error; 520 /* 521 * The algorithm requires two passes due to locking considerations. 522 * 523 * 1) Lookup the process and acquire the locks in the required order. 524 * 2) If enough cpusets have not been allocated release the locks and 525 * allocate them. Loop. 526 */ 527 LIST_INIT(&freelist); 528 LIST_INIT(&droplist); 529 nfree = 0; 530 for (;;) { 531 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset); 532 if (error) 533 goto out; 534 if (nfree >= p->p_numthreads) 535 break; 536 threads = p->p_numthreads; 537 PROC_UNLOCK(p); 538 for (; nfree < threads; nfree++) { 539 nset = uma_zalloc(cpuset_zone, M_WAITOK); 540 LIST_INSERT_HEAD(&freelist, nset, cs_link); 541 } 542 } 543 PROC_LOCK_ASSERT(p, MA_OWNED); 544 /* 545 * Now that the appropriate locks are held and we have enough cpusets, 546 * make sure the operation will succeed before applying changes. The 547 * proc lock prevents td_cpuset from changing between calls. 548 */ 549 error = 0; 550 FOREACH_THREAD_IN_PROC(p, td) { 551 thread_lock(td); 552 tdset = td->td_cpuset; 553 /* 554 * Verify that a new mask doesn't specify cpus outside of 555 * the set the thread is a member of. 556 */ 557 if (mask) { 558 if (tdset->cs_id == CPUSET_INVALID) 559 tdset = tdset->cs_parent; 560 if (!CPU_SUBSET(&tdset->cs_mask, mask)) 561 error = EDEADLK; 562 /* 563 * Verify that a new set won't leave an existing thread 564 * mask without a cpu to run on. It can, however, restrict 565 * the set. 566 */ 567 } else if (tdset->cs_id == CPUSET_INVALID) { 568 if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask)) 569 error = EDEADLK; 570 } 571 thread_unlock(td); 572 if (error) 573 goto unlock_out; 574 } 575 /* 576 * Replace each thread's cpuset while using deferred release. We 577 * must do this because the thread lock must be held while operating 578 * on the thread and this limits the type of operations allowed. 579 */ 580 FOREACH_THREAD_IN_PROC(p, td) { 581 thread_lock(td); 582 /* 583 * If we presently have an anonymous set or are applying a 584 * mask we must create an anonymous shadow set. That is 585 * either parented to our existing base or the supplied set. 586 * 587 * If we have a base set with no anonymous shadow we simply 588 * replace it outright. 589 */ 590 tdset = td->td_cpuset; 591 if (tdset->cs_id == CPUSET_INVALID || mask) { 592 nset = LIST_FIRST(&freelist); 593 LIST_REMOVE(nset, cs_link); 594 if (mask) 595 error = cpuset_shadow(tdset, nset, mask); 596 else 597 error = _cpuset_create(nset, set, 598 &tdset->cs_mask, CPUSET_INVALID); 599 if (error) { 600 LIST_INSERT_HEAD(&freelist, nset, cs_link); 601 thread_unlock(td); 602 break; 603 } 604 } else 605 nset = cpuset_ref(set); 606 cpuset_rel_defer(&droplist, tdset); 607 td->td_cpuset = nset; 608 sched_affinity(td); 609 thread_unlock(td); 610 } 611unlock_out: 612 PROC_UNLOCK(p); 613out: 614 while ((nset = LIST_FIRST(&droplist)) != NULL) 615 cpuset_rel_complete(nset); 616 while ((nset = LIST_FIRST(&freelist)) != NULL) { 617 LIST_REMOVE(nset, cs_link); 618 uma_zfree(cpuset_zone, nset); 619 } 620 return (error); 621} 622 623/* 624 * Calculate the ffs() of the cpuset. 625 */ 626int 627cpusetobj_ffs(const cpuset_t *set) 628{ 629 size_t i; 630 int cbit; 631 632 cbit = 0; 633 for (i = 0; i < _NCPUWORDS; i++) { 634 if (set->__bits[i] != 0) { 635 cbit = ffsl(set->__bits[i]); 636 cbit += i * _NCPUBITS; 637 break; 638 } 639 } 640 return (cbit); 641} 642 643/* 644 * Return a string representing a valid layout for a cpuset_t object. 645 * It expects an incoming buffer at least sized as CPUSETBUFSIZ. 646 */ 647char * 648cpusetobj_strprint(char *buf, const cpuset_t *set) 649{ 650 char *tbuf; 651 size_t i, bytesp, bufsiz; 652 653 tbuf = buf; 654 bytesp = 0; 655 bufsiz = CPUSETBUFSIZ; 656 657 for (i = 0; i < (_NCPUWORDS - 1); i++) { 658 bytesp = snprintf(tbuf, bufsiz, "%lx,", set->__bits[i]); 659 bufsiz -= bytesp; 660 tbuf += bytesp; 661 } 662 snprintf(tbuf, bufsiz, "%lx", set->__bits[_NCPUWORDS - 1]); 663 return (buf); 664} 665 666/* 667 * Build a valid cpuset_t object from a string representation. 668 * It expects an incoming buffer at least sized as CPUSETBUFSIZ. 669 */ 670int 671cpusetobj_strscan(cpuset_t *set, const char *buf) 672{ 673 u_int nwords; 674 int i, ret; 675 676 if (strlen(buf) > CPUSETBUFSIZ - 1) 677 return (-1); 678 679 /* Allow to pass a shorter version of the mask when necessary. */ 680 nwords = 1; 681 for (i = 0; buf[i] != '\0'; i++) 682 if (buf[i] == ',') 683 nwords++; 684 if (nwords > _NCPUWORDS) 685 return (-1); 686 687 CPU_ZERO(set); 688 for (i = 0; i < (nwords - 1); i++) { 689 ret = sscanf(buf, "%lx,", &set->__bits[i]); 690 if (ret == 0 || ret == -1) 691 return (-1); 692 buf = strstr(buf, ","); 693 if (buf == NULL) 694 return (-1); 695 buf++; 696 } 697 ret = sscanf(buf, "%lx", &set->__bits[nwords - 1]); 698 if (ret == 0 || ret == -1) 699 return (-1); 700 return (0); 701} 702 703/* 704 * Apply an anonymous mask to a single thread. 705 */ 706int 707cpuset_setthread(lwpid_t id, cpuset_t *mask) 708{ 709 struct cpuset *nset; 710 struct cpuset *set; 711 struct thread *td; 712 struct proc *p; 713 int error; 714 715 nset = uma_zalloc(cpuset_zone, M_WAITOK); 716 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set); 717 if (error) 718 goto out; 719 set = NULL; 720 thread_lock(td); 721 error = cpuset_shadow(td->td_cpuset, nset, mask); 722 if (error == 0) { 723 set = td->td_cpuset; 724 td->td_cpuset = nset; 725 sched_affinity(td); 726 nset = NULL; 727 } 728 thread_unlock(td); 729 PROC_UNLOCK(p); 730 if (set) 731 cpuset_rel(set); 732out: 733 if (nset) 734 uma_zfree(cpuset_zone, nset); 735 return (error); 736} 737 738/* 739 * Creates the cpuset for thread0. We make two sets: 740 * 741 * 0 - The root set which should represent all valid processors in the 742 * system. It is initially created with a mask of all processors 743 * because we don't know what processors are valid until cpuset_init() 744 * runs. This set is immutable. 745 * 1 - The default set which all processes are a member of until changed. 746 * This allows an administrator to move all threads off of given cpus to 747 * dedicate them to high priority tasks or save power etc. 748 */ 749struct cpuset * 750cpuset_thread0(void) 751{ 752 struct cpuset *set; 753 int error; 754 755 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL, 756 NULL, NULL, UMA_ALIGN_PTR, 0); 757 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE); 758 /* 759 * Create the root system set for the whole machine. Doesn't use 760 * cpuset_create() due to NULL parent. 761 */ 762 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); 763 CPU_FILL(&set->cs_mask); 764 LIST_INIT(&set->cs_children); 765 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 766 set->cs_ref = 1; 767 set->cs_flags = CPU_SET_ROOT; 768 cpuset_zero = set; 769 cpuset_root = &set->cs_mask; 770 /* 771 * Now derive a default, modifiable set from that to give out. 772 */ 773 set = uma_zalloc(cpuset_zone, M_WAITOK); 774 error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1); 775 KASSERT(error == 0, ("Error creating default set: %d\n", error)); 776 /* 777 * Initialize the unit allocator. 0 and 1 are allocated above. 778 */ 779 cpuset_unr = new_unrhdr(2, INT_MAX, NULL); 780 781 return (set); 782} 783 784/* 785 * Create a cpuset, which would be cpuset_create() but 786 * mark the new 'set' as root. 787 * 788 * We are not going to reparent the td to it. Use cpuset_setproc_update_set() 789 * for that. 790 * 791 * In case of no error, returns the set in *setp locked with a reference. 792 */ 793int 794cpuset_create_root(struct prison *pr, struct cpuset **setp) 795{ 796 struct cpuset *set; 797 int error; 798 799 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__)); 800 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__)); 801 802 error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask); 803 if (error) 804 return (error); 805 806 KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data", 807 __func__, __LINE__)); 808 809 /* Mark the set as root. */ 810 set = *setp; 811 set->cs_flags |= CPU_SET_ROOT; 812 813 return (0); 814} 815 816int 817cpuset_setproc_update_set(struct proc *p, struct cpuset *set) 818{ 819 int error; 820 821 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__)); 822 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__)); 823 824 cpuset_ref(set); 825 error = cpuset_setproc(p->p_pid, set, NULL); 826 if (error) 827 return (error); 828 cpuset_rel(set); 829 return (0); 830} 831 832/* 833 * This is called once the final set of system cpus is known. Modifies 834 * the root set and all children and mark the root read-only. 835 */ 836static void 837cpuset_init(void *arg) 838{ 839 cpuset_t mask; 840 841 mask = all_cpus; 842 if (cpuset_modify(cpuset_zero, &mask)) 843 panic("Can't set initial cpuset mask.\n"); 844 cpuset_zero->cs_flags |= CPU_SET_RDONLY; 845} 846SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL); 847 848#ifndef _SYS_SYSPROTO_H_ 849struct cpuset_args { 850 cpusetid_t *setid; 851}; 852#endif 853int 854sys_cpuset(struct thread *td, struct cpuset_args *uap) 855{ 856 struct cpuset *root; 857 struct cpuset *set; 858 int error; 859 860 thread_lock(td); 861 root = cpuset_refroot(td->td_cpuset); 862 thread_unlock(td); 863 error = cpuset_create(&set, root, &root->cs_mask); 864 cpuset_rel(root); 865 if (error) 866 return (error); 867 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id)); 868 if (error == 0) 869 error = cpuset_setproc(-1, set, NULL); 870 cpuset_rel(set); 871 return (error); 872} 873 874#ifndef _SYS_SYSPROTO_H_ 875struct cpuset_setid_args { 876 cpuwhich_t which; 877 id_t id; 878 cpusetid_t setid; 879}; 880#endif 881int 882sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap) 883{ 884 struct cpuset *set; 885 int error; 886 887 /* 888 * Presently we only support per-process sets. 889 */ 890 if (uap->which != CPU_WHICH_PID) 891 return (EINVAL); 892 set = cpuset_lookup(uap->setid, td); 893 if (set == NULL) 894 return (ESRCH); 895 error = cpuset_setproc(uap->id, set, NULL); 896 cpuset_rel(set); 897 return (error); 898} 899 900#ifndef _SYS_SYSPROTO_H_ 901struct cpuset_getid_args { 902 cpulevel_t level; 903 cpuwhich_t which; 904 id_t id; 905 cpusetid_t *setid; 906}; 907#endif 908int 909sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap) 910{ 911 struct cpuset *nset; 912 struct cpuset *set; 913 struct thread *ttd; 914 struct proc *p; 915 cpusetid_t id; 916 int error; 917 918 if (uap->level == CPU_LEVEL_WHICH && uap->which != CPU_WHICH_CPUSET) 919 return (EINVAL); 920 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 921 if (error) 922 return (error); 923 switch (uap->which) { 924 case CPU_WHICH_TID: 925 case CPU_WHICH_PID: 926 thread_lock(ttd); 927 set = cpuset_refbase(ttd->td_cpuset); 928 thread_unlock(ttd); 929 PROC_UNLOCK(p); 930 break; 931 case CPU_WHICH_CPUSET: 932 case CPU_WHICH_JAIL: 933 break; 934 case CPU_WHICH_IRQ: 935 return (EINVAL); 936 } 937 switch (uap->level) { 938 case CPU_LEVEL_ROOT: 939 nset = cpuset_refroot(set); 940 cpuset_rel(set); 941 set = nset; 942 break; 943 case CPU_LEVEL_CPUSET: 944 break; 945 case CPU_LEVEL_WHICH: 946 break; 947 } 948 id = set->cs_id; 949 cpuset_rel(set); 950 if (error == 0) 951 error = copyout(&id, uap->setid, sizeof(id)); 952 953 return (error); 954} 955 956#ifndef _SYS_SYSPROTO_H_ 957struct cpuset_getaffinity_args { 958 cpulevel_t level; 959 cpuwhich_t which; 960 id_t id; 961 size_t cpusetsize; 962 cpuset_t *mask; 963}; 964#endif 965int 966sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap) 967{ 968 struct thread *ttd; 969 struct cpuset *nset; 970 struct cpuset *set; 971 struct proc *p; 972 cpuset_t *mask; 973 int error; 974 size_t size; 975 976 if (uap->cpusetsize < sizeof(cpuset_t) || 977 uap->cpusetsize > CPU_MAXSIZE / NBBY) 978 return (ERANGE); 979 size = uap->cpusetsize; 980 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO); 981 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 982 if (error) 983 goto out; 984 switch (uap->level) { 985 case CPU_LEVEL_ROOT: 986 case CPU_LEVEL_CPUSET: 987 switch (uap->which) { 988 case CPU_WHICH_TID: 989 case CPU_WHICH_PID: 990 thread_lock(ttd); 991 set = cpuset_ref(ttd->td_cpuset); 992 thread_unlock(ttd); 993 break; 994 case CPU_WHICH_CPUSET: 995 case CPU_WHICH_JAIL: 996 break; 997 case CPU_WHICH_IRQ: 998 error = EINVAL; 999 goto out; 1000 } 1001 if (uap->level == CPU_LEVEL_ROOT) 1002 nset = cpuset_refroot(set); 1003 else 1004 nset = cpuset_refbase(set); 1005 CPU_COPY(&nset->cs_mask, mask); 1006 cpuset_rel(nset); 1007 break; 1008 case CPU_LEVEL_WHICH: 1009 switch (uap->which) { 1010 case CPU_WHICH_TID: 1011 thread_lock(ttd); 1012 CPU_COPY(&ttd->td_cpuset->cs_mask, mask); 1013 thread_unlock(ttd); 1014 break; 1015 case CPU_WHICH_PID: 1016 FOREACH_THREAD_IN_PROC(p, ttd) { 1017 thread_lock(ttd); 1018 CPU_OR(mask, &ttd->td_cpuset->cs_mask); 1019 thread_unlock(ttd); 1020 } 1021 break; 1022 case CPU_WHICH_CPUSET: 1023 case CPU_WHICH_JAIL: 1024 CPU_COPY(&set->cs_mask, mask); 1025 break; 1026 case CPU_WHICH_IRQ: 1027 error = intr_getaffinity(uap->id, mask); 1028 break; 1029 } 1030 break; 1031 default: 1032 error = EINVAL; 1033 break; 1034 } 1035 if (set) 1036 cpuset_rel(set); 1037 if (p) 1038 PROC_UNLOCK(p); 1039 if (error == 0) 1040 error = copyout(mask, uap->mask, size); 1041out: 1042 free(mask, M_TEMP); 1043 return (error); 1044} 1045 1046#ifndef _SYS_SYSPROTO_H_ 1047struct cpuset_setaffinity_args { 1048 cpulevel_t level; 1049 cpuwhich_t which; 1050 id_t id; 1051 size_t cpusetsize; 1052 const cpuset_t *mask; 1053}; 1054#endif 1055int 1056sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap) 1057{ 1058 struct cpuset *nset; 1059 struct cpuset *set; 1060 struct thread *ttd; 1061 struct proc *p; 1062 cpuset_t *mask; 1063 int error; 1064 1065 if (uap->cpusetsize < sizeof(cpuset_t) || 1066 uap->cpusetsize > CPU_MAXSIZE / NBBY) 1067 return (ERANGE); 1068 mask = malloc(uap->cpusetsize, M_TEMP, M_WAITOK | M_ZERO); 1069 error = copyin(uap->mask, mask, uap->cpusetsize); 1070 if (error) 1071 goto out; 1072 /* 1073 * Verify that no high bits are set. 1074 */ 1075 if (uap->cpusetsize > sizeof(cpuset_t)) { 1076 char *end; 1077 char *cp; 1078 1079 end = cp = (char *)&mask->__bits; 1080 end += uap->cpusetsize; 1081 cp += sizeof(cpuset_t); 1082 while (cp != end) 1083 if (*cp++ != 0) { 1084 error = EINVAL; 1085 goto out; 1086 } 1087 1088 } 1089 switch (uap->level) { 1090 case CPU_LEVEL_ROOT: 1091 case CPU_LEVEL_CPUSET: 1092 error = cpuset_which(uap->which, uap->id, &p, &ttd, &set); 1093 if (error) 1094 break; 1095 switch (uap->which) { 1096 case CPU_WHICH_TID: 1097 case CPU_WHICH_PID: 1098 thread_lock(ttd); 1099 set = cpuset_ref(ttd->td_cpuset); 1100 thread_unlock(ttd); 1101 PROC_UNLOCK(p); 1102 break; 1103 case CPU_WHICH_CPUSET: 1104 case CPU_WHICH_JAIL: 1105 break; 1106 case CPU_WHICH_IRQ: 1107 error = EINVAL; 1108 goto out; 1109 } 1110 if (uap->level == CPU_LEVEL_ROOT) 1111 nset = cpuset_refroot(set); 1112 else 1113 nset = cpuset_refbase(set); 1114 error = cpuset_modify(nset, mask); 1115 cpuset_rel(nset); 1116 cpuset_rel(set); 1117 break; 1118 case CPU_LEVEL_WHICH: 1119 switch (uap->which) { 1120 case CPU_WHICH_TID: 1121 error = cpuset_setthread(uap->id, mask); 1122 break; 1123 case CPU_WHICH_PID: 1124 error = cpuset_setproc(uap->id, NULL, mask); 1125 break; 1126 case CPU_WHICH_CPUSET: 1127 case CPU_WHICH_JAIL: 1128 error = cpuset_which(uap->which, uap->id, &p, 1129 &ttd, &set); 1130 if (error == 0) { 1131 error = cpuset_modify(set, mask); 1132 cpuset_rel(set); 1133 } 1134 break; 1135 case CPU_WHICH_IRQ: 1136 error = intr_setaffinity(uap->id, mask); 1137 break; 1138 default: 1139 error = EINVAL; 1140 break; 1141 } 1142 break; 1143 default: 1144 error = EINVAL; 1145 break; 1146 } 1147out: 1148 free(mask, M_TEMP); 1149 return (error); 1150} 1151 1152#ifdef DDB 1153DB_SHOW_COMMAND(cpusets, db_show_cpusets) 1154{ 1155 struct cpuset *set; 1156 int cpu, once; 1157 1158 LIST_FOREACH(set, &cpuset_ids, cs_link) { 1159 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n", 1160 set, set->cs_id, set->cs_ref, set->cs_flags, 1161 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0); 1162 db_printf(" mask="); 1163 for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) { 1164 if (CPU_ISSET(cpu, &set->cs_mask)) { 1165 if (once == 0) { 1166 db_printf("%d", cpu); 1167 once = 1; 1168 } else 1169 db_printf(",%d", cpu); 1170 } 1171 } 1172 db_printf("\n"); 1173 if (db_pager_quit) 1174 break; 1175 } 1176} 1177#endif /* DDB */ 1178