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: stable/11/sys/kern/kern_cpuset.c 333338 2018-05-07 21:42:22Z shurd $"); 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/capsicum.h> 51#include <sys/cpuset.h> 52#include <sys/sx.h> 53#include <sys/queue.h> 54#include <sys/libkern.h> 55#include <sys/limits.h> 56#include <sys/bus.h> 57#include <sys/interrupt.h> 58 59#include <vm/uma.h> 60#include <vm/vm.h> 61#include <vm/vm_page.h> 62#include <vm/vm_param.h> 63#include <vm/vm_phys.h> 64 65#ifdef DDB 66#include <ddb/ddb.h> 67#endif /* DDB */ 68 69/* 70 * cpusets provide a mechanism for creating and manipulating sets of 71 * processors for the purpose of constraining the scheduling of threads to 72 * specific processors. 73 * 74 * Each process belongs to an identified set, by default this is set 1. Each 75 * thread may further restrict the cpus it may run on to a subset of this 76 * named set. This creates an anonymous set which other threads and processes 77 * may not join by number. 78 * 79 * The named set is referred to herein as the 'base' set to avoid ambiguity. 80 * This set is usually a child of a 'root' set while the anonymous set may 81 * simply be referred to as a mask. In the syscall api these are referred to 82 * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here. 83 * 84 * Threads inherit their set from their creator whether it be anonymous or 85 * not. This means that anonymous sets are immutable because they may be 86 * shared. To modify an anonymous set a new set is created with the desired 87 * mask and the same parent as the existing anonymous set. This gives the 88 * illusion of each thread having a private mask. 89 * 90 * Via the syscall apis a user may ask to retrieve or modify the root, base, 91 * or mask that is discovered via a pid, tid, or setid. Modifying a set 92 * modifies all numbered and anonymous child sets to comply with the new mask. 93 * Modifying a pid or tid's mask applies only to that tid but must still 94 * exist within the assigned parent set. 95 * 96 * A thread may not be assigned to a group separate from other threads in 97 * the process. This is to remove ambiguity when the setid is queried with 98 * a pid argument. There is no other technical limitation. 99 * 100 * This somewhat complex arrangement is intended to make it easy for 101 * applications to query available processors and bind their threads to 102 * specific processors while also allowing administrators to dynamically 103 * reprovision by changing sets which apply to groups of processes. 104 * 105 * A simple application should not concern itself with sets at all and 106 * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id 107 * meaning 'curthread'. It may query available cpus for that tid with a 108 * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...). 109 */ 110static uma_zone_t cpuset_zone; 111static struct mtx cpuset_lock; 112static struct setlist cpuset_ids; 113static struct unrhdr *cpuset_unr; 114static struct cpuset *cpuset_zero, *cpuset_default; 115 116/* Return the size of cpuset_t at the kernel level */ 117SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD, 118 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)"); 119 120cpuset_t *cpuset_root; 121cpuset_t cpuset_domain[MAXMEMDOM]; 122 123/* 124 * Acquire a reference to a cpuset, all pointers must be tracked with refs. 125 */ 126struct cpuset * 127cpuset_ref(struct cpuset *set) 128{ 129 130 refcount_acquire(&set->cs_ref); 131 return (set); 132} 133 134/* 135 * Walks up the tree from 'set' to find the root. Returns the root 136 * referenced. 137 */ 138static struct cpuset * 139cpuset_refroot(struct cpuset *set) 140{ 141 142 for (; set->cs_parent != NULL; set = set->cs_parent) 143 if (set->cs_flags & CPU_SET_ROOT) 144 break; 145 cpuset_ref(set); 146 147 return (set); 148} 149 150/* 151 * Find the first non-anonymous set starting from 'set'. Returns this set 152 * referenced. May return the passed in set with an extra ref if it is 153 * not anonymous. 154 */ 155static struct cpuset * 156cpuset_refbase(struct cpuset *set) 157{ 158 159 if (set->cs_id == CPUSET_INVALID) 160 set = set->cs_parent; 161 cpuset_ref(set); 162 163 return (set); 164} 165 166/* 167 * Release a reference in a context where it is safe to allocate. 168 */ 169void 170cpuset_rel(struct cpuset *set) 171{ 172 cpusetid_t id; 173 174 if (refcount_release(&set->cs_ref) == 0) 175 return; 176 mtx_lock_spin(&cpuset_lock); 177 LIST_REMOVE(set, cs_siblings); 178 id = set->cs_id; 179 if (id != CPUSET_INVALID) 180 LIST_REMOVE(set, cs_link); 181 mtx_unlock_spin(&cpuset_lock); 182 cpuset_rel(set->cs_parent); 183 uma_zfree(cpuset_zone, set); 184 if (id != CPUSET_INVALID) 185 free_unr(cpuset_unr, id); 186} 187 188/* 189 * Deferred release must be used when in a context that is not safe to 190 * allocate/free. This places any unreferenced sets on the list 'head'. 191 */ 192static void 193cpuset_rel_defer(struct setlist *head, struct cpuset *set) 194{ 195 196 if (refcount_release(&set->cs_ref) == 0) 197 return; 198 mtx_lock_spin(&cpuset_lock); 199 LIST_REMOVE(set, cs_siblings); 200 if (set->cs_id != CPUSET_INVALID) 201 LIST_REMOVE(set, cs_link); 202 LIST_INSERT_HEAD(head, set, cs_link); 203 mtx_unlock_spin(&cpuset_lock); 204} 205 206/* 207 * Complete a deferred release. Removes the set from the list provided to 208 * cpuset_rel_defer. 209 */ 210static void 211cpuset_rel_complete(struct cpuset *set) 212{ 213 LIST_REMOVE(set, cs_link); 214 cpuset_rel(set->cs_parent); 215 uma_zfree(cpuset_zone, set); 216} 217 218/* 219 * Find a set based on an id. Returns it with a ref. 220 */ 221static struct cpuset * 222cpuset_lookup(cpusetid_t setid, struct thread *td) 223{ 224 struct cpuset *set; 225 226 if (setid == CPUSET_INVALID) 227 return (NULL); 228 mtx_lock_spin(&cpuset_lock); 229 LIST_FOREACH(set, &cpuset_ids, cs_link) 230 if (set->cs_id == setid) 231 break; 232 if (set) 233 cpuset_ref(set); 234 mtx_unlock_spin(&cpuset_lock); 235 236 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__)); 237 if (set != NULL && jailed(td->td_ucred)) { 238 struct cpuset *jset, *tset; 239 240 jset = td->td_ucred->cr_prison->pr_cpuset; 241 for (tset = set; tset != NULL; tset = tset->cs_parent) 242 if (tset == jset) 243 break; 244 if (tset == NULL) { 245 cpuset_rel(set); 246 set = NULL; 247 } 248 } 249 250 return (set); 251} 252 253/* 254 * Create a set in the space provided in 'set' with the provided parameters. 255 * The set is returned with a single ref. May return EDEADLK if the set 256 * will have no valid cpu based on restrictions from the parent. 257 */ 258static int 259_cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask, 260 cpusetid_t id) 261{ 262 263 if (!CPU_OVERLAP(&parent->cs_mask, mask)) 264 return (EDEADLK); 265 CPU_COPY(mask, &set->cs_mask); 266 LIST_INIT(&set->cs_children); 267 refcount_init(&set->cs_ref, 1); 268 set->cs_flags = 0; 269 mtx_lock_spin(&cpuset_lock); 270 CPU_AND(&set->cs_mask, &parent->cs_mask); 271 set->cs_id = id; 272 set->cs_parent = cpuset_ref(parent); 273 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings); 274 if (set->cs_id != CPUSET_INVALID) 275 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 276 mtx_unlock_spin(&cpuset_lock); 277 278 return (0); 279} 280 281/* 282 * Create a new non-anonymous set with the requested parent and mask. May 283 * return failures if the mask is invalid or a new number can not be 284 * allocated. 285 */ 286static int 287cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask) 288{ 289 struct cpuset *set; 290 cpusetid_t id; 291 int error; 292 293 id = alloc_unr(cpuset_unr); 294 if (id == -1) 295 return (ENFILE); 296 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK); 297 error = _cpuset_create(set, parent, mask, id); 298 if (error == 0) 299 return (0); 300 free_unr(cpuset_unr, id); 301 uma_zfree(cpuset_zone, set); 302 303 return (error); 304} 305 306/* 307 * Recursively check for errors that would occur from applying mask to 308 * the tree of sets starting at 'set'. Checks for sets that would become 309 * empty as well as RDONLY flags. 310 */ 311static int 312cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask) 313{ 314 struct cpuset *nset; 315 cpuset_t newmask; 316 int error; 317 318 mtx_assert(&cpuset_lock, MA_OWNED); 319 if (set->cs_flags & CPU_SET_RDONLY) 320 return (EPERM); 321 if (check_mask) { 322 if (!CPU_OVERLAP(&set->cs_mask, mask)) 323 return (EDEADLK); 324 CPU_COPY(&set->cs_mask, &newmask); 325 CPU_AND(&newmask, mask); 326 } else 327 CPU_COPY(mask, &newmask); 328 error = 0; 329 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 330 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0) 331 break; 332 return (error); 333} 334 335/* 336 * Applies the mask 'mask' without checking for empty sets or permissions. 337 */ 338static void 339cpuset_update(struct cpuset *set, cpuset_t *mask) 340{ 341 struct cpuset *nset; 342 343 mtx_assert(&cpuset_lock, MA_OWNED); 344 CPU_AND(&set->cs_mask, mask); 345 LIST_FOREACH(nset, &set->cs_children, cs_siblings) 346 cpuset_update(nset, &set->cs_mask); 347 348 return; 349} 350 351/* 352 * Modify the set 'set' to use a copy of the mask provided. Apply this new 353 * mask to restrict all children in the tree. Checks for validity before 354 * applying the changes. 355 */ 356static int 357cpuset_modify(struct cpuset *set, cpuset_t *mask) 358{ 359 struct cpuset *root; 360 int error; 361 362 error = priv_check(curthread, PRIV_SCHED_CPUSET); 363 if (error) 364 return (error); 365 /* 366 * In case we are called from within the jail 367 * we do not allow modifying the dedicated root 368 * cpuset of the jail but may still allow to 369 * change child sets. 370 */ 371 if (jailed(curthread->td_ucred) && 372 set->cs_flags & CPU_SET_ROOT) 373 return (EPERM); 374 /* 375 * Verify that we have access to this set of 376 * cpus. 377 */ 378 root = set->cs_parent; 379 if (root && !CPU_SUBSET(&root->cs_mask, mask)) 380 return (EINVAL); 381 mtx_lock_spin(&cpuset_lock); 382 error = cpuset_testupdate(set, mask, 0); 383 if (error) 384 goto out; 385 CPU_COPY(mask, &set->cs_mask); 386 cpuset_update(set, mask); 387out: 388 mtx_unlock_spin(&cpuset_lock); 389 390 return (error); 391} 392 393/* 394 * Resolve the 'which' parameter of several cpuset apis. 395 * 396 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also 397 * checks for permission via p_cansched(). 398 * 399 * For WHICH_SET returns a valid set with a new reference. 400 * 401 * -1 may be supplied for any argument to mean the current proc/thread or 402 * the base set of the current thread. May fail with ESRCH/EPERM. 403 */ 404int 405cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp, 406 struct cpuset **setp) 407{ 408 struct cpuset *set; 409 struct thread *td; 410 struct proc *p; 411 int error; 412 413 *pp = p = NULL; 414 *tdp = td = NULL; 415 *setp = set = NULL; 416 switch (which) { 417 case CPU_WHICH_PID: 418 if (id == -1) { 419 PROC_LOCK(curproc); 420 p = curproc; 421 break; 422 } 423 if ((p = pfind(id)) == NULL) 424 return (ESRCH); 425 break; 426 case CPU_WHICH_TID: 427 if (id == -1) { 428 PROC_LOCK(curproc); 429 p = curproc; 430 td = curthread; 431 break; 432 } 433 td = tdfind(id, -1); 434 if (td == NULL) 435 return (ESRCH); 436 p = td->td_proc; 437 break; 438 case CPU_WHICH_CPUSET: 439 if (id == -1) { 440 thread_lock(curthread); 441 set = cpuset_refbase(curthread->td_cpuset); 442 thread_unlock(curthread); 443 } else 444 set = cpuset_lookup(id, curthread); 445 if (set) { 446 *setp = set; 447 return (0); 448 } 449 return (ESRCH); 450 case CPU_WHICH_JAIL: 451 { 452 /* Find `set' for prison with given id. */ 453 struct prison *pr; 454 455 sx_slock(&allprison_lock); 456 pr = prison_find_child(curthread->td_ucred->cr_prison, id); 457 sx_sunlock(&allprison_lock); 458 if (pr == NULL) 459 return (ESRCH); 460 cpuset_ref(pr->pr_cpuset); 461 *setp = pr->pr_cpuset; 462 mtx_unlock(&pr->pr_mtx); 463 return (0); 464 } 465 case CPU_WHICH_IRQ: 466 case CPU_WHICH_DOMAIN: 467 return (0); 468 default: 469 return (EINVAL); 470 } 471 error = p_cansched(curthread, p); 472 if (error) { 473 PROC_UNLOCK(p); 474 return (error); 475 } 476 if (td == NULL) 477 td = FIRST_THREAD_IN_PROC(p); 478 *pp = p; 479 *tdp = td; 480 return (0); 481} 482 483/* 484 * Create an anonymous set with the provided mask in the space provided by 485 * 'fset'. If the passed in set is anonymous we use its parent otherwise 486 * the new set is a child of 'set'. 487 */ 488static int 489cpuset_shadow(struct cpuset *set, struct cpuset *fset, const cpuset_t *mask) 490{ 491 struct cpuset *parent; 492 493 if (set->cs_id == CPUSET_INVALID) 494 parent = set->cs_parent; 495 else 496 parent = set; 497 if (!CPU_SUBSET(&parent->cs_mask, mask)) 498 return (EDEADLK); 499 return (_cpuset_create(fset, parent, mask, CPUSET_INVALID)); 500} 501 502/* 503 * Handle two cases for replacing the base set or mask of an entire process. 504 * 505 * 1) Set is non-null and mask is null. This reparents all anonymous sets 506 * to the provided set and replaces all non-anonymous td_cpusets with the 507 * provided set. 508 * 2) Mask is non-null and set is null. This replaces or creates anonymous 509 * sets for every thread with the existing base as a parent. 510 * 511 * This is overly complicated because we can't allocate while holding a 512 * spinlock and spinlocks must be held while changing and examining thread 513 * state. 514 */ 515static int 516cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask) 517{ 518 struct setlist freelist; 519 struct setlist droplist; 520 struct cpuset *tdset; 521 struct cpuset *nset; 522 struct thread *td; 523 struct proc *p; 524 int threads; 525 int nfree; 526 int error; 527 528 /* 529 * The algorithm requires two passes due to locking considerations. 530 * 531 * 1) Lookup the process and acquire the locks in the required order. 532 * 2) If enough cpusets have not been allocated release the locks and 533 * allocate them. Loop. 534 */ 535 LIST_INIT(&freelist); 536 LIST_INIT(&droplist); 537 nfree = 0; 538 for (;;) { 539 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset); 540 if (error) 541 goto out; 542 if (nfree >= p->p_numthreads) 543 break; 544 threads = p->p_numthreads; 545 PROC_UNLOCK(p); 546 for (; nfree < threads; nfree++) { 547 nset = uma_zalloc(cpuset_zone, M_WAITOK); 548 LIST_INSERT_HEAD(&freelist, nset, cs_link); 549 } 550 } 551 PROC_LOCK_ASSERT(p, MA_OWNED); 552 /* 553 * Now that the appropriate locks are held and we have enough cpusets, 554 * make sure the operation will succeed before applying changes. The 555 * proc lock prevents td_cpuset from changing between calls. 556 */ 557 error = 0; 558 FOREACH_THREAD_IN_PROC(p, td) { 559 thread_lock(td); 560 tdset = td->td_cpuset; 561 /* 562 * Verify that a new mask doesn't specify cpus outside of 563 * the set the thread is a member of. 564 */ 565 if (mask) { 566 if (tdset->cs_id == CPUSET_INVALID) 567 tdset = tdset->cs_parent; 568 if (!CPU_SUBSET(&tdset->cs_mask, mask)) 569 error = EDEADLK; 570 /* 571 * Verify that a new set won't leave an existing thread 572 * mask without a cpu to run on. It can, however, restrict 573 * the set. 574 */ 575 } else if (tdset->cs_id == CPUSET_INVALID) { 576 if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask)) 577 error = EDEADLK; 578 } 579 thread_unlock(td); 580 if (error) 581 goto unlock_out; 582 } 583 /* 584 * Replace each thread's cpuset while using deferred release. We 585 * must do this because the thread lock must be held while operating 586 * on the thread and this limits the type of operations allowed. 587 */ 588 FOREACH_THREAD_IN_PROC(p, td) { 589 thread_lock(td); 590 /* 591 * If we presently have an anonymous set or are applying a 592 * mask we must create an anonymous shadow set. That is 593 * either parented to our existing base or the supplied set. 594 * 595 * If we have a base set with no anonymous shadow we simply 596 * replace it outright. 597 */ 598 tdset = td->td_cpuset; 599 if (tdset->cs_id == CPUSET_INVALID || mask) { 600 nset = LIST_FIRST(&freelist); 601 LIST_REMOVE(nset, cs_link); 602 if (mask) 603 error = cpuset_shadow(tdset, nset, mask); 604 else 605 error = _cpuset_create(nset, set, 606 &tdset->cs_mask, CPUSET_INVALID); 607 if (error) { 608 LIST_INSERT_HEAD(&freelist, nset, cs_link); 609 thread_unlock(td); 610 break; 611 } 612 } else 613 nset = cpuset_ref(set); 614 cpuset_rel_defer(&droplist, tdset); 615 td->td_cpuset = nset; 616 sched_affinity(td); 617 thread_unlock(td); 618 } 619unlock_out: 620 PROC_UNLOCK(p); 621out: 622 while ((nset = LIST_FIRST(&droplist)) != NULL) 623 cpuset_rel_complete(nset); 624 while ((nset = LIST_FIRST(&freelist)) != NULL) { 625 LIST_REMOVE(nset, cs_link); 626 uma_zfree(cpuset_zone, nset); 627 } 628 return (error); 629} 630 631/* 632 * Return a string representing a valid layout for a cpuset_t object. 633 * It expects an incoming buffer at least sized as CPUSETBUFSIZ. 634 */ 635char * 636cpusetobj_strprint(char *buf, const cpuset_t *set) 637{ 638 char *tbuf; 639 size_t i, bytesp, bufsiz; 640 641 tbuf = buf; 642 bytesp = 0; 643 bufsiz = CPUSETBUFSIZ; 644 645 for (i = 0; i < (_NCPUWORDS - 1); i++) { 646 bytesp = snprintf(tbuf, bufsiz, "%lx,", set->__bits[i]); 647 bufsiz -= bytesp; 648 tbuf += bytesp; 649 } 650 snprintf(tbuf, bufsiz, "%lx", set->__bits[_NCPUWORDS - 1]); 651 return (buf); 652} 653 654/* 655 * Build a valid cpuset_t object from a string representation. 656 * It expects an incoming buffer at least sized as CPUSETBUFSIZ. 657 */ 658int 659cpusetobj_strscan(cpuset_t *set, const char *buf) 660{ 661 u_int nwords; 662 int i, ret; 663 664 if (strlen(buf) > CPUSETBUFSIZ - 1) 665 return (-1); 666 667 /* Allow to pass a shorter version of the mask when necessary. */ 668 nwords = 1; 669 for (i = 0; buf[i] != '\0'; i++) 670 if (buf[i] == ',') 671 nwords++; 672 if (nwords > _NCPUWORDS) 673 return (-1); 674 675 CPU_ZERO(set); 676 for (i = 0; i < (nwords - 1); i++) { 677 ret = sscanf(buf, "%lx,", &set->__bits[i]); 678 if (ret == 0 || ret == -1) 679 return (-1); 680 buf = strstr(buf, ","); 681 if (buf == NULL) 682 return (-1); 683 buf++; 684 } 685 ret = sscanf(buf, "%lx", &set->__bits[nwords - 1]); 686 if (ret == 0 || ret == -1) 687 return (-1); 688 return (0); 689} 690 691/* 692 * Apply an anonymous mask to a single thread. 693 */ 694int 695cpuset_setthread(lwpid_t id, cpuset_t *mask) 696{ 697 struct cpuset *nset; 698 struct cpuset *set; 699 struct thread *td; 700 struct proc *p; 701 int error; 702 703 nset = uma_zalloc(cpuset_zone, M_WAITOK); 704 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set); 705 if (error) 706 goto out; 707 set = NULL; 708 thread_lock(td); 709 error = cpuset_shadow(td->td_cpuset, nset, mask); 710 if (error == 0) { 711 set = td->td_cpuset; 712 td->td_cpuset = nset; 713 sched_affinity(td); 714 nset = NULL; 715 } 716 thread_unlock(td); 717 PROC_UNLOCK(p); 718 if (set) 719 cpuset_rel(set); 720out: 721 if (nset) 722 uma_zfree(cpuset_zone, nset); 723 return (error); 724} 725 726/* 727 * Apply new cpumask to the ithread. 728 */ 729int 730cpuset_setithread(lwpid_t id, int cpu) 731{ 732 struct cpuset *nset, *rset; 733 struct cpuset *parent, *old_set; 734 struct thread *td; 735 struct proc *p; 736 cpusetid_t cs_id; 737 cpuset_t mask; 738 int error; 739 740 nset = uma_zalloc(cpuset_zone, M_WAITOK); 741 rset = uma_zalloc(cpuset_zone, M_WAITOK); 742 cs_id = CPUSET_INVALID; 743 744 CPU_ZERO(&mask); 745 if (cpu == NOCPU) 746 CPU_COPY(cpuset_root, &mask); 747 else 748 CPU_SET(cpu, &mask); 749 750 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &old_set); 751 if (error != 0 || ((cs_id = alloc_unr(cpuset_unr)) == CPUSET_INVALID)) 752 goto out; 753 754 /* cpuset_which() returns with PROC_LOCK held. */ 755 old_set = td->td_cpuset; 756 757 if (cpu == NOCPU) { 758 759 /* 760 * roll back to default set. We're not using cpuset_shadow() 761 * here because we can fail CPU_SUBSET() check. This can happen 762 * if default set does not contain all CPUs. 763 */ 764 error = _cpuset_create(nset, cpuset_default, &mask, 765 CPUSET_INVALID); 766 767 goto applyset; 768 } 769 770 if (old_set->cs_id == 1 || (old_set->cs_id == CPUSET_INVALID && 771 old_set->cs_parent->cs_id == 1)) { 772 773 /* 774 * Current set is either default (1) or 775 * shadowed version of default set. 776 * 777 * Allocate new root set to be able to shadow it 778 * with any mask. 779 */ 780 error = _cpuset_create(rset, cpuset_zero, 781 &cpuset_zero->cs_mask, cs_id); 782 if (error != 0) { 783 PROC_UNLOCK(p); 784 goto out; 785 } 786 rset->cs_flags |= CPU_SET_ROOT; 787 parent = rset; 788 rset = NULL; 789 cs_id = CPUSET_INVALID; 790 } else { 791 /* Assume existing set was already allocated by previous call */ 792 parent = old_set; 793 old_set = NULL; 794 } 795 796 error = cpuset_shadow(parent, nset, &mask); 797applyset: 798 if (error == 0) { 799 thread_lock(td); 800 td->td_cpuset = nset; 801 sched_affinity(td); 802 thread_unlock(td); 803 nset = NULL; 804 } else 805 old_set = NULL; 806 PROC_UNLOCK(p); 807 if (old_set != NULL) 808 cpuset_rel(old_set); 809out: 810 if (nset != NULL) 811 uma_zfree(cpuset_zone, nset); 812 if (rset != NULL) 813 uma_zfree(cpuset_zone, rset); 814 if (cs_id != CPUSET_INVALID) 815 free_unr(cpuset_unr, cs_id); 816 return (error); 817} 818 819 820/* 821 * Creates system-wide cpusets and the cpuset for thread0 including two 822 * sets: 823 * 824 * 0 - The root set which should represent all valid processors in the 825 * system. It is initially created with a mask of all processors 826 * because we don't know what processors are valid until cpuset_init() 827 * runs. This set is immutable. 828 * 1 - The default set which all processes are a member of until changed. 829 * This allows an administrator to move all threads off of given cpus to 830 * dedicate them to high priority tasks or save power etc. 831 */ 832struct cpuset * 833cpuset_thread0(void) 834{ 835 struct cpuset *set; 836 int error, i; 837 838 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL, 839 NULL, NULL, UMA_ALIGN_PTR, 0); 840 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE); 841 842 /* 843 * Create the root system set for the whole machine. Doesn't use 844 * cpuset_create() due to NULL parent. 845 */ 846 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); 847 CPU_FILL(&set->cs_mask); 848 LIST_INIT(&set->cs_children); 849 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); 850 set->cs_ref = 1; 851 set->cs_flags = CPU_SET_ROOT; 852 cpuset_zero = set; 853 cpuset_root = &set->cs_mask; 854 855 /* 856 * Now derive a default, modifiable set from that to give out. 857 */ 858 set = uma_zalloc(cpuset_zone, M_WAITOK); 859 error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1); 860 KASSERT(error == 0, ("Error creating default set: %d\n", error)); 861 cpuset_default = set; 862 863 /* 864 * Initialize the unit allocator. 0 and 1 are allocated above. 865 */ 866 cpuset_unr = new_unrhdr(2, INT_MAX, NULL); 867 868 /* 869 * If MD code has not initialized per-domain cpusets, place all 870 * CPUs in domain 0. 871 */ 872 for (i = 0; i < MAXMEMDOM; i++) 873 if (!CPU_EMPTY(&cpuset_domain[i])) 874 goto domains_set; 875 CPU_COPY(&all_cpus, &cpuset_domain[0]); 876domains_set: 877 878 return (set); 879} 880 881/* 882 * Create a cpuset, which would be cpuset_create() but 883 * mark the new 'set' as root. 884 * 885 * We are not going to reparent the td to it. Use cpuset_setproc_update_set() 886 * for that. 887 * 888 * In case of no error, returns the set in *setp locked with a reference. 889 */ 890int 891cpuset_create_root(struct prison *pr, struct cpuset **setp) 892{ 893 struct cpuset *set; 894 int error; 895 896 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__)); 897 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__)); 898 899 error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask); 900 if (error) 901 return (error); 902 903 KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data", 904 __func__, __LINE__)); 905 906 /* Mark the set as root. */ 907 set = *setp; 908 set->cs_flags |= CPU_SET_ROOT; 909 910 return (0); 911} 912 913int 914cpuset_setproc_update_set(struct proc *p, struct cpuset *set) 915{ 916 int error; 917 918 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__)); 919 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__)); 920 921 cpuset_ref(set); 922 error = cpuset_setproc(p->p_pid, set, NULL); 923 if (error) 924 return (error); 925 cpuset_rel(set); 926 return (0); 927} 928 929/* 930 * This is called once the final set of system cpus is known. Modifies 931 * the root set and all children and mark the root read-only. 932 */ 933static void 934cpuset_init(void *arg) 935{ 936 cpuset_t mask; 937 938 mask = all_cpus; 939 if (cpuset_modify(cpuset_zero, &mask)) 940 panic("Can't set initial cpuset mask.\n"); 941 cpuset_zero->cs_flags |= CPU_SET_RDONLY; 942} 943SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL); 944 945#ifndef _SYS_SYSPROTO_H_ 946struct cpuset_args { 947 cpusetid_t *setid; 948}; 949#endif 950int 951sys_cpuset(struct thread *td, struct cpuset_args *uap) 952{ 953 struct cpuset *root; 954 struct cpuset *set; 955 int error; 956 957 thread_lock(td); 958 root = cpuset_refroot(td->td_cpuset); 959 thread_unlock(td); 960 error = cpuset_create(&set, root, &root->cs_mask); 961 cpuset_rel(root); 962 if (error) 963 return (error); 964 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id)); 965 if (error == 0) 966 error = cpuset_setproc(-1, set, NULL); 967 cpuset_rel(set); 968 return (error); 969} 970 971#ifndef _SYS_SYSPROTO_H_ 972struct cpuset_setid_args { 973 cpuwhich_t which; 974 id_t id; 975 cpusetid_t setid; 976}; 977#endif 978int 979sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap) 980{ 981 982 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid)); 983} 984 985int 986kern_cpuset_setid(struct thread *td, cpuwhich_t which, 987 id_t id, cpusetid_t setid) 988{ 989 struct cpuset *set; 990 int error; 991 992 /* 993 * Presently we only support per-process sets. 994 */ 995 if (which != CPU_WHICH_PID) 996 return (EINVAL); 997 set = cpuset_lookup(setid, td); 998 if (set == NULL) 999 return (ESRCH); 1000 error = cpuset_setproc(id, set, NULL); 1001 cpuset_rel(set); 1002 return (error); 1003} 1004 1005#ifndef _SYS_SYSPROTO_H_ 1006struct cpuset_getid_args { 1007 cpulevel_t level; 1008 cpuwhich_t which; 1009 id_t id; 1010 cpusetid_t *setid; 1011}; 1012#endif 1013int 1014sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap) 1015{ 1016 1017 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id, 1018 uap->setid)); 1019} 1020 1021int 1022kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which, 1023 id_t id, cpusetid_t *setid) 1024{ 1025 struct cpuset *nset; 1026 struct cpuset *set; 1027 struct thread *ttd; 1028 struct proc *p; 1029 cpusetid_t tmpid; 1030 int error; 1031 1032 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET) 1033 return (EINVAL); 1034 error = cpuset_which(which, id, &p, &ttd, &set); 1035 if (error) 1036 return (error); 1037 switch (which) { 1038 case CPU_WHICH_TID: 1039 case CPU_WHICH_PID: 1040 thread_lock(ttd); 1041 set = cpuset_refbase(ttd->td_cpuset); 1042 thread_unlock(ttd); 1043 PROC_UNLOCK(p); 1044 break; 1045 case CPU_WHICH_CPUSET: 1046 case CPU_WHICH_JAIL: 1047 break; 1048 case CPU_WHICH_IRQ: 1049 case CPU_WHICH_DOMAIN: 1050 return (EINVAL); 1051 } 1052 switch (level) { 1053 case CPU_LEVEL_ROOT: 1054 nset = cpuset_refroot(set); 1055 cpuset_rel(set); 1056 set = nset; 1057 break; 1058 case CPU_LEVEL_CPUSET: 1059 break; 1060 case CPU_LEVEL_WHICH: 1061 break; 1062 } 1063 tmpid = set->cs_id; 1064 cpuset_rel(set); 1065 if (error == 0) 1066 error = copyout(&tmpid, setid, sizeof(tmpid)); 1067 1068 return (error); 1069} 1070 1071#ifndef _SYS_SYSPROTO_H_ 1072struct cpuset_getaffinity_args { 1073 cpulevel_t level; 1074 cpuwhich_t which; 1075 id_t id; 1076 size_t cpusetsize; 1077 cpuset_t *mask; 1078}; 1079#endif 1080int 1081sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap) 1082{ 1083 1084 return (kern_cpuset_getaffinity(td, uap->level, uap->which, 1085 uap->id, uap->cpusetsize, uap->mask)); 1086} 1087 1088int 1089kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, 1090 id_t id, size_t cpusetsize, cpuset_t *maskp) 1091{ 1092 struct thread *ttd; 1093 struct cpuset *nset; 1094 struct cpuset *set; 1095 struct proc *p; 1096 cpuset_t *mask; 1097 int error; 1098 size_t size; 1099 1100 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) 1101 return (ERANGE); 1102 /* In Capability mode, you can only get your own CPU set. */ 1103 if (IN_CAPABILITY_MODE(td)) { 1104 if (level != CPU_LEVEL_WHICH) 1105 return (ECAPMODE); 1106 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) 1107 return (ECAPMODE); 1108 if (id != -1) 1109 return (ECAPMODE); 1110 } 1111 size = cpusetsize; 1112 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO); 1113 error = cpuset_which(which, id, &p, &ttd, &set); 1114 if (error) 1115 goto out; 1116 switch (level) { 1117 case CPU_LEVEL_ROOT: 1118 case CPU_LEVEL_CPUSET: 1119 switch (which) { 1120 case CPU_WHICH_TID: 1121 case CPU_WHICH_PID: 1122 thread_lock(ttd); 1123 set = cpuset_ref(ttd->td_cpuset); 1124 thread_unlock(ttd); 1125 break; 1126 case CPU_WHICH_CPUSET: 1127 case CPU_WHICH_JAIL: 1128 break; 1129 case CPU_WHICH_IRQ: 1130 case CPU_WHICH_INTRHANDLER: 1131 case CPU_WHICH_ITHREAD: 1132 case CPU_WHICH_DOMAIN: 1133 error = EINVAL; 1134 goto out; 1135 } 1136 if (level == CPU_LEVEL_ROOT) 1137 nset = cpuset_refroot(set); 1138 else 1139 nset = cpuset_refbase(set); 1140 CPU_COPY(&nset->cs_mask, mask); 1141 cpuset_rel(nset); 1142 break; 1143 case CPU_LEVEL_WHICH: 1144 switch (which) { 1145 case CPU_WHICH_TID: 1146 thread_lock(ttd); 1147 CPU_COPY(&ttd->td_cpuset->cs_mask, mask); 1148 thread_unlock(ttd); 1149 break; 1150 case CPU_WHICH_PID: 1151 FOREACH_THREAD_IN_PROC(p, ttd) { 1152 thread_lock(ttd); 1153 CPU_OR(mask, &ttd->td_cpuset->cs_mask); 1154 thread_unlock(ttd); 1155 } 1156 break; 1157 case CPU_WHICH_CPUSET: 1158 case CPU_WHICH_JAIL: 1159 CPU_COPY(&set->cs_mask, mask); 1160 break; 1161 case CPU_WHICH_IRQ: 1162 case CPU_WHICH_INTRHANDLER: 1163 case CPU_WHICH_ITHREAD: 1164 error = intr_getaffinity(id, which, mask); 1165 break; 1166 case CPU_WHICH_DOMAIN: 1167 if (id < 0 || id >= MAXMEMDOM) 1168 error = ESRCH; 1169 else 1170 CPU_COPY(&cpuset_domain[id], mask); 1171 break; 1172 } 1173 break; 1174 default: 1175 error = EINVAL; 1176 break; 1177 } 1178 if (set) 1179 cpuset_rel(set); 1180 if (p) 1181 PROC_UNLOCK(p); 1182 if (error == 0) 1183 error = copyout(mask, maskp, size); 1184out: 1185 free(mask, M_TEMP); 1186 return (error); 1187} 1188 1189#ifndef _SYS_SYSPROTO_H_ 1190struct cpuset_setaffinity_args { 1191 cpulevel_t level; 1192 cpuwhich_t which; 1193 id_t id; 1194 size_t cpusetsize; 1195 const cpuset_t *mask; 1196}; 1197#endif 1198int 1199sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap) 1200{ 1201 1202 return (kern_cpuset_setaffinity(td, uap->level, uap->which, 1203 uap->id, uap->cpusetsize, uap->mask)); 1204} 1205 1206int 1207kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, 1208 id_t id, size_t cpusetsize, const cpuset_t *maskp) 1209{ 1210 struct cpuset *nset; 1211 struct cpuset *set; 1212 struct thread *ttd; 1213 struct proc *p; 1214 cpuset_t *mask; 1215 int error; 1216 1217 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) 1218 return (ERANGE); 1219 /* In Capability mode, you can only set your own CPU set. */ 1220 if (IN_CAPABILITY_MODE(td)) { 1221 if (level != CPU_LEVEL_WHICH) 1222 return (ECAPMODE); 1223 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) 1224 return (ECAPMODE); 1225 if (id != -1) 1226 return (ECAPMODE); 1227 } 1228 mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO); 1229 error = copyin(maskp, mask, cpusetsize); 1230 if (error) 1231 goto out; 1232 /* 1233 * Verify that no high bits are set. 1234 */ 1235 if (cpusetsize > sizeof(cpuset_t)) { 1236 char *end; 1237 char *cp; 1238 1239 end = cp = (char *)&mask->__bits; 1240 end += cpusetsize; 1241 cp += sizeof(cpuset_t); 1242 while (cp != end) 1243 if (*cp++ != 0) { 1244 error = EINVAL; 1245 goto out; 1246 } 1247 1248 } 1249 switch (level) { 1250 case CPU_LEVEL_ROOT: 1251 case CPU_LEVEL_CPUSET: 1252 error = cpuset_which(which, id, &p, &ttd, &set); 1253 if (error) 1254 break; 1255 switch (which) { 1256 case CPU_WHICH_TID: 1257 case CPU_WHICH_PID: 1258 thread_lock(ttd); 1259 set = cpuset_ref(ttd->td_cpuset); 1260 thread_unlock(ttd); 1261 PROC_UNLOCK(p); 1262 break; 1263 case CPU_WHICH_CPUSET: 1264 case CPU_WHICH_JAIL: 1265 break; 1266 case CPU_WHICH_IRQ: 1267 case CPU_WHICH_INTRHANDLER: 1268 case CPU_WHICH_ITHREAD: 1269 case CPU_WHICH_DOMAIN: 1270 error = EINVAL; 1271 goto out; 1272 } 1273 if (level == CPU_LEVEL_ROOT) 1274 nset = cpuset_refroot(set); 1275 else 1276 nset = cpuset_refbase(set); 1277 error = cpuset_modify(nset, mask); 1278 cpuset_rel(nset); 1279 cpuset_rel(set); 1280 break; 1281 case CPU_LEVEL_WHICH: 1282 switch (which) { 1283 case CPU_WHICH_TID: 1284 error = cpuset_setthread(id, mask); 1285 break; 1286 case CPU_WHICH_PID: 1287 error = cpuset_setproc(id, NULL, mask); 1288 break; 1289 case CPU_WHICH_CPUSET: 1290 case CPU_WHICH_JAIL: 1291 error = cpuset_which(which, id, &p, &ttd, &set); 1292 if (error == 0) { 1293 error = cpuset_modify(set, mask); 1294 cpuset_rel(set); 1295 } 1296 break; 1297 case CPU_WHICH_IRQ: 1298 case CPU_WHICH_INTRHANDLER: 1299 case CPU_WHICH_ITHREAD: 1300 error = intr_setaffinity(id, which, mask); 1301 break; 1302 default: 1303 error = EINVAL; 1304 break; 1305 } 1306 break; 1307 default: 1308 error = EINVAL; 1309 break; 1310 } 1311out: 1312 free(mask, M_TEMP); 1313 return (error); 1314} 1315 1316#ifdef DDB 1317void 1318ddb_display_cpuset(const cpuset_t *set) 1319{ 1320 int cpu, once; 1321 1322 for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) { 1323 if (CPU_ISSET(cpu, set)) { 1324 if (once == 0) { 1325 db_printf("%d", cpu); 1326 once = 1; 1327 } else 1328 db_printf(",%d", cpu); 1329 } 1330 } 1331 if (once == 0) 1332 db_printf("<none>"); 1333} 1334 1335DB_SHOW_COMMAND(cpusets, db_show_cpusets) 1336{ 1337 struct cpuset *set; 1338 1339 LIST_FOREACH(set, &cpuset_ids, cs_link) { 1340 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n", 1341 set, set->cs_id, set->cs_ref, set->cs_flags, 1342 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0); 1343 db_printf(" mask="); 1344 ddb_display_cpuset(&set->cs_mask); 1345 db_printf("\n"); 1346 if (db_pager_quit) 1347 break; 1348 } 1349} 1350#endif /* DDB */ 1351