1 /*-
2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3 *
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 *
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
37 *
38 * Permission to use, copy, modify and distribute this software and
39 * its documentation is hereby granted, provided that both the copyright
40 * notice and this permission notice appear in all copies of the
41 * software, derivative works or modified versions, and any portions
42 * thereof, and that both notices appear in supporting documentation.
43 *
44 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
45 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
46 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
47 *
48 * Carnegie Mellon requests users of this software to return to
49 *
50 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
51 * School of Computer Science
52 * Carnegie Mellon University
53 * Pittsburgh PA 15213-3890
54 *
55 * any improvements or extensions that they make and grant Carnegie the
56 * rights to redistribute these changes.
57 */
58
59 #include "opt_vm.h"
60 #include "opt_kstack_pages.h"
61 #include "opt_kstack_max_pages.h"
62 #include "opt_kstack_usage_prof.h"
63
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/asan.h>
67 #include <sys/domainset.h>
68 #include <sys/limits.h>
69 #include <sys/lock.h>
70 #include <sys/malloc.h>
71 #include <sys/msan.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/racct.h>
75 #include <sys/refcount.h>
76 #include <sys/resourcevar.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sf_buf.h>
80 #include <sys/shm.h>
81 #include <sys/smp.h>
82 #include <sys/vmmeter.h>
83 #include <sys/vmem.h>
84 #include <sys/sx.h>
85 #include <sys/sysctl.h>
86 #include <sys/kernel.h>
87 #include <sys/ktr.h>
88 #include <sys/unistd.h>
89
90 #include <vm/uma.h>
91 #include <vm/vm.h>
92 #include <vm/vm_param.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_domainset.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pagequeue.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_radix.h>
102 #include <vm/vm_extern.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_phys.h>
105
106 #include <machine/cpu.h>
107
108 #if VM_NRESERVLEVEL > 1
109 #define KVA_KSTACK_QUANTUM_SHIFT (VM_LEVEL_1_ORDER + VM_LEVEL_0_ORDER + \
110 PAGE_SHIFT)
111 #elif VM_NRESERVLEVEL > 0
112 #define KVA_KSTACK_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
113 #else
114 #define KVA_KSTACK_QUANTUM_SHIFT (8 + PAGE_SHIFT)
115 #endif
116 #define KVA_KSTACK_QUANTUM (1ul << KVA_KSTACK_QUANTUM_SHIFT)
117
118 /*
119 * MPSAFE
120 *
121 * WARNING! This code calls vm_map_check_protection() which only checks
122 * the associated vm_map_entry range. It does not determine whether the
123 * contents of the memory is actually readable or writable. In most cases
124 * just checking the vm_map_entry is sufficient within the kernel's address
125 * space.
126 */
127 bool
kernacc(void * addr,int len,int rw)128 kernacc(void *addr, int len, int rw)
129 {
130 boolean_t rv;
131 vm_offset_t saddr, eaddr;
132 vm_prot_t prot;
133
134 KASSERT((rw & ~VM_PROT_ALL) == 0,
135 ("illegal ``rw'' argument to kernacc (%x)\n", rw));
136
137 if ((vm_offset_t)addr + len > vm_map_max(kernel_map) ||
138 (vm_offset_t)addr + len < (vm_offset_t)addr)
139 return (false);
140
141 prot = rw;
142 saddr = trunc_page((vm_offset_t)addr);
143 eaddr = round_page((vm_offset_t)addr + len);
144 vm_map_lock_read(kernel_map);
145 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
146 vm_map_unlock_read(kernel_map);
147 return (rv == TRUE);
148 }
149
150 /*
151 * MPSAFE
152 *
153 * WARNING! This code calls vm_map_check_protection() which only checks
154 * the associated vm_map_entry range. It does not determine whether the
155 * contents of the memory is actually readable or writable. vmapbuf(),
156 * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
157 * used in conjunction with this call.
158 */
159 bool
useracc(void * addr,int len,int rw)160 useracc(void *addr, int len, int rw)
161 {
162 boolean_t rv;
163 vm_prot_t prot;
164 vm_map_t map;
165
166 KASSERT((rw & ~VM_PROT_ALL) == 0,
167 ("illegal ``rw'' argument to useracc (%x)\n", rw));
168 prot = rw;
169 map = &curproc->p_vmspace->vm_map;
170 if ((vm_offset_t)addr + len > vm_map_max(map) ||
171 (vm_offset_t)addr + len < (vm_offset_t)addr) {
172 return (false);
173 }
174 vm_map_lock_read(map);
175 rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
176 round_page((vm_offset_t)addr + len), prot);
177 vm_map_unlock_read(map);
178 return (rv == TRUE);
179 }
180
181 int
vslock(void * addr,size_t len)182 vslock(void *addr, size_t len)
183 {
184 vm_offset_t end, last, start;
185 vm_size_t npages;
186 int error;
187
188 last = (vm_offset_t)addr + len;
189 start = trunc_page((vm_offset_t)addr);
190 end = round_page(last);
191 if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
192 return (EINVAL);
193 npages = atop(end - start);
194 if (npages > vm_page_max_user_wired)
195 return (ENOMEM);
196 error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
197 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
198 if (error == KERN_SUCCESS) {
199 curthread->td_vslock_sz += len;
200 return (0);
201 }
202
203 /*
204 * Return EFAULT on error to match copy{in,out}() behaviour
205 * rather than returning ENOMEM like mlock() would.
206 */
207 return (EFAULT);
208 }
209
210 void
vsunlock(void * addr,size_t len)211 vsunlock(void *addr, size_t len)
212 {
213
214 /* Rely on the parameter sanity checks performed by vslock(). */
215 MPASS(curthread->td_vslock_sz >= len);
216 curthread->td_vslock_sz -= len;
217 (void)vm_map_unwire(&curproc->p_vmspace->vm_map,
218 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
219 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
220 }
221
222 /*
223 * Pin the page contained within the given object at the given offset. If the
224 * page is not resident, allocate and load it using the given object's pager.
225 * Return the pinned page if successful; otherwise, return NULL.
226 */
227 static vm_page_t
vm_imgact_hold_page(vm_object_t object,vm_ooffset_t offset)228 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
229 {
230 vm_page_t m;
231 vm_pindex_t pindex;
232
233 pindex = OFF_TO_IDX(offset);
234 (void)vm_page_grab_valid_unlocked(&m, object, pindex,
235 VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
236 return (m);
237 }
238
239 /*
240 * Return a CPU private mapping to the page at the given offset within the
241 * given object. The page is pinned before it is mapped.
242 */
243 struct sf_buf *
vm_imgact_map_page(vm_object_t object,vm_ooffset_t offset)244 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
245 {
246 vm_page_t m;
247
248 m = vm_imgact_hold_page(object, offset);
249 if (m == NULL)
250 return (NULL);
251 sched_pin();
252 return (sf_buf_alloc(m, SFB_CPUPRIVATE));
253 }
254
255 /*
256 * Destroy the given CPU private mapping and unpin the page that it mapped.
257 */
258 void
vm_imgact_unmap_page(struct sf_buf * sf)259 vm_imgact_unmap_page(struct sf_buf *sf)
260 {
261 vm_page_t m;
262
263 m = sf_buf_page(sf);
264 sf_buf_free(sf);
265 sched_unpin();
266 vm_page_unwire(m, PQ_ACTIVE);
267 }
268
269 void
vm_sync_icache(vm_map_t map,vm_offset_t va,vm_offset_t sz)270 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
271 {
272
273 pmap_sync_icache(map->pmap, va, sz);
274 }
275
276 static vm_object_t kstack_object;
277 static vm_object_t kstack_alt_object;
278 static uma_zone_t kstack_cache;
279 static int kstack_cache_size;
280 static vmem_t *vmd_kstack_arena[MAXMEMDOM];
281
282 static vm_pindex_t vm_kstack_pindex(vm_offset_t ks, int npages);
283 static vm_object_t vm_thread_kstack_size_to_obj(int npages);
284 static int vm_thread_stack_back(vm_offset_t kaddr, vm_page_t ma[], int npages,
285 int req_class, int domain);
286
287 static int
sysctl_kstack_cache_size(SYSCTL_HANDLER_ARGS)288 sysctl_kstack_cache_size(SYSCTL_HANDLER_ARGS)
289 {
290 int error, oldsize;
291
292 oldsize = kstack_cache_size;
293 error = sysctl_handle_int(oidp, arg1, arg2, req);
294 if (error == 0 && req->newptr && oldsize != kstack_cache_size)
295 uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
296 return (error);
297 }
298 SYSCTL_PROC(_vm, OID_AUTO, kstack_cache_size,
299 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &kstack_cache_size, 0,
300 sysctl_kstack_cache_size, "IU", "Maximum number of cached kernel stacks");
301
302 /*
303 * Allocate a virtual address range from a domain kstack arena, following
304 * the specified NUMA policy.
305 */
306 static vm_offset_t
vm_thread_alloc_kstack_kva(vm_size_t size,int domain)307 vm_thread_alloc_kstack_kva(vm_size_t size, int domain)
308 {
309 #ifndef __ILP32__
310 int rv;
311 vmem_t *arena;
312 vm_offset_t addr = 0;
313
314 size = round_page(size);
315 /* Allocate from the kernel arena for non-standard kstack sizes. */
316 if (size != ptoa(kstack_pages + KSTACK_GUARD_PAGES)) {
317 arena = vm_dom[domain].vmd_kernel_arena;
318 } else {
319 arena = vmd_kstack_arena[domain];
320 }
321 rv = vmem_alloc(arena, size, M_BESTFIT | M_NOWAIT, &addr);
322 if (rv == ENOMEM)
323 return (0);
324 KASSERT(atop(addr - VM_MIN_KERNEL_ADDRESS) %
325 (kstack_pages + KSTACK_GUARD_PAGES) == 0,
326 ("%s: allocated kstack KVA not aligned to multiple of kstack size",
327 __func__));
328
329 return (addr);
330 #else
331 return (kva_alloc(size));
332 #endif
333 }
334
335 /*
336 * Release a region of kernel virtual memory
337 * allocated from the kstack arena.
338 */
339 static __noinline void
vm_thread_free_kstack_kva(vm_offset_t addr,vm_size_t size,int domain)340 vm_thread_free_kstack_kva(vm_offset_t addr, vm_size_t size, int domain)
341 {
342 vmem_t *arena;
343
344 size = round_page(size);
345 #ifdef __ILP32__
346 arena = kernel_arena;
347 #else
348 arena = vmd_kstack_arena[domain];
349 if (size != ptoa(kstack_pages + KSTACK_GUARD_PAGES)) {
350 arena = vm_dom[domain].vmd_kernel_arena;
351 }
352 #endif
353 vmem_free(arena, addr, size);
354 }
355
356 static vmem_size_t
vm_thread_kstack_import_quantum(void)357 vm_thread_kstack_import_quantum(void)
358 {
359 #ifndef __ILP32__
360 /*
361 * The kstack_quantum is larger than KVA_QUANTUM to account
362 * for holes induced by guard pages.
363 */
364 return (KVA_KSTACK_QUANTUM * (kstack_pages + KSTACK_GUARD_PAGES));
365 #else
366 return (KVA_KSTACK_QUANTUM);
367 #endif
368 }
369
370 /*
371 * Import KVA from a parent arena into the kstack arena. Imports must be
372 * a multiple of kernel stack pages + guard pages in size.
373 *
374 * Kstack VA allocations need to be aligned so that the linear KVA pindex
375 * is divisible by the total number of kstack VA pages. This is necessary to
376 * make vm_kstack_pindex work properly.
377 *
378 * We import a multiple of KVA_KSTACK_QUANTUM-sized region from the parent
379 * arena. The actual size used by the kstack arena is one kstack smaller to
380 * allow for the necessary alignment adjustments to be made.
381 */
382 static int
vm_thread_kstack_arena_import(void * arena,vmem_size_t size,int flags,vmem_addr_t * addrp)383 vm_thread_kstack_arena_import(void *arena, vmem_size_t size, int flags,
384 vmem_addr_t *addrp)
385 {
386 int error, rem;
387 size_t kpages = kstack_pages + KSTACK_GUARD_PAGES;
388
389 KASSERT(atop(size) % kpages == 0,
390 ("%s: Size %jd is not a multiple of kstack pages (%d)", __func__,
391 (intmax_t)size, (int)kpages));
392
393 error = vmem_xalloc(arena, vm_thread_kstack_import_quantum(),
394 KVA_KSTACK_QUANTUM, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags,
395 addrp);
396 if (error) {
397 return (error);
398 }
399
400 rem = atop(*addrp - VM_MIN_KERNEL_ADDRESS) % kpages;
401 if (rem != 0) {
402 /* Bump addr to next aligned address */
403 *addrp = *addrp + (kpages - rem) * PAGE_SIZE;
404 }
405
406 return (0);
407 }
408
409 /*
410 * Release KVA from a parent arena into the kstack arena. Released imports must
411 * be a multiple of kernel stack pages + guard pages in size.
412 */
413 static void
vm_thread_kstack_arena_release(void * arena,vmem_addr_t addr,vmem_size_t size)414 vm_thread_kstack_arena_release(void *arena, vmem_addr_t addr, vmem_size_t size)
415 {
416 int rem;
417 size_t kpages __diagused = kstack_pages + KSTACK_GUARD_PAGES;
418
419 KASSERT(size % kpages == 0,
420 ("%s: Size %jd is not a multiple of kstack pages (%d)", __func__,
421 (intmax_t)size, (int)kpages));
422
423 KASSERT((addr - VM_MIN_KERNEL_ADDRESS) % kpages == 0,
424 ("%s: Address %p is not properly aligned (%p)", __func__,
425 (void *)addr, (void *)VM_MIN_KERNEL_ADDRESS));
426 /*
427 * If the address is not KVA_KSTACK_QUANTUM-aligned we have to decrement
428 * it to account for the shift in kva_import_kstack.
429 */
430 rem = addr % KVA_KSTACK_QUANTUM;
431 if (rem) {
432 KASSERT(rem <= ptoa(kpages),
433 ("%s: rem > kpages (%d), (%d)", __func__, rem,
434 (int)kpages));
435 addr -= rem;
436 }
437 vmem_xfree(arena, addr, vm_thread_kstack_import_quantum());
438 }
439
440 /*
441 * Create the kernel stack for a new thread.
442 */
443 static vm_offset_t
vm_thread_stack_create(struct domainset * ds,int pages,int flags)444 vm_thread_stack_create(struct domainset *ds, int pages, int flags)
445 {
446 vm_page_t ma[KSTACK_MAX_PAGES];
447 struct vm_domainset_iter di;
448 int req;
449 vm_offset_t ks;
450 int domain, i;
451
452 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
453 req = malloc2vm_flags(flags);
454 do {
455 /*
456 * Get a kernel virtual address for this thread's kstack.
457 */
458 ks = vm_thread_alloc_kstack_kva(ptoa(pages + KSTACK_GUARD_PAGES),
459 domain);
460 if (ks == 0)
461 continue;
462 ks += ptoa(KSTACK_GUARD_PAGES);
463
464 /*
465 * Allocate physical pages to back the stack.
466 */
467 if (vm_thread_stack_back(ks, ma, pages, req, domain) != 0) {
468 vm_thread_free_kstack_kva(ks - ptoa(KSTACK_GUARD_PAGES),
469 ptoa(pages + KSTACK_GUARD_PAGES), domain);
470 continue;
471 }
472 if (KSTACK_GUARD_PAGES != 0) {
473 pmap_qremove(ks - ptoa(KSTACK_GUARD_PAGES),
474 KSTACK_GUARD_PAGES);
475 }
476 for (i = 0; i < pages; i++)
477 vm_page_valid(ma[i]);
478 pmap_qenter(ks, ma, pages);
479 return (ks);
480 } while (vm_domainset_iter_policy(&di, &domain) == 0);
481
482 return (0);
483 }
484
485 static __noinline void
vm_thread_stack_dispose(vm_offset_t ks,int pages)486 vm_thread_stack_dispose(vm_offset_t ks, int pages)
487 {
488 vm_page_t m;
489 vm_pindex_t pindex;
490 int i, domain;
491 vm_object_t obj = vm_thread_kstack_size_to_obj(pages);
492
493 pindex = vm_kstack_pindex(ks, pages);
494 domain = vm_phys_domain(vtophys(ks));
495 pmap_qremove(ks, pages);
496 VM_OBJECT_WLOCK(obj);
497 for (i = 0; i < pages; i++) {
498 m = vm_page_lookup(obj, pindex + i);
499 if (m == NULL)
500 panic("%s: kstack already missing?", __func__);
501 KASSERT(vm_page_domain(m) == domain,
502 ("%s: page %p domain mismatch, expected %d got %d",
503 __func__, m, domain, vm_page_domain(m)));
504 vm_page_xbusy_claim(m);
505 vm_page_unwire_noq(m);
506 vm_page_free(m);
507 }
508 VM_OBJECT_WUNLOCK(obj);
509 kasan_mark((void *)ks, ptoa(pages), ptoa(pages), 0);
510 vm_thread_free_kstack_kva(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
511 ptoa(pages + KSTACK_GUARD_PAGES), domain);
512 }
513
514 /*
515 * Allocate the kernel stack for a new thread.
516 */
517 int
vm_thread_new(struct thread * td,int pages)518 vm_thread_new(struct thread *td, int pages)
519 {
520 vm_offset_t ks;
521 u_short ks_domain;
522
523 /* Bounds check */
524 if (pages <= 1)
525 pages = kstack_pages;
526 else if (pages > KSTACK_MAX_PAGES)
527 pages = KSTACK_MAX_PAGES;
528
529 ks = 0;
530 if (pages == kstack_pages && kstack_cache != NULL)
531 ks = (vm_offset_t)uma_zalloc(kstack_cache, M_NOWAIT);
532 if (ks == 0)
533 ks = vm_thread_stack_create(DOMAINSET_PREF(PCPU_GET(domain)),
534 pages, M_NOWAIT);
535 if (ks == 0)
536 return (0);
537
538 ks_domain = vm_phys_domain(vtophys(ks));
539 KASSERT(ks_domain >= 0 && ks_domain < vm_ndomains,
540 ("%s: invalid domain for kstack %p", __func__, (void *)ks));
541 td->td_kstack = ks;
542 td->td_kstack_pages = pages;
543 td->td_kstack_domain = ks_domain;
544 return (1);
545 }
546
547 /*
548 * Dispose of a thread's kernel stack.
549 */
550 void
vm_thread_dispose(struct thread * td)551 vm_thread_dispose(struct thread *td)
552 {
553 vm_offset_t ks;
554 int pages;
555
556 pages = td->td_kstack_pages;
557 ks = td->td_kstack;
558 td->td_kstack = 0;
559 td->td_kstack_pages = 0;
560 td->td_kstack_domain = MAXMEMDOM;
561 if (pages == kstack_pages) {
562 kasan_mark((void *)ks, 0, ptoa(pages), KASAN_KSTACK_FREED);
563 uma_zfree(kstack_cache, (void *)ks);
564 } else {
565 vm_thread_stack_dispose(ks, pages);
566 }
567 }
568
569 /*
570 * Calculate kstack pindex.
571 *
572 * Uses a non-identity mapping if guard pages are
573 * active to avoid pindex holes in the kstack object.
574 */
575 static vm_pindex_t
vm_kstack_pindex(vm_offset_t ks,int kpages)576 vm_kstack_pindex(vm_offset_t ks, int kpages)
577 {
578 vm_pindex_t pindex = atop(ks - VM_MIN_KERNEL_ADDRESS);
579
580 #ifdef __ILP32__
581 return (pindex);
582 #else
583 /*
584 * Return the linear pindex if guard pages aren't active or if we are
585 * allocating a non-standard kstack size.
586 */
587 if (KSTACK_GUARD_PAGES == 0 || kpages != kstack_pages) {
588 return (pindex);
589 }
590 KASSERT(pindex % (kpages + KSTACK_GUARD_PAGES) >= KSTACK_GUARD_PAGES,
591 ("%s: Attempting to calculate kstack guard page pindex", __func__));
592
593 return (pindex -
594 (pindex / (kpages + KSTACK_GUARD_PAGES) + 1) * KSTACK_GUARD_PAGES);
595 #endif
596 }
597
598 /*
599 * Allocate physical pages, following the specified NUMA policy, to back a
600 * kernel stack.
601 */
602 static int
vm_thread_stack_back(vm_offset_t ks,vm_page_t ma[],int npages,int req_class,int domain)603 vm_thread_stack_back(vm_offset_t ks, vm_page_t ma[], int npages, int req_class,
604 int domain)
605 {
606 struct pctrie_iter pages;
607 vm_object_t obj = vm_thread_kstack_size_to_obj(npages);
608 vm_pindex_t pindex;
609 vm_page_t m;
610 int n;
611
612 pindex = vm_kstack_pindex(ks, npages);
613
614 vm_page_iter_init(&pages, obj);
615 VM_OBJECT_WLOCK(obj);
616 for (n = 0; n < npages; ma[n++] = m) {
617 m = vm_page_grab_iter(obj, pindex + n,
618 VM_ALLOC_NOCREAT | VM_ALLOC_WIRED, &pages);
619 if (m != NULL)
620 continue;
621 m = vm_page_alloc_domain_iter(obj, pindex + n,
622 domain, req_class | VM_ALLOC_WIRED, &pages);
623 if (m != NULL)
624 continue;
625 for (int i = 0; i < n; i++) {
626 m = ma[i];
627 (void)vm_page_unwire_noq(m);
628 vm_page_free(m);
629 }
630 break;
631 }
632 VM_OBJECT_WUNLOCK(obj);
633 return (n < npages ? ENOMEM : 0);
634 }
635
636 static vm_object_t
vm_thread_kstack_size_to_obj(int npages)637 vm_thread_kstack_size_to_obj(int npages)
638 {
639 return (npages == kstack_pages ? kstack_object : kstack_alt_object);
640 }
641
642 static int
kstack_import(void * arg,void ** store,int cnt,int domain,int flags)643 kstack_import(void *arg, void **store, int cnt, int domain, int flags)
644 {
645 struct domainset *ds;
646 int i;
647
648 if (domain == UMA_ANYDOMAIN)
649 ds = DOMAINSET_RR();
650 else
651 ds = DOMAINSET_PREF(domain);
652
653 for (i = 0; i < cnt; i++) {
654 store[i] = (void *)vm_thread_stack_create(ds, kstack_pages,
655 flags);
656 if (store[i] == NULL)
657 break;
658 }
659 return (i);
660 }
661
662 static void
kstack_release(void * arg,void ** store,int cnt)663 kstack_release(void *arg, void **store, int cnt)
664 {
665 vm_offset_t ks;
666 int i;
667
668 for (i = 0; i < cnt; i++) {
669 ks = (vm_offset_t)store[i];
670 vm_thread_stack_dispose(ks, kstack_pages);
671 }
672 }
673
674 static void
kstack_cache_init(void * null)675 kstack_cache_init(void *null)
676 {
677 vm_size_t kstack_quantum;
678 int domain;
679
680 kstack_object = vm_object_allocate(OBJT_PHYS,
681 atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS));
682 kstack_cache = uma_zcache_create("kstack_cache",
683 kstack_pages * PAGE_SIZE, NULL, NULL, NULL, NULL,
684 kstack_import, kstack_release, NULL,
685 UMA_ZONE_FIRSTTOUCH);
686 kstack_cache_size = imax(128, mp_ncpus * 4);
687 uma_zone_set_maxcache(kstack_cache, kstack_cache_size);
688
689 kstack_alt_object = vm_object_allocate(OBJT_PHYS,
690 atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS));
691
692 kstack_quantum = vm_thread_kstack_import_quantum();
693 /*
694 * Reduce size used by the kstack arena to allow for
695 * alignment adjustments in vm_thread_kstack_arena_import.
696 */
697 kstack_quantum -= (kstack_pages + KSTACK_GUARD_PAGES) * PAGE_SIZE;
698 /*
699 * Create the kstack_arena for each domain and set kernel_arena as
700 * parent.
701 */
702 for (domain = 0; domain < vm_ndomains; domain++) {
703 vmd_kstack_arena[domain] = vmem_create("kstack arena", 0, 0,
704 PAGE_SIZE, 0, M_WAITOK);
705 KASSERT(vmd_kstack_arena[domain] != NULL,
706 ("%s: failed to create domain %d kstack_arena", __func__,
707 domain));
708 vmem_set_import(vmd_kstack_arena[domain],
709 vm_thread_kstack_arena_import,
710 vm_thread_kstack_arena_release,
711 vm_dom[domain].vmd_kernel_arena, kstack_quantum);
712 }
713 }
714 SYSINIT(vm_kstacks, SI_SUB_KMEM, SI_ORDER_ANY, kstack_cache_init, NULL);
715
716 #ifdef KSTACK_USAGE_PROF
717 /*
718 * Track maximum stack used by a thread in kernel.
719 */
720 static int max_kstack_used;
721
722 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
723 &max_kstack_used, 0,
724 "Maximum stack depth used by a thread in kernel");
725
726 void
intr_prof_stack_use(struct thread * td,struct trapframe * frame)727 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
728 {
729 vm_offset_t stack_top;
730 vm_offset_t current;
731 int used, prev_used;
732
733 /*
734 * Testing for interrupted kernel mode isn't strictly
735 * needed. It optimizes the execution, since interrupts from
736 * usermode will have only the trap frame on the stack.
737 */
738 if (TRAPF_USERMODE(frame))
739 return;
740
741 stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
742 current = (vm_offset_t)(uintptr_t)&stack_top;
743
744 /*
745 * Try to detect if interrupt is using kernel thread stack.
746 * Hardware could use a dedicated stack for interrupt handling.
747 */
748 if (stack_top <= current || current < td->td_kstack)
749 return;
750
751 used = stack_top - current;
752 for (;;) {
753 prev_used = max_kstack_used;
754 if (prev_used >= used)
755 break;
756 if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
757 break;
758 }
759 }
760 #endif /* KSTACK_USAGE_PROF */
761
762 /*
763 * Implement fork's actions on an address space.
764 * Here we arrange for the address space to be copied or referenced,
765 * allocate a user struct (pcb and kernel stack), then call the
766 * machine-dependent layer to fill those in and make the new process
767 * ready to run. The new process is set up so that it returns directly
768 * to user mode to avoid stack copying and relocation problems.
769 */
770 int
vm_forkproc(struct thread * td,struct proc * p2,struct thread * td2,struct vmspace * vm2,int flags)771 vm_forkproc(struct thread *td, struct proc *p2, struct thread *td2,
772 struct vmspace *vm2, int flags)
773 {
774 struct proc *p1 = td->td_proc;
775 struct domainset *dset;
776 int error;
777
778 if ((flags & RFPROC) == 0) {
779 /*
780 * Divorce the memory, if it is shared, essentially
781 * this changes shared memory amongst threads, into
782 * COW locally.
783 */
784 if ((flags & RFMEM) == 0) {
785 error = vmspace_unshare(p1);
786 if (error)
787 return (error);
788 }
789 cpu_fork(td, p2, td2, flags);
790 return (0);
791 }
792
793 if (flags & RFMEM) {
794 p2->p_vmspace = p1->p_vmspace;
795 refcount_acquire(&p1->p_vmspace->vm_refcnt);
796 }
797 dset = td2->td_domain.dr_policy;
798 while (vm_page_count_severe_set(&dset->ds_mask)) {
799 vm_wait_doms(&dset->ds_mask, 0);
800 }
801
802 if ((flags & RFMEM) == 0) {
803 p2->p_vmspace = vm2;
804 if (p1->p_vmspace->vm_shm)
805 shmfork(p1, p2);
806 }
807
808 /*
809 * cpu_fork will copy and update the pcb, set up the kernel stack,
810 * and make the child ready to run.
811 */
812 cpu_fork(td, p2, td2, flags);
813 return (0);
814 }
815
816 /*
817 * Called after process has been wait(2)'ed upon and is being reaped.
818 * The idea is to reclaim resources that we could not reclaim while
819 * the process was still executing.
820 */
821 void
vm_waitproc(struct proc * p)822 vm_waitproc(struct proc *p)
823 {
824
825 vmspace_exitfree(p); /* and clean-out the vmspace */
826 }
827