1 /*-
2 * Copyright (c) 2004 Tim J. Robbins
3 * Copyright (c) 2002 Doug Rabson
4 * Copyright (c) 2000 Marcel Moolenaar
5 * Copyright (c) 1994-1995 Søren Schmidt
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, this list of conditions and the following disclaimer
13 * in this position and unchanged.
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 * 3. The name of the author may not be used to endorse or promote products
18 * derived from this software without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
21 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
22 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
23 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
24 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
29 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 */
31
32 #include <sys/fcntl.h>
33 #include <sys/file.h>
34 #include <sys/ktr.h>
35 #include <sys/lock.h>
36 #include <sys/malloc.h>
37 #include <sys/mman.h>
38 #include <sys/proc.h>
39 #include <sys/resourcevar.h>
40 #include <sys/rwlock.h>
41 #include <sys/syscallsubr.h>
42 #include <sys/sysent.h>
43 #include <sys/sysproto.h>
44
45 #include <vm/pmap.h>
46 #include <vm/vm_extern.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_object.h>
49
50 #include <compat/linux/linux_emul.h>
51 #include <compat/linux/linux_mmap.h>
52 #include <compat/linux/linux_persona.h>
53 #include <compat/linux/linux_util.h>
54
55 #define STACK_SIZE (2 * 1024 * 1024)
56 #define GUARD_SIZE (4 * PAGE_SIZE)
57
58 #if defined(__amd64__)
59 static void linux_fixup_prot(struct thread *td, int *prot);
60 #endif
61
62 static int
linux_mmap_check_fp(struct file * fp,int flags,int prot,int maxprot)63 linux_mmap_check_fp(struct file *fp, int flags, int prot, int maxprot)
64 {
65
66 /* Linux mmap() just fails for O_WRONLY files */
67 if ((fp->f_flag & FREAD) == 0)
68 return (EACCES);
69
70 return (0);
71 }
72
73 int
linux_mmap_common(struct thread * td,uintptr_t addr,size_t len,int prot,int flags,int fd,off_t pos)74 linux_mmap_common(struct thread *td, uintptr_t addr, size_t len, int prot,
75 int flags, int fd, off_t pos)
76 {
77 struct mmap_req mr, mr_fixed;
78 struct proc *p = td->td_proc;
79 struct vmspace *vms = td->td_proc->p_vmspace;
80 int bsd_flags, error;
81
82 LINUX_CTR6(mmap2, "0x%lx, %ld, %ld, 0x%08lx, %ld, 0x%lx",
83 addr, len, prot, flags, fd, pos);
84
85 error = 0;
86 bsd_flags = 0;
87
88 /*
89 * Linux mmap(2):
90 * You must specify exactly one of MAP_SHARED and MAP_PRIVATE
91 */
92 if (!((flags & LINUX_MAP_SHARED) ^ (flags & LINUX_MAP_PRIVATE)))
93 return (EINVAL);
94
95 if (flags & LINUX_MAP_SHARED)
96 bsd_flags |= MAP_SHARED;
97 if (flags & LINUX_MAP_PRIVATE)
98 bsd_flags |= MAP_PRIVATE;
99 if (flags & LINUX_MAP_FIXED)
100 bsd_flags |= MAP_FIXED;
101 if (flags & LINUX_MAP_ANON) {
102 /* Enforce pos to be on page boundary, then ignore. */
103 if ((pos & PAGE_MASK) != 0)
104 return (EINVAL);
105 pos = 0;
106 bsd_flags |= MAP_ANON;
107 } else
108 bsd_flags |= MAP_NOSYNC;
109 if (flags & LINUX_MAP_GROWSDOWN)
110 bsd_flags |= MAP_STACK;
111
112 #if defined(__amd64__)
113 /*
114 * According to the Linux mmap(2) man page, "MAP_32BIT flag
115 * is ignored when MAP_FIXED is set."
116 */
117 if ((flags & LINUX_MAP_32BIT) && (flags & LINUX_MAP_FIXED) == 0)
118 bsd_flags |= MAP_32BIT;
119
120 /*
121 * PROT_READ, PROT_WRITE, or PROT_EXEC implies PROT_READ and PROT_EXEC
122 * on Linux/i386 if the binary requires executable stack.
123 * We do this only for IA32 emulation as on native i386 this is does not
124 * make sense without PAE.
125 *
126 * XXX. Linux checks that the file system is not mounted with noexec.
127 */
128 linux_fixup_prot(td, &prot);
129 #endif
130
131 /* Linux does not check file descriptor when MAP_ANONYMOUS is set. */
132 fd = (bsd_flags & MAP_ANON) ? -1 : fd;
133 if (flags & LINUX_MAP_GROWSDOWN) {
134 /*
135 * The Linux MAP_GROWSDOWN option does not limit auto
136 * growth of the region. Linux mmap with this option
137 * takes as addr the initial BOS, and as len, the initial
138 * region size. It can then grow down from addr without
139 * limit. However, Linux threads has an implicit internal
140 * limit to stack size of STACK_SIZE. Its just not
141 * enforced explicitly in Linux. But, here we impose
142 * a limit of (STACK_SIZE - GUARD_SIZE) on the stack
143 * region, since we can do this with our mmap.
144 *
145 * Our mmap with MAP_STACK takes addr as the maximum
146 * downsize limit on BOS, and as len the max size of
147 * the region. It then maps the top SGROWSIZ bytes,
148 * and auto grows the region down, up to the limit
149 * in addr.
150 *
151 * If we don't use the MAP_STACK option, the effect
152 * of this code is to allocate a stack region of a
153 * fixed size of (STACK_SIZE - GUARD_SIZE).
154 */
155
156 if ((caddr_t)addr + len > vms->vm_maxsaddr) {
157 /*
158 * Some Linux apps will attempt to mmap
159 * thread stacks near the top of their
160 * address space. If their TOS is greater
161 * than vm_maxsaddr, vm_map_growstack()
162 * will confuse the thread stack with the
163 * process stack and deliver a SEGV if they
164 * attempt to grow the thread stack past their
165 * current stacksize rlimit. To avoid this,
166 * adjust vm_maxsaddr upwards to reflect
167 * the current stacksize rlimit rather
168 * than the maximum possible stacksize.
169 * It would be better to adjust the
170 * mmap'ed region, but some apps do not check
171 * mmap's return value.
172 */
173 PROC_LOCK(p);
174 vms->vm_maxsaddr = (char *)round_page(vms->vm_stacktop) -
175 lim_cur_proc(p, RLIMIT_STACK);
176 PROC_UNLOCK(p);
177 }
178
179 /*
180 * This gives us our maximum stack size and a new BOS.
181 * If we're using VM_STACK, then mmap will just map
182 * the top SGROWSIZ bytes, and let the stack grow down
183 * to the limit at BOS. If we're not using VM_STACK
184 * we map the full stack, since we don't have a way
185 * to autogrow it.
186 */
187 if (len <= STACK_SIZE - GUARD_SIZE) {
188 addr = addr - (STACK_SIZE - GUARD_SIZE - len);
189 len = STACK_SIZE - GUARD_SIZE;
190 }
191 }
192
193 /*
194 * FreeBSD is free to ignore the address hint if MAP_FIXED wasn't
195 * passed. However, some Linux applications, like the ART runtime,
196 * depend on the hint. If the MAP_FIXED wasn't passed, but the
197 * address is not zero, try with MAP_FIXED and MAP_EXCL first,
198 * and fall back to the normal behaviour if that fails.
199 */
200 mr = (struct mmap_req) {
201 .mr_hint = addr,
202 .mr_len = len,
203 .mr_prot = prot,
204 .mr_flags = bsd_flags,
205 .mr_fd = fd,
206 .mr_pos = pos,
207 .mr_check_fp_fn = linux_mmap_check_fp,
208 };
209 if (addr != 0 && (bsd_flags & MAP_FIXED) == 0 &&
210 (bsd_flags & MAP_EXCL) == 0) {
211 mr_fixed = mr;
212 mr_fixed.mr_flags |= MAP_FIXED | MAP_EXCL;
213 error = kern_mmap(td, &mr_fixed);
214 if (error == 0)
215 goto out;
216 }
217
218 error = kern_mmap(td, &mr);
219 out:
220 LINUX_CTR2(mmap2, "return: %d (%p)", error, td->td_retval[0]);
221
222 return (error);
223 }
224
225 int
linux_mprotect_common(struct thread * td,uintptr_t addr,size_t len,int prot)226 linux_mprotect_common(struct thread *td, uintptr_t addr, size_t len, int prot)
227 {
228 int flags = 0;
229
230 /* XXX Ignore PROT_GROWSUP for now. */
231 prot &= ~LINUX_PROT_GROWSUP;
232 if ((prot & ~(LINUX_PROT_GROWSDOWN | PROT_READ | PROT_WRITE |
233 PROT_EXEC)) != 0)
234 return (EINVAL);
235 if ((prot & LINUX_PROT_GROWSDOWN) != 0) {
236 prot &= ~LINUX_PROT_GROWSDOWN;
237 flags |= VM_MAP_PROTECT_GROWSDOWN;
238 }
239
240 #if defined(__amd64__)
241 linux_fixup_prot(td, &prot);
242 #endif
243 return (kern_mprotect(td, addr, len, prot, flags));
244 }
245
246 /*
247 * Implement Linux madvise(MADV_DONTNEED), which has unusual semantics: for
248 * anonymous memory, pages in the range are immediately discarded.
249 */
250 static int
linux_madvise_dontneed(struct thread * td,vm_offset_t start,vm_offset_t end)251 linux_madvise_dontneed(struct thread *td, vm_offset_t start, vm_offset_t end)
252 {
253 vm_map_t map;
254 vm_map_entry_t entry;
255 vm_object_t backing_object, object;
256 vm_offset_t estart, eend;
257 vm_pindex_t pstart, pend;
258 int error;
259
260 map = &td->td_proc->p_vmspace->vm_map;
261
262 if (!vm_map_range_valid(map, start, end))
263 return (EINVAL);
264 start = trunc_page(start);
265 end = round_page(end);
266
267 error = 0;
268 vm_map_lock_read(map);
269 if (!vm_map_lookup_entry(map, start, &entry))
270 entry = vm_map_entry_succ(entry);
271 for (; entry->start < end; entry = vm_map_entry_succ(entry)) {
272 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
273 continue;
274
275 if (entry->wired_count != 0) {
276 error = EINVAL;
277 break;
278 }
279
280 object = entry->object.vm_object;
281 if (object == NULL)
282 continue;
283 if ((object->flags & (OBJ_UNMANAGED | OBJ_FICTITIOUS)) != 0)
284 continue;
285
286 pstart = OFF_TO_IDX(entry->offset);
287 if (start > entry->start) {
288 pstart += atop(start - entry->start);
289 estart = start;
290 } else {
291 estart = entry->start;
292 }
293 pend = OFF_TO_IDX(entry->offset) +
294 atop(entry->end - entry->start);
295 if (entry->end > end) {
296 pend -= atop(entry->end - end);
297 eend = end;
298 } else {
299 eend = entry->end;
300 }
301
302 if ((object->flags & (OBJ_ANON | OBJ_ONEMAPPING)) ==
303 (OBJ_ANON | OBJ_ONEMAPPING)) {
304 /*
305 * Singly-mapped anonymous memory is discarded. This
306 * does not match Linux's semantics when the object
307 * belongs to a shadow chain of length > 1, since
308 * subsequent faults may retrieve pages from an
309 * intermediate anonymous object. However, handling
310 * this case correctly introduces a fair bit of
311 * complexity.
312 */
313 VM_OBJECT_WLOCK(object);
314 if ((object->flags & OBJ_ONEMAPPING) != 0) {
315 vm_object_collapse(object);
316 vm_object_page_remove(object, pstart, pend, 0);
317 backing_object = object->backing_object;
318 if (backing_object != NULL &&
319 (backing_object->flags & OBJ_ANON) != 0)
320 linux_msg(td,
321 "possibly incorrect MADV_DONTNEED");
322 VM_OBJECT_WUNLOCK(object);
323 continue;
324 }
325 VM_OBJECT_WUNLOCK(object);
326 }
327
328 /*
329 * Handle shared mappings. Remove them outright instead of
330 * calling pmap_advise(), for consistency with Linux.
331 */
332 pmap_remove(map->pmap, estart, eend);
333 vm_object_madvise(object, pstart, pend, MADV_DONTNEED);
334 }
335 vm_map_unlock_read(map);
336
337 return (error);
338 }
339
340 int
linux_madvise_common(struct thread * td,uintptr_t addr,size_t len,int behav)341 linux_madvise_common(struct thread *td, uintptr_t addr, size_t len, int behav)
342 {
343
344 switch (behav) {
345 case LINUX_MADV_NORMAL:
346 return (kern_madvise(td, addr, len, MADV_NORMAL));
347 case LINUX_MADV_RANDOM:
348 return (kern_madvise(td, addr, len, MADV_RANDOM));
349 case LINUX_MADV_SEQUENTIAL:
350 return (kern_madvise(td, addr, len, MADV_SEQUENTIAL));
351 case LINUX_MADV_WILLNEED:
352 return (kern_madvise(td, addr, len, MADV_WILLNEED));
353 case LINUX_MADV_DONTNEED:
354 return (linux_madvise_dontneed(td, addr, addr + len));
355 case LINUX_MADV_FREE:
356 return (kern_madvise(td, addr, len, MADV_FREE));
357 case LINUX_MADV_REMOVE:
358 linux_msg(curthread, "unsupported madvise MADV_REMOVE");
359 return (EINVAL);
360 case LINUX_MADV_DONTFORK:
361 return (kern_minherit(td, addr, len, INHERIT_NONE));
362 case LINUX_MADV_DOFORK:
363 return (kern_minherit(td, addr, len, INHERIT_COPY));
364 case LINUX_MADV_MERGEABLE:
365 linux_msg(curthread, "unsupported madvise MADV_MERGEABLE");
366 return (EINVAL);
367 case LINUX_MADV_UNMERGEABLE:
368 /* We don't merge anyway. */
369 return (0);
370 case LINUX_MADV_HUGEPAGE:
371 /* Ignored; on FreeBSD huge pages are always on. */
372 return (0);
373 case LINUX_MADV_NOHUGEPAGE:
374 #if 0
375 /*
376 * Don't warn - Firefox uses it a lot, and in real Linux it's
377 * an optional feature.
378 */
379 linux_msg(curthread, "unsupported madvise MADV_NOHUGEPAGE");
380 #endif
381 return (EINVAL);
382 case LINUX_MADV_DONTDUMP:
383 return (kern_madvise(td, addr, len, MADV_NOCORE));
384 case LINUX_MADV_DODUMP:
385 return (kern_madvise(td, addr, len, MADV_CORE));
386 case LINUX_MADV_WIPEONFORK:
387 return (kern_minherit(td, addr, len, INHERIT_ZERO));
388 case LINUX_MADV_KEEPONFORK:
389 return (kern_minherit(td, addr, len, INHERIT_COPY));
390 case LINUX_MADV_HWPOISON:
391 linux_msg(curthread, "unsupported madvise MADV_HWPOISON");
392 return (EINVAL);
393 case LINUX_MADV_SOFT_OFFLINE:
394 linux_msg(curthread, "unsupported madvise MADV_SOFT_OFFLINE");
395 return (EINVAL);
396 case -1:
397 /*
398 * -1 is sometimes used as a dummy value to detect simplistic
399 * madvise(2) stub implementations. This safeguard is used by
400 * BoringSSL, for example, before assuming MADV_WIPEONFORK is
401 * safe to use. Don't produce an "unsupported" error message
402 * for this special dummy value, which is unlikely to be used
403 * by any new advisory behavior feature.
404 */
405 return (EINVAL);
406 default:
407 linux_msg(curthread, "unsupported madvise behav %d", behav);
408 return (EINVAL);
409 }
410 }
411
412 #if defined(__amd64__)
413 static void
linux_fixup_prot(struct thread * td,int * prot)414 linux_fixup_prot(struct thread *td, int *prot)
415 {
416 struct linux_pemuldata *pem;
417
418 if (SV_PROC_FLAG(td->td_proc, SV_ILP32) && *prot & PROT_READ) {
419 pem = pem_find(td->td_proc);
420 if (pem->persona & LINUX_READ_IMPLIES_EXEC)
421 *prot |= PROT_EXEC;
422 }
423
424 }
425 #endif
426