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 * $FreeBSD$ 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/capsicum.h> 38 #include <sys/file.h> 39 #include <sys/imgact.h> 40 #include <sys/ktr.h> 41 #include <sys/lock.h> 42 #include <sys/mman.h> 43 #include <sys/proc.h> 44 #include <sys/resourcevar.h> 45 #include <sys/rwlock.h> 46 #include <sys/syscallsubr.h> 47 #include <sys/sysent.h> 48 #include <sys/sysproto.h> 49 50 #include <vm/pmap.h> 51 #include <vm/vm_extern.h> 52 #include <vm/vm_map.h> 53 #include <vm/vm_object.h> 54 55 #include <compat/linux/linux_emul.h> 56 #include <compat/linux/linux_mmap.h> 57 #include <compat/linux/linux_persona.h> 58 #include <compat/linux/linux_util.h> 59 60 #define STACK_SIZE (2 * 1024 * 1024) 61 #define GUARD_SIZE (4 * PAGE_SIZE) 62 63 #if defined(__amd64__) 64 static void linux_fixup_prot(struct thread *td, int *prot); 65 #endif 66 67 static int 68 linux_mmap_check_fp(struct file *fp, int flags, int prot, int maxprot) 69 { 70 71 /* Linux mmap() just fails for O_WRONLY files */ 72 if ((fp->f_flag & FREAD) == 0) 73 return (EACCES); 74 75 return (0); 76 } 77 78 int 79 linux_mmap_common(struct thread *td, uintptr_t addr, size_t len, int prot, 80 int flags, int fd, off_t pos) 81 { 82 struct mmap_req mr, mr_fixed; 83 struct proc *p = td->td_proc; 84 struct vmspace *vms = td->td_proc->p_vmspace; 85 int bsd_flags, error; 86 struct file *fp; 87 88 LINUX_CTR6(mmap2, "0x%lx, %ld, %ld, 0x%08lx, %ld, 0x%lx", 89 addr, len, prot, flags, fd, pos); 90 91 error = 0; 92 bsd_flags = 0; 93 fp = NULL; 94 95 /* 96 * Linux mmap(2): 97 * You must specify exactly one of MAP_SHARED and MAP_PRIVATE 98 */ 99 if (!((flags & LINUX_MAP_SHARED) ^ (flags & LINUX_MAP_PRIVATE))) 100 return (EINVAL); 101 102 if (flags & LINUX_MAP_SHARED) 103 bsd_flags |= MAP_SHARED; 104 if (flags & LINUX_MAP_PRIVATE) 105 bsd_flags |= MAP_PRIVATE; 106 if (flags & LINUX_MAP_FIXED) 107 bsd_flags |= MAP_FIXED; 108 if (flags & LINUX_MAP_ANON) { 109 /* Enforce pos to be on page boundary, then ignore. */ 110 if ((pos & PAGE_MASK) != 0) 111 return (EINVAL); 112 pos = 0; 113 bsd_flags |= MAP_ANON; 114 } else 115 bsd_flags |= MAP_NOSYNC; 116 if (flags & LINUX_MAP_GROWSDOWN) 117 bsd_flags |= MAP_STACK; 118 119 #if defined(__amd64__) 120 /* 121 * According to the Linux mmap(2) man page, "MAP_32BIT flag 122 * is ignored when MAP_FIXED is set." 123 */ 124 if ((flags & LINUX_MAP_32BIT) && (flags & LINUX_MAP_FIXED) == 0) 125 bsd_flags |= MAP_32BIT; 126 127 /* 128 * PROT_READ, PROT_WRITE, or PROT_EXEC implies PROT_READ and PROT_EXEC 129 * on Linux/i386 if the binary requires executable stack. 130 * We do this only for IA32 emulation as on native i386 this is does not 131 * make sense without PAE. 132 * 133 * XXX. Linux checks that the file system is not mounted with noexec. 134 */ 135 linux_fixup_prot(td, &prot); 136 #endif 137 138 /* Linux does not check file descriptor when MAP_ANONYMOUS is set. */ 139 fd = (bsd_flags & MAP_ANON) ? -1 : fd; 140 if (flags & LINUX_MAP_GROWSDOWN) { 141 /* 142 * The Linux MAP_GROWSDOWN option does not limit auto 143 * growth of the region. Linux mmap with this option 144 * takes as addr the initial BOS, and as len, the initial 145 * region size. It can then grow down from addr without 146 * limit. However, Linux threads has an implicit internal 147 * limit to stack size of STACK_SIZE. Its just not 148 * enforced explicitly in Linux. But, here we impose 149 * a limit of (STACK_SIZE - GUARD_SIZE) on the stack 150 * region, since we can do this with our mmap. 151 * 152 * Our mmap with MAP_STACK takes addr as the maximum 153 * downsize limit on BOS, and as len the max size of 154 * the region. It then maps the top SGROWSIZ bytes, 155 * and auto grows the region down, up to the limit 156 * in addr. 157 * 158 * If we don't use the MAP_STACK option, the effect 159 * of this code is to allocate a stack region of a 160 * fixed size of (STACK_SIZE - GUARD_SIZE). 161 */ 162 163 if ((caddr_t)addr + len > vms->vm_maxsaddr) { 164 /* 165 * Some Linux apps will attempt to mmap 166 * thread stacks near the top of their 167 * address space. If their TOS is greater 168 * than vm_maxsaddr, vm_map_growstack() 169 * will confuse the thread stack with the 170 * process stack and deliver a SEGV if they 171 * attempt to grow the thread stack past their 172 * current stacksize rlimit. To avoid this, 173 * adjust vm_maxsaddr upwards to reflect 174 * the current stacksize rlimit rather 175 * than the maximum possible stacksize. 176 * It would be better to adjust the 177 * mmap'ed region, but some apps do not check 178 * mmap's return value. 179 */ 180 PROC_LOCK(p); 181 vms->vm_maxsaddr = (char *)p->p_sysent->sv_usrstack - 182 lim_cur_proc(p, RLIMIT_STACK); 183 PROC_UNLOCK(p); 184 } 185 186 /* 187 * This gives us our maximum stack size and a new BOS. 188 * If we're using VM_STACK, then mmap will just map 189 * the top SGROWSIZ bytes, and let the stack grow down 190 * to the limit at BOS. If we're not using VM_STACK 191 * we map the full stack, since we don't have a way 192 * to autogrow it. 193 */ 194 if (len <= STACK_SIZE - GUARD_SIZE) { 195 addr = addr - (STACK_SIZE - GUARD_SIZE - len); 196 len = STACK_SIZE - GUARD_SIZE; 197 } 198 } 199 200 /* 201 * FreeBSD is free to ignore the address hint if MAP_FIXED wasn't 202 * passed. However, some Linux applications, like the ART runtime, 203 * depend on the hint. If the MAP_FIXED wasn't passed, but the 204 * address is not zero, try with MAP_FIXED and MAP_EXCL first, 205 * and fall back to the normal behaviour if that fails. 206 */ 207 mr = (struct mmap_req) { 208 .mr_hint = addr, 209 .mr_len = len, 210 .mr_prot = prot, 211 .mr_flags = bsd_flags, 212 .mr_fd = fd, 213 .mr_pos = pos, 214 .mr_check_fp_fn = linux_mmap_check_fp, 215 }; 216 if (addr != 0 && (bsd_flags & MAP_FIXED) == 0 && 217 (bsd_flags & MAP_EXCL) == 0) { 218 mr_fixed = mr; 219 mr_fixed.mr_flags |= MAP_FIXED | MAP_EXCL; 220 error = kern_mmap(td, &mr_fixed); 221 if (error == 0) 222 goto out; 223 } 224 225 error = kern_mmap(td, &mr); 226 out: 227 LINUX_CTR2(mmap2, "return: %d (%p)", error, td->td_retval[0]); 228 229 return (error); 230 } 231 232 int 233 linux_mprotect_common(struct thread *td, uintptr_t addr, size_t len, int prot) 234 { 235 236 /* XXX Ignore PROT_GROWSDOWN and PROT_GROWSUP for now. */ 237 prot &= ~(LINUX_PROT_GROWSDOWN | LINUX_PROT_GROWSUP); 238 if ((prot & ~(PROT_READ | PROT_WRITE | PROT_EXEC)) != 0) 239 return (EINVAL); 240 241 #if defined(__amd64__) 242 linux_fixup_prot(td, &prot); 243 #endif 244 return (kern_mprotect(td, addr, len, prot)); 245 } 246 247 /* 248 * Implement Linux madvise(MADV_DONTNEED), which has unusual semantics: for 249 * anonymous memory, pages in the range are immediately discarded. 250 */ 251 static int 252 linux_madvise_dontneed(struct thread *td, vm_offset_t start, vm_offset_t end) 253 { 254 vm_map_t map; 255 vm_map_entry_t entry; 256 vm_object_t backing_object, object; 257 vm_offset_t estart, eend; 258 vm_pindex_t pstart, pend; 259 int error; 260 261 map = &td->td_proc->p_vmspace->vm_map; 262 263 if (!vm_map_range_valid(map, start, end)) 264 return (EINVAL); 265 start = trunc_page(start); 266 end = round_page(end); 267 268 error = 0; 269 vm_map_lock_read(map); 270 if (!vm_map_lookup_entry(map, start, &entry)) 271 entry = vm_map_entry_succ(entry); 272 for (; entry->start < end; entry = vm_map_entry_succ(entry)) { 273 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 274 continue; 275 276 if (entry->wired_count != 0) { 277 error = EINVAL; 278 break; 279 } 280 281 object = entry->object.vm_object; 282 if (object == NULL) 283 continue; 284 if ((object->flags & (OBJ_UNMANAGED | OBJ_FICTITIOUS)) != 0) 285 continue; 286 287 pstart = OFF_TO_IDX(entry->offset); 288 if (start > entry->start) { 289 pstart += atop(start - entry->start); 290 estart = start; 291 } else { 292 estart = entry->start; 293 } 294 pend = OFF_TO_IDX(entry->offset) + 295 atop(entry->end - entry->start); 296 if (entry->end > end) { 297 pend -= atop(entry->end - end); 298 eend = end; 299 } else { 300 eend = entry->end; 301 } 302 303 if ((object->flags & (OBJ_ANON | OBJ_ONEMAPPING)) == 304 (OBJ_ANON | OBJ_ONEMAPPING)) { 305 /* 306 * Singly-mapped anonymous memory is discarded. This 307 * does not match Linux's semantics when the object 308 * belongs to a shadow chain of length > 1, since 309 * subsequent faults may retrieve pages from an 310 * intermediate anonymous object. However, handling 311 * this case correctly introduces a fair bit of 312 * complexity. 313 */ 314 VM_OBJECT_WLOCK(object); 315 if ((object->flags & OBJ_ONEMAPPING) != 0) { 316 vm_object_collapse(object); 317 vm_object_page_remove(object, pstart, pend, 0); 318 backing_object = object->backing_object; 319 if (backing_object != NULL && 320 (backing_object->flags & OBJ_ANON) != 0) 321 linux_msg(td, 322 "possibly incorrect MADV_DONTNEED"); 323 VM_OBJECT_WUNLOCK(object); 324 continue; 325 } 326 VM_OBJECT_WUNLOCK(object); 327 } 328 329 /* 330 * Handle shared mappings. Remove them outright instead of 331 * calling pmap_advise(), for consistency with Linux. 332 */ 333 pmap_remove(map->pmap, estart, eend); 334 vm_object_madvise(object, pstart, pend, MADV_DONTNEED); 335 } 336 vm_map_unlock_read(map); 337 338 return (error); 339 } 340 341 int 342 linux_madvise_common(struct thread *td, uintptr_t addr, size_t len, int behav) 343 { 344 345 switch (behav) { 346 case LINUX_MADV_NORMAL: 347 return (kern_madvise(td, addr, len, MADV_NORMAL)); 348 case LINUX_MADV_RANDOM: 349 return (kern_madvise(td, addr, len, MADV_RANDOM)); 350 case LINUX_MADV_SEQUENTIAL: 351 return (kern_madvise(td, addr, len, MADV_SEQUENTIAL)); 352 case LINUX_MADV_WILLNEED: 353 return (kern_madvise(td, addr, len, MADV_WILLNEED)); 354 case LINUX_MADV_DONTNEED: 355 return (linux_madvise_dontneed(td, addr, addr + len)); 356 case LINUX_MADV_FREE: 357 return (kern_madvise(td, addr, len, MADV_FREE)); 358 case LINUX_MADV_REMOVE: 359 linux_msg(curthread, "unsupported madvise MADV_REMOVE"); 360 return (EINVAL); 361 case LINUX_MADV_DONTFORK: 362 return (kern_minherit(td, addr, len, INHERIT_NONE)); 363 case LINUX_MADV_DOFORK: 364 return (kern_minherit(td, addr, len, INHERIT_COPY)); 365 case LINUX_MADV_MERGEABLE: 366 linux_msg(curthread, "unsupported madvise MADV_MERGEABLE"); 367 return (EINVAL); 368 case LINUX_MADV_UNMERGEABLE: 369 /* We don't merge anyway. */ 370 return (0); 371 case LINUX_MADV_HUGEPAGE: 372 /* Ignored; on FreeBSD huge pages are always on. */ 373 return (0); 374 case LINUX_MADV_NOHUGEPAGE: 375 #if 0 376 /* 377 * Don't warn - Firefox uses it a lot, and in real Linux it's 378 * an optional feature. 379 */ 380 linux_msg(curthread, "unsupported madvise MADV_NOHUGEPAGE"); 381 #endif 382 return (EINVAL); 383 case LINUX_MADV_DONTDUMP: 384 return (kern_madvise(td, addr, len, MADV_NOCORE)); 385 case LINUX_MADV_DODUMP: 386 return (kern_madvise(td, addr, len, MADV_CORE)); 387 case LINUX_MADV_WIPEONFORK: 388 return (kern_minherit(td, addr, len, INHERIT_ZERO)); 389 case LINUX_MADV_KEEPONFORK: 390 return (kern_minherit(td, addr, len, INHERIT_COPY)); 391 case LINUX_MADV_HWPOISON: 392 linux_msg(curthread, "unsupported madvise MADV_HWPOISON"); 393 return (EINVAL); 394 case LINUX_MADV_SOFT_OFFLINE: 395 linux_msg(curthread, "unsupported madvise MADV_SOFT_OFFLINE"); 396 return (EINVAL); 397 case -1: 398 /* 399 * -1 is sometimes used as a dummy value to detect simplistic 400 * madvise(2) stub implementations. This safeguard is used by 401 * BoringSSL, for example, before assuming MADV_WIPEONFORK is 402 * safe to use. Don't produce an "unsupported" error message 403 * for this special dummy value, which is unlikely to be used 404 * by any new advisory behavior feature. 405 */ 406 return (EINVAL); 407 default: 408 linux_msg(curthread, "unsupported madvise behav %d", behav); 409 return (EINVAL); 410 } 411 } 412 413 #if defined(__amd64__) 414 static void 415 linux_fixup_prot(struct thread *td, int *prot) 416 { 417 struct linux_pemuldata *pem; 418 419 if (SV_PROC_FLAG(td->td_proc, SV_ILP32) && *prot & PROT_READ) { 420 pem = pem_find(td->td_proc); 421 if (pem->persona & LINUX_READ_IMPLIES_EXEC) 422 *prot |= PROT_EXEC; 423 } 424 425 } 426 #endif 427