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/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include <sys/fcntl.h> 36 #include <sys/file.h> 37 #include <sys/ktr.h> 38 #include <sys/lock.h> 39 #include <sys/mman.h> 40 #include <sys/proc.h> 41 #include <sys/resourcevar.h> 42 #include <sys/rwlock.h> 43 #include <sys/syscallsubr.h> 44 #include <sys/sysent.h> 45 #include <sys/sysproto.h> 46 47 #include <vm/pmap.h> 48 #include <vm/vm_extern.h> 49 #include <vm/vm_map.h> 50 #include <vm/vm_object.h> 51 52 #include <compat/linux/linux_emul.h> 53 #include <compat/linux/linux_mmap.h> 54 #include <compat/linux/linux_persona.h> 55 #include <compat/linux/linux_util.h> 56 57 #define STACK_SIZE (2 * 1024 * 1024) 58 #define GUARD_SIZE (4 * PAGE_SIZE) 59 60 #if defined(__amd64__) 61 static void linux_fixup_prot(struct thread *td, int *prot); 62 #endif 63 64 static int 65 linux_mmap_check_fp(struct file *fp, int flags, int prot, int maxprot) 66 { 67 68 /* Linux mmap() just fails for O_WRONLY files */ 69 if ((fp->f_flag & FREAD) == 0) 70 return (EACCES); 71 72 return (0); 73 } 74 75 int 76 linux_mmap_common(struct thread *td, uintptr_t addr, size_t len, int prot, 77 int flags, int fd, off_t pos) 78 { 79 struct mmap_req mr, mr_fixed; 80 struct proc *p = td->td_proc; 81 struct vmspace *vms = td->td_proc->p_vmspace; 82 int bsd_flags, error; 83 84 LINUX_CTR6(mmap2, "0x%lx, %ld, %ld, 0x%08lx, %ld, 0x%lx", 85 addr, len, prot, flags, fd, pos); 86 87 error = 0; 88 bsd_flags = 0; 89 90 /* 91 * Linux mmap(2): 92 * You must specify exactly one of MAP_SHARED and MAP_PRIVATE 93 */ 94 if (!((flags & LINUX_MAP_SHARED) ^ (flags & LINUX_MAP_PRIVATE))) 95 return (EINVAL); 96 97 if (flags & LINUX_MAP_SHARED) 98 bsd_flags |= MAP_SHARED; 99 if (flags & LINUX_MAP_PRIVATE) 100 bsd_flags |= MAP_PRIVATE; 101 if (flags & LINUX_MAP_FIXED) 102 bsd_flags |= MAP_FIXED; 103 if (flags & LINUX_MAP_ANON) { 104 /* Enforce pos to be on page boundary, then ignore. */ 105 if ((pos & PAGE_MASK) != 0) 106 return (EINVAL); 107 pos = 0; 108 bsd_flags |= MAP_ANON; 109 } else 110 bsd_flags |= MAP_NOSYNC; 111 if (flags & LINUX_MAP_GROWSDOWN) 112 bsd_flags |= MAP_STACK; 113 114 #if defined(__amd64__) 115 /* 116 * According to the Linux mmap(2) man page, "MAP_32BIT flag 117 * is ignored when MAP_FIXED is set." 118 */ 119 if ((flags & LINUX_MAP_32BIT) && (flags & LINUX_MAP_FIXED) == 0) 120 bsd_flags |= MAP_32BIT; 121 122 /* 123 * PROT_READ, PROT_WRITE, or PROT_EXEC implies PROT_READ and PROT_EXEC 124 * on Linux/i386 if the binary requires executable stack. 125 * We do this only for IA32 emulation as on native i386 this is does not 126 * make sense without PAE. 127 * 128 * XXX. Linux checks that the file system is not mounted with noexec. 129 */ 130 linux_fixup_prot(td, &prot); 131 #endif 132 133 /* Linux does not check file descriptor when MAP_ANONYMOUS is set. */ 134 fd = (bsd_flags & MAP_ANON) ? -1 : fd; 135 if (flags & LINUX_MAP_GROWSDOWN) { 136 /* 137 * The Linux MAP_GROWSDOWN option does not limit auto 138 * growth of the region. Linux mmap with this option 139 * takes as addr the initial BOS, and as len, the initial 140 * region size. It can then grow down from addr without 141 * limit. However, Linux threads has an implicit internal 142 * limit to stack size of STACK_SIZE. Its just not 143 * enforced explicitly in Linux. But, here we impose 144 * a limit of (STACK_SIZE - GUARD_SIZE) on the stack 145 * region, since we can do this with our mmap. 146 * 147 * Our mmap with MAP_STACK takes addr as the maximum 148 * downsize limit on BOS, and as len the max size of 149 * the region. It then maps the top SGROWSIZ bytes, 150 * and auto grows the region down, up to the limit 151 * in addr. 152 * 153 * If we don't use the MAP_STACK option, the effect 154 * of this code is to allocate a stack region of a 155 * fixed size of (STACK_SIZE - GUARD_SIZE). 156 */ 157 158 if ((caddr_t)addr + len > vms->vm_maxsaddr) { 159 /* 160 * Some Linux apps will attempt to mmap 161 * thread stacks near the top of their 162 * address space. If their TOS is greater 163 * than vm_maxsaddr, vm_map_growstack() 164 * will confuse the thread stack with the 165 * process stack and deliver a SEGV if they 166 * attempt to grow the thread stack past their 167 * current stacksize rlimit. To avoid this, 168 * adjust vm_maxsaddr upwards to reflect 169 * the current stacksize rlimit rather 170 * than the maximum possible stacksize. 171 * It would be better to adjust the 172 * mmap'ed region, but some apps do not check 173 * mmap's return value. 174 */ 175 PROC_LOCK(p); 176 vms->vm_maxsaddr = (char *)round_page(vms->vm_stacktop) - 177 lim_cur_proc(p, RLIMIT_STACK); 178 PROC_UNLOCK(p); 179 } 180 181 /* 182 * This gives us our maximum stack size and a new BOS. 183 * If we're using VM_STACK, then mmap will just map 184 * the top SGROWSIZ bytes, and let the stack grow down 185 * to the limit at BOS. If we're not using VM_STACK 186 * we map the full stack, since we don't have a way 187 * to autogrow it. 188 */ 189 if (len <= STACK_SIZE - GUARD_SIZE) { 190 addr = addr - (STACK_SIZE - GUARD_SIZE - len); 191 len = STACK_SIZE - GUARD_SIZE; 192 } 193 } 194 195 /* 196 * FreeBSD is free to ignore the address hint if MAP_FIXED wasn't 197 * passed. However, some Linux applications, like the ART runtime, 198 * depend on the hint. If the MAP_FIXED wasn't passed, but the 199 * address is not zero, try with MAP_FIXED and MAP_EXCL first, 200 * and fall back to the normal behaviour if that fails. 201 */ 202 mr = (struct mmap_req) { 203 .mr_hint = addr, 204 .mr_len = len, 205 .mr_prot = prot, 206 .mr_flags = bsd_flags, 207 .mr_fd = fd, 208 .mr_pos = pos, 209 .mr_check_fp_fn = linux_mmap_check_fp, 210 }; 211 if (addr != 0 && (bsd_flags & MAP_FIXED) == 0 && 212 (bsd_flags & MAP_EXCL) == 0) { 213 mr_fixed = mr; 214 mr_fixed.mr_flags |= MAP_FIXED | MAP_EXCL; 215 error = kern_mmap(td, &mr_fixed); 216 if (error == 0) 217 goto out; 218 } 219 220 error = kern_mmap(td, &mr); 221 out: 222 LINUX_CTR2(mmap2, "return: %d (%p)", error, td->td_retval[0]); 223 224 return (error); 225 } 226 227 int 228 linux_mprotect_common(struct thread *td, uintptr_t addr, size_t len, int prot) 229 { 230 int flags = 0; 231 232 /* XXX Ignore PROT_GROWSUP for now. */ 233 prot &= ~LINUX_PROT_GROWSUP; 234 if ((prot & ~(LINUX_PROT_GROWSDOWN | PROT_READ | PROT_WRITE | 235 PROT_EXEC)) != 0) 236 return (EINVAL); 237 if ((prot & LINUX_PROT_GROWSDOWN) != 0) { 238 prot &= ~LINUX_PROT_GROWSDOWN; 239 flags |= VM_MAP_PROTECT_GROWSDOWN; 240 } 241 242 #if defined(__amd64__) 243 linux_fixup_prot(td, &prot); 244 #endif 245 return (kern_mprotect(td, addr, len, prot, flags)); 246 } 247 248 /* 249 * Implement Linux madvise(MADV_DONTNEED), which has unusual semantics: for 250 * anonymous memory, pages in the range are immediately discarded. 251 */ 252 static int 253 linux_madvise_dontneed(struct thread *td, vm_offset_t start, vm_offset_t end) 254 { 255 vm_map_t map; 256 vm_map_entry_t entry; 257 vm_object_t backing_object, object; 258 vm_offset_t estart, eend; 259 vm_pindex_t pstart, pend; 260 int error; 261 262 map = &td->td_proc->p_vmspace->vm_map; 263 264 if (!vm_map_range_valid(map, start, end)) 265 return (EINVAL); 266 start = trunc_page(start); 267 end = round_page(end); 268 269 error = 0; 270 vm_map_lock_read(map); 271 if (!vm_map_lookup_entry(map, start, &entry)) 272 entry = vm_map_entry_succ(entry); 273 for (; entry->start < end; entry = vm_map_entry_succ(entry)) { 274 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 275 continue; 276 277 if (entry->wired_count != 0) { 278 error = EINVAL; 279 break; 280 } 281 282 object = entry->object.vm_object; 283 if (object == NULL) 284 continue; 285 if ((object->flags & (OBJ_UNMANAGED | OBJ_FICTITIOUS)) != 0) 286 continue; 287 288 pstart = OFF_TO_IDX(entry->offset); 289 if (start > entry->start) { 290 pstart += atop(start - entry->start); 291 estart = start; 292 } else { 293 estart = entry->start; 294 } 295 pend = OFF_TO_IDX(entry->offset) + 296 atop(entry->end - entry->start); 297 if (entry->end > end) { 298 pend -= atop(entry->end - end); 299 eend = end; 300 } else { 301 eend = entry->end; 302 } 303 304 if ((object->flags & (OBJ_ANON | OBJ_ONEMAPPING)) == 305 (OBJ_ANON | OBJ_ONEMAPPING)) { 306 /* 307 * Singly-mapped anonymous memory is discarded. This 308 * does not match Linux's semantics when the object 309 * belongs to a shadow chain of length > 1, since 310 * subsequent faults may retrieve pages from an 311 * intermediate anonymous object. However, handling 312 * this case correctly introduces a fair bit of 313 * complexity. 314 */ 315 VM_OBJECT_WLOCK(object); 316 if ((object->flags & OBJ_ONEMAPPING) != 0) { 317 vm_object_collapse(object); 318 vm_object_page_remove(object, pstart, pend, 0); 319 backing_object = object->backing_object; 320 if (backing_object != NULL && 321 (backing_object->flags & OBJ_ANON) != 0) 322 linux_msg(td, 323 "possibly incorrect MADV_DONTNEED"); 324 VM_OBJECT_WUNLOCK(object); 325 continue; 326 } 327 VM_OBJECT_WUNLOCK(object); 328 } 329 330 /* 331 * Handle shared mappings. Remove them outright instead of 332 * calling pmap_advise(), for consistency with Linux. 333 */ 334 pmap_remove(map->pmap, estart, eend); 335 vm_object_madvise(object, pstart, pend, MADV_DONTNEED); 336 } 337 vm_map_unlock_read(map); 338 339 return (error); 340 } 341 342 int 343 linux_madvise_common(struct thread *td, uintptr_t addr, size_t len, int behav) 344 { 345 346 switch (behav) { 347 case LINUX_MADV_NORMAL: 348 return (kern_madvise(td, addr, len, MADV_NORMAL)); 349 case LINUX_MADV_RANDOM: 350 return (kern_madvise(td, addr, len, MADV_RANDOM)); 351 case LINUX_MADV_SEQUENTIAL: 352 return (kern_madvise(td, addr, len, MADV_SEQUENTIAL)); 353 case LINUX_MADV_WILLNEED: 354 return (kern_madvise(td, addr, len, MADV_WILLNEED)); 355 case LINUX_MADV_DONTNEED: 356 return (linux_madvise_dontneed(td, addr, addr + len)); 357 case LINUX_MADV_FREE: 358 return (kern_madvise(td, addr, len, MADV_FREE)); 359 case LINUX_MADV_REMOVE: 360 linux_msg(curthread, "unsupported madvise MADV_REMOVE"); 361 return (EINVAL); 362 case LINUX_MADV_DONTFORK: 363 return (kern_minherit(td, addr, len, INHERIT_NONE)); 364 case LINUX_MADV_DOFORK: 365 return (kern_minherit(td, addr, len, INHERIT_COPY)); 366 case LINUX_MADV_MERGEABLE: 367 linux_msg(curthread, "unsupported madvise MADV_MERGEABLE"); 368 return (EINVAL); 369 case LINUX_MADV_UNMERGEABLE: 370 /* We don't merge anyway. */ 371 return (0); 372 case LINUX_MADV_HUGEPAGE: 373 /* Ignored; on FreeBSD huge pages are always on. */ 374 return (0); 375 case LINUX_MADV_NOHUGEPAGE: 376 #if 0 377 /* 378 * Don't warn - Firefox uses it a lot, and in real Linux it's 379 * an optional feature. 380 */ 381 linux_msg(curthread, "unsupported madvise MADV_NOHUGEPAGE"); 382 #endif 383 return (EINVAL); 384 case LINUX_MADV_DONTDUMP: 385 return (kern_madvise(td, addr, len, MADV_NOCORE)); 386 case LINUX_MADV_DODUMP: 387 return (kern_madvise(td, addr, len, MADV_CORE)); 388 case LINUX_MADV_WIPEONFORK: 389 return (kern_minherit(td, addr, len, INHERIT_ZERO)); 390 case LINUX_MADV_KEEPONFORK: 391 return (kern_minherit(td, addr, len, INHERIT_COPY)); 392 case LINUX_MADV_HWPOISON: 393 linux_msg(curthread, "unsupported madvise MADV_HWPOISON"); 394 return (EINVAL); 395 case LINUX_MADV_SOFT_OFFLINE: 396 linux_msg(curthread, "unsupported madvise MADV_SOFT_OFFLINE"); 397 return (EINVAL); 398 case -1: 399 /* 400 * -1 is sometimes used as a dummy value to detect simplistic 401 * madvise(2) stub implementations. This safeguard is used by 402 * BoringSSL, for example, before assuming MADV_WIPEONFORK is 403 * safe to use. Don't produce an "unsupported" error message 404 * for this special dummy value, which is unlikely to be used 405 * by any new advisory behavior feature. 406 */ 407 return (EINVAL); 408 default: 409 linux_msg(curthread, "unsupported madvise behav %d", behav); 410 return (EINVAL); 411 } 412 } 413 414 #if defined(__amd64__) 415 static void 416 linux_fixup_prot(struct thread *td, int *prot) 417 { 418 struct linux_pemuldata *pem; 419 420 if (SV_PROC_FLAG(td->td_proc, SV_ILP32) && *prot & PROT_READ) { 421 pem = pem_find(td->td_proc); 422 if (pem->persona & LINUX_READ_IMPLIES_EXEC) 423 *prot |= PROT_EXEC; 424 } 425 426 } 427 #endif 428