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