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