1 use core::{intrinsics::unlikely, slice::from_raw_parts}; 2 3 use alloc::sync::Arc; 4 use log::error; 5 use system_error::SystemError; 6 7 use crate::{ 8 arch::MMArch, 9 driver::base::block::SeekFrom, 10 ipc::shm::ShmFlags, 11 libs::align::{check_aligned, page_align_up}, 12 mm::MemoryManagementArch, 13 syscall::Syscall, 14 }; 15 16 use super::{ 17 allocator::page_frame::{PageFrameCount, VirtPageFrame}, 18 ucontext::{AddressSpace, DEFAULT_MMAP_MIN_ADDR}, 19 verify_area, MsFlags, VirtAddr, VmFlags, 20 }; 21 22 bitflags! { 23 /// Memory protection flags 24 pub struct ProtFlags: u64 { 25 const PROT_NONE = 0x0; 26 const PROT_READ = 0x1; 27 const PROT_WRITE = 0x2; 28 const PROT_EXEC = 0x4; 29 } 30 31 /// Memory mapping flags 32 pub struct MapFlags: u64 { 33 const MAP_NONE = 0x0; 34 /// share changes 35 const MAP_SHARED = 0x1; 36 /// changes are private 37 const MAP_PRIVATE = 0x2; 38 /// Interpret addr exactly 39 const MAP_FIXED = 0x10; 40 /// don't use a file 41 const MAP_ANONYMOUS = 0x20; 42 // linux-6.1-rc5/include/uapi/asm-generic/mman.h#7 43 /// stack-like segment 44 const MAP_GROWSDOWN = 0x100; 45 /// ETXTBSY 46 const MAP_DENYWRITE = 0x800; 47 /// Mark it as an executable 48 const MAP_EXECUTABLE = 0x1000; 49 /// Pages are locked 50 const MAP_LOCKED = 0x2000; 51 /// don't check for reservations 52 const MAP_NORESERVE = 0x4000; 53 /// populate (prefault) pagetables 54 const MAP_POPULATE = 0x8000; 55 /// do not block on IO 56 const MAP_NONBLOCK = 0x10000; 57 /// give out an address that is best suited for process/thread stacks 58 const MAP_STACK = 0x20000; 59 /// create a huge page mapping 60 const MAP_HUGETLB = 0x40000; 61 /// perform synchronous page faults for the mapping 62 const MAP_SYNC = 0x80000; 63 /// MAP_FIXED which doesn't unmap underlying mapping 64 const MAP_FIXED_NOREPLACE = 0x100000; 65 66 /// For anonymous mmap, memory could be uninitialized 67 const MAP_UNINITIALIZED = 0x4000000; 68 } 69 70 /// Memory mremapping flags 71 pub struct MremapFlags: u8 { 72 const MREMAP_MAYMOVE = 1; 73 const MREMAP_FIXED = 2; 74 const MREMAP_DONTUNMAP = 4; 75 } 76 77 78 pub struct MadvFlags: u64 { 79 /// 默认行为，系统会进行一定的预读和预写，适用于一般读取场景 80 const MADV_NORMAL = 0; 81 /// 随机访问模式，系统会尽量最小化数据读取量，适用于随机访问的场景 82 const MADV_RANDOM = 1; 83 /// 顺序访问模式，系统会进行积极的预读，访问后的页面可以尽快释放，适用于顺序读取场景 84 const MADV_SEQUENTIAL = 2; 85 /// 通知系统预读某些页面，用于应用程序提前准备数据 86 const MADV_WILLNEED = 3; 87 /// 通知系统应用程序不再需要某些页面，内核可以释放相关资源 88 const MADV_DONTNEED = 4; 89 90 /// 将指定范围的页面标记为延迟释放，真正的释放会延迟至内存压力发生时 91 const MADV_FREE = 8; 92 /// 应用程序请求释放指定范围的页面和相关的后备存储 93 const MADV_REMOVE = 9; 94 /// 在 fork 时排除指定区域 95 const MADV_DONTFORK = 10; 96 /// 取消 MADV_DONTFORK 的效果，不再在 fork 时排除指定区域 97 const MADV_DOFORK = 11; 98 /// 模拟内存硬件错误，触发内存错误处理器处理 99 const MADV_HWPOISON = 100; 100 /// 尝试软下线指定的内存范围 101 const MADV_SOFT_OFFLINE = 101; 102 103 /// 应用程序建议内核尝试合并指定范围内内容相同的页面 104 const MADV_MERGEABLE = 12; 105 /// 取消 MADV_MERGEABLE 的效果，不再合并页面 106 const MADV_UNMERGEABLE = 13; 107 108 /// 应用程序希望将指定范围以透明大页方式支持 109 const MADV_HUGEPAGE = 14; 110 /// 将指定范围标记为不值得用透明大页支持 111 const MADV_NOHUGEPAGE = 15; 112 113 /// 应用程序请求在核心转储时排除指定范围内的页面 114 const MADV_DONTDUMP = 16; 115 /// 取消 MADV_DONTDUMP 的效果，不再排除核心转储时的页面 116 const MADV_DODUMP = 17; 117 118 /// 在 fork 时将子进程的该区域内存填充为零 119 const MADV_WIPEONFORK = 18; 120 /// 取消 `MADV_WIPEONFORK` 的效果，不再在 fork 时填充子进程的内存 121 const MADV_KEEPONFORK = 19; 122 123 /// 应用程序不会立刻使用这些内存，内核将页面设置为非活动状态以便在内存压力发生时轻松回收 124 const MADV_COLD = 20; 125 /// 应用程序不会立刻使用这些内存，内核立即将这些页面换出 126 const MADV_PAGEOUT = 21; 127 128 /// 预先填充页面表，可读，通过触发读取故障 129 const MADV_POPULATE_READ = 22; 130 /// 预先填充页面表，可写，通过触发写入故障 131 const MADV_POPULATE_WRITE = 23; 132 133 /// 与 `MADV_DONTNEED` 类似，会将被锁定的页面释放 134 const MADV_DONTNEED_LOCKED = 24; 135 136 /// 同步将页面合并为新的透明大页 137 const MADV_COLLAPSE = 25; 138 139 } 140 } 141 142 impl From<MapFlags> for VmFlags { 143 fn from(map_flags: MapFlags) -> Self { 144 let mut vm_flags = VmFlags::VM_NONE; 145 146 if map_flags.contains(MapFlags::MAP_GROWSDOWN) { 147 vm_flags |= VmFlags::VM_GROWSDOWN; 148 } 149 150 if map_flags.contains(MapFlags::MAP_LOCKED) { 151 vm_flags |= VmFlags::VM_LOCKED; 152 } 153 154 if map_flags.contains(MapFlags::MAP_SYNC) { 155 vm_flags |= VmFlags::VM_SYNC; 156 } 157 158 if map_flags.contains(MapFlags::MAP_SHARED) { 159 vm_flags |= VmFlags::VM_SHARED; 160 } 161 162 vm_flags 163 } 164 } 165 166 impl From<ProtFlags> for VmFlags { 167 fn from(prot_flags: ProtFlags) -> Self { 168 let mut vm_flags = VmFlags::VM_NONE; 169 170 if prot_flags.contains(ProtFlags::PROT_READ) { 171 vm_flags |= VmFlags::VM_READ; 172 } 173 174 if prot_flags.contains(ProtFlags::PROT_WRITE) { 175 vm_flags |= VmFlags::VM_WRITE; 176 } 177 178 if prot_flags.contains(ProtFlags::PROT_EXEC) { 179 vm_flags |= VmFlags::VM_EXEC; 180 } 181 182 vm_flags 183 } 184 } 185 186 impl From<ShmFlags> for VmFlags { 187 fn from(shm_flags: ShmFlags) -> Self { 188 let mut vm_flags = VmFlags::VM_NONE; 189 190 if shm_flags.contains(ShmFlags::SHM_RDONLY) { 191 vm_flags |= VmFlags::VM_READ; 192 } else { 193 vm_flags |= VmFlags::VM_READ | VmFlags::VM_WRITE; 194 } 195 196 if shm_flags.contains(ShmFlags::SHM_EXEC) { 197 vm_flags |= VmFlags::VM_EXEC; 198 } 199 200 if shm_flags.contains(ShmFlags::SHM_HUGETLB) { 201 vm_flags |= VmFlags::VM_HUGETLB; 202 } 203 204 vm_flags 205 } 206 } 207 208 impl From<VmFlags> for MapFlags { 209 fn from(value: VmFlags) -> Self { 210 let mut map_flags = MapFlags::MAP_NONE; 211 212 if value.contains(VmFlags::VM_GROWSDOWN) { 213 map_flags |= MapFlags::MAP_GROWSDOWN; 214 } 215 216 if value.contains(VmFlags::VM_LOCKED) { 217 map_flags |= MapFlags::MAP_LOCKED; 218 } 219 220 if value.contains(VmFlags::VM_SYNC) { 221 map_flags |= MapFlags::MAP_SYNC; 222 } 223 224 if value.contains(VmFlags::VM_MAYSHARE) { 225 map_flags |= MapFlags::MAP_SHARED; 226 } 227 228 map_flags 229 } 230 } 231 232 impl From<VmFlags> for ProtFlags { 233 fn from(value: VmFlags) -> Self { 234 let mut prot_flags = ProtFlags::PROT_NONE; 235 236 if value.contains(VmFlags::VM_READ) { 237 prot_flags |= ProtFlags::PROT_READ; 238 } 239 240 if value.contains(VmFlags::VM_WRITE) { 241 prot_flags |= ProtFlags::PROT_WRITE; 242 } 243 244 if value.contains(VmFlags::VM_EXEC) { 245 prot_flags |= ProtFlags::PROT_EXEC; 246 } 247 248 prot_flags 249 } 250 } 251 252 impl Syscall { 253 pub fn brk(new_addr: VirtAddr) -> Result<VirtAddr, SystemError> { 254 // debug!("brk: new_addr={:?}", new_addr); 255 let address_space = AddressSpace::current()?; 256 let mut address_space = address_space.write(); 257 258 if new_addr < address_space.brk_start || new_addr >= MMArch::USER_END_VADDR { 259 return Ok(address_space.brk); 260 } 261 if new_addr == address_space.brk { 262 return Ok(address_space.brk); 263 } 264 265 unsafe { 266 address_space 267 .set_brk(VirtAddr::new(page_align_up(new_addr.data()))) 268 .ok(); 269 270 return Ok(address_space.sbrk(0).unwrap()); 271 } 272 } 273 274 pub fn sbrk(incr: isize) -> Result<VirtAddr, SystemError> { 275 let address_space = AddressSpace::current()?; 276 assert!(address_space.read().user_mapper.utable.is_current()); 277 let mut address_space = address_space.write(); 278 let r = unsafe { address_space.sbrk(incr) }; 279 280 return r; 281 } 282 283 /// ## mmap系统调用 284 /// 285 /// 该函数的实现参考了Linux内核的实现，但是并不完全相同。因为有些功能咱们还没实现 286 /// 287 /// ## 参数 288 /// 289 /// - `start_vaddr`：映射的起始地址 290 /// - `len`：映射的长度 291 /// - `prot`：保护标志 292 /// - `flags`：映射标志 293 /// - `fd`：文件描述符（暂时不支持） 294 /// - `offset`：文件偏移量 （暂时不支持） 295 /// 296 /// ## 返回值 297 /// 298 /// 成功时返回映射的起始地址，失败时返回错误码 299 pub fn mmap( 300 start_vaddr: VirtAddr, 301 len: usize, 302 prot_flags: usize, 303 map_flags: usize, 304 fd: i32, 305 offset: usize, 306 ) -> Result<usize, SystemError> { 307 let map_flags = MapFlags::from_bits_truncate(map_flags as u64); 308 let prot_flags = ProtFlags::from_bits_truncate(prot_flags as u64); 309 310 if start_vaddr < VirtAddr::new(DEFAULT_MMAP_MIN_ADDR) 311 && map_flags.contains(MapFlags::MAP_FIXED) 312 { 313 error!( 314 "mmap: MAP_FIXED is not supported for address below {}", 315 DEFAULT_MMAP_MIN_ADDR 316 ); 317 return Err(SystemError::EINVAL); 318 } 319 320 // 暂时不支持巨页映射 321 if map_flags.contains(MapFlags::MAP_HUGETLB) { 322 error!("mmap: not support huge page mapping"); 323 return Err(SystemError::ENOSYS); 324 } 325 let current_address_space = AddressSpace::current()?; 326 let start_page = if map_flags.contains(MapFlags::MAP_ANONYMOUS) { 327 // 匿名映射 328 current_address_space.write().map_anonymous( 329 start_vaddr, 330 len, 331 prot_flags, 332 map_flags, 333 true, 334 false, 335 )? 336 } else { 337 // 文件映射 338 current_address_space.write().file_mapping( 339 start_vaddr, 340 len, 341 prot_flags, 342 map_flags, 343 fd, 344 offset, 345 true, 346 false, 347 )? 348 }; 349 350 return Ok(start_page.virt_address().data()); 351 } 352 353 /// ## mremap系统调用 354 /// 355 /// 356 /// ## 参数 357 /// 358 /// - `old_vaddr`：原映射的起始地址 359 /// - `old_len`：原映射的长度 360 /// - `new_len`：重新映射的长度 361 /// - `mremap_flags`：重映射标志 362 /// - `new_vaddr`：重新映射的起始地址 363 /// 364 /// ## 返回值 365 /// 366 /// 成功时返回重映射的起始地址，失败时返回错误码 367 pub fn mremap( 368 old_vaddr: VirtAddr, 369 old_len: usize, 370 new_len: usize, 371 mremap_flags: MremapFlags, 372 new_vaddr: VirtAddr, 373 ) -> Result<usize, SystemError> { 374 // 需要重映射到新内存区域的情况下，必须包含MREMAP_MAYMOVE并且指定新地址 375 if mremap_flags.contains(MremapFlags::MREMAP_FIXED) 376 && (!mremap_flags.contains(MremapFlags::MREMAP_MAYMOVE) 377 || new_vaddr == VirtAddr::new(0)) 378 { 379 return Err(SystemError::EINVAL); 380 } 381 382 // 不取消旧映射的情况下，必须包含MREMAP_MAYMOVE并且新内存大小等于旧内存大小 383 if mremap_flags.contains(MremapFlags::MREMAP_DONTUNMAP) 384 && (!mremap_flags.contains(MremapFlags::MREMAP_MAYMOVE) || old_len != new_len) 385 { 386 return Err(SystemError::EINVAL); 387 } 388 389 // 旧内存地址必须对齐 390 if !old_vaddr.check_aligned(MMArch::PAGE_SIZE) { 391 return Err(SystemError::EINVAL); 392 } 393 394 // 将old_len、new_len 对齐页面大小 395 let old_len = page_align_up(old_len); 396 let new_len = page_align_up(new_len); 397 398 // 不允许重映射内存区域大小为0 399 if new_len == 0 { 400 return Err(SystemError::EINVAL); 401 } 402 403 let current_address_space = AddressSpace::current()?; 404 let vma = current_address_space.read().mappings.contains(old_vaddr); 405 if vma.is_none() { 406 return Err(SystemError::EINVAL); 407 } 408 let vma = vma.unwrap(); 409 let vm_flags = *vma.lock_irqsave().vm_flags(); 410 411 // 暂时不支持巨页映射 412 if vm_flags.contains(VmFlags::VM_HUGETLB) { 413 error!("mmap: not support huge page mapping"); 414 return Err(SystemError::ENOSYS); 415 } 416 417 // 缩小旧内存映射区域 418 if old_len > new_len { 419 Self::munmap(old_vaddr + new_len, old_len - new_len)?; 420 return Ok(old_vaddr.data()); 421 } 422 423 // 重映射到新内存区域 424 let r = current_address_space.write().mremap( 425 old_vaddr, 426 old_len, 427 new_len, 428 mremap_flags, 429 new_vaddr, 430 vm_flags, 431 )?; 432 433 if !mremap_flags.contains(MremapFlags::MREMAP_DONTUNMAP) { 434 Self::munmap(old_vaddr, old_len)?; 435 } 436 437 return Ok(r.data()); 438 } 439 440 /// ## munmap系统调用 441 /// 442 /// ## 参数 443 /// 444 /// - `start_vaddr`：取消映射的起始地址（已经对齐到页） 445 /// - `len`：取消映射的字节数(已经对齐到页) 446 /// 447 /// ## 返回值 448 /// 449 /// 成功时返回0，失败时返回错误码 450 pub fn munmap(start_vaddr: VirtAddr, len: usize) -> Result<usize, SystemError> { 451 assert!(start_vaddr.check_aligned(MMArch::PAGE_SIZE)); 452 assert!(check_aligned(len, MMArch::PAGE_SIZE)); 453 454 if unlikely(verify_area(start_vaddr, len).is_err()) { 455 return Err(SystemError::EINVAL); 456 } 457 if unlikely(len == 0) { 458 return Err(SystemError::EINVAL); 459 } 460 461 let current_address_space: Arc<AddressSpace> = AddressSpace::current()?; 462 let start_frame = VirtPageFrame::new(start_vaddr); 463 let page_count = PageFrameCount::new(len / MMArch::PAGE_SIZE); 464 465 current_address_space 466 .write() 467 .munmap(start_frame, page_count) 468 .map_err(|_| SystemError::EINVAL)?; 469 470 return Ok(0); 471 } 472 473 /// ## mprotect系统调用 474 /// 475 /// ## 参数 476 /// 477 /// - `start_vaddr`：起始地址(已经对齐到页) 478 /// - `len`：长度(已经对齐到页) 479 /// - `prot_flags`：保护标志 480 pub fn mprotect( 481 start_vaddr: VirtAddr, 482 len: usize, 483 prot_flags: usize, 484 ) -> Result<usize, SystemError> { 485 assert!(start_vaddr.check_aligned(MMArch::PAGE_SIZE)); 486 assert!(check_aligned(len, MMArch::PAGE_SIZE)); 487 488 if unlikely(verify_area(start_vaddr, len).is_err()) { 489 return Err(SystemError::EINVAL); 490 } 491 if unlikely(len == 0) { 492 return Err(SystemError::EINVAL); 493 } 494 495 let prot_flags = ProtFlags::from_bits(prot_flags as u64).ok_or(SystemError::EINVAL)?; 496 497 let current_address_space: Arc<AddressSpace> = AddressSpace::current()?; 498 let start_frame = VirtPageFrame::new(start_vaddr); 499 let page_count = PageFrameCount::new(len / MMArch::PAGE_SIZE); 500 501 current_address_space 502 .write() 503 .mprotect(start_frame, page_count, prot_flags) 504 .map_err(|_| SystemError::EINVAL)?; 505 return Ok(0); 506 } 507 508 /// ## madvise系统调用 509 /// 510 /// ## 参数 511 /// 512 /// - `start_vaddr`：起始地址(已经对齐到页) 513 /// - `len`：长度(已经对齐到页) 514 /// - `madv_flags`：建议标志 515 pub fn madvise( 516 start_vaddr: VirtAddr, 517 len: usize, 518 madv_flags: usize, 519 ) -> Result<usize, SystemError> { 520 if !start_vaddr.check_aligned(MMArch::PAGE_SIZE) || !check_aligned(len, MMArch::PAGE_SIZE) { 521 return Err(SystemError::EINVAL); 522 } 523 524 if unlikely(verify_area(start_vaddr, len).is_err()) { 525 return Err(SystemError::EINVAL); 526 } 527 if unlikely(len == 0) { 528 return Err(SystemError::EINVAL); 529 } 530 531 let madv_flags = MadvFlags::from_bits(madv_flags as u64).ok_or(SystemError::EINVAL)?; 532 533 let current_address_space: Arc<AddressSpace> = AddressSpace::current()?; 534 let start_frame = VirtPageFrame::new(start_vaddr); 535 let page_count = PageFrameCount::new(len / MMArch::PAGE_SIZE); 536 537 current_address_space 538 .write() 539 .madvise(start_frame, page_count, madv_flags) 540 .map_err(|_| SystemError::EINVAL)?; 541 return Ok(0); 542 } 543 544 /// ## msync系统调用 545 /// 546 /// ## 参数 547 /// 548 /// - `start`：起始地址(已经对齐到页) 549 /// - `len`：长度(已经对齐到页) 550 /// - `flags`：标志 551 pub fn msync(start: VirtAddr, len: usize, flags: usize) -> Result<usize, SystemError> { 552 if !start.check_aligned(MMArch::PAGE_SIZE) || !check_aligned(len, MMArch::PAGE_SIZE) { 553 return Err(SystemError::EINVAL); 554 } 555 556 if unlikely(verify_area(start, len).is_err()) { 557 return Err(SystemError::EINVAL); 558 } 559 if unlikely(len == 0) { 560 return Err(SystemError::EINVAL); 561 } 562 563 let mut start = start.data(); 564 let end = start + len; 565 let flags = MsFlags::from_bits_truncate(flags); 566 let mut unmapped_error = Ok(0); 567 568 if !flags.intersects(MsFlags::MS_ASYNC | MsFlags::MS_INVALIDATE | MsFlags::MS_SYNC) { 569 return Err(SystemError::EINVAL); 570 } 571 572 if flags.contains(MsFlags::MS_ASYNC | MsFlags::MS_SYNC) { 573 return Err(SystemError::EINVAL); 574 } 575 576 if end < start { 577 return Err(SystemError::ENOMEM); 578 } 579 580 if start == end { 581 return Ok(0); 582 } 583 584 let current_address_space = AddressSpace::current()?; 585 let mut err = Err(SystemError::ENOMEM); 586 let mut next_vma = current_address_space 587 .read() 588 .mappings 589 .find_nearest(VirtAddr::new(start)); 590 loop { 591 if let Some(vma) = next_vma.clone() { 592 let guard = vma.lock_irqsave(); 593 let vm_start = guard.region().start().data(); 594 let vm_end = guard.region().end().data(); 595 if start < vm_start { 596 if flags == MsFlags::MS_ASYNC { 597 break; 598 } 599 start = vm_start; 600 if start >= vm_end { 601 break; 602 } 603 unmapped_error = Err(SystemError::ENOMEM); 604 } 605 let vm_flags = *guard.vm_flags(); 606 if flags.contains(MsFlags::MS_INVALIDATE) && vm_flags.contains(VmFlags::VM_LOCKED) { 607 err = Err(SystemError::EBUSY); 608 break; 609 } 610 let file = guard.vm_file(); 611 let fstart = (start - vm_start) 612 + (guard.file_page_offset().unwrap_or(0) << MMArch::PAGE_SHIFT); 613 let fend = fstart + (core::cmp::min(end, vm_end) - start) - 1; 614 let old_start = start; 615 start = vm_end; 616 // log::info!("flags: {:?}", flags); 617 // log::info!("vm_flags: {:?}", vm_flags); 618 // log::info!("file: {:?}", file); 619 if flags.contains(MsFlags::MS_SYNC) && vm_flags.contains(VmFlags::VM_SHARED) { 620 if let Some(file) = file { 621 let old_pos = file.lseek(SeekFrom::SeekCurrent(0)).unwrap(); 622 file.lseek(SeekFrom::SeekSet(fstart as i64)).unwrap(); 623 err = file.write(len, unsafe { 624 from_raw_parts(old_start as *mut u8, fend - fstart + 1) 625 }); 626 file.lseek(SeekFrom::SeekSet(old_pos as i64)).unwrap(); 627 if err.is_err() { 628 break; 629 } else if start >= end { 630 err = unmapped_error; 631 break; 632 } 633 next_vma = current_address_space 634 .read() 635 .mappings 636 .find_nearest(VirtAddr::new(start)); 637 } 638 } else { 639 if start >= end { 640 err = unmapped_error; 641 break; 642 } 643 next_vma = current_address_space 644 .read() 645 .mappings 646 .find_nearest(VirtAddr::new(vm_end)); 647 } 648 } else { 649 return Err(SystemError::ENOMEM); 650 } 651 } 652 return err; 653 } 654 } 655