/// @Author: longjin@dragonos.org /// @Author: kongweichao@dragonos.org /// @Date: 2023-03-28 16:03:47 /// @FilePath: /DragonOS/kernel/src/mm/allocator/buddy.rs /// @Description: 伙伴分配器 use crate::arch::MMArch; use crate::mm::allocator::bump::BumpAllocator; use crate::mm::allocator::page_frame::{FrameAllocator, PageFrameCount, PageFrameUsage}; use crate::mm::{MemoryManagementArch, PhysAddr, VirtAddr}; use crate::{kdebug, kwarn}; use core::cmp::{max, min}; use core::fmt::Debug; use core::intrinsics::{likely, unlikely}; use core::{marker::PhantomData, mem}; // 一个全局变量MAX_ORDER,用来表示buddy算法的最大阶数 [MIN_ORDER, MAX_ORDER)左闭右开区间 const MAX_ORDER: usize = 31; // 4KB const MIN_ORDER: usize = 12; /// 保存buddy算法中每一页存放的BuddyEntry的信息,占据每个页的起始位置 #[derive(Debug)] pub struct PageList { // 页存放entry的数量 entry_num: usize, // 下一个页面的地址 next_page: PhysAddr, phantom: PhantomData, } impl Clone for PageList { fn clone(&self) -> Self { Self { entry_num: self.entry_num, next_page: self.next_page, phantom: PhantomData, } } } impl PageList { #[allow(dead_code)] fn empty() -> Self { Self { entry_num: 0, next_page: PhysAddr::new(0), phantom: PhantomData, } } fn new(entry_num: usize, next_page: PhysAddr) -> Self { Self { entry_num, next_page, phantom: PhantomData, } } } /// @brief: 用来表示 buddy 算法中的一个 buddy 块,整体存放在area的头部 // 这种方式会出现对齐问题 // #[repr(packed)] #[repr(C)] #[derive(Debug)] pub struct BuddyAllocator { // 存放每个阶的空闲“链表”的头部地址 free_area: [PhysAddr; (MAX_ORDER - MIN_ORDER) as usize], phantom: PhantomData, } impl BuddyAllocator { const BUDDY_ENTRIES: usize = // 定义一个变量记录buddy表的大小 (A::PAGE_SIZE - mem::size_of::>()) / mem::size_of::(); pub unsafe fn new(mut bump_allocator: BumpAllocator) -> Option { let initial_free_pages = bump_allocator.usage().free(); kdebug!("Free pages before init buddy: {:?}", initial_free_pages); kdebug!("Buddy entries: {}", Self::BUDDY_ENTRIES); // 最高阶的链表页数 let max_order_linked_list_page_num = max( 1, (((initial_free_pages.data() * A::PAGE_SIZE) >> (MAX_ORDER - 1)) + Self::BUDDY_ENTRIES - 1) / Self::BUDDY_ENTRIES, ); let mut free_area: [PhysAddr; (MAX_ORDER - MIN_ORDER) as usize] = [PhysAddr::new(0); (MAX_ORDER - MIN_ORDER) as usize]; // Buddy初始占用的空间从bump分配 for f in free_area.iter_mut() { let curr_page = bump_allocator.allocate_one(); // 保存每个阶的空闲链表的头部地址 *f = curr_page.unwrap(); // 清空当前页 core::ptr::write_bytes(MMArch::phys_2_virt(*f)?.data() as *mut u8, 0, A::PAGE_SIZE); let page_list: PageList = PageList::new(0, PhysAddr::new(0)); Self::write_page(*f, page_list); } // 分配最高阶的链表页 for _ in 1..max_order_linked_list_page_num { let curr_page = bump_allocator.allocate_one().unwrap(); // 清空当前页 core::ptr::write_bytes( MMArch::phys_2_virt(curr_page)?.data() as *mut u8, 0, A::PAGE_SIZE, ); let page_list: PageList = PageList::new(0, free_area[Self::order2index((MAX_ORDER - 1) as u8)]); Self::write_page(curr_page, page_list); free_area[Self::order2index((MAX_ORDER - 1) as u8)] = curr_page; } let initial_bump_offset = bump_allocator.offset(); let pages_to_buddy = bump_allocator.usage().free(); kdebug!("pages_to_buddy {:?}", pages_to_buddy); // kdebug!("initial_bump_offset {:#x}", initial_bump_offset); let mut paddr = initial_bump_offset; let mut remain_pages = pages_to_buddy; // 设置entry,这里假设了bump_allocator当前offset之后,所有的area的地址是连续的. // TODO: 这里需要修改,按照area来处理 for i in MIN_ORDER..MAX_ORDER { // kdebug!("i {i}, remain pages={}", remain_pages.data()); if remain_pages.data() < (1 << (i - MIN_ORDER)) { break; } assert!(paddr & ((1 << i) - 1) == 0); if likely(i != MAX_ORDER - 1) { // 要填写entry if paddr & (1 << i) != 0 { let page_list_paddr: PhysAddr = free_area[Self::order2index(i as u8)]; let mut page_list: PageList = Self::read_page(page_list_paddr); A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num), paddr, ); page_list.entry_num += 1; Self::write_page(page_list_paddr, page_list); paddr += 1 << i; remain_pages -= 1 << (i - MIN_ORDER); }; } else { // 往最大的阶数的链表中添加entry(注意要考虑到最大阶数的链表可能有多页) // 断言剩余页面数量是MAX_ORDER-1阶的整数倍 let mut entries = (remain_pages.data() * A::PAGE_SIZE) >> i; let mut page_list_paddr: PhysAddr = free_area[Self::order2index(i as u8)]; let block_size = 1usize << i; if entries > Self::BUDDY_ENTRIES { // 在第一页填写一些entries let num = entries % Self::BUDDY_ENTRIES; entries -= num; let mut page_list: PageList = Self::read_page(page_list_paddr); for _j in 0..num { A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num), paddr, ); page_list.entry_num += 1; paddr += block_size; remain_pages -= 1 << (i - MIN_ORDER); } page_list_paddr = page_list.next_page; Self::write_page(page_list_paddr, page_list); assert!(!page_list_paddr.is_null()); } while entries > 0 { let mut page_list: PageList = Self::read_page(page_list_paddr); for _ in 0..Self::BUDDY_ENTRIES { A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num), paddr, ); page_list.entry_num += 1; paddr += block_size; remain_pages -= 1 << (i - MIN_ORDER); entries -= 1; if entries == 0 { break; } } page_list_paddr = page_list.next_page; Self::write_page(page_list_paddr, page_list); if likely(entries > 0) { assert!(!page_list_paddr.is_null()); } } } } let mut remain_bytes = remain_pages.data() * A::PAGE_SIZE; assert!(remain_bytes < (1 << MAX_ORDER - 1)); for i in (MIN_ORDER..MAX_ORDER).rev() { if remain_bytes >= (1 << i) { assert!(paddr & ((1 << i) - 1) == 0); let page_list_paddr: PhysAddr = free_area[Self::order2index(i as u8)]; let mut page_list: PageList = Self::read_page(page_list_paddr); A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num), paddr, ); page_list.entry_num += 1; Self::write_page(page_list_paddr, page_list); paddr += 1 << i; remain_bytes -= 1 << i; } } assert!(remain_bytes == 0); assert!(paddr == initial_bump_offset + pages_to_buddy.data() * A::PAGE_SIZE); // Self::print_free_area(free_area); let allocator = Self { free_area, phantom: PhantomData, }; Some(allocator) } /// 获取第j个entry的虚拟地址, /// j从0开始计数 pub fn entry_virt_addr(base_addr: PhysAddr, j: usize) -> VirtAddr { let entry_virt_addr = unsafe { A::phys_2_virt(Self::entry_addr(base_addr, j)) }; return entry_virt_addr.unwrap(); } pub fn entry_addr(base_addr: PhysAddr, j: usize) -> PhysAddr { let entry_addr = base_addr + mem::size_of::>() + j * mem::size_of::(); return entry_addr; } pub fn read_page(addr: PhysAddr) -> T { let page_list = unsafe { A::read(A::phys_2_virt(addr).unwrap()) }; return page_list; } pub fn write_page(curr_page: PhysAddr, page_list: PageList) { // 把物理地址转换为虚拟地址 let virt_addr = unsafe { A::phys_2_virt(curr_page) }; let virt_addr = virt_addr.unwrap(); unsafe { A::write(virt_addr, page_list) }; } /// 从order转换为free_area的下标 /// /// # 参数 /// /// - `order` - order /// /// # 返回值 /// /// free_area的下标 #[inline] fn order2index(order: u8) -> usize { (order as usize - MIN_ORDER) as usize } /// 从空闲链表的开头,取出1个指定阶数的伙伴块,如果没有,则返回None /// /// ## 参数 /// /// - `order` - 伙伴块的阶数 fn pop_front(&mut self, order: u8) -> Option { let mut alloc_in_specific_order = |spec_order: u8| { // 先尝试在order阶的“空闲链表”的开头位置分配一个伙伴块 let mut page_list_addr = self.free_area[Self::order2index(spec_order)]; let mut page_list: PageList = Self::read_page(page_list_addr); // 循环删除头部的空闲链表页 while page_list.entry_num == 0 { let next_page_list_addr = page_list.next_page; // 找完了,都是空的 if next_page_list_addr.is_null() { return None; } if !next_page_list_addr.is_null() { // 此时page_list已经没有空闲伙伴块了,又因为非唯一页,需要删除该page_list self.free_area[Self::order2index(spec_order)] = next_page_list_addr; drop(page_list); // kdebug!("FREE: page_list_addr={:b}", page_list_addr.data()); unsafe { self.buddy_free(page_list_addr, MMArch::PAGE_SHIFT as u8); } } // 由于buddy_free可能导致首部的链表页发生变化,因此需要重新读取 let next_page_list_addr = self.free_area[Self::order2index(spec_order)]; assert!(!next_page_list_addr.is_null()); page_list = Self::read_page(next_page_list_addr); page_list_addr = next_page_list_addr; } // 有空闲页面,直接分配 if page_list.entry_num > 0 { let entry: PhysAddr = unsafe { A::read(Self::entry_virt_addr( page_list_addr, page_list.entry_num - 1, )) }; // 清除该entry unsafe { A::write( Self::entry_virt_addr(page_list_addr, page_list.entry_num - 1), PhysAddr::new(0), ) }; if entry.is_null() { panic!( "entry is null, entry={:?}, order={}, entry_num = {}", entry, spec_order, page_list.entry_num - 1 ); } // kdebug!("entry={entry:?}"); // 更新page_list的entry_num page_list.entry_num -= 1; let tmp_current_entry_num = page_list.entry_num; if page_list.entry_num == 0 { if !page_list.next_page.is_null() { // 此时page_list已经没有空闲伙伴块了,又因为非唯一页,需要删除该page_list self.free_area[Self::order2index(spec_order)] = page_list.next_page; drop(page_list); unsafe { self.buddy_free(page_list_addr, MMArch::PAGE_SHIFT as u8) }; } else { Self::write_page(page_list_addr, page_list); } } else { // 若entry_num不为0,说明该page_list还有空闲伙伴块,需要更新该page_list // 把更新后的page_list写回 Self::write_page(page_list_addr, page_list.clone()); } // 检测entry 是否对齐 if !entry.check_aligned(1 << spec_order) { panic!("entry={:?} is not aligned, spec_order={spec_order}, page_list.entry_num={}", entry, tmp_current_entry_num); } return Some(entry); } return None; }; let result: Option = alloc_in_specific_order(order as u8); // kdebug!("result={:?}", result); if result.is_some() { return result; } // 尝试从更大的链表中分裂 let mut current_order = (order + 1) as usize; let mut x: Option = None; while current_order < MAX_ORDER { x = alloc_in_specific_order(current_order as u8); // kdebug!("current_order={:?}", current_order); if x.is_some() { break; } current_order += 1; } // kdebug!("x={:?}", x); // 如果找到一个大的块,就进行分裂 if x.is_some() { // 分裂到order阶 while current_order > order as usize { current_order -= 1; // 把后面那半块放回空闲链表 let buddy = *x.as_ref().unwrap() + (1 << current_order); // kdebug!("x={:?}, buddy={:?}", x, buddy); // kdebug!("current_order={:?}, buddy={:?}", current_order, buddy); unsafe { self.buddy_free(buddy, current_order as u8) }; } return x; } return None; } /// 从伙伴系统中分配count个页面 /// /// ## 参数 /// /// - `count`:需要分配的页面数 /// /// ## 返回值 /// /// 返回分配的页面的物理地址和页面数 fn buddy_alloc(&mut self, count: PageFrameCount) -> Option<(PhysAddr, PageFrameCount)> { assert!(count.data().is_power_of_two()); // 计算需要分配的阶数 let mut order = log2(count.data() as usize); if count.data() & ((1 << order) - 1) != 0 { order += 1; } let order = (order + MIN_ORDER) as u8; if order as usize >= MAX_ORDER { return None; } // kdebug!("buddy_alloc: order = {}", order); // 获取该阶数的一个空闲页面 let free_addr = self.pop_front(order); // kdebug!( // "buddy_alloc: order = {}, free_addr = {:?}", // order, // free_addr // ); return free_addr .map(|addr| (addr, PageFrameCount::new(1 << (order as usize - MIN_ORDER)))); } /// 释放一个块 /// /// ## 参数 /// /// - `base` - 块的起始地址 /// - `order` - 块的阶数 unsafe fn buddy_free(&mut self, mut base: PhysAddr, order: u8) { // kdebug!("buddy_free: base = {:?}, order = {}", base, order); let mut order = order as usize; while order < MAX_ORDER { // 检测地址是否合法 if base.data() & ((1 << (order)) - 1) != 0 { panic!( "buddy_free: base is not aligned, base = {:#x}, order = {}", base.data(), order ); } // 在链表中寻找伙伴块 // 伙伴块的地址是base ^ (1 << order) let buddy_addr = PhysAddr::new(base.data() ^ (1 << order)); let first_page_list_paddr = self.free_area[Self::order2index(order as u8)]; let mut page_list_paddr = first_page_list_paddr; let mut page_list: PageList = Self::read_page(page_list_paddr); let first_page_list = page_list.clone(); let mut buddy_entry_virt_vaddr = None; let mut buddy_entry_page_list_paddr = None; // 除非order是最大的,否则尝试查找伙伴块 if likely(order != MAX_ORDER - 1) { 'outer: loop { for i in 0..page_list.entry_num { let entry_virt_addr = Self::entry_virt_addr(page_list_paddr, i); let entry: PhysAddr = unsafe { A::read(entry_virt_addr) }; if entry == buddy_addr { // 找到了伙伴块,记录该entry相关信息,然后退出查找 buddy_entry_virt_vaddr = Some(entry_virt_addr); buddy_entry_page_list_paddr = Some(page_list_paddr); break 'outer; } } if page_list.next_page.is_null() { break; } page_list_paddr = page_list.next_page; page_list = Self::read_page(page_list_paddr); } } // 如果没有找到伙伴块 if buddy_entry_virt_vaddr.is_none() { assert!( page_list.entry_num <= Self::BUDDY_ENTRIES, "buddy_free: page_list.entry_num > Self::BUDDY_ENTRIES" ); // 当前第一个page_list没有空间了 if first_page_list.entry_num == Self::BUDDY_ENTRIES { // 如果当前order是最小的,那么就把这个块当作新的page_list使用 let new_page_list_addr = if order == MIN_ORDER { base } else { // 否则分配新的page_list // 请注意,分配之后,有可能当前的entry_num会减1(伙伴块分裂),造成出现整个链表为null的entry数量为Self::BUDDY_ENTRIES+1的情况 // 但是不影响,我们在后面插入链表项的时候,会处理这种情况,检查链表中的第2个页是否有空位 self.buddy_alloc(PageFrameCount::new(1)) .expect("buddy_alloc failed: no enough memory") .0 }; // 清空这个页面 core::ptr::write_bytes( A::phys_2_virt(new_page_list_addr) .expect( "Buddy free: failed to get virt address of [new_page_list_addr]", ) .as_ptr::(), 0, 1 << order, ); assert!( first_page_list_paddr == self.free_area[Self::order2index(order as u8)] ); // 初始化新的page_list let new_page_list = PageList::new(0, first_page_list_paddr); Self::write_page(new_page_list_addr, new_page_list); self.free_area[Self::order2index(order as u8)] = new_page_list_addr; } // 由于上面可能更新了第一个链表页,因此需要重新获取这个值 let first_page_list_paddr = self.free_area[Self::order2index(order as u8)]; let first_page_list: PageList = Self::read_page(first_page_list_paddr); // 检查第二个page_list是否有空位 let second_page_list = if first_page_list.next_page.is_null() { None } else { Some(Self::read_page::>(first_page_list.next_page)) }; let (paddr, mut page_list) = if let Some(second) = second_page_list { // 第二个page_list有空位 // 应当符合之前的假设:还有1个空位 assert!(second.entry_num == Self::BUDDY_ENTRIES - 1); (first_page_list.next_page, second) } else { // 在第一个page list中分配 (first_page_list_paddr, first_page_list) }; // kdebug!("to write entry, page_list_base={paddr:?}, page_list.entry_num={}, value={base:?}", page_list.entry_num); assert!(page_list.entry_num < Self::BUDDY_ENTRIES); // 把要归还的块,写入到链表项中 unsafe { A::write(Self::entry_virt_addr(paddr, page_list.entry_num), base) } page_list.entry_num += 1; Self::write_page(paddr, page_list); return; } else { // 如果找到了伙伴块,合并,向上递归 // 伙伴块所在的表项的虚拟地址 let buddy_entry_virt_addr = buddy_entry_virt_vaddr.unwrap(); // 伙伴块所在的page_list的物理地址 let buddy_entry_page_list_paddr = buddy_entry_page_list_paddr.unwrap(); let mut page_list_paddr = self.free_area[Self::order2index(order as u8)]; let mut page_list = Self::read_page::>(page_list_paddr); // 找第一个有空闲块的链表页。跳过空闲链表页。不进行回收的原因是担心出现死循环 while page_list.entry_num == 0 { if page_list.next_page.is_null() { panic!( "buddy_free: page_list.entry_num == 0 && page_list.next_page.is_null()" ); } page_list_paddr = page_list.next_page; page_list = Self::read_page(page_list_paddr); } // 如果伙伴块不在第一个链表页,则把第一个链表中的某个空闲块替换到伙伴块的位置 if page_list_paddr != buddy_entry_page_list_paddr { let entry: PhysAddr = unsafe { A::read(Self::entry_virt_addr( page_list_paddr, page_list.entry_num - 1, )) }; // 把这个空闲块写入到伙伴块的位置 unsafe { A::write(buddy_entry_virt_addr, entry); } // 设置刚才那个entry为空 unsafe { A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num - 1), PhysAddr::new(0), ); } // 更新当前链表页的统计数据 page_list.entry_num -= 1; Self::write_page(page_list_paddr, page_list); } else { // 伙伴块所在的链表页就是第一个链表页 let last_entry: PhysAddr = unsafe { A::read(Self::entry_virt_addr( page_list_paddr, page_list.entry_num - 1, )) }; // 如果最后一个空闲块不是伙伴块,则把最后一个空闲块移动到伙伴块的位置 // 否则后面的操作也将删除这个伙伴块 if last_entry != buddy_addr { unsafe { A::write(buddy_entry_virt_addr, last_entry); A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num - 1), PhysAddr::new(0), ); } } else { unsafe { A::write( Self::entry_virt_addr(page_list_paddr, page_list.entry_num - 1), PhysAddr::new(0), ); } } // 更新当前链表页的统计数据 page_list.entry_num -= 1; Self::write_page(page_list_paddr, page_list); } } base = min(base, buddy_addr); order += 1; } // 走到这一步,order应该为MAX_ORDER-1 assert!(order == MAX_ORDER - 1); } } impl FrameAllocator for BuddyAllocator { unsafe fn allocate(&mut self, count: PageFrameCount) -> Option<(PhysAddr, PageFrameCount)> { return self.buddy_alloc(count); } /// 释放一个块 /// /// ## 参数 /// /// - `base` - 块的起始地址 /// - `count` - 块的页数(必须是2的幂) /// /// ## Panic /// /// 如果count不是2的幂,会panic unsafe fn free(&mut self, base: PhysAddr, count: PageFrameCount) { // 要求count是2的幂 if unlikely(!count.data().is_power_of_two()) { kwarn!("buddy free: count is not power of two"); } let mut order = log2(count.data() as usize); if count.data() & ((1 << order) - 1) != 0 { order += 1; } let order = (order + MIN_ORDER) as u8; // kdebug!("free: base={:?}, count={:?}", base, count); self.buddy_free(base, order); } unsafe fn usage(&self) -> PageFrameUsage { todo!("BuddyAllocator::usage") } } /// 一个用于计算整数的对数的函数,会向下取整。(由于内核不能进行浮点运算,因此需要这个函数) fn log2(x: usize) -> usize { let leading_zeros = x.leading_zeros() as usize; let log2x = 63 - leading_zeros; return log2x; }