From: Jian J Wang Date: Tue, 28 Nov 2017 13:49:31 +0000 (+0800) Subject: MdeModulePkg/DxeIpl: Mark page table as read-only X-Git-Tag: edk2-stable201903~2860 X-Git-Url: https://git.proxmox.com/?p=mirror_edk2.git;a=commitdiff_plain;h=2ac1730bf2a51d6d0483347a0218e1768d6d7992 MdeModulePkg/DxeIpl: Mark page table as read-only This patch will set the memory pages used for page table as read-only memory after the paging is setup. CR0.WP must set to let it take into effect. A simple page table memory management mechanism, page table pool concept, is introduced to simplify the page table memory allocation and protection. It will also help to reduce the potential recursive "split" action during updating memory paging attributes. The basic idea is to allocate a bunch of continuous pages of memory in advance as one or more page table pools, and all future page tables consumption will happen in those pool instead of system memory. If the page pool is reserved at the boundary of 2MB page and with same size of 2MB page, there's no page granularity "split" operation will be needed, because the memory of new page tables (if needed) will be usually in the same page as target page table you're working on. And since we have centralized page tables (a few 2MB pages), it's easier to protect them by changing their attributes to be read-only once and for all. There's no need to apply the protection for new page tables any more as long as the pool has free pages available. Once current page table pool has been used up, one can allocate another 2MB memory pool and just set this new 2MB memory block to be read-only instead of setting the new page tables one page by one page. Two new PCDs PcdPageTablePoolUnitSize and PcdPageTablePoolAlignment are used to specify the size and alignment for page table pool. For IA32 processor 0x200000 (2MB) is the only choice for both of them to meet the requirement of page table pool. Laszlo (lersek@redhat.com) did a regression test on QEMU virtual platform with one middle version of this series patch. The details can be found at https://lists.01.org/pipermail/edk2-devel/2017-December/018625.html There're a few changes after his work. Cc: Jiewen Yao Cc: Star Zeng Cc: Eric Dong Cc: Ruiyu Ni Contributed-under: TianoCore Contribution Agreement 1.1 Signed-off-by: Jian J Wang Reviewed-by: Jiewen Yao --- diff --git a/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c b/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c index 441096ad0f..0ec60893ee 100644 --- a/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c +++ b/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c @@ -99,7 +99,7 @@ Create4GPageTablesIa32Pae ( NumberOfPdpEntriesNeeded = (UINT32) LShiftU64 (1, (PhysicalAddressBits - 30)); TotalPagesNum = NumberOfPdpEntriesNeeded + 1; - PageAddress = (UINTN) AllocatePages (TotalPagesNum); + PageAddress = (UINTN) AllocatePageTableMemory (TotalPagesNum); ASSERT (PageAddress != 0); PageMap = (VOID *) PageAddress; @@ -149,6 +149,12 @@ Create4GPageTablesIa32Pae ( ); } + // + // Protect the page table by marking the memory used for page table to be + // read-only. + // + EnablePageTableProtection ((UINTN)PageMap, FALSE); + return (UINTN) PageMap; } diff --git a/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.c b/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.c index 7f63144510..26116e420c 100644 --- a/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.c +++ b/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.c @@ -31,6 +31,11 @@ WITHOUT WARRANTIES OR REPRESENTATIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED. #include "DxeIpl.h" #include "VirtualMemory.h" +// +// Global variable to keep track current available memory used as page table. +// +PAGE_TABLE_POOL *mPageTablePool = NULL; + /** Clear legacy memory located at the first 4K-page, if available. @@ -150,6 +155,110 @@ ToSplitPageTable ( return FALSE; } +/** + Initialize a buffer pool for page table use only. + + To reduce the potential split operation on page table, the pages reserved for + page table should be allocated in the times of PAGE_TABLE_POOL_UNIT_PAGES and + at the boundary of PAGE_TABLE_POOL_ALIGNMENT. So the page pool is always + initialized with number of pages greater than or equal to the given PoolPages. + + Once the pages in the pool are used up, this method should be called again to + reserve at least another PAGE_TABLE_POOL_UNIT_PAGES. But usually this won't + happen in practice. + + @param PoolPages The least page number of the pool to be created. + + @retval TRUE The pool is initialized successfully. + @retval FALSE The memory is out of resource. +**/ +BOOLEAN +InitializePageTablePool ( + IN UINTN PoolPages + ) +{ + VOID *Buffer; + + // + // Always reserve at least PAGE_TABLE_POOL_UNIT_PAGES, including one page for + // header. + // + PoolPages += 1; // Add one page for header. + PoolPages = ((PoolPages - 1) / PAGE_TABLE_POOL_UNIT_PAGES + 1) * + PAGE_TABLE_POOL_UNIT_PAGES; + Buffer = AllocateAlignedPages (PoolPages, PAGE_TABLE_POOL_ALIGNMENT); + if (Buffer == NULL) { + DEBUG ((DEBUG_ERROR, "ERROR: Out of aligned pages\r\n")); + return FALSE; + } + + // + // Link all pools into a list for easier track later. + // + if (mPageTablePool == NULL) { + mPageTablePool = Buffer; + mPageTablePool->NextPool = mPageTablePool; + } else { + ((PAGE_TABLE_POOL *)Buffer)->NextPool = mPageTablePool->NextPool; + mPageTablePool->NextPool = Buffer; + mPageTablePool = Buffer; + } + + // + // Reserve one page for pool header. + // + mPageTablePool->FreePages = PoolPages - 1; + mPageTablePool->Offset = EFI_PAGES_TO_SIZE (1); + + return TRUE; +} + +/** + This API provides a way to allocate memory for page table. + + This API can be called more than once to allocate memory for page tables. + + Allocates the number of 4KB pages and returns a pointer to the allocated + buffer. The buffer returned is aligned on a 4KB boundary. + + If Pages is 0, then NULL is returned. + If there is not enough memory remaining to satisfy the request, then NULL is + returned. + + @param Pages The number of 4 KB pages to allocate. + + @return A pointer to the allocated buffer or NULL if allocation fails. + +**/ +VOID * +AllocatePageTableMemory ( + IN UINTN Pages + ) +{ + VOID *Buffer; + + if (Pages == 0) { + return NULL; + } + + // + // Renew the pool if necessary. + // + if (mPageTablePool == NULL || + Pages > mPageTablePool->FreePages) { + if (!InitializePageTablePool (Pages)) { + return NULL; + } + } + + Buffer = (UINT8 *)mPageTablePool + mPageTablePool->Offset; + + mPageTablePool->Offset += EFI_PAGES_TO_SIZE (Pages); + mPageTablePool->FreePages -= Pages; + + return Buffer; +} + /** Split 2M page to 4K. @@ -177,7 +286,7 @@ Split2MPageTo4K ( // AddressEncMask = PcdGet64 (PcdPteMemoryEncryptionAddressOrMask) & PAGING_1G_ADDRESS_MASK_64; - PageTableEntry = AllocatePages (1); + PageTableEntry = AllocatePageTableMemory (1); ASSERT (PageTableEntry != NULL); // @@ -238,7 +347,7 @@ Split1GPageTo2M ( // AddressEncMask = PcdGet64 (PcdPteMemoryEncryptionAddressOrMask) & PAGING_1G_ADDRESS_MASK_64; - PageDirectoryEntry = AllocatePages (1); + PageDirectoryEntry = AllocatePageTableMemory (1); ASSERT (PageDirectoryEntry != NULL); // @@ -265,6 +374,184 @@ Split1GPageTo2M ( } } +/** + Set one page of page table pool memory to be read-only. + + @param[in] PageTableBase Base address of page table (CR3). + @param[in] Address Start address of a page to be set as read-only. + @param[in] Level4Paging Level 4 paging flag. + +**/ +VOID +SetPageTablePoolReadOnly ( + IN UINTN PageTableBase, + IN EFI_PHYSICAL_ADDRESS Address, + IN BOOLEAN Level4Paging + ) +{ + UINTN Index; + UINTN EntryIndex; + UINT64 AddressEncMask; + EFI_PHYSICAL_ADDRESS PhysicalAddress; + UINT64 *PageTable; + UINT64 *NewPageTable; + UINT64 PageAttr; + UINT64 LevelSize[5]; + UINT64 LevelMask[5]; + UINTN LevelShift[5]; + UINTN Level; + UINT64 PoolUnitSize; + + ASSERT (PageTableBase != 0); + + // + // Since the page table is always from page table pool, which is always + // located at the boundary of PcdPageTablePoolAlignment, we just need to + // set the whole pool unit to be read-only. + // + Address = Address & PAGE_TABLE_POOL_ALIGN_MASK; + + LevelShift[1] = PAGING_L1_ADDRESS_SHIFT; + LevelShift[2] = PAGING_L2_ADDRESS_SHIFT; + LevelShift[3] = PAGING_L3_ADDRESS_SHIFT; + LevelShift[4] = PAGING_L4_ADDRESS_SHIFT; + + LevelMask[1] = PAGING_4K_ADDRESS_MASK_64; + LevelMask[2] = PAGING_2M_ADDRESS_MASK_64; + LevelMask[3] = PAGING_1G_ADDRESS_MASK_64; + LevelMask[4] = PAGING_1G_ADDRESS_MASK_64; + + LevelSize[1] = SIZE_4KB; + LevelSize[2] = SIZE_2MB; + LevelSize[3] = SIZE_1GB; + LevelSize[4] = SIZE_512GB; + + AddressEncMask = PcdGet64 (PcdPteMemoryEncryptionAddressOrMask) & + PAGING_1G_ADDRESS_MASK_64; + PageTable = (UINT64 *)(UINTN)PageTableBase; + PoolUnitSize = PAGE_TABLE_POOL_UNIT_SIZE; + + for (Level = (Level4Paging) ? 4 : 3; Level > 0; --Level) { + Index = ((UINTN)RShiftU64 (Address, LevelShift[Level])); + Index &= PAGING_PAE_INDEX_MASK; + + PageAttr = PageTable[Index]; + if ((PageAttr & IA32_PG_PS) == 0) { + // + // Go to next level of table. + // + PageTable = (UINT64 *)(UINTN)(PageAttr & ~AddressEncMask & + PAGING_4K_ADDRESS_MASK_64); + continue; + } + + if (PoolUnitSize >= LevelSize[Level]) { + // + // Clear R/W bit if current page granularity is not larger than pool unit + // size. + // + if ((PageAttr & IA32_PG_RW) != 0) { + while (PoolUnitSize > 0) { + // + // PAGE_TABLE_POOL_UNIT_SIZE and PAGE_TABLE_POOL_ALIGNMENT are fit in + // one page (2MB). Then we don't need to update attributes for pages + // crossing page directory. ASSERT below is for that purpose. + // + ASSERT (Index < EFI_PAGE_SIZE/sizeof (UINT64)); + + PageTable[Index] &= ~(UINT64)IA32_PG_RW; + PoolUnitSize -= LevelSize[Level]; + + ++Index; + } + } + + break; + + } else { + // + // The smaller granularity of page must be needed. + // + NewPageTable = AllocatePageTableMemory (1); + ASSERT (NewPageTable != NULL); + + PhysicalAddress = PageAttr & LevelMask[Level]; + for (EntryIndex = 0; + EntryIndex < EFI_PAGE_SIZE/sizeof (UINT64); + ++EntryIndex) { + NewPageTable[EntryIndex] = PhysicalAddress | AddressEncMask | + IA32_PG_P | IA32_PG_RW; + if (Level > 1) { + NewPageTable[EntryIndex] |= IA32_PG_PS; + } + PhysicalAddress += LevelSize[Level]; + } + + PageTable[Index] = (UINT64)(UINTN)NewPageTable | AddressEncMask | + IA32_PG_P | IA32_PG_RW; + PageTable = NewPageTable; + } + } +} + +/** + Prevent the memory pages used for page table from been overwritten. + + @param[in] PageTableBase Base address of page table (CR3). + @param[in] Level4Paging Level 4 paging flag. + +**/ +VOID +EnablePageTableProtection ( + IN UINTN PageTableBase, + IN BOOLEAN Level4Paging + ) +{ + PAGE_TABLE_POOL *HeadPool; + PAGE_TABLE_POOL *Pool; + UINT64 PoolSize; + EFI_PHYSICAL_ADDRESS Address; + + if (mPageTablePool == NULL) { + return; + } + + // + // Disable write protection, because we need to mark page table to be write + // protected. + // + AsmWriteCr0 (AsmReadCr0() & ~CR0_WP); + + // + // SetPageTablePoolReadOnly might update mPageTablePool. It's safer to + // remember original one in advance. + // + HeadPool = mPageTablePool; + Pool = HeadPool; + do { + Address = (EFI_PHYSICAL_ADDRESS)(UINTN)Pool; + PoolSize = Pool->Offset + EFI_PAGES_TO_SIZE (Pool->FreePages); + + // + // The size of one pool must be multiple of PAGE_TABLE_POOL_UNIT_SIZE, which + // is one of page size of the processor (2MB by default). Let's apply the + // protection to them one by one. + // + while (PoolSize > 0) { + SetPageTablePoolReadOnly(PageTableBase, Address, Level4Paging); + Address += PAGE_TABLE_POOL_UNIT_SIZE; + PoolSize -= PAGE_TABLE_POOL_UNIT_SIZE; + } + + Pool = Pool->NextPool; + } while (Pool != HeadPool); + + // + // Enable write protection, after page table attribute updated. + // + AsmWriteCr0 (AsmReadCr0() | CR0_WP); +} + /** Allocates and fills in the Page Directory and Page Table Entries to establish a 1:1 Virtual to Physical mapping. @@ -360,7 +647,7 @@ CreateIdentityMappingPageTables ( } else { TotalPagesNum = NumberOfPml4EntriesNeeded + 1; } - BigPageAddress = (UINTN) AllocatePages (TotalPagesNum); + BigPageAddress = (UINTN) AllocatePageTableMemory (TotalPagesNum); ASSERT (BigPageAddress != 0); // @@ -455,6 +742,12 @@ CreateIdentityMappingPageTables ( ); } + // + // Protect the page table by marking the memory used for page table to be + // read-only. + // + EnablePageTableProtection ((UINTN)PageMap, TRUE); + if (PcdGetBool (PcdSetNxForStack)) { EnableExecuteDisableBit (); } diff --git a/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.h b/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.h index 26a2100f0b..e7959b2cf0 100644 --- a/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.h +++ b/MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.h @@ -148,11 +148,37 @@ typedef union { #pragma pack() +#define CR0_WP BIT16 + #define IA32_PG_P BIT0 #define IA32_PG_RW BIT1 +#define IA32_PG_PS BIT7 + +#define PAGING_PAE_INDEX_MASK 0x1FF +#define PAGING_4K_ADDRESS_MASK_64 0x000FFFFFFFFFF000ull +#define PAGING_2M_ADDRESS_MASK_64 0x000FFFFFFFE00000ull #define PAGING_1G_ADDRESS_MASK_64 0x000FFFFFC0000000ull +#define PAGING_L1_ADDRESS_SHIFT 12 +#define PAGING_L2_ADDRESS_SHIFT 21 +#define PAGING_L3_ADDRESS_SHIFT 30 +#define PAGING_L4_ADDRESS_SHIFT 39 + +#define PAGING_PML4E_NUMBER 4 + +#define PAGE_TABLE_POOL_ALIGNMENT BASE_2MB +#define PAGE_TABLE_POOL_UNIT_SIZE SIZE_2MB +#define PAGE_TABLE_POOL_UNIT_PAGES EFI_SIZE_TO_PAGES (PAGE_TABLE_POOL_UNIT_SIZE) +#define PAGE_TABLE_POOL_ALIGN_MASK \ + (~(EFI_PHYSICAL_ADDRESS)(PAGE_TABLE_POOL_ALIGNMENT - 1)) + +typedef struct { + VOID *NextPool; + UINTN Offset; + UINTN FreePages; +} PAGE_TABLE_POOL; + /** Enable Execute Disable Bit. @@ -252,4 +278,39 @@ IsNullDetectionEnabled ( VOID ); +/** + Prevent the memory pages used for page table from been overwritten. + + @param[in] PageTableBase Base address of page table (CR3). + @param[in] Level4Paging Level 4 paging flag. + +**/ +VOID +EnablePageTableProtection ( + IN UINTN PageTableBase, + IN BOOLEAN Level4Paging + ); + +/** + This API provides a way to allocate memory for page table. + + This API can be called more than once to allocate memory for page tables. + + Allocates the number of 4KB pages and returns a pointer to the allocated + buffer. The buffer returned is aligned on a 4KB boundary. + + If Pages is 0, then NULL is returned. + If there is not enough memory remaining to satisfy the request, then NULL is + returned. + + @param Pages The number of 4 KB pages to allocate. + + @return A pointer to the allocated buffer or NULL if allocation fails. + +**/ +VOID * +AllocatePageTableMemory ( + IN UINTN Pages + ); + #endif