Hi,
> -----Original Message-----
> From: Ni, Ruiyu
> Sent: Wednesday, December 06, 2017 6:05 PM
> To: Wang, Jian J <jian.j.w...@intel.com>; edk2-devel@lists.01.org
> Cc: Yao, Jiewen <jiewen....@intel.com>; Dong, Eric <eric.d...@intel.com>;
> Zeng, Star <star.z...@intel.com>
> Subject: Re: [edk2] [PATCH v3 1/2] MdeModulePkg/DxeIpl: Mark page table as
> read-only
>
> On 12/5/2017 4:16 PM, Jian J Wang wrote:
> >> v3:
> >> Remove the public definition of PAGE_TABLE_POOL_HEADER but keep
> similar
> >> concept locally. CpuDxe has its own page table pool.
> >
> >> v2:
> >> Introduce page table pool to ease the page table memory allocation and
> >> protection, which replaces the direct calling of AllocatePages().
> >
> > 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.
> >
> > Cc: Jiewen Yao <jiewen....@intel.com>
> > Cc: Star Zeng <star.z...@intel.com>
> > Cc: Eric Dong <eric.d...@intel.com>
> > Cc: Ruiyu Ni <ruiyu...@intel.com>
> > Contributed-under: TianoCore Contribution Agreement 1.1
> > Signed-off-by: Jian J Wang <jian.j.w...@intel.com>
> > ---
> > MdeModulePkg/Core/DxeIplPeim/DxeIpl.h | 34 +++
> > MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c | 8 +-
> > MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.c | 301
> ++++++++++++++++++++++-
> > MdeModulePkg/Core/DxeIplPeim/X64/VirtualMemory.h | 26 ++
> > 4 files changed, 365 insertions(+), 4 deletions(-)
> >
> > diff --git a/MdeModulePkg/Core/DxeIplPeim/DxeIpl.h
> b/MdeModulePkg/Core/DxeIplPeim/DxeIpl.h
> > index f3aabdb7e0..9dc80b1508 100644
> > --- a/MdeModulePkg/Core/DxeIplPeim/DxeIpl.h
> > +++ b/MdeModulePkg/Core/DxeIplPeim/DxeIpl.h
> > @@ -265,4 +265,38 @@ IsNullDetectionEnabled (
> > VOID
> > );
> >
> > +/**
> > + Prevent the memory pages used for page table from been overwritten.
> > +
> > + @param[in] PageTableBase Base address of page table (CR3).
> > +
> > +**/
> > +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
> > diff --git a/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c
> b/MdeModulePkg/Core/DxeIplPeim/Ia32/DxeLoadFunc.c
> > index 5649265367..13fff28e93 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 29b6205e88..4ef3521224 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.
> >
> > @@ -117,6 +122,112 @@ EnableExecuteDisableBit (
> > AsmWriteMsr64 (0xC0000080, MsrRegisters);
> > }
> >
> > +/**
> > + 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 512 (= SIZE_2MB) and at
> > the
> > + boundary of SIZE_2MB. 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 512 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 (512).
> > + //
> > + if (PoolPages <= PAGE_TABLE_POOL_UNIT_PAGES) {
> > + PoolPages = PAGE_TABLE_POOL_UNIT_PAGES;
> > + } else {
> > + PoolPages = ((PoolPages + PAGE_TABLE_POOL_UNIT_PAGES - 1) /
> > + PAGE_TABLE_POOL_UNIT_PAGES) *
> PAGE_TABLE_POOL_UNIT_PAGES;
>
> What does the above calculation mean?
> If PoolPages equals to 513, new PoolPages = ((513 + 512) % 512) * 512
> = 512.
>
> Maybe the logic is trying to make PoolPages multiple of 512?
> If it's an internal function, just assert when PoolPages is not multiple
> of 512 should be good enough.
>
I want to always get multiple times of 512 pages which is large than requested
pages.
Because there's a chance (most likely at the first time of requesting page
tables)
that more than 512 pages of page table will be requested, and we don't know for
sure
how many page tables are required at each boot on different platform, I think
it is safer
to let the InitiliazePageTablePool to calculate a right number pages to reserve
in pool.
In addition, I think it'd be better not to expose the pool details to its
users, who just need
to request desired number of pages for page table.
And you're right there's error in the calculation. Thanks for catching it.
> > + }
> > +
> > + Buffer = AllocateAlignedPages (PoolPages,
> PAGE_TABLE_POOL_ALIGNMENT);
>
> PAGE_TABLE_POOL_UNIT_PAGES and PAGE_TABLE_POOL_ALIGNMENT actually
> do
> have some relationship between. Maybe just use
> EFI_SIZE_TO_PAGES (SIZE_2MB) to replace PAGE_TABLE_POOL_UNIT_PAGES
> and
> SIZE_2MB to replace PAGE_TABLE_POOL_ALIGNMENT?
>
For IA processor, 2MB is actually our only choice for the pool unit size and
alignment.
So I don't have strong opinion for this. Anyway works for me.
> > + 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) {
>
> Great to use FreePages to track the remaining free pages.
> It's possible that when FreePages equals to 3, Pages
> equals to 4, the 3 pages are wasted.
> But I think it should be fine. Not a big waste of memory.
>
Yes and No. Theoretically, pages will be wasted in your case. That's for sure.
But in actual current code, there's no chance to request more than 1 page
each time after the first page table memory request. So your case won't
happen.
Anyway, I think we don't need a complicated memory management here.
I agree that a little bit memory waste is acceptable.
> > + 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.
> >
> > @@ -144,7 +255,7 @@ Split2MPageTo4K (
> > //
> > AddressEncMask = PcdGet64 (PcdPteMemoryEncryptionAddressOrMask) &
> PAGING_1G_ADDRESS_MASK_64;
> >
> > - PageTableEntry = AllocatePages (1);
> > + PageTableEntry = AllocatePageTableMemory (1);
> > ASSERT (PageTableEntry != NULL);
> >
> > //
> > @@ -204,7 +315,7 @@ Split1GPageTo2M (
> > //
> > AddressEncMask = PcdGet64 (PcdPteMemoryEncryptionAddressOrMask) &
> PAGING_1G_ADDRESS_MASK_64;
> >
> > - PageDirectoryEntry = AllocatePages (1);
> > + PageDirectoryEntry = AllocatePageTableMemory (1);
> > ASSERT (PageDirectoryEntry != NULL);
> >
> > //
> > @@ -234,6 +345,184 @@ 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.
> > @@ -329,7 +618,7 @@ CreateIdentityMappingPageTables (
> > } else {
> > TotalPagesNum = NumberOfPml4EntriesNeeded + 1;
> > }
> > - BigPageAddress = (UINTN) AllocatePages (TotalPagesNum);
> > + BigPageAddress = (UINTN) AllocatePageTableMemory (TotalPagesNum);
> > ASSERT (BigPageAddress != 0);
> >
> > //
> > @@ -430,6 +719,12 @@ 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 7c9bb49e3e..9f14ac6007 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 (SIZE_2MB)
> > +#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.
> >
> >
>
>
> --
> Thanks,
> Ray
_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel
