On 7 September 2015 at 10:41, Heyi Guo <[email protected]> wrote: > > > On 09/06/2015 07:43 PM, Ard Biesheuvel wrote: >> >> On 6 September 2015 at 10:15, Heyi Guo <[email protected]> wrote: >>> >>> The code has a simple bug on calculating aligned page table address. >>> We need to add alignment - 1 to allocated address first and then mask >>> the unaligned bits. >>> >> Nice find! >> >>> Contributed-under: TianoCore Contribution Agreement 1.0 >>> Signed-off-by: Heyi Guo <[email protected]> >>> Cc: Leif Lindholm <[email protected]> >>> Cc: Ard Biesheuvel <[email protected]> >>> --- >>> ArmPkg/Library/ArmLib/AArch64/AArch64Mmu.c | 4 ++-- >>> 1 file changed, 2 insertions(+), 2 deletions(-) >>> >>> diff --git a/ArmPkg/Library/ArmLib/AArch64/AArch64Mmu.c >>> b/ArmPkg/Library/ArmLib/AArch64/AArch64Mmu.c >>> index 3d58d5d..4db4bbe 100644 >>> --- a/ArmPkg/Library/ArmLib/AArch64/AArch64Mmu.c >>> +++ b/ArmPkg/Library/ArmLib/AArch64/AArch64Mmu.c >>> @@ -381,7 +381,7 @@ GetBlockEntryListFromAddress ( >>> if (TranslationTable == NULL) { >>> return NULL; >>> } >>> - TranslationTable = (UINT64*)((UINTN)TranslationTable & >>> TT_ADDRESS_MASK_DESCRIPTION_TABLE); >>> + TranslationTable = (UINT64*)(((UINTN)TranslationTable + >>> TT_ALIGNMENT_DESCRIPTION_TABLE - 1) & TT_ADDRESS_MASK_DESCRIPTION_TABLE); >>> >> Could we do (TranslationTable | (TT_ALIGNMENT_DESCRIPTION_TABLE - 1)) >> + 1 instead of using two different symbolic constants? >> >>> // Populate the newly created lower level table >>> SubTableBlockEntry = TranslationTable; >>> @@ -409,7 +409,7 @@ GetBlockEntryListFromAddress ( >>> if (TranslationTable == NULL) { >>> return NULL; >>> } >>> - TranslationTable = (UINT64*)((UINTN)TranslationTable & >>> TT_ADDRESS_MASK_DESCRIPTION_TABLE); >>> + TranslationTable = (UINT64*)(((UINTN)TranslationTable + >>> TT_ALIGNMENT_DESCRIPTION_TABLE - 1) & TT_ADDRESS_MASK_DESCRIPTION_TABLE); >>> >>> ZeroMem (TranslationTable, TT_ENTRY_COUNT * sizeof(UINT64)); >>> >> I think you missed another instance at line 625 >> >> So what is hidden from view by all these symbolic constants is that we >> are aligning the return value of AllocatePages() to 4 KB, which is >> rather pointless, and is actually resulting in every allocation to >> waste 4 KB. But that also means the buggy rounding code never produces >> incorrect values, making the severity of this bug low. >> >> So instead, could we get a comprehensive patch that: >> - replaces all three instances with a call to a static function that >> does the allocation and the rounding >> - apply the rounding only if the alignment exceeds 4 KB >> - fix the rounding logic. > > > Good suggestion to make code clean; thanks :) >
Just as a note, you could implement something like: Table = AllocatePages (size + rounding) RoundedTable = (Table | (Alignment - 1)) +1 FreePages (Table, RoundedTable - Table) FreePages(<the unused pages at the top>) to immediately free the pages that you will not use after rounding. -- Ard. _______________________________________________ edk2-devel mailing list [email protected] https://lists.01.org/mailman/listinfo/edk2-devel
