Hi Ming, Could you please raise this as a PR on https://github.com/tianocore/edk2 ?
Best Regards, Leif On Fri, 23 Oct 2026 at 09:33, Ming Huang via groups.io <[email protected]> wrote: > > The DstBuffer adjustment in FfsProcessSection() was added in the > BeagleBoard era (2ef2b01e07), when the decompressed section stream > placed the encapsulated section data directly after the 4/8 byte > EFI_COMMON_SECTION_HEADER. Bumping DstBuffer forward by > EFI_PAGE_SIZE - sizeof (EFI_COMMON_SECTION_HEADER) put the section > data back on a page boundary at the cost of one extra page. > > For FV images this compensation is obsolete and harmful: > > - GenFds aligns the FV_IMAGE section data inside the stream to the > FvAlignment recorded in the inner FV header, by emitting a RAW pad > section in front of the FV_IMAGE section. With a page aligned > DstBuffer, the decompressed FV therefore already starts at a > FvAlignment-aligned offset (e.g. DstBuffer + 0x10 for a typical > FvAlignment = 16 build). > > - The forward bump instead moves the FV to > DstBuffer + EFI_PAGE_SIZE - sizeof (EFI_COMMON_SECTION_HEADER) + pad, > which is 12 bytes off for FvAlignment = 16. FfsProcessFvFile() then > takes its misalignment fallback and AllocateAlignedPages() a second > buffer to CopyMem() the entire FV, doubling the temporary RAM > footprint of the decompressed DXEFV. > > On platforms with a modest SEC temporary RAM pool this is fatal: a > RISC-V platform with a 16 MiB pool and a ~9 MiB LZMA compressed DXEFV > fails to boot with EFI_OUT_OF_RESOURCES from DecompressFirstFv(), > because the redundant aligned copy no longer fits. The same FDF > pattern (FvAlignment = 16 DXEFV wrapped in the LZMA GUIDED section) > is used by OvmfPkg/RiscVVirt, which gets away with it only because > of its more generous memory layout, and pays a full extra copy of > the FV on every boot. > > Remove the adjustment, and the extra page that was reserved for it, > so the decompressed FV keeps its natural, build-time guaranteed > alignment. Verified on the affected RISC-V platform: the FV now > lands 16-byte aligned, the fallback copy is not taken, and boot > succeeds in both RELEASE and DEBUG builds. > > Cc: Leif Lindholm <[email protected]> > Cc: Ard Biesheuvel <[email protected]> > Cc: Abner Chang <[email protected]> > Signed-off-by: Ming Huang <[email protected]> > --- > EmbeddedPkg/Library/PrePiLib/FwVol.c | 13 +++++-------- > 1 file changed, 5 insertions(+), 8 deletions(-) > > diff --git a/EmbeddedPkg/Library/PrePiLib/FwVol.c > b/EmbeddedPkg/Library/PrePiLib/FwVol.c > index 69c493187b..287c2b8062 100644 > --- a/EmbeddedPkg/Library/PrePiLib/FwVol.c > +++ b/EmbeddedPkg/Library/PrePiLib/FwVol.c > @@ -449,19 +449,16 @@ FfsProcessSection ( > } > > // > - // Allocate destination buffer, extra one page for adjustment > + // Allocate destination buffer. The buffer is page aligned, and the > + // decompressed section stream keeps the alignment of the embedded FV > + // (GenFds pads the FV_IMAGE section data to the FvAlignment recorded > in > + // the inner FV header), so no further adjustment is needed here. > // > - DstBuffer = (VOID *)(UINTN)AllocatePages (EFI_SIZE_TO_PAGES > (DstBufferSize) + 1); > + DstBuffer = (VOID *)(UINTN)AllocatePages (EFI_SIZE_TO_PAGES > (DstBufferSize)); > if (DstBuffer == NULL) { > return EFI_OUT_OF_RESOURCES; > } > > - // > - // DstBuffer still is one section. Adjust DstBuffer offset, skip EFI > section header > - // to make section data at page alignment. > - // > - DstBuffer = (UINT8 *)DstBuffer + EFI_PAGE_SIZE - SectionHeaderSize; > - > // > // Call decompress function > // > -- > 2.17.1 > > > > > > -=-=-=-=-=-=-=-=-=-=-=- Groups.io Links: You receive all messages sent to this group. View/Reply Online (#122217): https://edk2.groups.io/g/devel/message/122217 Mute This Topic: https://groups.io/mt/121897317/21656 Group Owner: [email protected] Unsubscribe: https://edk2.groups.io/g/devel/unsub [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
