git: u4b: Do not assume page-aligned busdma segments in the DMA page table

Aaron LI aly at crater.dragonflybsd.org
Wed Nov 18 05:12:57 PST 2026


commit caceac66aca06ff7a052ee179993e5ec55ccf8ae
Author: David Bidner <david at crap.solutions>
Date:   Mon Nov 2 16:36:38 2026 +0100

    u4b: Do not assume page-aligned busdma segments in the DMA page table
    
    usb_pc_common_mem_cb() walked the busdma segment list assuming every
    segment after the first begins on a page boundary (rem reset to 0 and
    the next entry always taken from the new segment's base address).
    busdma splits segments at physical page boundaries, at physical
    discontinuities and at maxsegsz; the first two begin a new segment on
    a page boundary, but when a buffer starts mid-page and consecutive
    pages are physically adjacent, the maxsegsz split ends a segment in
    the middle of a page and its continuation begins mid-page.  When that
    happened, the walk recorded the continuation's frame a second time
    under the next page index, shifting every later entry by one page.
    
    usbd_get_page() then handed the host controller wrong bus addresses
    for all pages after the duplicated entry, so multi-page USB storage
    reads returned the content of the following page and left the tail of
    the transfer with unrelated memory.  This needs a buffer whose first
    byte is not page aligned together with such a mid-page split, which is
    why 1 KiB and 2 KiB block filesystems on USB storage were affected
    while page-aligned buffers and single page transfers were not.
    
    Index the page table by buffer byte position instead: entry j is the
    frame containing buffer byte j * USB_PAGE_SIZE - rem, located in
    whichever segment covers it.  The rem adjustment in usbd_get_page()
    already matches this layout.
    
    FreeBSD's usb_busdma.c has the same walk.
    
    Verified in QEMU with usb-storage: ext2 1 KiB, 2 KiB and 4 KiB
    block and UFS reads, ext2 1 KiB reads over UHCI, raw device reads and
    writes with sub-page buffers, and write-heavy remount cycles; the
    corrupt remount reads are gone.  On real xHCI hardware with a USB
    stick (ext2 1 KiB), the corrupt remount reads occur with the old walk
    and not with this one.
    
    Bug: https://bugs.dragonflybsd.org/issues/3460

Summary of changes:
 sys/bus/u4b/usb_busdma.c | 33 +++++++++++++++++++++------------
 1 file changed, 21 insertions(+), 12 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/caceac66aca06ff7a052ee179993e5ec55ccf8ae


-- 
DragonFly BSD source repository


More information about the Commits mailing list