[lkml]   [2011]   [Feb]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: ARM unaligned MMIO access with attribute((packed))
On Wed, Feb 2, 2011 at 5:00 PM, Arnd Bergmann <> wrote:
> As noticed by Peter Maydell, the EHCI device driver in Linux gets
> miscompiled by some versions of arm-gcc (still need to find out which)
> due to a combination of problems:
> 1. In include/linux/usb/ehci_def.h, struct ehci_caps is defined
> with __attribute__((packed)), for no good reason. This is clearly
> a bug and needs to get fixed, but other drivers have the same bug
> and it used to work. The attribute forces byte access on all members
> accessed through pointer dereference, which is not allowed on
> MMIO accesses in general. The specific code triggering the problem
> in Peter's case is in ehci-omap.c:
>        omap->ehci->regs = hcd->regs
>                + HC_LENGTH(readl(&omap->ehci->caps->hc_capbase));
> 2. The ARM version of the readl() function is implemented as a macro
> doing a direct pointer dereference with a typecast:
> #define __raw_readl(a)          (__chk_io_ptr(a), *(volatile unsigned int __force   *)(a))
> #define readl_relaxed(c) ({ u32 __v = le32_to_cpu((__force __le32) \
>                                        __raw_readl(__mem_pci(c))); __v; })
> #define readl(c)                ({ u32 __v = readl_relaxed(c); __iormb(); __v; })
> On other architectures, readl() is implemented using an inline assembly
> specifically to prevent gcc from issuing anything but a single 32-bit
> load instruction. readl() only makes sense on aligned memory, so in case
> of a misaligned pointer argument, it should cause a trap anyway.
> 3. gcc does not seem to clearly define what happens during a cast between
> aligned an packed pointers. In this case, the original pointer is packed
> (byte aligned), while the access is done through a 32-bit aligned
> volatile unsigned int pointer. In gcc-4.4, casting from "unsigned int
> __attribute__((packed))" to "volatile unsigned int" resulted in a 32-bit
> aligned access, while casting to "unsigned int" (without volatile) resulted
> in four byte accesses. gcc-4.5 seems to have changed this to always do
> a byte access in both cases, but still does not document the behavior.
> (need to confirm this).
> I would suggest fixing this by:
> 1. auditing all uses of __attribute__((packed)) in the Linux USB code
> and other drivers, removing the ones that are potentially harmful.
> 2. Changing the ARM MMIO functions to use inline assembly instead of
> direct pointer dereference.
> 3. Documenting the gcc behavior as undefined.

The pointer conversions already invoke undefined behavior as specified by the
C standard (

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2011-02-02 17:53    [W:0.103 / U:2.400 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site