[Qemu-devel] Small patch for getdents syscall

[Qemu-devel] Small patch for getdents syscall

[Henry Wertz]

I have a trivial, 1-line patch for getdents function; due to use of
unsigned long, the struct on 64-bit and 32-bit systems is not the same.
qemu is aware of this, however it currently only checks for a 32-bit target
on 64-bit host case; in my case, I'm running 64-bit target on 32-bit host
(x86-64 binaries on a 32-bit ARM).

My use case for this patch, I have Android Studio on my (32-bit ARM)
chromebook -- thank goodness I got the 4GB model!  Since Android Studio is
java, that worked; but it calls aapt from build tools.  Up through 23.x.x
(which is build for 32-bit x86) it worked.  With 25.0.0 (and 24.x.x
version) aapt is now built for 64-bit x86-64; it would give an odd error
about a directory being invalid.  By comparing strace between qemu and real
x86-64, I found getdents was behaving differently.  WIth this patch aapt
completes.

I really wasn't sure if this would be better this way ("if target != host")
or if something like  "if (target==32 && host==64) || (target==64 &&
host==32)" would be preferrable.
---------------------------------------------------

Signed-off-by: Henry Wertz <hwertz10@gmail.com>


*** linux-user/syscall.c~       2017-03-04 10:31:14.000000000 -0600
--- linux-user/syscall.c        2017-03-07 17:08:24.615399116 -0600
***************
*** 9913,9921 ****
  #endif
  #ifdef TARGET_NR_getdents
      case TARGET_NR_getdents:
  #ifdef __NR_getdents
! #if TARGET_ABI_BITS == 32 && HOST_LONG_BITS == 64
          {
              struct target_dirent *target_dirp;
              struct linux_dirent *dirp;
              abi_long count = arg3;
--- 9913,9921 ----
  #endif
  #ifdef TARGET_NR_getdents
      case TARGET_NR_getdents:
  #ifdef __NR_getdents
! #if TARGET_ABI_BITS != HOST_LONG_BITS
          {
              struct target_dirent *target_dirp;
              struct linux_dirent *dirp;
              abi_long count = arg3;

Re: [Qemu-devel] Small patch for getdents syscall

[Peter Maydell]

On 8 March 2017 at 00:40, Henry Wertz <hwertz10@gmail.com> wrote:
> I have a trivial, 1-line patch for getdents function; due to use of
> unsigned long, the struct on 64-bit and 32-bit systems is not the same.
> qemu is aware of this, however it currently only checks for a 32-bit target
> on 64-bit host case; in my case, I'm running 64-bit target on 32-bit host
> (x86-64 binaries on a 32-bit ARM).
>
> My use case for this patch, I have Android Studio on my (32-bit ARM)
> chromebook -- thank goodness I got the 4GB model!  Since Android Studio is
> java, that worked; but it calls aapt from build tools.  Up through 23.x.x
> (which is build for 32-bit x86) it worked.  With 25.0.0 (and 24.x.x
> version) aapt is now built for 64-bit x86-64; it would give an odd error
> about a directory being invalid.  By comparing strace between qemu and real
> x86-64, I found getdents was behaving differently.  WIth this patch aapt
> completes.
>
> I really wasn't sure if this would be better this way ("if target != host")
> or if something like  "if (target==32 && host==64) || (target==64 &&
> host==32)" would be preferrable.
> ---------------------------------------------------
>
> Signed-off-by: Henry Wertz <hwertz10@gmail.com>
>
> *** linux-user/syscall.c~       2017-03-04 10:31:14.000000000 -0600
> --- linux-user/syscall.c        2017-03-07 17:08:24.615399116 -0600
> ***************
> *** 9913,9921 ****
>   #endif
>   #ifdef TARGET_NR_getdents
>       case TARGET_NR_getdents:
>   #ifdef __NR_getdents
> ! #if TARGET_ABI_BITS == 32 && HOST_LONG_BITS == 64
>           {
>               struct target_dirent *target_dirp;
>               struct linux_dirent *dirp;
>               abi_long count = arg3;
> --- 9913,9921 ----
>   #endif
>   #ifdef TARGET_NR_getdents
>       case TARGET_NR_getdents:
>   #ifdef __NR_getdents
> ! #if TARGET_ABI_BITS != HOST_LONG_BITS
>           {
>               struct target_dirent *target_dirp;
>               struct linux_dirent *dirp;
>               abi_long count = arg3;

There is a theoretically better way to handle the 64-bit
guest on 32-bit host case: since here you know that the
guest dirent struct is bigger than the host dirent
struct you could read them directly into the guest buffer
and then convert in-place rather than having to malloc
a temporary buffer. The conversion code is a little
bit awkward though because you have to make sure you don't
overwrite host data you haven't read yet, so I think
you would need to start at the end of the buffer and
work backwards. This is basically fixing the #else
part of this code to work for guest > host as well
as for guest == host.

However, since 64-bit-on-32 is not exactly a very common
use case these days I think your patch is an OK fix for
things.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>

PS: we prefer 'unified' format diffs; git format-patch
should produce those by default.

thanks
-- PMM

Re: [Qemu-devel] Small patch for getdents syscall

[Henry Wertz]

On Mar 13, 2017 5:49 AM, "Peter Maydell" <peter.maydell@linaro.org> wrote:
>
> On 8 March 2017 at 00:40, Henry Wertz <hwertz10@gmail.com> wrote:
> > I have a trivial, 1-line patch for getdents function; due to
> >...
> There is a theoretically better way to handle the 64-bit
> guest on 32-bit host case: since here you know that the
> guest dirent struct is bigger than the host dirent
> struct you could read them directly into the guest buffer
> and then convert in-place rather than having to malloc
> a temporary buffer. The conversion code is a little
> bit awkward though because you have to make sure you don't
> overwrite host data you haven't read yet, so I think
> you would need to start at the end of the buffer and
> work backwards. This is basically fixing the #else
> part of this code to work for guest > host as well
> as for guest == host.

That does sound like it'd be better (avoiding one unnecessary malloc at
least.)

>
> However, since 64-bit-on-32 is not exactly a very common
> use case these days I think your patch is an OK fix for
> things.
I'd guess this use might (slightly) increase if more x86 binary software
begins to be shipped x86-64 instead.  That said, this particular code path
may still barely be used; I would have expected most (especially 64-bit)
software would call getdents64 instead of getdents, and was surprised to
find aapt uses getdents instead.

> Reviewed-by: Peter Maydell <peter.maydell@linaro.org>

>
> PS: we prefer 'unified' format diffs; git format-patch
> should produce those by default.
>
> thanks
> -- PMM

OK I'll make sure to do that next time.

Thanks!

-- Henry Wertz