| Message ID | 20181024023831.2529-1-jcmvbkbc@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | linux-user/flatload: fix initial stack pointer alignment | expand |
On 24/10/2018 03:38, Max Filippov wrote: > Stack pointer alignment code incorrectly adds stack_size to sp instead > of subtracting it. It also does not take flat_argvp_envp_on_stack() into > account when calculating stack_size. This results in initial stack > pointer misalignment with certain set of command line arguments and > environment variables and correct alignment for the same binary with a > different set of arguments. This misalignment causes failures in the > following tests in the testsuite of gcc built for xtensa uclinux: > > gcc.dg/torture/vshuf-v64qi.c > gcc.dg/torture/vshuf-v8sf.c > gcc.dg/torture/vshuf-v8si.c > > Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> > --- > linux-user/flatload.c | 8 +++++--- > 1 file changed, 5 insertions(+), 3 deletions(-) > > diff --git a/linux-user/flatload.c b/linux-user/flatload.c > index 2eefe55e5000..1893966b5b30 100644 > --- a/linux-user/flatload.c > +++ b/linux-user/flatload.c > @@ -771,10 +771,12 @@ int load_flt_binary(struct linux_binprm *bprm, struct image_info *info) > /* Enforce final stack alignment of 16 bytes. This is sufficient > for all current targets, and excess alignment is harmless. */ > stack_len = bprm->envc + bprm->argc + 2; > - stack_len += 3; /* argc, arvg, argp */ > + stack_len += flat_argvp_envp_on_stack() ? 2 : 0; /* arvg, argp */ > + stack_len += 1; /* argc */> stack_len *= sizeof(abi_ulong); > - if ((sp + stack_len) & 15) > - sp -= 16 - ((sp + stack_len) & 15); > + if ((sp - stack_len) & 15) { > + sp -= ((sp - stack_len) & 15); > + } If I understand correctly the purpose, I think it could be clearer like: sp = (sp - stack_len) & ~15; Thanks, Laurent
On Wed, Oct 24, 2018 at 4:35 AM Laurent Vivier <laurent@vivier.eu> wrote: > > diff --git a/linux-user/flatload.c b/linux-user/flatload.c > > index 2eefe55e5000..1893966b5b30 100644 > > --- a/linux-user/flatload.c > > +++ b/linux-user/flatload.c > > - sp -= 16 - ((sp + stack_len) & 15); > > + if ((sp - stack_len) & 15) { > > + sp -= ((sp - stack_len) & 15); > > + } > > If I understand correctly the purpose, I think it could be clearer like: > > sp = (sp - stack_len) & ~15; Yes, you're right. I'll send v2.
On Wed, Oct 24, 2018 at 10:19 AM Max Filippov <jcmvbkbc@gmail.com> wrote: > > On Wed, Oct 24, 2018 at 4:35 AM Laurent Vivier <laurent@vivier.eu> wrote: > > > diff --git a/linux-user/flatload.c b/linux-user/flatload.c > > > index 2eefe55e5000..1893966b5b30 100644 > > > --- a/linux-user/flatload.c > > > +++ b/linux-user/flatload.c > > > > - sp -= 16 - ((sp + stack_len) & 15); > > > + if ((sp - stack_len) & 15) { > > > + sp -= ((sp - stack_len) & 15); > > > + } > > > > If I understand correctly the purpose, I think it could be clearer like: > > > > sp = (sp - stack_len) & ~15; > > Yes, you're right. I'll send v2. Well, not exactly. The sp is not moved down stack_len + alignment bytes, only the alignment bytes. I'll send v2 anyway.
diff --git a/linux-user/flatload.c b/linux-user/flatload.c index 2eefe55e5000..1893966b5b30 100644 --- a/linux-user/flatload.c +++ b/linux-user/flatload.c @@ -771,10 +771,12 @@ int load_flt_binary(struct linux_binprm *bprm, struct image_info *info) /* Enforce final stack alignment of 16 bytes. This is sufficient for all current targets, and excess alignment is harmless. */ stack_len = bprm->envc + bprm->argc + 2; - stack_len += 3; /* argc, arvg, argp */ + stack_len += flat_argvp_envp_on_stack() ? 2 : 0; /* arvg, argp */ + stack_len += 1; /* argc */ stack_len *= sizeof(abi_ulong); - if ((sp + stack_len) & 15) - sp -= 16 - ((sp + stack_len) & 15); + if ((sp - stack_len) & 15) { + sp -= ((sp - stack_len) & 15); + } sp = loader_build_argptr(bprm->envc, bprm->argc, sp, p, flat_argvp_envp_on_stack());
Stack pointer alignment code incorrectly adds stack_size to sp instead of subtracting it. It also does not take flat_argvp_envp_on_stack() into account when calculating stack_size. This results in initial stack pointer misalignment with certain set of command line arguments and environment variables and correct alignment for the same binary with a different set of arguments. This misalignment causes failures in the following tests in the testsuite of gcc built for xtensa uclinux: gcc.dg/torture/vshuf-v64qi.c gcc.dg/torture/vshuf-v8sf.c gcc.dg/torture/vshuf-v8si.c Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> --- linux-user/flatload.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)