| Message ID | 20191021225100.17990-1-guillem@hadrons.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] aio: Fix io_pgetevents() struct __compat_aio_sigset layout | expand |
On Tue, Oct 22, 2019 at 12:49 AM Guillem Jover <guillem@hadrons.org> wrote: > > This type is used to pass the sigset_t from userland to the kernel, > but it was using the kernel native pointer type for the member > representing the compat userland pointer to the userland sigset_t. > > This messes up the layout, and makes the kernel eat up both the > userland pointer and the size into the kernel pointer, and then > reads garbage into the kernel sigsetsize. Which makes the sigset_t > size consistency check fail, and consequently the syscall always > returns -EINVAL. > > This breaks both libaio and strace on 32-bit userland running on 64-bit > kernels. And there are apparently no users in the wild of the current > broken layout (at least according to codesearch.debian.org and a brief > check over github.com search). So it looks safe to fix this directly > in the kernel, instead of either letting userland deal with this > permanently with the additional overhead or trying to make the syscall > infer what layout userland used, even though this is also being worked > around in libaio to temporarily cope with kernels that have not yet > been fixed. > > We use a proper compat_uptr_t instead of a compat_sigset_t pointer. > > Fixes: 7a074e96dee6 ("aio: implement io_pgetevents") > Signed-off-by: Guillem Jover <guillem@hadrons.org> When resending a patch that has already been reviewed, please add the tags you received so they get picked up into the final changeset as well: Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Jeff Moyer <jmoyer@redhat.com> Let's make sure this also gets added to stable kernels Cc: <stable@vger.kernel.org> # v4.18+ Finally (if you like) Reviewed-by: Arnd Bergmann <arnd@arndb.de> Arnd
diff --git a/fs/aio.c b/fs/aio.c index 01e0fb9ae45a..0d9a559d488c 100644 --- a/fs/aio.c +++ b/fs/aio.c @@ -2179,7 +2179,7 @@ SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id, #ifdef CONFIG_COMPAT struct __compat_aio_sigset { - compat_sigset_t __user *sigmask; + compat_uptr_t sigmask; compat_size_t sigsetsize; }; @@ -2193,7 +2193,7 @@ COMPAT_SYSCALL_DEFINE6(io_pgetevents, struct old_timespec32 __user *, timeout, const struct __compat_aio_sigset __user *, usig) { - struct __compat_aio_sigset ksig = { NULL, }; + struct __compat_aio_sigset ksig = { 0, }; struct timespec64 t; bool interrupted; int ret; @@ -2204,7 +2204,7 @@ COMPAT_SYSCALL_DEFINE6(io_pgetevents, if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) return -EFAULT; - ret = set_compat_user_sigmask(ksig.sigmask, ksig.sigsetsize); + ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize); if (ret) return ret; @@ -2228,7 +2228,7 @@ COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, struct __kernel_timespec __user *, timeout, const struct __compat_aio_sigset __user *, usig) { - struct __compat_aio_sigset ksig = { NULL, }; + struct __compat_aio_sigset ksig = { 0, }; struct timespec64 t; bool interrupted; int ret; @@ -2239,7 +2239,7 @@ COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64, if (usig && copy_from_user(&ksig, usig, sizeof(ksig))) return -EFAULT; - ret = set_compat_user_sigmask(ksig.sigmask, ksig.sigsetsize); + ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize); if (ret) return ret;
This type is used to pass the sigset_t from userland to the kernel, but it was using the kernel native pointer type for the member representing the compat userland pointer to the userland sigset_t. This messes up the layout, and makes the kernel eat up both the userland pointer and the size into the kernel pointer, and then reads garbage into the kernel sigsetsize. Which makes the sigset_t size consistency check fail, and consequently the syscall always returns -EINVAL. This breaks both libaio and strace on 32-bit userland running on 64-bit kernels. And there are apparently no users in the wild of the current broken layout (at least according to codesearch.debian.org and a brief check over github.com search). So it looks safe to fix this directly in the kernel, instead of either letting userland deal with this permanently with the additional overhead or trying to make the syscall infer what layout userland used, even though this is also being worked around in libaio to temporarily cope with kernels that have not yet been fixed. We use a proper compat_uptr_t instead of a compat_sigset_t pointer. Fixes: 7a074e96dee6 ("aio: implement io_pgetevents") Signed-off-by: Guillem Jover <guillem@hadrons.org> --- fs/aio.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)