On Thu, Apr 18, 2019 at 12:03:07AM +0200, Arnd Bergmann wrote:
On Wed, Apr 17, 2019 at 11:13 PM Al Viro viro@zeniv.linux.org.uk wrote:
On Tue, Apr 16, 2019 at 10:19:40PM +0200, Arnd Bergmann wrote:
diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c index c708400fff4a..04252c3492ee 100644 --- a/drivers/net/ppp/ppp_generic.c +++ b/drivers/net/ppp/ppp_generic.c @@ -899,6 +899,7 @@ static const struct file_operations ppp_device_fops = { .write = ppp_write, .poll = ppp_poll, .unlocked_ioctl = ppp_ioctl,
.compat_ioctl = ppp_ioctl,Oh? What happens on e.g. s390 with something like PPPIOCNEWUNIT? Current kernel: * no ->compat_ioctl() * ->unlock_ioctl() is present * found by compat_ioctl_check_table() * pass (unsigned long)compat_ptr(arg) to do_vfs_ioctl() * pass that to ppp_ioctl() * pass that to ppp_unattached_ioctl() * fetch int from (int __user *)compat_ptr(arg)
With your patch: * call ppp_ioctl() * pass arg to ppp_unattached_ioctl() * fetch int from (int __user *)arg
AFAICS, that's broken...
Correct. I had added this patch to the series from an older set of patches (predating the compat_ptr_ioctl() series) , and did not check for this issue again. When I originally created the patch, I assumed that even on s390 it would be no problem.
Looking at that ppp_ioctl(), pointer to arch-independent type or ignored: PPPIOCNEWUNIT PPPIOCATTACH PPPIOCATTCHAN PPPIOCSMRU PPPIOCSFLAGS PPPIOCGFLAGS PPPIOCGUNIT PPPIOCSDEBUG PPPIOCSMAXCID PPPIOCCONNECT PPPIOCGDEBUG PPPIOCSMAXCID PPPIOCSMRRU PPPIOCDETACH PPPIOCDISCONN PPPIOCGASYNCMAP PPPIOCSASYNCMAP PPPIOCGRASYNCMAP PPPIOCSRASYNCMAP PPPIOCGXASYNCMAP PPPIOCSXASYNCMAP PPPIOCGNPMODE PPPIOCSNPMODE pointer to struct ppp_option_data (with further pointer-chasing in it): PPPIOCSCOMPRESS pointer to struct ppp_idle: PPPIOCGIDLE pointer to struct sock_filter (with hidden pointer-chasing, AFAICS): PPPIOCSPASS PPPIOCSACTIVE
Pretty much all of them take pointers. What's more, reaction to unknown is -ENOTTY, not -ENOIOCTLCM, so that patch will have prevent the translated ones from reaching do_ioctl_trans()
Good point, this patch sequence does break bisection.
What am I missing here? Why not simply do
compat_ppp_ioctl() { PPPIOCSCOMPRESS32 => deal with it PPPIOCGIDLE32 => deal with it PPPIOCSPASS32 / PPPIOCSACTIVE32 => deal with it default: pass compat_ptr(arg) to ppp_ioctl() and be done with that }
with BPF-related bits (both compat and native) taken to e.g. net/core/bpf-ppp.c, picked by both generic and isdn? IDGI...
I was trying to unify the native and compat code paths as much as possible here. Handling the four PPPIO*32 commands in compat_ppp_ioctl would either require duplicating large chunks of ppp_ioctl, or keeping the extra compat_alloc_user_space() copy from the existing implementation.
I'll try to come up with a different way to structure the patches.
Huh? Instead of case PPPIOCSCOMPRESS: err = ppp_set_compress(ppp, arg); break; in native, have struct ppp_option_data data; ... case PPPIOCSCOMPRESS: if (copy_from_user(&data, argp, sizeof(data))) err = -EFAULT; else err = ppp_set_compress(ppp, &data); break;
while in compat do struct ppp_option_data32 data32;
case PPPIOCSCOMPRESS32: if (copy_from_user(&data32, argp, sizeof(data32))) err = -EFAULT; else err = ppp_set_compress(ppp, &(struct ppp_option_data){ .ptr = compat_ptr(data32.ptr), .length = data32.length, .transmit = data32.transmit }); break;
PPPIOCGIDLE is small to start with - not a lot to copy there. And as for the filters... What you need is something like
struct bpf_prog *get_ppp_bpf(struct sock_fprog __user *p) { struct sock_fprog uprog; struct sock_fprog_kern fprog; struct sock_filter *code = NULL; struct bpf_prog *res; int err;
if (copy_from_user(&uprog, p, sizeof(uprog))) return ERR_PTR(-EFAULT);
if (!uprog.len) return NULL;
fprog.len = uprog.len * sizeof(struct sock_filter); code = fprog.filter = memdup_user(uprog.filter, fprog.len); if (IS_ERR(code)) return ERR_CAST(code);
err = bpf_prog_create(&res, &fprog); kfree(code);
if (err) return ERR_PTR(err); return res; } in net/core/ppp-filter.c (or in net/core/filter.c, for that matter, under appropriate ifdef)) with obvious compat counterpart next to it. Hell, turn the above into
static struct bpf_prog *__get_ppp_bpf(struct sock_fprog *kp) { struct sock_fprog_kern fprog; struct sock_filter *code = NULL; struct bpf_prog *res; int err;
if (!kp->len) return NULL;
fprog.len = kp->len * sizeof(struct sock_filter); code = fprog.filter = memdup_user(kp->filter, fprog.len); if (IS_ERR(code)) return ERR_CAST(code);
err = bpf_prog_create(&res, &fprog); kfree(code);
if (err) return ERR_PTR(err); return res; }
struct bpf_prog *get_ppp_bpf(struct sock_fprog __user *p) { struct sock_fprog uprog;
if (copy_from_user(&uprog, p, sizeof(uprog))) return ERR_PTR(-EFAULT); return __get_ppp_bpf(&uprog); }
struct bpf_prog *compat_get_ppp_bpf(struct sock_fprog32 __user *p) { struct sock_fprog uprog32;
if (copy_from_user(&uprog32, p, sizeof(uprog32))) return ERR_PTR(-EFAULT); return __get_ppp_bpf(&(struct sock_fprog){ .len = uprog32.len, .filter = compat_ptr(uprog32.filter)}); }
Then in native ioctl do case PPPIOCSPASS: case PPPIOCSACTIVE: { struct bpf_prog *filter = get_bpf_ppp(argp);
if (IS_ERR(filter)) { err = PTR_ERR(filter); } else { struct bpf_prog **which; if (cmd == PPPIOCSPASS) which = &ppp->pass_filter; else which = &ppp->active_filter; ppp_lock(ppp); if (*which) bpf_prog_destroy(*which); *which = filter; ppp_unlock(ppp); err = 0; } break; } in native and similar in compat, with get_bpf_ppp() replaced with call of compat_get_bpf_ppp() and ioctl numbers obviously adjusted. All there is to it... Helpers obviously shared with isdn and yes, all crap gone from fs/compat_ioctl.c...
Why would you want to duplicate large chunks of anything? The above is not even compile-tested, but... I can put together a patch if you wish. Or am I missing something here?